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) 2018,2019,2020,2021, 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/math/vec.h"
  55 #include "gromacs/math/vecdump.h"
  56 #include "gromacs/mdlib/forcerec_threading.h"
  57 #include "gromacs/mdtypes/commrec.h"
  58 #include "gromacs/mdtypes/enerdata.h"
  59 #include "gromacs/mdtypes/forceoutput.h"
  60 #include "gromacs/mdtypes/forcerec.h"
  61 #include "gromacs/mdtypes/inputrec.h"
  62 #include "gromacs/mdtypes/interaction_const.h"
  63 #include "gromacs/mdtypes/md_enums.h"
  64 #include "gromacs/mdtypes/mdatom.h"
  65 #include "gromacs/mdtypes/simulation_workload.h"
  66 #include "gromacs/pbcutil/ishift.h"
  67 #include "gromacs/pbcutil/pbc.h"
  68 #include "gromacs/timing/wallcycle.h"
  69 #include "gromacs/utility/exceptions.h"
  70 #include "gromacs/utility/fatalerror.h"
  71 #include "gromacs/utility/smalloc.h"
  72
  73 using gmx::ArrayRef;
  74 using gmx::RVec;
  75
  76 static void clearEwaldThreadOutput(ewald_corr_thread_t* ewc_t)
  77 {
  78     ewc_t->Vcorr_q                                        = 0;
  79     ewc_t->Vcorr_lj                                       = 0;
  80     ewc_t->dvdl[FreeEnergyPerturbationCouplingType::Coul] = 0;
  81     ewc_t->dvdl[FreeEnergyPerturbationCouplingType::Vdw]  = 0;
  82     clear_mat(ewc_t->vir_q);
  83     clear_mat(ewc_t->vir_lj);
  84 }
  85
  86 static void reduceEwaldThreadOuput(int nthreads, gmx::ArrayRef<ewald_corr_thread_t> ewc_t)
  87 {
  88     ewald_corr_thread_t& dest = ewc_t[0];
  89
  90     for (int t = 1; t < nthreads; t++)
  91     {
  92         dest.Vcorr_q += ewc_t[t].Vcorr_q;
  93         dest.Vcorr_lj += ewc_t[t].Vcorr_lj;
  94         dest.dvdl[FreeEnergyPerturbationCouplingType::Coul] +=
  95                 ewc_t[t].dvdl[FreeEnergyPerturbationCouplingType::Coul];
  96         dest.dvdl[FreeEnergyPerturbationCouplingType::Vdw] +=
  97                 ewc_t[t].dvdl[FreeEnergyPerturbationCouplingType::Vdw];
  98         m_add(dest.vir_q, ewc_t[t].vir_q, dest.vir_q);
  99         m_add(dest.vir_lj, ewc_t[t].vir_lj, dest.vir_lj);
 100     }
 101 }
 102
 103 void calculateLongRangeNonbondeds(t_forcerec*                    fr,
 104                                   const t_inputrec&              ir,
 105                                   const t_commrec*               cr,
 106                                   t_nrnb*                        nrnb,
 107                                   gmx_wallcycle*                 wcycle,
 108                                   const t_mdatoms*               md,
 109                                   gmx::ArrayRef<const RVec>      coordinates,
 110                                   gmx::ForceWithVirial*          forceWithVirial,
 111                                   gmx_enerdata_t*                enerd,
 112                                   const matrix                   box,
 113                                   gmx::ArrayRef<const real>      lambda,
 114                                   gmx::ArrayRef<const gmx::RVec> mu_tot,
 115                                   const gmx::StepWorkload&       stepWork,
 116                                   const DDBalanceRegionHandler&  ddBalanceRegionHandler)
 117 {
 118     const bool computePmeOnCpu = (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
 119                                  && thisRankHasDuty(cr, DUTY_PME)
 120                                  && (pme_run_mode(fr->pmedata) == PmeRunMode::CPU);
 121
 122     const bool haveEwaldSurfaceTerm = haveEwaldSurfaceContribution(ir);
 123
 124     /* Do long-range electrostatics and/or LJ-PME
 125      * and compute PME surface terms when necessary.
 126      */
 127     if ((computePmeOnCpu || fr->ic->eeltype == CoulombInteractionType::Ewald || haveEwaldSurfaceTerm)
 128         && stepWork.computeNonbondedForces)
 129     {
 130         int  status = 0;
 131         real Vlr_q = 0, Vlr_lj = 0;
 132
 133         /* We reduce all virial, dV/dlambda and energy contributions, except
 134          * for the reciprocal energies (Vlr_q, Vlr_lj) into the same struct.
 135          */
 136         ewald_corr_thread_t& ewaldOutput = fr->ewc_t[0];
 137         clearEwaldThreadOutput(&ewaldOutput);
 138
 139         if (EEL_PME_EWALD(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
 140         {
 141             /* Calculate the Ewald surface force and energy contributions, when necessary */
 142             if (haveEwaldSurfaceTerm)
 143             {
 144                 wallcycle_sub_start(wcycle, WallCycleSubCounter::EwaldCorrection);
 145
 146                 int nthreads = fr->nthread_ewc;
 147 #pragma omp parallel for num_threads(nthreads) schedule(static)
 148                 for (int t = 0; t < nthreads; t++)
 149                 {
 150                     try
 151                     {
 152                         ewald_corr_thread_t& ewc_t = fr->ewc_t[t];
 153                         if (t > 0)
 154                         {
 155                             clearEwaldThreadOutput(&ewc_t);
 156                         }
 157
 158                         /* Threading is only supported with the Verlet cut-off
 159                          * scheme and then only single particle forces (no
 160                          * exclusion forces) are calculated, so we can store
 161                          * the forces in the normal, single forceWithVirial->force_ array.
 162                          */
 163                         ewald_LRcorrection(
 164                                 md->homenr,
 165                                 cr,
 166                                 nthreads,
 167                                 t,
 168                                 *fr,
 169                                 ir,
 170                                 md->chargeA ? gmx::constArrayRefFromArray(md->chargeA, md->nr)
 171                                             : gmx::ArrayRef<const real>{},
 172                                 md->chargeB ? gmx::constArrayRefFromArray(md->chargeB, md->nr)
 173                                             : gmx::ArrayRef<const real>{},
 174                                 (md->nChargePerturbed != 0),
 175                                 coordinates,
 176                                 box,
 177                                 mu_tot,
 178                                 forceWithVirial->force_,
 179                                 &ewc_t.Vcorr_q,
 180                                 lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
 181                                 &ewc_t.dvdl[FreeEnergyPerturbationCouplingType::Coul]);
 182                     }
 183                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 184                 }
 185                 if (nthreads > 1)
 186                 {
 187                     reduceEwaldThreadOuput(nthreads, fr->ewc_t);
 188                 }
 189                 wallcycle_sub_stop(wcycle, WallCycleSubCounter::EwaldCorrection);
 190             }
 191
 192             if (EEL_PME_EWALD(fr->ic->eeltype) && fr->n_tpi == 0)
 193             {
 194                 /* This is not in a subcounter because it takes a
 195                    negligible and constant-sized amount of time */
 196                 ewaldOutput.Vcorr_q += ewald_charge_correction(
 197                         cr,
 198                         fr->ic->epsilon_r,
 199                         fr->ic->ewaldcoeff_q,
 200                         fr->qsum,
 201                         lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
 202                         box,
 203                         &ewaldOutput.dvdl[FreeEnergyPerturbationCouplingType::Coul],
 204                         ewaldOutput.vir_q);
 205             }
 206
 207             if (computePmeOnCpu)
 208             {
 209                 /* Do reciprocal PME for Coulomb and/or LJ. */
 210                 assert(fr->n_tpi >= 0);
 211                 if (fr->n_tpi == 0 || stepWork.stateChanged)
 212                 {
 213                     /* With domain decomposition we close the CPU side load
 214                      * balancing region here, because PME does global
 215                      * communication that acts as a global barrier.
 216                      */
 217                     ddBalanceRegionHandler.closeAfterForceComputationCpu();
 218
 219                     wallcycle_start(wcycle, WallCycleCounter::PmeMesh);
 220                     status = gmx_pme_do(
 221                             fr->pmedata,
 222                             gmx::constArrayRefFromArray(coordinates.data(), md->homenr - fr->n_tpi),
 223                             forceWithVirial->force_,
 224                             md->chargeA ? gmx::constArrayRefFromArray(md->chargeA, md->nr)
 225                                         : gmx::ArrayRef<const real>{},
 226                             md->chargeB ? gmx::constArrayRefFromArray(md->chargeB, md->nr)
 227                                         : gmx::ArrayRef<const real>{},
 228                             md->sqrt_c6A ? gmx::constArrayRefFromArray(md->sqrt_c6A, md->nr)
 229                                          : gmx::ArrayRef<const real>{},
 230                             md->sqrt_c6B ? gmx::constArrayRefFromArray(md->sqrt_c6B, md->nr)
 231                                          : gmx::ArrayRef<const real>{},
 232                             md->sigmaA ? gmx::constArrayRefFromArray(md->sigmaA, md->nr)
 233                                        : gmx::ArrayRef<const real>{},
 234                             md->sigmaB ? gmx::constArrayRefFromArray(md->sigmaB, md->nr)
 235                                        : gmx::ArrayRef<const real>{},
 236                             box,
 237                             cr,
 238                             DOMAINDECOMP(cr) ? dd_pme_maxshift_x(*cr->dd) : 0,
 239                             DOMAINDECOMP(cr) ? dd_pme_maxshift_y(*cr->dd) : 0,
 240                             nrnb,
 241                             wcycle,
 242                             ewaldOutput.vir_q,
 243                             ewaldOutput.vir_lj,
 244                             &Vlr_q,
 245                             &Vlr_lj,
 246                             lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
 247                             lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)],
 248                             &ewaldOutput.dvdl[FreeEnergyPerturbationCouplingType::Coul],
 249                             &ewaldOutput.dvdl[FreeEnergyPerturbationCouplingType::Vdw],
 250                             stepWork);
 251                     wallcycle_stop(wcycle, WallCycleCounter::PmeMesh);
 252                     if (status != 0)
 253                     {
 254                         gmx_fatal(FARGS, "Error %d in reciprocal PME routine", status);
 255                     }
 256
 257                     /* We should try to do as little computation after
 258                      * this as possible, because parallel PME synchronizes
 259                      * the nodes, so we want all load imbalance of the
 260                      * rest of the force calculation to be before the PME
 261                      * call.  DD load balancing is done on the whole time
 262                      * of the force call (without PME).
 263                      */
 264                 }
 265                 if (fr->n_tpi > 0)
 266                 {
 267                     /* Determine the PME grid energy of the test molecule
 268                      * with the PME grid potential of the other charges.
 269                      */
 270                     Vlr_q = gmx_pme_calc_energy(
 271                             fr->pmedata,
 272                             coordinates.subArray(md->homenr - fr->n_tpi, fr->n_tpi),
 273                             gmx::arrayRefFromArray(md->chargeA + md->homenr - fr->n_tpi, fr->n_tpi));
 274                 }
 275             }
 276         }
 277
 278         if (fr->ic->eeltype == CoulombInteractionType::Ewald)
 279         {
 280             Vlr_q = do_ewald(ir,
 281                              coordinates,
 282                              forceWithVirial->force_,
 283                              gmx::arrayRefFromArray(md->chargeA, md->nr),
 284                              gmx::arrayRefFromArray(md->chargeB, md->nr),
 285                              box,
 286                              cr,
 287                              md->homenr,
 288                              ewaldOutput.vir_q,
 289                              fr->ic->ewaldcoeff_q,
 290                              lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
 291                              &ewaldOutput.dvdl[FreeEnergyPerturbationCouplingType::Coul],
 292                              fr->ewald_table.get());
 293         }
 294
 295         /* Note that with separate PME nodes we get the real energies later */
 296         // TODO it would be simpler if we just accumulated a single
 297         // long-range virial contribution.
 298         forceWithVirial->addVirialContribution(ewaldOutput.vir_q);
 299         forceWithVirial->addVirialContribution(ewaldOutput.vir_lj);
 300         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Coul] +=
 301                 ewaldOutput.dvdl[FreeEnergyPerturbationCouplingType::Coul];
 302         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] +=
 303                 ewaldOutput.dvdl[FreeEnergyPerturbationCouplingType::Vdw];
 304         enerd->term[F_COUL_RECIP] = Vlr_q + ewaldOutput.Vcorr_q;
 305         enerd->term[F_LJ_RECIP]   = Vlr_lj + ewaldOutput.Vcorr_lj;
 306
 307         if (debug)
 308         {
 309             fprintf(debug,
 310                     "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
 311                     Vlr_q,
 312                     ewaldOutput.Vcorr_q,
 313                     enerd->term[F_COUL_RECIP]);
 314             pr_rvecs(debug, 0, "vir_el_recip after corr", ewaldOutput.vir_q, DIM);
 315             fprintf(debug,
 316                     "Vlr_lj: %g, Vcorr_lj = %g, Vlr_corr_lj = %g\n",
 317                     Vlr_lj,
 318                     ewaldOutput.Vcorr_lj,
 319                     enerd->term[F_LJ_RECIP]);
 320             pr_rvecs(debug, 0, "vir_lj_recip after corr", ewaldOutput.vir_lj, DIM);
 321         }
 322     }
 323
 324     if (debug)
 325     {
 326         print_nrnb(debug, nrnb);
 327     }
 328 }