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) 2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "force.h"
  40
  41 #include <cassert>
  42 #include <cmath>
  43 #include <cstring>
  44
  45 #include "gromacs/domdec/dlbtiming.h"
  46 #include "gromacs/domdec/domdec.h"
  47 #include "gromacs/domdec/domdec_struct.h"
  48 #include "gromacs/ewald/ewald.h"
  49 #include "gromacs/ewald/long_range_correction.h"
  50 #include "gromacs/ewald/pme.h"
  51 #include "gromacs/gmxlib/network.h"
  52 #include "gromacs/gmxlib/nrnb.h"
  53 #include "gromacs/listed_forces/listed_forces.h"
  54 #include "gromacs/math/vec.h"
  55 #include "gromacs/math/vecdump.h"
  56 #include "gromacs/mdlib/forcerec_threading.h"
  57 #include "gromacs/mdlib/qmmm.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/md_enums.h"
  66 #include "gromacs/mdtypes/mdatom.h"
  67 #include "gromacs/mdtypes/simulation_workload.h"
  68 #include "gromacs/pbcutil/ishift.h"
  69 #include "gromacs/pbcutil/mshift.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 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[efptCOUL] = 0;
  81     ewc_t->dvdl[efptVDW]  = 0;
  82     clear_mat(ewc_t->vir_q);
  83     clear_mat(ewc_t->vir_lj);
  84 }
  85
  86 static void reduceEwaldThreadOuput(int nthreads, 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[efptCOUL] += ewc_t[t].dvdl[efptCOUL];
  95         dest.dvdl[efptVDW]  += ewc_t[t].dvdl[efptVDW];
  96         m_add(dest.vir_q,  ewc_t[t].vir_q,  dest.vir_q);
  97         m_add(dest.vir_lj, ewc_t[t].vir_lj, dest.vir_lj);
  98     }
  99 }
 100
 101 void
 102 do_force_lowlevel(t_forcerec                               *fr,
 103                   const t_inputrec                         *ir,
 104                   const t_idef                             *idef,
 105                   const t_commrec                          *cr,
 106                   const gmx_multisim_t                     *ms,
 107                   t_nrnb                                   *nrnb,
 108                   gmx_wallcycle_t                           wcycle,
 109                   const t_mdatoms                          *md,
 110                   gmx::ArrayRefWithPadding<gmx::RVec>       coordinates,
 111                   history_t                                *hist,
 112                   gmx::ForceOutputs                        *forceOutputs,
 113                   gmx_enerdata_t                           *enerd,
 114                   t_fcdata                                 *fcd,
 115                   const matrix                              box,
 116                   const real                               *lambda,
 117                   const t_graph                            *graph,
 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     rvec *x               = as_rvec_array(coordinates.paddedArrayRef().data());
 124
 125     auto &forceWithVirial = forceOutputs->forceWithVirial();
 126
 127     /* do QMMM first if requested */
 128     if (fr->bQMMM)
 129     {
 130         enerd->term[F_EQM] = calculate_QMMM(cr, &forceOutputs->forceWithShiftForces(), fr);
 131     }
 132
 133     /* Call the short range functions all in one go. */
 134
 135     if (ir->nwall)
 136     {
 137         /* foreign lambda component for walls */
 138         real dvdl_walls = do_walls(*ir, *fr, box, *md, x,
 139                                    &forceWithVirial, lambda[efptVDW],
 140                                    enerd->grpp.ener[egLJSR].data(), nrnb);
 141         enerd->dvdl_lin[efptVDW] += dvdl_walls;
 142     }
 143
 144     /* Shift the coordinates. Must be done before listed forces and PPPM,
 145      * but is also necessary for SHAKE and update, therefore it can NOT
 146      * go when no listed forces have to be evaluated.
 147      *
 148      * The shifting and PBC code is deliberately not timed, since with
 149      * the Verlet scheme it only takes non-zero time with triclinic
 150      * boxes, and even then the time is around a factor of 100 less
 151      * than the next smallest counter.
 152      */
 153
 154
 155     /* Here sometimes we would not need to shift with NBFonly,
 156      * but we do so anyhow for consistency of the returned coordinates.
 157      */
 158     if (graph)
 159     {
 160         shift_self(graph, box, x);
 161         if (TRICLINIC(box))
 162         {
 163             inc_nrnb(nrnb, eNR_SHIFTX, 2*graph->nnodes);
 164         }
 165         else
 166         {
 167             inc_nrnb(nrnb, eNR_SHIFTX, graph->nnodes);
 168         }
 169     }
 170
 171     {
 172         t_pbc      pbc;
 173
 174         /* Check whether we need to take into account PBC in listed interactions. */
 175         const auto needPbcForListedForces = fr->bMolPBC && stepWork.computeListedForces && haveCpuListedForces(*fr, *idef, *fcd);
 176         if (needPbcForListedForces)
 177         {
 178             /* Since all atoms are in the rectangular or triclinic unit-cell,
 179              * only single box vector shifts (2 in x) are required.
 180              */
 181             set_pbc_dd(&pbc, fr->ePBC, DOMAINDECOMP(cr) ? cr->dd->nc : nullptr,
 182                        TRUE, box);
 183         }
 184
 185         do_force_listed(wcycle, box, ir->fepvals, cr, ms,
 186                         idef, x, hist,
 187                         forceOutputs,
 188                         fr, &pbc, graph, enerd, nrnb, lambda, md, fcd,
 189                         DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr,
 190                         stepWork);
 191     }
 192
 193     const bool computePmeOnCpu =
 194         (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype)) &&
 195         thisRankHasDuty(cr, DUTY_PME) &&
 196         (pme_run_mode(fr->pmedata) == PmeRunMode::CPU);
 197
 198     const bool haveEwaldSurfaceTerms =
 199         EEL_PME_EWALD(fr->ic->eeltype) &&
 200         (ir->ewald_geometry != eewg3D || ir->epsilon_surface != 0);
 201
 202     /* Do long-range electrostatics and/or LJ-PME
 203      * and compute PME surface terms when necessary.
 204      */
 205     if (computePmeOnCpu ||
 206         fr->ic->eeltype == eelEWALD ||
 207         haveEwaldSurfaceTerms)
 208     {
 209         int  status            = 0;
 210         real Vlr_q             = 0, Vlr_lj = 0;
 211
 212         /* We reduce all virial, dV/dlambda and energy contributions, except
 213          * for the reciprocal energies (Vlr_q, Vlr_lj) into the same struct.
 214          */
 215         ewald_corr_thread_t &ewaldOutput = fr->ewc_t[0];
 216         clearEwaldThreadOutput(&ewaldOutput);
 217
 218         if (EEL_PME_EWALD(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
 219         {
 220             /* Calculate Ewald surface terms, when necessary */
 221             if (haveEwaldSurfaceTerms)
 222             {
 223                 wallcycle_sub_start(wcycle, ewcsEWALD_CORRECTION);
 224
 225                 if (fr->n_tpi > 0)
 226                 {
 227                     gmx_fatal(FARGS, "TPI with PME currently only works in a 3D geometry with tin-foil boundary conditions");
 228                 }
 229
 230                 int nthreads = fr->nthread_ewc;
 231 #pragma omp parallel for num_threads(nthreads) schedule(static)
 232                 for (int t = 0; t < nthreads; t++)
 233                 {
 234                     try
 235                     {
 236                         ewald_corr_thread_t &ewc_t = fr->ewc_t[t];
 237                         if (t > 0)
 238                         {
 239                             clearEwaldThreadOutput(&ewc_t);
 240                         }
 241
 242                         /* Threading is only supported with the Verlet cut-off
 243                          * scheme and then only single particle forces (no
 244                          * exclusion forces) are calculated, so we can store
 245                          * the forces in the normal, single forceWithVirial->force_ array.
 246                          */
 247                         ewald_LRcorrection(md->homenr, cr, nthreads, t, fr, ir,
 248                                            md->chargeA, md->chargeB,
 249                                            (md->nChargePerturbed != 0),
 250                                            x, box, mu_tot,
 251                                            ir->ewald_geometry,
 252                                            ir->epsilon_surface,
 253                                            as_rvec_array(forceWithVirial.force_.data()),
 254                                            &ewc_t.Vcorr_q,
 255                                            lambda[efptCOUL],
 256                                            &ewc_t.dvdl[efptCOUL]);
 257                     }
 258                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 259                 }
 260                 if (nthreads > 1)
 261                 {
 262                     reduceEwaldThreadOuput(nthreads, fr->ewc_t);
 263                 }
 264                 wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
 265             }
 266
 267             if (EEL_PME_EWALD(fr->ic->eeltype) && fr->n_tpi == 0)
 268             {
 269                 /* This is not in a subcounter because it takes a
 270                    negligible and constant-sized amount of time */
 271                 ewaldOutput.Vcorr_q +=
 272                     ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
 273                                             &ewaldOutput.dvdl[efptCOUL],
 274                                             ewaldOutput.vir_q);
 275             }
 276
 277             if (computePmeOnCpu)
 278             {
 279                 /* Do reciprocal PME for Coulomb and/or LJ. */
 280                 assert(fr->n_tpi >= 0);
 281                 if (fr->n_tpi == 0 || stepWork.stateChanged)
 282                 {
 283                     int pme_flags = GMX_PME_SPREAD | GMX_PME_SOLVE;
 284
 285                     if (stepWork.computeForces)
 286                     {
 287                         pme_flags |= GMX_PME_CALC_F;
 288                     }
 289                     if (stepWork.computeVirial)
 290                     {
 291                         pme_flags |= GMX_PME_CALC_ENER_VIR;
 292                     }
 293                     if (fr->n_tpi > 0)
 294                     {
 295                         /* We don't calculate f, but we do want the potential */
 296                         pme_flags |= GMX_PME_CALC_POT;
 297                     }
 298
 299                     /* With domain decomposition we close the CPU side load
 300                      * balancing region here, because PME does global
 301                      * communication that acts as a global barrier.
 302                      */
 303                     ddBalanceRegionHandler.closeAfterForceComputationCpu();
 304
 305                     wallcycle_start(wcycle, ewcPMEMESH);
 306                     status = gmx_pme_do(fr->pmedata,
 307                                         gmx::constArrayRefFromArray(coordinates.unpaddedConstArrayRef().data(), md->homenr - fr->n_tpi),
 308                                         forceWithVirial.force_,
 309                                         md->chargeA, md->chargeB,
 310                                         md->sqrt_c6A, md->sqrt_c6B,
 311                                         md->sigmaA, md->sigmaB,
 312                                         box, cr,
 313                                         DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
 314                                         DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
 315                                         nrnb, wcycle,
 316                                         ewaldOutput.vir_q, ewaldOutput.vir_lj,
 317                                         &Vlr_q, &Vlr_lj,
 318                                         lambda[efptCOUL], lambda[efptVDW],
 319                                         &ewaldOutput.dvdl[efptCOUL],
 320                                         &ewaldOutput.dvdl[efptVDW],
 321                                         pme_flags);
 322                     wallcycle_stop(wcycle, ewcPMEMESH);
 323                     if (status != 0)
 324                     {
 325                         gmx_fatal(FARGS, "Error %d in reciprocal PME routine", status);
 326                     }
 327
 328                     /* We should try to do as little computation after
 329                      * this as possible, because parallel PME synchronizes
 330                      * the nodes, so we want all load imbalance of the
 331                      * rest of the force calculation to be before the PME
 332                      * call.  DD load balancing is done on the whole time
 333                      * of the force call (without PME).
 334                      */
 335                 }
 336                 if (fr->n_tpi > 0)
 337                 {
 338                     if (EVDW_PME(ir->vdwtype))
 339                     {
 340
 341                         gmx_fatal(FARGS, "Test particle insertion not implemented with LJ-PME");
 342                     }
 343                     /* Determine the PME grid energy of the test molecule
 344                      * with the PME grid potential of the other charges.
 345                      */
 346                     gmx_pme_calc_energy(fr->pmedata,
 347                                         coordinates.unpaddedConstArrayRef().subArray(md->homenr - fr->n_tpi, fr->n_tpi),
 348                                         gmx::arrayRefFromArray(md->chargeA + md->homenr - fr->n_tpi, fr->n_tpi),
 349                                         &Vlr_q);
 350                 }
 351             }
 352         }
 353
 354         if (fr->ic->eeltype == eelEWALD)
 355         {
 356             Vlr_q = do_ewald(ir, x, as_rvec_array(forceWithVirial.force_.data()),
 357                              md->chargeA, md->chargeB,
 358                              box, cr, md->homenr,
 359                              ewaldOutput.vir_q, fr->ic->ewaldcoeff_q,
 360                              lambda[efptCOUL], &ewaldOutput.dvdl[efptCOUL],
 361                              fr->ewald_table);
 362         }
 363
 364         /* Note that with separate PME nodes we get the real energies later */
 365         // TODO it would be simpler if we just accumulated a single
 366         // long-range virial contribution.
 367         forceWithVirial.addVirialContribution(ewaldOutput.vir_q);
 368         forceWithVirial.addVirialContribution(ewaldOutput.vir_lj);
 369         enerd->dvdl_lin[efptCOUL] += ewaldOutput.dvdl[efptCOUL];
 370         enerd->dvdl_lin[efptVDW]  += ewaldOutput.dvdl[efptVDW];
 371         enerd->term[F_COUL_RECIP]  = Vlr_q + ewaldOutput.Vcorr_q;
 372         enerd->term[F_LJ_RECIP]    = Vlr_lj + ewaldOutput.Vcorr_lj;
 373
 374         if (debug)
 375         {
 376             fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
 377                     Vlr_q, ewaldOutput.Vcorr_q, enerd->term[F_COUL_RECIP]);
 378             pr_rvecs(debug, 0, "vir_el_recip after corr", ewaldOutput.vir_q, DIM);
 379             rvec *fshift = as_rvec_array(forceOutputs->forceWithShiftForces().shiftForces().data());
 380             pr_rvecs(debug, 0, "fshift after LR Corrections", fshift, SHIFTS);
 381             fprintf(debug, "Vlr_lj: %g, Vcorr_lj = %g, Vlr_corr_lj = %g\n",
 382                     Vlr_lj, ewaldOutput.Vcorr_lj, enerd->term[F_LJ_RECIP]);
 383             pr_rvecs(debug, 0, "vir_lj_recip after corr", ewaldOutput.vir_lj, DIM);
 384         }
 385     }
 386
 387     if (debug)
 388     {
 389         print_nrnb(debug, nrnb);
 390     }
 391
 392     if (debug)
 393     {
 394         rvec *fshift = as_rvec_array(forceOutputs->forceWithShiftForces().shiftForces().data());
 395         pr_rvecs(debug, 0, "fshift after bondeds", fshift, SHIFTS);
 396     }
 397
 398 }