src/gromacs/mdlib/force.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017 by the GROMACS development team.
   7  * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
   8  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   9  * and including many others, as listed in the AUTHORS file in the
  10  * top-level source directory and at http://www.gromacs.org.
  11  *
  12  * GROMACS is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public License
  14  * as published by the Free Software Foundation; either version 2.1
  15  * of the License, or (at your option) any later version.
  16  *
  17  * GROMACS is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with GROMACS; if not, see
  24  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  25  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  26  *
  27  * If you want to redistribute modifications to GROMACS, please
  28  * consider that scientific software is very special. Version
  29  * control is crucial - bugs must be traceable. We will be happy to
  30  * consider code for inclusion in the official distribution, but
  31  * derived work must not be called official GROMACS. Details are found
  32  * in the README & COPYING files - if they are missing, get the
  33  * official version at http://www.gromacs.org.
  34  *
  35  * To help us fund GROMACS development, we humbly ask that you cite
  36  * the research papers on the package. Check out http://www.gromacs.org.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "force.h"
  41
  42 #include <cassert>
  43 #include <cmath>
  44 #include <cstring>
  45
  46 #include "gromacs/domdec/dlbtiming.h"
  47 #include "gromacs/domdec/domdec.h"
  48 #include "gromacs/domdec/domdec_struct.h"
  49 #include "gromacs/ewald/ewald.h"
  50 #include "gromacs/ewald/long_range_correction.h"
  51 #include "gromacs/ewald/pme.h"
  52 #include "gromacs/gmxlib/network.h"
  53 #include "gromacs/gmxlib/nrnb.h"
  54 #include "gromacs/listed_forces/listed_forces.h"
  55 #include "gromacs/math/vec.h"
  56 #include "gromacs/math/vecdump.h"
  57 #include "gromacs/mdlib/forcerec_threading.h"
  58 #include "gromacs/mdlib/rf_util.h"
  59 #include "gromacs/mdlib/wall.h"
  60 #include "gromacs/mdtypes/commrec.h"
  61 #include "gromacs/mdtypes/enerdata.h"
  62 #include "gromacs/mdtypes/forceoutput.h"
  63 #include "gromacs/mdtypes/forcerec.h"
  64 #include "gromacs/mdtypes/inputrec.h"
  65 #include "gromacs/mdtypes/interaction_const.h"
  66 #include "gromacs/mdtypes/md_enums.h"
  67 #include "gromacs/mdtypes/mdatom.h"
  68 #include "gromacs/mdtypes/simulation_workload.h"
  69 #include "gromacs/pbcutil/ishift.h"
  70 #include "gromacs/pbcutil/pbc.h"
  71 #include "gromacs/timing/wallcycle.h"
  72 #include "gromacs/utility/exceptions.h"
  73 #include "gromacs/utility/fatalerror.h"
  74 #include "gromacs/utility/smalloc.h"
  75
  76 using gmx::ArrayRef;
  77 using gmx::RVec;
  78
  79 static void clearEwaldThreadOutput(ewald_corr_thread_t* ewc_t)
  80 {
  81     ewc_t->Vcorr_q        = 0;
  82     ewc_t->Vcorr_lj       = 0;
  83     ewc_t->dvdl[efptCOUL] = 0;
  84     ewc_t->dvdl[efptVDW]  = 0;
  85     clear_mat(ewc_t->vir_q);
  86     clear_mat(ewc_t->vir_lj);
  87 }
  88
  89 static void reduceEwaldThreadOuput(int nthreads, ewald_corr_thread_t* ewc_t)
  90 {
  91     ewald_corr_thread_t& dest = ewc_t[0];
  92
  93     for (int t = 1; t < nthreads; t++)
  94     {
  95         dest.Vcorr_q += ewc_t[t].Vcorr_q;
  96         dest.Vcorr_lj += ewc_t[t].Vcorr_lj;
  97         dest.dvdl[efptCOUL] += ewc_t[t].dvdl[efptCOUL];
  98         dest.dvdl[efptVDW] += ewc_t[t].dvdl[efptVDW];
  99         m_add(dest.vir_q, ewc_t[t].vir_q, dest.vir_q);
 100         m_add(dest.vir_lj, ewc_t[t].vir_lj, dest.vir_lj);
 101     }
 102 }
 103
 104 void do_force_lowlevel(t_forcerec*                          fr,
 105                        const t_inputrec*                    ir,
 106                        const t_commrec*                     cr,
 107                        const gmx_multisim_t*                ms,
 108                        t_nrnb*                              nrnb,
 109                        gmx_wallcycle_t                      wcycle,
 110                        const t_mdatoms*                     md,
 111                        gmx::ArrayRefWithPadding<const RVec> coordinates,
 112                        ArrayRef<const RVec>                 xWholeMolecules,
 113                        history_t*                           hist,
 114                        gmx::ForceOutputs*                   forceOutputs,
 115                        gmx_enerdata_t*                      enerd,
 116                        const matrix                         box,
 117                        const real*                          lambda,
 118                        const rvec*                          mu_tot,
 119                        const gmx::StepWorkload&             stepWork,
 120                        const DDBalanceRegionHandler&        ddBalanceRegionHandler)
 121 {
 122     // TODO: Replace all uses of x by const coordinates
 123     const rvec* x = as_rvec_array(coordinates.paddedArrayRef().data());
 124
 125     auto& forceWithVirial = forceOutputs->forceWithVirial();
 126
 127     /* Call the short range functions all in one go. */
 128
 129     if (ir->nwall)
 130     {
 131         /* foreign lambda component for walls */
 132         real dvdl_walls = do_walls(*ir, *fr, box, *md, x, &forceWithVirial, lambda[efptVDW],
 133                                    enerd->grpp.ener[egLJSR].data(), nrnb);
 134         enerd->dvdl_lin[efptVDW] += dvdl_walls;
 135     }
 136
 137     {
 138         t_pbc pbc;
 139
 140         /* Check whether we need to take into account PBC in listed interactions. */
 141         ListedForces& listedForces           = fr->listedForces[0];
 142         const auto    needPbcForListedForces = fr->bMolPBC && stepWork.computeListedForces
 143                                             && listedForces.haveCpuListedForces(*fr->fcdata);
 144         if (needPbcForListedForces)
 145         {
 146             /* Since all atoms are in the rectangular or triclinic unit-cell,
 147              * only single box vector shifts (2 in x) are required.
 148              */
 149             set_pbc_dd(&pbc, fr->pbcType, DOMAINDECOMP(cr) ? cr->dd->numCells : nullptr, TRUE, box);
 150         }
 151
 152         listedForces.calculate(wcycle, box, ir->fepvals, cr, ms, x, xWholeMolecules, fr->fcdata.get(),
 153                                hist, forceOutputs, fr, &pbc, enerd, nrnb, lambda, md,
 154                                DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr, stepWork);
 155     }
 156
 157     const bool computePmeOnCpu = (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
 158                                  && thisRankHasDuty(cr, DUTY_PME)
 159                                  && (pme_run_mode(fr->pmedata) == PmeRunMode::CPU);
 160
 161     const bool haveEwaldSurfaceTerm = haveEwaldSurfaceContribution(*ir);
 162
 163     /* Do long-range electrostatics and/or LJ-PME
 164      * and compute PME surface terms when necessary.
 165      */
 166     if (computePmeOnCpu || fr->ic->eeltype == eelEWALD || haveEwaldSurfaceTerm)
 167     {
 168         int  status = 0;
 169         real Vlr_q = 0, Vlr_lj = 0;
 170
 171         /* We reduce all virial, dV/dlambda and energy contributions, except
 172          * for the reciprocal energies (Vlr_q, Vlr_lj) into the same struct.
 173          */
 174         ewald_corr_thread_t& ewaldOutput = fr->ewc_t[0];
 175         clearEwaldThreadOutput(&ewaldOutput);
 176
 177         if (EEL_PME_EWALD(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
 178         {
 179             /* Calculate the Ewald surface force and energy contributions, when necessary */
 180             if (haveEwaldSurfaceTerm)
 181             {
 182                 wallcycle_sub_start(wcycle, ewcsEWALD_CORRECTION);
 183
 184                 int nthreads = fr->nthread_ewc;
 185 #pragma omp parallel for num_threads(nthreads) schedule(static)
 186                 for (int t = 0; t < nthreads; t++)
 187                 {
 188                     try
 189                     {
 190                         ewald_corr_thread_t& ewc_t = fr->ewc_t[t];
 191                         if (t > 0)
 192                         {
 193                             clearEwaldThreadOutput(&ewc_t);
 194                         }
 195
 196                         /* Threading is only supported with the Verlet cut-off
 197                          * scheme and then only single particle forces (no
 198                          * exclusion forces) are calculated, so we can store
 199                          * the forces in the normal, single forceWithVirial->force_ array.
 200                          */
 201                         ewald_LRcorrection(md->homenr, cr, nthreads, t, *fr, *ir, md->chargeA,
 202                                            md->chargeB, (md->nChargePerturbed != 0), x, box, mu_tot,
 203                                            as_rvec_array(forceWithVirial.force_.data()),
 204                                            &ewc_t.Vcorr_q, lambda[efptCOUL], &ewc_t.dvdl[efptCOUL]);
 205                     }
 206                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 207                 }
 208                 if (nthreads > 1)
 209                 {
 210                     reduceEwaldThreadOuput(nthreads, fr->ewc_t);
 211                 }
 212                 wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
 213             }
 214
 215             if (EEL_PME_EWALD(fr->ic->eeltype) && fr->n_tpi == 0)
 216             {
 217                 /* This is not in a subcounter because it takes a
 218                    negligible and constant-sized amount of time */
 219                 ewaldOutput.Vcorr_q += ewald_charge_correction(
 220                         cr, fr, lambda[efptCOUL], box, &ewaldOutput.dvdl[efptCOUL], ewaldOutput.vir_q);
 221             }
 222
 223             if (computePmeOnCpu)
 224             {
 225                 /* Do reciprocal PME for Coulomb and/or LJ. */
 226                 assert(fr->n_tpi >= 0);
 227                 if (fr->n_tpi == 0 || stepWork.stateChanged)
 228                 {
 229                     /* With domain decomposition we close the CPU side load
 230                      * balancing region here, because PME does global
 231                      * communication that acts as a global barrier.
 232                      */
 233                     ddBalanceRegionHandler.closeAfterForceComputationCpu();
 234
 235                     wallcycle_start(wcycle, ewcPMEMESH);
 236                     status = gmx_pme_do(
 237                             fr->pmedata,
 238                             gmx::constArrayRefFromArray(coordinates.unpaddedConstArrayRef().data(),
 239                                                         md->homenr - fr->n_tpi),
 240                             forceWithVirial.force_, md->chargeA, md->chargeB, md->sqrt_c6A,
 241                             md->sqrt_c6B, md->sigmaA, md->sigmaB, box, cr,
 242                             DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
 243                             DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0, nrnb, wcycle,
 244                             ewaldOutput.vir_q, ewaldOutput.vir_lj, &Vlr_q, &Vlr_lj,
 245                             lambda[efptCOUL], lambda[efptVDW], &ewaldOutput.dvdl[efptCOUL],
 246                             &ewaldOutput.dvdl[efptVDW], stepWork);
 247                     wallcycle_stop(wcycle, ewcPMEMESH);
 248                     if (status != 0)
 249                     {
 250                         gmx_fatal(FARGS, "Error %d in reciprocal PME routine", status);
 251                     }
 252
 253                     /* We should try to do as little computation after
 254                      * this as possible, because parallel PME synchronizes
 255                      * the nodes, so we want all load imbalance of the
 256                      * rest of the force calculation to be before the PME
 257                      * call.  DD load balancing is done on the whole time
 258                      * of the force call (without PME).
 259                      */
 260                 }
 261                 if (fr->n_tpi > 0)
 262                 {
 263                     /* Determine the PME grid energy of the test molecule
 264                      * with the PME grid potential of the other charges.
 265                      */
 266                     gmx_pme_calc_energy(
 267                             fr->pmedata,
 268                             coordinates.unpaddedConstArrayRef().subArray(md->homenr - fr->n_tpi, fr->n_tpi),
 269                             gmx::arrayRefFromArray(md->chargeA + md->homenr - fr->n_tpi, fr->n_tpi),
 270                             &Vlr_q);
 271                 }
 272             }
 273         }
 274
 275         if (fr->ic->eeltype == eelEWALD)
 276         {
 277             Vlr_q = do_ewald(ir, x, as_rvec_array(forceWithVirial.force_.data()), md->chargeA,
 278                              md->chargeB, box, cr, md->homenr, ewaldOutput.vir_q, fr->ic->ewaldcoeff_q,
 279                              lambda[efptCOUL], &ewaldOutput.dvdl[efptCOUL], fr->ewald_table);
 280         }
 281
 282         /* Note that with separate PME nodes we get the real energies later */
 283         // TODO it would be simpler if we just accumulated a single
 284         // long-range virial contribution.
 285         forceWithVirial.addVirialContribution(ewaldOutput.vir_q);
 286         forceWithVirial.addVirialContribution(ewaldOutput.vir_lj);
 287         enerd->dvdl_lin[efptCOUL] += ewaldOutput.dvdl[efptCOUL];
 288         enerd->dvdl_lin[efptVDW] += ewaldOutput.dvdl[efptVDW];
 289         enerd->term[F_COUL_RECIP] = Vlr_q + ewaldOutput.Vcorr_q;
 290         enerd->term[F_LJ_RECIP]   = Vlr_lj + ewaldOutput.Vcorr_lj;
 291
 292         if (debug)
 293         {
 294             fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n", Vlr_q,
 295                     ewaldOutput.Vcorr_q, enerd->term[F_COUL_RECIP]);
 296             pr_rvecs(debug, 0, "vir_el_recip after corr", ewaldOutput.vir_q, DIM);
 297             rvec* fshift = as_rvec_array(forceOutputs->forceWithShiftForces().shiftForces().data());
 298             pr_rvecs(debug, 0, "fshift after LR Corrections", fshift, SHIFTS);
 299             fprintf(debug, "Vlr_lj: %g, Vcorr_lj = %g, Vlr_corr_lj = %g\n", Vlr_lj,
 300                     ewaldOutput.Vcorr_lj, enerd->term[F_LJ_RECIP]);
 301             pr_rvecs(debug, 0, "vir_lj_recip after corr", ewaldOutput.vir_lj, DIM);
 302         }
 303     }
 304
 305     if (debug)
 306     {
 307         print_nrnb(debug, nrnb);
 308     }
 309
 310     if (debug)
 311     {
 312         rvec* fshift = as_rvec_array(forceOutputs->forceWithShiftForces().shiftForces().data());
 313         pr_rvecs(debug, 0, "fshift after bondeds", fshift, SHIFTS);
 314     }
 315 }