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/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/force_flags.h"
80 #include "gromacs/mdlib/forcerec.h"
81 #include "gromacs/mdlib/gmx_omp_nthreads.h"
82 #include "gromacs/mdlib/update.h"
83 #include "gromacs/mdlib/vsite.h"
84 #include "gromacs/mdlib/wholemoleculetransform.h"
85 #include "gromacs/mdtypes/commrec.h"
86 #include "gromacs/mdtypes/enerdata.h"
87 #include "gromacs/mdtypes/forceoutput.h"
88 #include "gromacs/mdtypes/forcerec.h"
89 #include "gromacs/mdtypes/iforceprovider.h"
90 #include "gromacs/mdtypes/inputrec.h"
91 #include "gromacs/mdtypes/md_enums.h"
92 #include "gromacs/mdtypes/mdatom.h"
93 #include "gromacs/mdtypes/simulation_workload.h"
94 #include "gromacs/mdtypes/state.h"
95 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
96 #include "gromacs/nbnxm/gpu_data_mgmt.h"
97 #include "gromacs/nbnxm/nbnxm.h"
98 #include "gromacs/nbnxm/nbnxm_gpu.h"
99 #include "gromacs/pbcutil/ishift.h"
100 #include "gromacs/pbcutil/pbc.h"
101 #include "gromacs/pulling/pull.h"
102 #include "gromacs/pulling/pull_rotation.h"
103 #include "gromacs/timing/cyclecounter.h"
104 #include "gromacs/timing/gpu_timing.h"
105 #include "gromacs/timing/wallcycle.h"
106 #include "gromacs/timing/wallcyclereporting.h"
107 #include "gromacs/timing/walltime_accounting.h"
108 #include "gromacs/topology/topology.h"
109 #include "gromacs/utility/arrayref.h"
110 #include "gromacs/utility/basedefinitions.h"
111 #include "gromacs/utility/cstringutil.h"
112 #include "gromacs/utility/exceptions.h"
113 #include "gromacs/utility/fatalerror.h"
114 #include "gromacs/utility/fixedcapacityvector.h"
115 #include "gromacs/utility/gmxassert.h"
116 #include "gromacs/utility/gmxmpi.h"
117 #include "gromacs/utility/logger.h"
118 #include "gromacs/utility/smalloc.h"
119 #include "gromacs/utility/strconvert.h"
120 #include "gromacs/utility/sysinfo.h"
122 using gmx::AtomLocality;
123 using gmx::DomainLifetimeWorkload;
124 using gmx::ForceOutputs;
125 using gmx::InteractionLocality;
126 using gmx::SimulationWorkload;
127 using gmx::StepWorkload;
129 // TODO: this environment variable allows us to verify before release
130 // that on less common architectures the total cost of polling is not larger than
131 // a blocking wait (so polling does not introduce overhead when the static
132 // PME-first ordering would suffice).
133 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
135 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
137 const int end = forceToAdd.size();
139 int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
140 #pragma omp parallel for num_threads(nt) schedule(static)
141 for (int i = 0; i < end; i++)
143 rvec_inc(f[i], forceToAdd[i]);
147 static void calc_virial(int start,
150 const gmx::ForceWithShiftForces& forceWithShiftForces,
154 const t_forcerec* fr,
157 /* The short-range virial from surrounding boxes */
158 const rvec* fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
159 calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, pbcType == PbcType::Screw, box);
160 inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
162 /* Calculate partial virial, for local atoms only, based on short range.
163 * Total virial is computed in global_stat, called from do_md
165 const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
166 f_calc_vir(start, start + homenr, x, f, vir_part, box);
167 inc_nrnb(nrnb, eNR_VIRIAL, homenr);
171 pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
175 static void pull_potential_wrapper(const t_commrec* cr,
176 const t_inputrec* ir,
178 gmx::ArrayRef<const gmx::RVec> x,
179 gmx::ForceWithVirial* force,
180 const t_mdatoms* mdatoms,
181 gmx_enerdata_t* enerd,
185 gmx_wallcycle_t wcycle)
190 /* Calculate the center of mass forces, this requires communication,
191 * which is why pull_potential is called close to other communication.
193 wallcycle_start(wcycle, ewcPULLPOT);
194 set_pbc(&pbc, ir->pbcType, box);
196 enerd->term[F_COM_PULL] += pull_potential(pull_work, mdatoms, &pbc, cr, t, lambda[efptRESTRAINT],
197 as_rvec_array(x.data()), force, &dvdl);
198 enerd->dvdl_lin[efptRESTRAINT] += dvdl;
199 for (auto& dhdl : enerd->dhdlLambda)
203 wallcycle_stop(wcycle, ewcPULLPOT);
206 static void pme_receive_force_ener(t_forcerec* fr,
208 gmx::ForceWithVirial* forceWithVirial,
209 gmx_enerdata_t* enerd,
210 bool useGpuPmePpComms,
211 bool receivePmeForceToGpu,
212 gmx_wallcycle_t wcycle)
214 real e_q, e_lj, dvdl_q, dvdl_lj;
215 float cycles_ppdpme, cycles_seppme;
217 cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
218 dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
220 /* In case of node-splitting, the PP nodes receive the long-range
221 * forces, virial and energy from the PME nodes here.
223 wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
226 gmx_pme_receive_f(fr->pmePpCommGpu.get(), cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
227 useGpuPmePpComms, receivePmeForceToGpu, &cycles_seppme);
228 enerd->term[F_COUL_RECIP] += e_q;
229 enerd->term[F_LJ_RECIP] += e_lj;
230 enerd->dvdl_lin[efptCOUL] += dvdl_q;
231 enerd->dvdl_lin[efptVDW] += dvdl_lj;
233 for (auto& dhdl : enerd->dhdlLambda)
235 dhdl += dvdl_q + dvdl_lj;
240 dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
242 wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
245 static void print_large_forces(FILE* fp,
253 real force2Tolerance = gmx::square(forceTolerance);
254 gmx::index numNonFinite = 0;
255 for (int i = 0; i < md->homenr; i++)
257 real force2 = norm2(f[i]);
258 bool nonFinite = !std::isfinite(force2);
259 if (force2 >= force2Tolerance || nonFinite)
261 fprintf(fp, "step %" PRId64 " atom %6d x %8.3f %8.3f %8.3f force %12.5e\n", step,
262 ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
269 if (numNonFinite > 0)
271 /* Note that with MPI this fatal call on one rank might interrupt
272 * the printing on other ranks. But we can only avoid that with
273 * an expensive MPI barrier that we would need at each step.
275 gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
279 static void post_process_forces(const t_commrec* cr,
282 gmx_wallcycle_t wcycle,
283 const gmx_localtop_t* top,
286 ForceOutputs* forceOutputs,
288 const t_mdatoms* mdatoms,
289 const t_forcerec* fr,
290 const gmx_vsite_t* vsite,
291 const StepWorkload& stepWork)
293 rvec* f = as_rvec_array(forceOutputs->forceWithShiftForces().force().data());
295 if (fr->haveDirectVirialContributions)
297 auto& forceWithVirial = forceOutputs->forceWithVirial();
298 rvec* fDirectVir = as_rvec_array(forceWithVirial.force_.data());
302 /* Spread the mesh force on virtual sites to the other particles...
303 * This is parallellized. MPI communication is performed
304 * if the constructing atoms aren't local.
306 matrix virial = { { 0 } };
307 spread_vsite_f(vsite, x, fDirectVir, nullptr, stepWork.computeVirial, virial, nrnb,
308 top->idef, fr->pbcType, fr->bMolPBC, box, cr, wcycle);
309 forceWithVirial.addVirialContribution(virial);
312 if (stepWork.computeVirial)
314 /* Now add the forces, this is local */
315 sum_forces(f, forceWithVirial.force_);
317 /* Add the direct virial contributions */
319 forceWithVirial.computeVirial_,
320 "forceWithVirial should request virial computation when we request the virial");
321 m_add(vir_force, forceWithVirial.getVirial(), vir_force);
325 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
330 if (fr->print_force >= 0)
332 print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
336 static void do_nb_verlet(t_forcerec* fr,
337 const interaction_const_t* ic,
338 gmx_enerdata_t* enerd,
339 const StepWorkload& stepWork,
340 const InteractionLocality ilocality,
344 gmx_wallcycle_t wcycle)
346 if (!stepWork.computeNonbondedForces)
348 /* skip non-bonded calculation */
352 nonbonded_verlet_t* nbv = fr->nbv.get();
354 /* GPU kernel launch overhead is already timed separately */
355 if (fr->cutoff_scheme != ecutsVERLET)
357 gmx_incons("Invalid cut-off scheme passed!");
362 /* When dynamic pair-list pruning is requested, we need to prune
363 * at nstlistPrune steps.
365 if (nbv->isDynamicPruningStepCpu(step))
367 /* Prune the pair-list beyond fr->ic->rlistPrune using
368 * the current coordinates of the atoms.
370 wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
371 nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
372 wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
376 nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
379 static inline void clear_rvecs_omp(int n, rvec v[])
381 int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
383 /* Note that we would like to avoid this conditional by putting it
384 * into the omp pragma instead, but then we still take the full
385 * omp parallel for overhead (at least with gcc5).
389 for (int i = 0; i < n; i++)
396 #pragma omp parallel for num_threads(nth) schedule(static)
397 for (int i = 0; i < n; i++)
404 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
406 * \param groupOptions Group options, containing T-coupling options
408 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
410 real nrdfCoupled = 0;
411 real nrdfUncoupled = 0;
412 real kineticEnergy = 0;
413 for (int g = 0; g < groupOptions.ngtc; g++)
415 if (groupOptions.tau_t[g] >= 0)
417 nrdfCoupled += groupOptions.nrdf[g];
418 kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * BOLTZ;
422 nrdfUncoupled += groupOptions.nrdf[g];
426 /* This conditional with > also catches nrdf=0 */
427 if (nrdfCoupled > nrdfUncoupled)
429 return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
437 /*! \brief This routine checks that the potential energy is finite.
439 * Always checks that the potential energy is finite. If step equals
440 * inputrec.init_step also checks that the magnitude of the potential energy
441 * is reasonable. Terminates with a fatal error when a check fails.
442 * Note that passing this check does not guarantee finite forces,
443 * since those use slightly different arithmetics. But in most cases
444 * there is just a narrow coordinate range where forces are not finite
445 * and energies are finite.
447 * \param[in] step The step number, used for checking and printing
448 * \param[in] enerd The energy data; the non-bonded group energies need to be added to
449 * enerd.term[F_EPOT] before calling this routine \param[in] inputrec The input record
451 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
453 /* Threshold valid for comparing absolute potential energy against
454 * the kinetic energy. Normally one should not consider absolute
455 * potential energy values, but with a factor of one million
456 * we should never get false positives.
458 constexpr real c_thresholdFactor = 1e6;
460 bool energyIsNotFinite = !std::isfinite(enerd.term[F_EPOT]);
461 real averageKineticEnergy = 0;
462 /* We only check for large potential energy at the initial step,
463 * because that is by far the most likely step for this too occur
464 * and because computing the average kinetic energy is not free.
465 * Note: nstcalcenergy >> 1 often does not allow to catch large energies
466 * before they become NaN.
468 if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
470 averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
473 if (energyIsNotFinite
474 || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
479 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
480 "contributions to the energy are %g and %g, respectively. A %s potential energy "
481 "can be caused by overlapping interactions in bonded interactions or very large%s "
482 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
483 "configuration, incorrect interactions or parameters in the topology.",
484 step, enerd.term[F_EPOT], energyIsNotFinite ? "not finite" : "extremely high",
485 enerd.term[F_LJ], enerd.term[F_COUL_SR],
486 energyIsNotFinite ? "non-finite" : "very high", energyIsNotFinite ? " or Nan" : "");
490 /*! \brief Return true if there are special forces computed this step.
492 * The conditionals exactly correspond to those in computeSpecialForces().
494 static bool haveSpecialForces(const t_inputrec& inputrec,
495 const gmx::ForceProviders& forceProviders,
496 const pull_t* pull_work,
497 const bool computeForces,
501 return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
502 (inputrec.bPull && pull_have_potential(pull_work)) || // pull
503 inputrec.bRot || // enforced rotation
504 (ed != nullptr) || // flooding
505 (inputrec.bIMD && computeForces)); // IMD
508 /*! \brief Compute forces and/or energies for special algorithms
510 * The intention is to collect all calls to algorithms that compute
511 * forces on local atoms only and that do not contribute to the local
512 * virial sum (but add their virial contribution separately).
513 * Eventually these should likely all become ForceProviders.
514 * Within this function the intention is to have algorithms that do
515 * global communication at the end, so global barriers within the MD loop
516 * are as close together as possible.
518 * \param[in] fplog The log file
519 * \param[in] cr The communication record
520 * \param[in] inputrec The input record
521 * \param[in] awh The Awh module (nullptr if none in use).
522 * \param[in] enforcedRotation Enforced rotation module.
523 * \param[in] imdSession The IMD session
524 * \param[in] pull_work The pull work structure.
525 * \param[in] step The current MD step
526 * \param[in] t The current time
527 * \param[in,out] wcycle Wallcycle accounting struct
528 * \param[in,out] forceProviders Pointer to a list of force providers
529 * \param[in] box The unit cell
530 * \param[in] x The coordinates
531 * \param[in] mdatoms Per atom properties
532 * \param[in] lambda Array of free-energy lambda values
533 * \param[in] stepWork Step schedule flags
534 * \param[in,out] forceWithVirial Force and virial buffers
535 * \param[in,out] enerd Energy buffer
536 * \param[in,out] ed Essential dynamics pointer
537 * \param[in] didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
539 * \todo Remove didNeighborSearch, which is used incorrectly.
540 * \todo Convert all other algorithms called here to ForceProviders.
542 static void computeSpecialForces(FILE* fplog,
544 const t_inputrec* inputrec,
546 gmx_enfrot* enforcedRotation,
547 gmx::ImdSession* imdSession,
551 gmx_wallcycle_t wcycle,
552 gmx::ForceProviders* forceProviders,
554 gmx::ArrayRef<const gmx::RVec> x,
555 const t_mdatoms* mdatoms,
557 const StepWorkload& stepWork,
558 gmx::ForceWithVirial* forceWithVirial,
559 gmx_enerdata_t* enerd,
561 bool didNeighborSearch)
563 /* NOTE: Currently all ForceProviders only provide forces.
564 * When they also provide energies, remove this conditional.
566 if (stepWork.computeForces)
568 gmx::ForceProviderInput forceProviderInput(x, *mdatoms, t, box, *cr);
569 gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
571 /* Collect forces from modules */
572 forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
575 if (inputrec->bPull && pull_have_potential(pull_work))
577 pull_potential_wrapper(cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work,
582 enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
583 inputrec->pbcType, *mdatoms, box, forceWithVirial, t, step, wcycle, fplog);
587 rvec* f = as_rvec_array(forceWithVirial->force_.data());
589 /* Add the forces from enforced rotation potentials (if any) */
592 wallcycle_start(wcycle, ewcROTadd);
593 enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
594 wallcycle_stop(wcycle, ewcROTadd);
599 /* Note that since init_edsam() is called after the initialization
600 * of forcerec, edsam doesn't request the noVirSum force buffer.
601 * Thus if no other algorithm (e.g. PME) requires it, the forces
602 * here will contribute to the virial.
604 do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
607 /* Add forces from interactive molecular dynamics (IMD), if any */
608 if (inputrec->bIMD && stepWork.computeForces)
610 imdSession->applyForces(f);
614 /*! \brief Launch the prepare_step and spread stages of PME GPU.
616 * \param[in] pmedata The PME structure
617 * \param[in] box The box matrix
618 * \param[in] stepWork Step schedule flags
619 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in
620 * the device memory. \param[in] wcycle The wallcycle structure
622 static inline void launchPmeGpuSpread(gmx_pme_t* pmedata,
624 const StepWorkload& stepWork,
625 GpuEventSynchronizer* xReadyOnDevice,
626 gmx_wallcycle_t wcycle)
628 pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
629 pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle);
632 /*! \brief Launch the FFT and gather stages of PME GPU
634 * This function only implements setting the output forces (no accumulation).
636 * \param[in] pmedata The PME structure
637 * \param[in] wcycle The wallcycle structure
638 * \param[in] stepWork Step schedule flags
640 static void launchPmeGpuFftAndGather(gmx_pme_t* pmedata, gmx_wallcycle_t wcycle, const gmx::StepWorkload& stepWork)
642 pme_gpu_launch_complex_transforms(pmedata, wcycle, stepWork);
643 pme_gpu_launch_gather(pmedata, wcycle);
647 * Polling wait for either of the PME or nonbonded GPU tasks.
649 * Instead of a static order in waiting for GPU tasks, this function
650 * polls checking which of the two tasks completes first, and does the
651 * associated force buffer reduction overlapped with the other task.
652 * By doing that, unlike static scheduling order, it can always overlap
653 * one of the reductions, regardless of the GPU task completion order.
655 * \param[in] nbv Nonbonded verlet structure
656 * \param[in,out] pmedata PME module data
657 * \param[in,out] forceOutputs Output buffer for the forces and virial
658 * \param[in,out] enerd Energy data structure results are reduced into
659 * \param[in] stepWork Step schedule flags
660 * \param[in] wcycle The wallcycle structure
662 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
664 gmx::ForceOutputs* forceOutputs,
665 gmx_enerdata_t* enerd,
666 const StepWorkload& stepWork,
667 gmx_wallcycle_t wcycle)
669 bool isPmeGpuDone = false;
670 bool isNbGpuDone = false;
673 gmx::ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
674 gmx::ForceWithVirial& forceWithVirial = forceOutputs->forceWithVirial();
676 gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
678 while (!isPmeGpuDone || !isNbGpuDone)
682 GpuTaskCompletion completionType =
683 (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
684 isPmeGpuDone = pme_gpu_try_finish_task(pmedata, stepWork, wcycle, &forceWithVirial,
685 enerd, completionType);
690 GpuTaskCompletion completionType =
691 (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
692 isNbGpuDone = Nbnxm::gpu_try_finish_task(
693 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
694 enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(),
695 completionType, wcycle);
699 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShiftForces.force());
705 /*! \brief Set up the different force buffers; also does clearing.
707 * \param[in] fr force record pointer
708 * \param[in] pull_work The pull work object.
709 * \param[in] inputrec input record
710 * \param[in] force force array
711 * \param[in] stepWork Step schedule flags
712 * \param[out] wcycle wallcycle recording structure
714 * \returns Cleared force output structure
716 static ForceOutputs setupForceOutputs(t_forcerec* fr,
718 const t_inputrec& inputrec,
719 gmx::ArrayRefWithPadding<gmx::RVec> force,
720 const StepWorkload& stepWork,
721 gmx_wallcycle_t wcycle)
723 wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
725 /* NOTE: We assume fr->shiftForces is all zeros here */
726 gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial, fr->shiftForces);
728 if (stepWork.computeForces)
730 /* Clear the short- and long-range forces */
731 clear_rvecs_omp(fr->natoms_force_constr, as_rvec_array(forceWithShiftForces.force().data()));
734 /* If we need to compute the virial, we might need a separate
735 * force buffer for algorithms for which the virial is calculated
736 * directly, such as PME. Otherwise, forceWithVirial uses the
737 * the same force (f in legacy calls) buffer as other algorithms.
739 const bool useSeparateForceWithVirialBuffer =
740 (stepWork.computeForces && (stepWork.computeVirial && fr->haveDirectVirialContributions));
741 /* forceWithVirial uses the local atom range only */
742 gmx::ForceWithVirial forceWithVirial(useSeparateForceWithVirialBuffer ? fr->forceBufferForDirectVirialContributions
743 : force.unpaddedArrayRef(),
744 stepWork.computeVirial);
746 if (useSeparateForceWithVirialBuffer)
748 /* TODO: update comment
749 * We only compute forces on local atoms. Note that vsites can
750 * spread to non-local atoms, but that part of the buffer is
751 * cleared separately in the vsite spreading code.
753 clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
756 if (inputrec.bPull && pull_have_constraint(pull_work))
758 clear_pull_forces(pull_work);
761 wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
763 return ForceOutputs(forceWithShiftForces, forceWithVirial);
767 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
769 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec& inputrec,
770 const t_forcerec& fr,
771 const pull_t* pull_work,
773 const InteractionDefinitions& idef,
775 const t_mdatoms& mdatoms,
776 const SimulationWorkload& simulationWork,
777 const StepWorkload& stepWork)
779 DomainLifetimeWorkload domainWork;
780 // Note that haveSpecialForces is constant over the whole run
781 domainWork.haveSpecialForces =
782 haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
783 domainWork.haveCpuBondedWork = haveCpuBondeds(fr);
784 domainWork.haveGpuBondedWork = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
785 domainWork.haveRestraintsWork = haveRestraints(idef, fcd);
786 domainWork.haveCpuListedForceWork = haveCpuListedForces(fr, idef, fcd);
787 // Note that haveFreeEnergyWork is constant over the whole run
788 domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
789 // We assume we have local force work if there are CPU
790 // force tasks including PME or nonbondeds.
791 domainWork.haveCpuLocalForceWork =
792 domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
793 || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
794 || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
799 /*! \brief Set up force flag stuct from the force bitmask.
801 * \param[in] legacyFlags Force bitmask flags used to construct the new flags
802 * \param[in] isNonbondedOn Global override, if false forces to turn off all nonbonded calculation.
803 * \param[in] simulationWork Simulation workload description.
804 * \param[in] rankHasPmeDuty If this rank computes PME.
806 * \returns New Stepworkload description.
808 static StepWorkload setupStepWorkload(const int legacyFlags,
809 const bool isNonbondedOn,
810 const SimulationWorkload& simulationWork,
811 const bool rankHasPmeDuty)
814 flags.stateChanged = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
815 flags.haveDynamicBox = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
816 flags.doNeighborSearch = ((legacyFlags & GMX_FORCE_NS) != 0);
817 flags.computeVirial = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
818 flags.computeEnergy = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
819 flags.computeForces = ((legacyFlags & GMX_FORCE_FORCES) != 0);
820 flags.computeListedForces = ((legacyFlags & GMX_FORCE_LISTED) != 0);
821 flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
822 flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
824 if (simulationWork.useGpuBufferOps)
826 GMX_ASSERT(simulationWork.useGpuNonbonded,
827 "Can only offload buffer ops if nonbonded computation is also offloaded");
829 flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
830 // on virial steps the CPU reduction path is taken
831 flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
832 flags.useGpuPmeFReduction = flags.useGpuFBufferOps
833 && (simulationWork.useGpuPme
834 && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication));
840 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
842 * TODO: eliminate \p useGpuPmeOnThisRank when this is
843 * incorporated in DomainLifetimeWorkload.
845 static void launchGpuEndOfStepTasks(nonbonded_verlet_t* nbv,
846 gmx::GpuBonded* gpuBonded,
848 gmx_enerdata_t* enerd,
849 const gmx::MdrunScheduleWorkload& runScheduleWork,
850 bool useGpuPmeOnThisRank,
852 gmx_wallcycle_t wcycle)
854 if (runScheduleWork.simulationWork.useGpuNonbonded)
856 /* Launch pruning before buffer clearing because the API overhead of the
857 * clear kernel launches can leave the GPU idle while it could be running
860 if (nbv->isDynamicPruningStepGpu(step))
862 nbv->dispatchPruneKernelGpu(step);
865 /* now clear the GPU outputs while we finish the step on the CPU */
866 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
867 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
868 Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
869 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
870 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
873 if (useGpuPmeOnThisRank)
875 pme_gpu_reinit_computation(pmedata, wcycle);
878 if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
880 // in principle this should be included in the DD balancing region,
881 // but generally it is infrequent so we'll omit it for the sake of
883 gpuBonded->waitAccumulateEnergyTerms(enerd);
885 gpuBonded->clearEnergies();
889 //! \brief Data structure to hold dipole-related data and staging arrays
892 //! Dipole staging for fast summing over MPI
893 gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
894 //! Dipole staging for states A and B (index 0 and 1 resp.)
895 gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
899 static void reduceAndUpdateMuTot(DipoleData* dipoleData,
901 const bool haveFreeEnergy,
902 gmx::ArrayRef<const real> lambda,
904 const DDBalanceRegionHandler& ddBalanceRegionHandler)
908 gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
909 ddBalanceRegionHandler.reopenRegionCpu();
911 for (int i = 0; i < 2; i++)
913 for (int j = 0; j < DIM; j++)
915 dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
921 copy_rvec(dipoleData->muStateAB[0], muTotal);
925 for (int j = 0; j < DIM; j++)
927 muTotal[j] = (1.0 - lambda[efptCOUL]) * dipoleData->muStateAB[0][j]
928 + lambda[efptCOUL] * dipoleData->muStateAB[1][j];
933 void do_force(FILE* fplog,
935 const gmx_multisim_t* ms,
936 const t_inputrec* inputrec,
938 gmx_enfrot* enforcedRotation,
939 gmx::ImdSession* imdSession,
943 gmx_wallcycle_t wcycle,
944 const gmx_localtop_t* top,
946 gmx::ArrayRefWithPadding<gmx::RVec> x,
948 gmx::ArrayRefWithPadding<gmx::RVec> force,
950 const t_mdatoms* mdatoms,
951 gmx_enerdata_t* enerd,
953 gmx::ArrayRef<real> lambda,
955 gmx::MdrunScheduleWorkload* runScheduleWork,
956 const gmx_vsite_t* vsite,
961 const DDBalanceRegionHandler& ddBalanceRegionHandler)
963 GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
964 "The size of the force buffer should be at least the number of atoms to compute "
967 nonbonded_verlet_t* nbv = fr->nbv.get();
968 interaction_const_t* ic = fr->ic;
969 gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
971 const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
974 runScheduleWork->stepWork = setupStepWorkload(legacyFlags, fr->bNonbonded, simulationWork,
975 thisRankHasDuty(cr, DUTY_PME));
976 const StepWorkload& stepWork = runScheduleWork->stepWork;
979 const bool useGpuPmeOnThisRank = simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME);
981 /* At a search step we need to start the first balancing region
982 * somewhere early inside the step after communication during domain
983 * decomposition (and not during the previous step as usual).
985 if (stepWork.doNeighborSearch)
987 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
990 clear_mat(vir_force);
992 if (fr->pbcType != PbcType::No)
994 /* Compute shift vectors every step,
995 * because of pressure coupling or box deformation!
997 if (stepWork.haveDynamicBox && stepWork.stateChanged)
999 calc_shifts(box, fr->shift_vec);
1002 const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1003 const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1006 put_atoms_in_box_omp(fr->pbcType, box, x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1007 gmx_omp_nthreads_get(emntDefault));
1008 inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1012 nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1015 const bool pmeSendCoordinatesFromGpu =
1016 simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1017 const bool reinitGpuPmePpComms =
1018 simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1021 const auto localXReadyOnDevice = (stateGpu != nullptr)
1022 ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1023 AtomLocality::Local, simulationWork, stepWork)
1027 // If coordinates are to be sent to PME task from CPU memory, perform that send here.
1028 // Otherwise the send will occur after H2D coordinate transfer.
1029 if (!thisRankHasDuty(cr, DUTY_PME) && !pmeSendCoordinatesFromGpu)
1031 /* Send particle coordinates to the pme nodes */
1032 if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1034 GMX_RELEASE_ASSERT(false,
1035 "GPU update and separate PME ranks are only supported with GPU "
1036 "direct communication!");
1037 // TODO: when this code-path becomes supported add:
1038 // stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1041 gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1042 lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1043 step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1044 pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1046 #endif /* GMX_MPI */
1048 // Coordinates on the device are needed if PME or BufferOps are offloaded.
1049 // The local coordinates can be copied right away.
1050 // NOTE: Consider moving this copy to right after they are updated and constrained,
1051 // if the later is not offloaded.
1052 if (useGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1054 if (stepWork.doNeighborSearch)
1056 // TODO refactor this to do_md, after partitioning.
1057 stateGpu->reinit(mdatoms->homenr,
1058 cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1059 if (useGpuPmeOnThisRank)
1061 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1062 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1065 // We need to copy coordinates when:
1066 // 1. Update is not offloaded
1067 // 2. The buffers were reinitialized on search step
1068 if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1070 GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1071 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1075 // TODO Update this comment when introducing SimulationWorkload
1077 // The conditions for gpuHaloExchange e.g. using GPU buffer
1078 // operations were checked before construction, so here we can
1079 // just use it and assert upon any conditions.
1080 const bool ddUsesGpuDirectCommunication =
1081 ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange.empty()));
1082 GMX_ASSERT(!ddUsesGpuDirectCommunication || stepWork.useGpuXBufferOps,
1083 "Must use coordinate buffer ops with GPU halo exchange");
1084 const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && stepWork.useGpuFBufferOps;
1086 // Copy coordinate from the GPU if update is on the GPU and there
1087 // are forces to be computed on the CPU, or for the computation of
1088 // virial, or if host-side data will be transferred from this task
1089 // to a remote task for halo exchange or PME-PP communication. At
1090 // search steps the current coordinates are already on the host,
1091 // hence copy is not needed.
1092 const bool haveHostPmePpComms =
1093 !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1094 const bool haveHostHaloExchangeComms = havePPDomainDecomposition(cr) && !ddUsesGpuDirectCommunication;
1096 bool gmx_used_in_debug haveCopiedXFromGpu = false;
1097 if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1098 && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1099 || haveHostPmePpComms || haveHostHaloExchangeComms))
1101 GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1102 stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1103 haveCopiedXFromGpu = true;
1107 // If coordinates are to be sent to PME task from GPU memory, perform that send here.
1108 // Otherwise the send will occur before the H2D coordinate transfer.
1109 if (pmeSendCoordinatesFromGpu)
1111 /* Send particle coordinates to the pme nodes */
1112 gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1113 lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1114 step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1115 pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1117 #endif /* GMX_MPI */
1119 if (useGpuPmeOnThisRank)
1121 launchPmeGpuSpread(fr->pmedata, box, stepWork, localXReadyOnDevice, wcycle);
1124 /* do gridding for pair search */
1125 if (stepWork.doNeighborSearch)
1127 if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1129 fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1133 // - vzero is constant, do we need to pass it?
1134 // - box_diag should be passed directly to nbnxn_put_on_grid
1140 box_diag[XX] = box[XX][XX];
1141 box_diag[YY] = box[YY][YY];
1142 box_diag[ZZ] = box[ZZ][ZZ];
1144 wallcycle_start(wcycle, ewcNS);
1145 if (!DOMAINDECOMP(cr))
1147 wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1148 nbnxn_put_on_grid(nbv, box, 0, vzero, box_diag, nullptr, { 0, mdatoms->homenr }, -1,
1149 fr->cginfo, x.unpaddedArrayRef(), 0, nullptr);
1150 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1154 wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1155 nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1156 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1159 nbv->setAtomProperties(*mdatoms, fr->cginfo);
1161 wallcycle_stop(wcycle, ewcNS);
1163 /* initialize the GPU nbnxm atom data and bonded data structures */
1164 if (simulationWork.useGpuNonbonded)
1166 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1168 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1169 Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1170 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1174 /* Now we put all atoms on the grid, we can assign bonded
1175 * interactions to the GPU, where the grid order is
1176 * needed. Also the xq, f and fshift device buffers have
1177 * been reallocated if needed, so the bonded code can
1178 * learn about them. */
1179 // TODO the xq, f, and fshift buffers are now shared
1180 // resources, so they should be maintained by a
1181 // higher-level object than the nb module.
1182 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(
1183 nbv->getGridIndices(), top->idef, Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1184 Nbnxm::gpu_get_f(nbv->gpu_nbv), Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1186 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1189 // Need to run after the GPU-offload bonded interaction lists
1190 // are set up to be able to determine whether there is bonded work.
1191 runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1192 *inputrec, *fr, pull_work, ed, top->idef, *fcd, *mdatoms, simulationWork, stepWork);
1194 wallcycle_start_nocount(wcycle, ewcNS);
1195 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1196 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1197 nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1199 nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1201 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1202 wallcycle_stop(wcycle, ewcNS);
1204 if (stepWork.useGpuXBufferOps)
1206 nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1208 // For force buffer ops, we use the below conditon rather than
1209 // useGpuFBufferOps to ensure that init is performed even if this
1210 // NS step is also a virial step (on which f buf ops are deactivated).
1211 if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1213 GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1214 nbv->atomdata_init_add_nbat_f_to_f_gpu(stateGpu->fReducedOnDevice());
1217 else if (!EI_TPI(inputrec->eI))
1219 if (stepWork.useGpuXBufferOps)
1221 GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1222 nbv->convertCoordinatesGpu(AtomLocality::Local, false, stateGpu->getCoordinates(),
1223 localXReadyOnDevice);
1227 if (simulationWork.useGpuUpdate)
1229 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1230 GMX_ASSERT(haveCopiedXFromGpu,
1231 "a wait should only be triggered if copy has been scheduled");
1232 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1234 nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1238 const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1240 if (simulationWork.useGpuNonbonded)
1242 ddBalanceRegionHandler.openBeforeForceComputationGpu();
1244 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1246 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1247 Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1248 if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1250 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1252 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1253 // with X buffer ops offloaded to the GPU on all but the search steps
1255 // bonded work not split into separate local and non-local, so with DD
1256 // we can only launch the kernel after non-local coordinates have been received.
1257 if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1259 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1260 fr->gpuBonded->launchKernel(fr, stepWork, box);
1261 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1264 /* launch local nonbonded work on GPU */
1265 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1266 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1267 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1268 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1271 if (useGpuPmeOnThisRank)
1273 // In PME GPU and mixed mode we launch FFT / gather after the
1274 // X copy/transform to allow overlap as well as after the GPU NB
1275 // launch to avoid FFT launch overhead hijacking the CPU and delaying
1276 // the nonbonded kernel.
1277 launchPmeGpuFftAndGather(fr->pmedata, wcycle, stepWork);
1280 /* Communicate coordinates and sum dipole if necessary +
1281 do non-local pair search */
1282 if (havePPDomainDecomposition(cr))
1284 if (stepWork.doNeighborSearch)
1286 // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1287 wallcycle_start_nocount(wcycle, ewcNS);
1288 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1289 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1290 nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1292 nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1293 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1294 wallcycle_stop(wcycle, ewcNS);
1295 // TODO refactor this GPU halo exchange re-initialisation
1296 // to location in do_md where GPU halo exchange is
1297 // constructed at partitioning, after above stateGpu
1298 // re-initialization has similarly been refactored
1299 if (ddUsesGpuDirectCommunication)
1301 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1306 if (ddUsesGpuDirectCommunication)
1308 // The following must be called after local setCoordinates (which records an event
1309 // when the coordinate data has been copied to the device).
1310 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1312 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1314 // non-local part of coordinate buffer must be copied back to host for CPU work
1315 stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1320 // Note: GPU update + DD without direct communication is not supported,
1321 // a waitCoordinatesReadyOnHost() should be issued if it will be.
1322 GMX_ASSERT(!simulationWork.useGpuUpdate,
1323 "GPU update is not supported with CPU halo exchange");
1324 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1327 if (stepWork.useGpuXBufferOps)
1329 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1331 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1333 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false, stateGpu->getCoordinates(),
1334 stateGpu->getCoordinatesReadyOnDeviceEvent(
1335 AtomLocality::NonLocal, simulationWork, stepWork));
1339 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1343 if (simulationWork.useGpuNonbonded)
1345 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1347 if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1349 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1350 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1351 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1354 if (domainWork.haveGpuBondedWork)
1356 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1357 fr->gpuBonded->launchKernel(fr, stepWork, box);
1358 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1361 /* launch non-local nonbonded tasks on GPU */
1362 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1363 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1365 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1367 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1371 if (simulationWork.useGpuNonbonded)
1373 /* launch D2H copy-back F */
1374 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1375 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1377 if (havePPDomainDecomposition(cr))
1379 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1381 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1382 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1384 if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1386 fr->gpuBonded->launchEnergyTransfer();
1388 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1391 gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1392 if (fr->wholeMoleculeTransform)
1394 xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1397 DipoleData dipoleData;
1399 if (simulationWork.computeMuTot)
1401 const int start = 0;
1403 /* Calculate total (local) dipole moment in a temporary common array.
1404 * This makes it possible to sum them over nodes faster.
1406 gmx::ArrayRef<const gmx::RVec> xRef =
1407 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1408 calc_mu(start, mdatoms->homenr, xRef, mdatoms->chargeA, mdatoms->chargeB,
1409 mdatoms->nChargePerturbed, dipoleData.muStaging[0], dipoleData.muStaging[1]);
1411 reduceAndUpdateMuTot(&dipoleData, cr, (fr->efep != efepNO), lambda, muTotal, ddBalanceRegionHandler);
1414 /* Reset energies */
1415 reset_enerdata(enerd);
1416 /* Clear the shift forces */
1417 // TODO: This should be linked to the shift force buffer in use, or cleared before use instead
1418 for (gmx::RVec& elem : fr->shiftForces)
1420 elem = { 0.0_real, 0.0_real, 0.0_real };
1423 if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1425 wallcycle_start(wcycle, ewcPPDURINGPME);
1426 dd_force_flop_start(cr->dd, nrnb);
1429 // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1430 // this wait ensures that the D2H transfer is complete.
1431 if ((simulationWork.useGpuUpdate)
1432 && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1434 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1439 wallcycle_start(wcycle, ewcROT);
1440 do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step,
1441 stepWork.doNeighborSearch);
1442 wallcycle_stop(wcycle, ewcROT);
1445 /* Start the force cycle counter.
1446 * Note that a different counter is used for dynamic load balancing.
1448 wallcycle_start(wcycle, ewcFORCE);
1450 // Set up and clear force outputs.
1451 // We use std::move to keep the compiler happy, it has no effect.
1452 ForceOutputs forceOut =
1453 setupForceOutputs(fr, pull_work, *inputrec, std::move(force), stepWork, wcycle);
1455 /* We calculate the non-bonded forces, when done on the CPU, here.
1456 * We do this before calling do_force_lowlevel, because in that
1457 * function, the listed forces are calculated before PME, which
1458 * does communication. With this order, non-bonded and listed
1459 * force calculation imbalance can be balanced out by the domain
1460 * decomposition load balancing.
1463 const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1465 if (!useOrEmulateGpuNb)
1467 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1470 if (fr->efep != efepNO)
1472 /* Calculate the local and non-local free energy interactions here.
1473 * Happens here on the CPU both with and without GPU.
1475 nbv->dispatchFreeEnergyKernel(InteractionLocality::Local, fr,
1476 as_rvec_array(x.unpaddedArrayRef().data()),
1477 &forceOut.forceWithShiftForces(), *mdatoms, inputrec->fepvals,
1478 lambda.data(), enerd, stepWork, nrnb);
1480 if (havePPDomainDecomposition(cr))
1482 nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal, fr,
1483 as_rvec_array(x.unpaddedArrayRef().data()),
1484 &forceOut.forceWithShiftForces(), *mdatoms,
1485 inputrec->fepvals, lambda.data(), enerd, stepWork, nrnb);
1489 if (!useOrEmulateGpuNb)
1491 if (havePPDomainDecomposition(cr))
1493 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1497 if (stepWork.computeForces)
1499 /* Add all the non-bonded force to the normal force array.
1500 * This can be split into a local and a non-local part when overlapping
1501 * communication with calculation with domain decomposition.
1503 wallcycle_stop(wcycle, ewcFORCE);
1504 nbv->atomdata_add_nbat_f_to_f(AtomLocality::All, forceOut.forceWithShiftForces().force());
1505 wallcycle_start_nocount(wcycle, ewcFORCE);
1508 /* If there are multiple fshift output buffers we need to reduce them */
1509 if (stepWork.computeVirial)
1511 /* This is not in a subcounter because it takes a
1512 negligible and constant-sized amount of time */
1513 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1514 forceOut.forceWithShiftForces().shiftForces());
1518 // TODO Force flags should include haveFreeEnergyWork for this domain
1519 if (ddUsesGpuDirectCommunication && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1521 /* Wait for non-local coordinate data to be copied from device */
1522 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1524 /* Compute the bonded and non-bonded energies and optionally forces */
1525 do_force_lowlevel(fr, inputrec, top->idef, cr, ms, nrnb, wcycle, mdatoms, x, xWholeMolecules,
1526 hist, &forceOut, enerd, fcd, box, lambda.data(),
1527 as_rvec_array(dipoleData.muStateAB), stepWork, ddBalanceRegionHandler);
1529 wallcycle_stop(wcycle, ewcFORCE);
1531 computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation, imdSession, pull_work, step, t,
1532 wcycle, fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1533 stepWork, &forceOut.forceWithVirial(), enerd, ed, stepWork.doNeighborSearch);
1536 // Will store the amount of cycles spent waiting for the GPU that
1537 // will be later used in the DLB accounting.
1538 float cycles_wait_gpu = 0;
1539 if (useOrEmulateGpuNb)
1541 auto& forceWithShiftForces = forceOut.forceWithShiftForces();
1543 /* wait for non-local forces (or calculate in emulation mode) */
1544 if (havePPDomainDecomposition(cr))
1546 if (simulationWork.useGpuNonbonded)
1548 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
1549 nbv->gpu_nbv, stepWork, AtomLocality::NonLocal, enerd->grpp.ener[egLJSR].data(),
1550 enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(), wcycle);
1554 wallcycle_start_nocount(wcycle, ewcFORCE);
1555 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1556 step, nrnb, wcycle);
1557 wallcycle_stop(wcycle, ewcFORCE);
1560 if (stepWork.useGpuFBufferOps)
1562 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1564 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1565 // condition The bonded and free energy CPU tasks can have non-local force
1566 // contributions which are a dependency for the GPU force reduction.
1567 bool haveNonLocalForceContribInCpuBuffer =
1568 domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1570 if (haveNonLocalForceContribInCpuBuffer)
1572 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(),
1573 AtomLocality::NonLocal);
1574 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
1575 AtomLocality::NonLocal, stepWork.useGpuFBufferOps));
1578 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::NonLocal, stateGpu->getForces(),
1579 pme_gpu_get_device_f(fr->pmedata), dependencyList,
1580 false, haveNonLocalForceContribInCpuBuffer);
1581 if (!useGpuForcesHaloExchange)
1583 // copy from GPU input for dd_move_f()
1584 stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(),
1585 AtomLocality::NonLocal);
1590 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1594 if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1596 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1601 if (havePPDomainDecomposition(cr))
1603 /* We are done with the CPU compute.
1604 * We will now communicate the non-local forces.
1605 * If we use a GPU this will overlap with GPU work, so in that case
1606 * we do not close the DD force balancing region here.
1608 ddBalanceRegionHandler.closeAfterForceComputationCpu();
1610 if (stepWork.computeForces)
1613 if (useGpuForcesHaloExchange)
1615 if (domainWork.haveCpuLocalForceWork)
1617 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::Local);
1619 communicateGpuHaloForces(*cr, domainWork.haveCpuLocalForceWork);
1623 if (stepWork.useGpuFBufferOps)
1625 stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1627 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1632 // With both nonbonded and PME offloaded a GPU on the same rank, we use
1633 // an alternating wait/reduction scheme.
1634 bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
1635 && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
1636 if (alternateGpuWait)
1638 alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd, stepWork, wcycle);
1641 if (!alternateGpuWait && useGpuPmeOnThisRank)
1643 pme_gpu_wait_and_reduce(fr->pmedata, stepWork, wcycle, &forceOut.forceWithVirial(), enerd);
1646 /* Wait for local GPU NB outputs on the non-alternating wait path */
1647 if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1649 /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1650 * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1651 * but even with a step of 0.1 ms the difference is less than 1%
1654 const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1655 const float waitCycles = Nbnxm::gpu_wait_finish_task(
1656 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
1657 enerd->grpp.ener[egCOULSR].data(), forceOut.forceWithShiftForces().shiftForces(), wcycle);
1659 if (ddBalanceRegionHandler.useBalancingRegion())
1661 DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1662 if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1664 /* We measured few cycles, it could be that the kernel
1665 * and transfer finished earlier and there was no actual
1666 * wait time, only API call overhead.
1667 * Then the actual time could be anywhere between 0 and
1668 * cycles_wait_est. We will use half of cycles_wait_est.
1670 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1672 ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1676 if (fr->nbv->emulateGpu())
1678 // NOTE: emulation kernel is not included in the balancing region,
1679 // but emulation mode does not target performance anyway
1680 wallcycle_start_nocount(wcycle, ewcFORCE);
1681 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1682 DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes, step, nrnb, wcycle);
1683 wallcycle_stop(wcycle, ewcFORCE);
1686 // If on GPU PME-PP comms or GPU update path, receive forces from PME before GPU buffer ops
1687 // TODO refactor this and unify with below default-path call to the same function
1688 if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME)
1689 && (simulationWork.useGpuPmePpCommunication || simulationWork.useGpuUpdate))
1691 /* In case of node-splitting, the PP nodes receive the long-range
1692 * forces, virial and energy from the PME nodes here.
1694 pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1695 simulationWork.useGpuPmePpCommunication,
1696 stepWork.useGpuPmeFReduction, wcycle);
1700 /* Do the nonbonded GPU (or emulation) force buffer reduction
1701 * on the non-alternating path. */
1702 if (useOrEmulateGpuNb && !alternateGpuWait)
1704 // TODO simplify the below conditionals. Pass buffer and sync pointers at init stage rather than here. Unify getter fns for sameGPU/otherGPU cases.
1706 stepWork.useGpuPmeFReduction
1707 ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1708 : // PME force buffer on same GPU
1709 fr->pmePpCommGpu->getGpuForceStagingPtr()) // buffer received from other GPU
1710 : nullptr; // PME reduction not active on GPU
1712 GpuEventSynchronizer* const pmeSynchronizer =
1713 stepWork.useGpuPmeFReduction
1714 ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1715 : // PME force buffer on same GPU
1716 static_cast<GpuEventSynchronizer*>(
1717 fr->pmePpCommGpu->getForcesReadySynchronizer())) // buffer received from other GPU
1718 : nullptr; // PME reduction not active on GPU
1720 gmx::FixedCapacityVector<GpuEventSynchronizer*, 3> dependencyList;
1722 if (stepWork.useGpuPmeFReduction)
1724 dependencyList.push_back(pmeSynchronizer);
1727 gmx::ArrayRef<gmx::RVec> forceWithShift = forceOut.forceWithShiftForces().force();
1729 if (stepWork.useGpuFBufferOps)
1731 // Flag to specify whether the CPU force buffer has contributions to
1732 // local atoms. This depends on whether there are CPU-based force tasks
1733 // or when DD is active the halo exchange has resulted in contributions
1734 // from the non-local part.
1735 const bool haveLocalForceContribInCpuBuffer =
1736 (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1738 // TODO: move these steps as early as possible:
1739 // - CPU f H2D should be as soon as all CPU-side forces are done
1740 // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1741 // before the next CPU task that consumes the forces: vsite spread or update)
1742 // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1743 // of the halo exchange. In that case the copy is instead performed above, before the exchange.
1744 // These should be unified.
1745 if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1747 // Note: AtomLocality::All is used for the non-DD case because, as in this
1748 // case copyForcesToGpu() uses a separate stream, it allows overlap of
1749 // CPU force H2D with GPU force tasks on all streams including those in the
1750 // local stream which would otherwise be implicit dependencies for the
1751 // transfer and would not overlap.
1752 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
1754 stateGpu->copyForcesToGpu(forceWithShift, locality);
1755 dependencyList.push_back(
1756 stateGpu->getForcesReadyOnDeviceEvent(locality, stepWork.useGpuFBufferOps));
1758 if (useGpuForcesHaloExchange)
1760 dependencyList.push_back(cr->dd->gpuHaloExchange[0]->getForcesReadyOnDeviceEvent());
1762 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::Local, stateGpu->getForces(), pmeForcePtr,
1763 dependencyList, stepWork.useGpuPmeFReduction,
1764 haveLocalForceContribInCpuBuffer);
1765 // Copy forces to host if they are needed for update or if virtual sites are enabled.
1766 // If there are vsites, we need to copy forces every step to spread vsite forces on host.
1767 // TODO: When the output flags will be included in step workload, this copy can be combined with the
1768 // copy call done in sim_utils(...) for the output.
1769 // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
1770 // they should not be copied in do_md(...) for the output.
1771 if (!simulationWork.useGpuUpdate || vsite)
1773 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
1774 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1779 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
1783 launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork,
1784 useGpuPmeOnThisRank, step, wcycle);
1786 if (DOMAINDECOMP(cr))
1788 dd_force_flop_stop(cr->dd, nrnb);
1791 if (stepWork.computeForces)
1793 rvec* f = as_rvec_array(forceOut.forceWithShiftForces().force().data());
1795 /* If we have NoVirSum forces, but we do not calculate the virial,
1796 * we sum fr->f_novirsum=forceOut.f later.
1798 if (vsite && !(fr->haveDirectVirialContributions && !stepWork.computeVirial))
1800 rvec* fshift = as_rvec_array(forceOut.forceWithShiftForces().shiftForces().data());
1801 spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fshift, FALSE,
1802 nullptr, nrnb, top->idef, fr->pbcType, fr->bMolPBC, box, cr, wcycle);
1805 if (stepWork.computeVirial)
1807 /* Calculation of the virial must be done after vsites! */
1808 calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()),
1809 forceOut.forceWithShiftForces(), vir_force, box, nrnb, fr, inputrec->pbcType);
1813 // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
1814 if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
1815 && !simulationWork.useGpuUpdate)
1817 /* In case of node-splitting, the PP nodes receive the long-range
1818 * forces, virial and energy from the PME nodes here.
1820 pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1821 simulationWork.useGpuPmePpCommunication, false, wcycle);
1824 if (stepWork.computeForces)
1826 post_process_forces(cr, step, nrnb, wcycle, top, box, as_rvec_array(x.unpaddedArrayRef().data()),
1827 &forceOut, vir_force, mdatoms, fr, vsite, stepWork);
1830 if (stepWork.computeEnergy)
1832 /* Sum the potential energy terms from group contributions */
1833 sum_epot(&(enerd->grpp), enerd->term);
1835 if (!EI_TPI(inputrec->eI))
1837 checkPotentialEnergyValidity(step, *enerd, *inputrec);
1841 /* In case we don't have constraints and are using GPUs, the next balancing
1842 * region starts here.
1843 * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1844 * virial calculation and COM pulling, is not thus not included in
1845 * the balance timing, which is ok as most tasks do communication.
1847 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);