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 "config.h"
  42
  43 #include <cassert>
  44 #include <cmath>
  45 #include <cstring>
  46
  47 #include "gromacs/domdec/dlbtiming.h"
  48 #include "gromacs/domdec/domdec.h"
  49 #include "gromacs/domdec/domdec_struct.h"
  50 #include "gromacs/ewald/ewald.h"
  51 #include "gromacs/ewald/long_range_correction.h"
  52 #include "gromacs/ewald/pme.h"
  53 #include "gromacs/gmxlib/network.h"
  54 #include "gromacs/gmxlib/nrnb.h"
  55 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  56 #include "gromacs/listed_forces/listed_forces.h"
  57 #include "gromacs/math/vec.h"
  58 #include "gromacs/math/vecdump.h"
  59 #include "gromacs/mdlib/force_flags.h"
  60 #include "gromacs/mdlib/forcerec_threading.h"
  61 #include "gromacs/mdlib/ns.h"
  62 #include "gromacs/mdlib/qmmm.h"
  63 #include "gromacs/mdlib/rf_util.h"
  64 #include "gromacs/mdlib/wall.h"
  65 #include "gromacs/mdtypes/commrec.h"
  66 #include "gromacs/mdtypes/enerdata.h"
  67 #include "gromacs/mdtypes/forceoutput.h"
  68 #include "gromacs/mdtypes/forcerec.h"
  69 #include "gromacs/mdtypes/inputrec.h"
  70 #include "gromacs/mdtypes/md_enums.h"
  71 #include "gromacs/pbcutil/ishift.h"
  72 #include "gromacs/pbcutil/mshift.h"
  73 #include "gromacs/pbcutil/pbc.h"
  74 #include "gromacs/timing/wallcycle.h"
  75 #include "gromacs/utility/cstringutil.h"
  76 #include "gromacs/utility/exceptions.h"
  77 #include "gromacs/utility/fatalerror.h"
  78 #include "gromacs/utility/smalloc.h"
  79
  80 void ns(FILE                      *fp,
  81         t_forcerec                *fr,
  82         matrix                     box,
  83         const SimulationGroups    *groups,
  84         gmx_localtop_t            *top,
  85         const t_mdatoms           *md,
  86         const t_commrec           *cr,
  87         t_nrnb                    *nrnb,
  88         gmx_bool                   bFillGrid)
  89 {
  90     int     nsearch;
  91
  92
  93     if (!fr->ns->nblist_initialized)
  94     {
  95         init_neighbor_list(fp, fr, md->homenr);
  96     }
  97
  98     nsearch = search_neighbours(fp, fr, box, top, groups, cr, nrnb, md,
  99                                 bFillGrid);
 100     if (debug)
 101     {
 102         fprintf(debug, "nsearch = %d\n", nsearch);
 103     }
 104
 105     /* Check whether we have to do dynamic load balancing */
 106     /*if ((nsb->nstDlb > 0) && (mod(step,nsb->nstDlb) == 0))
 107        count_nb(cr,nsb,&(top->blocks[ebCGS]),nns,fr->nlr,
 108        &(top->idef),opts->ngener);
 109      */
 110     if (fr->ns->dump_nl > 0)
 111     {
 112         dump_nblist(fp, cr, fr, fr->ns->dump_nl);
 113     }
 114 }
 115
 116 static void clearEwaldThreadOutput(ewald_corr_thread_t *ewc_t)
 117 {
 118     ewc_t->Vcorr_q        = 0;
 119     ewc_t->Vcorr_lj       = 0;
 120     ewc_t->dvdl[efptCOUL] = 0;
 121     ewc_t->dvdl[efptVDW]  = 0;
 122     clear_mat(ewc_t->vir_q);
 123     clear_mat(ewc_t->vir_lj);
 124 }
 125
 126 static void reduceEwaldThreadOuput(int nthreads, ewald_corr_thread_t *ewc_t)
 127 {
 128     ewald_corr_thread_t &dest = ewc_t[0];
 129
 130     for (int t = 1; t < nthreads; t++)
 131     {
 132         dest.Vcorr_q        += ewc_t[t].Vcorr_q;
 133         dest.Vcorr_lj       += ewc_t[t].Vcorr_lj;
 134         dest.dvdl[efptCOUL] += ewc_t[t].dvdl[efptCOUL];
 135         dest.dvdl[efptVDW]  += ewc_t[t].dvdl[efptVDW];
 136         m_add(dest.vir_q,  ewc_t[t].vir_q,  dest.vir_q);
 137         m_add(dest.vir_lj, ewc_t[t].vir_lj, dest.vir_lj);
 138     }
 139 }
 140
 141 void do_force_lowlevel(t_forcerec                   *fr,
 142                        const t_inputrec             *ir,
 143                        const t_idef                 *idef,
 144                        const t_commrec              *cr,
 145                        const gmx_multisim_t         *ms,
 146                        t_nrnb                       *nrnb,
 147                        gmx_wallcycle_t               wcycle,
 148                        const t_mdatoms              *md,
 149                        rvec                          x[],
 150                        history_t                    *hist,
 151                        rvec                         *forceForUseWithShiftForces,
 152                        gmx::ForceWithVirial         *forceWithVirial,
 153                        gmx_enerdata_t               *enerd,
 154                        t_fcdata                     *fcd,
 155                        matrix                        box,
 156                        t_lambda                     *fepvals,
 157                        real                         *lambda,
 158                        const t_graph                *graph,
 159                        const t_blocka               *excl,
 160                        rvec                          mu_tot[],
 161                        int                           flags,
 162                        const DDBalanceRegionHandler &ddBalanceRegionHandler)
 163 {
 164     int         i, j;
 165     int         donb_flags;
 166     int         pme_flags;
 167     real        dvdl_dum[efptNR], dvdl_nb[efptNR];
 168
 169 #if GMX_MPI
 170     double  t0 = 0.0, t1, t2, t3; /* time measurement for coarse load balancing */
 171 #endif
 172
 173     /* reset free energy components */
 174     for (i = 0; i < efptNR; i++)
 175     {
 176         dvdl_nb[i]  = 0;
 177         dvdl_dum[i] = 0;
 178     }
 179
 180     /* do QMMM first if requested */
 181     if (fr->bQMMM)
 182     {
 183         enerd->term[F_EQM] = calculate_QMMM(cr, forceForUseWithShiftForces, fr);
 184     }
 185
 186     /* Call the short range functions all in one go. */
 187
 188 #if GMX_MPI
 189     /*#define TAKETIME ((cr->npmenodes) && (fr->timesteps < 12))*/
 190 #define TAKETIME FALSE
 191     if (TAKETIME)
 192     {
 193         MPI_Barrier(cr->mpi_comm_mygroup);
 194         t0 = MPI_Wtime();
 195     }
 196 #endif
 197
 198     if (ir->nwall)
 199     {
 200         /* foreign lambda component for walls */
 201         real dvdl_walls = do_walls(*ir, *fr, box, *md, x,
 202                                    forceWithVirial, lambda[efptVDW],
 203                                    enerd->grpp.ener[egLJSR], nrnb);
 204         enerd->dvdl_lin[efptVDW] += dvdl_walls;
 205     }
 206
 207     /* We only do non-bonded calculation with group scheme here, the verlet
 208      * calls are done from do_force_cutsVERLET(). */
 209     if (fr->cutoff_scheme == ecutsGROUP && (flags & GMX_FORCE_NONBONDED))
 210     {
 211         donb_flags = 0;
 212         /* Add short-range interactions */
 213         donb_flags |= GMX_NONBONDED_DO_SR;
 214
 215         /* Currently all group scheme kernels always calculate (shift-)forces */
 216         if (flags & GMX_FORCE_FORCES)
 217         {
 218             donb_flags |= GMX_NONBONDED_DO_FORCE;
 219         }
 220         if (flags & GMX_FORCE_VIRIAL)
 221         {
 222             donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 223         }
 224         if (flags & GMX_FORCE_ENERGY)
 225         {
 226             donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 227         }
 228
 229         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 230         do_nonbonded(fr, x, forceForUseWithShiftForces, md, excl,
 231                      &enerd->grpp, nrnb,
 232                      lambda, dvdl_nb, -1, -1, donb_flags);
 233
 234         /* If we do foreign lambda and we have soft-core interactions
 235          * we have to recalculate the (non-linear) energies contributions.
 236          */
 237         if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
 238         {
 239             for (i = 0; i < enerd->n_lambda; i++)
 240             {
 241                 real lam_i[efptNR];
 242
 243                 for (j = 0; j < efptNR; j++)
 244                 {
 245                     lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
 246                 }
 247                 reset_foreign_enerdata(enerd);
 248                 do_nonbonded(fr, x, forceForUseWithShiftForces, md, excl,
 249                              &(enerd->foreign_grpp), nrnb,
 250                              lam_i, dvdl_dum, -1, -1,
 251                              (donb_flags & ~GMX_NONBONDED_DO_FORCE) | GMX_NONBONDED_DO_FOREIGNLAMBDA);
 252                 sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
 253                 enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
 254             }
 255         }
 256         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 257     }
 258
 259 #if GMX_MPI
 260     if (TAKETIME)
 261     {
 262         t1          = MPI_Wtime();
 263         fr->t_fnbf += t1-t0;
 264     }
 265 #endif
 266
 267     if (fepvals->sc_alpha != 0)
 268     {
 269         enerd->dvdl_nonlin[efptVDW] += dvdl_nb[efptVDW];
 270     }
 271     else
 272     {
 273         enerd->dvdl_lin[efptVDW] += dvdl_nb[efptVDW];
 274     }
 275
 276     if (fepvals->sc_alpha != 0)
 277
 278     /* even though coulomb part is linear, we already added it, beacuse we
 279        need to go through the vdw calculation anyway */
 280     {
 281         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 282     }
 283     else
 284     {
 285         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 286     }
 287
 288     if (debug)
 289     {
 290         pr_rvecs(debug, 0, "fshift after SR", fr->fshift, SHIFTS);
 291     }
 292
 293     /* Shift the coordinates. Must be done before listed forces and PPPM,
 294      * but is also necessary for SHAKE and update, therefore it can NOT
 295      * go when no listed forces have to be evaluated.
 296      *
 297      * The shifting and PBC code is deliberately not timed, since with
 298      * the Verlet scheme it only takes non-zero time with triclinic
 299      * boxes, and even then the time is around a factor of 100 less
 300      * than the next smallest counter.
 301      */
 302
 303
 304     /* Here sometimes we would not need to shift with NBFonly,
 305      * but we do so anyhow for consistency of the returned coordinates.
 306      */
 307     if (graph)
 308     {
 309         shift_self(graph, box, x);
 310         if (TRICLINIC(box))
 311         {
 312             inc_nrnb(nrnb, eNR_SHIFTX, 2*graph->nnodes);
 313         }
 314         else
 315         {
 316             inc_nrnb(nrnb, eNR_SHIFTX, graph->nnodes);
 317         }
 318     }
 319
 320     t_pbc      pbc;
 321     const bool useRfWithGroupScheme = (fr->cutoff_scheme == ecutsGROUP) && EEL_RF(fr->ic->eeltype);
 322     /* Check whether we need to take into account PBC in the following force tasks:
 323      * listed interactions or when correcting for exclusions (for the Group scheme with RF) */
 324     if (fr->bMolPBC)
 325     {
 326         const auto needPbcForListedForces = bool(flags & GMX_FORCE_LISTED) && haveCpuListedForces(*fr, *idef, *fcd);
 327         if (needPbcForListedForces || useRfWithGroupScheme)
 328         {
 329             /* Since all atoms are in the rectangular or triclinic unit-cell,
 330              * only single box vector shifts (2 in x) are required.
 331              */
 332             set_pbc_dd(&pbc, fr->ePBC, DOMAINDECOMP(cr) ? cr->dd->nc : nullptr,
 333                        TRUE, box);
 334         }
 335     }
 336
 337     do_force_listed(wcycle, box, ir->fepvals, cr, ms,
 338                     idef, x, hist,
 339                     forceForUseWithShiftForces, forceWithVirial,
 340                     fr, &pbc, graph, enerd, nrnb, lambda, md, fcd,
 341                     DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr,
 342                     flags);
 343
 344
 345     /* Do long-range electrostatics and/or LJ-PME, including related short-range
 346      * corrections.
 347      */
 348     if (EEL_FULL(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
 349     {
 350         int  status            = 0;
 351         real Vlr_q             = 0, Vlr_lj = 0;
 352
 353         /* We reduce all virial, dV/dlambda and energy contributions, except
 354          * for the reciprocal energies (Vlr_q, Vlr_lj) into the same struct.
 355          */
 356         ewald_corr_thread_t &ewaldOutput = fr->ewc_t[0];
 357         clearEwaldThreadOutput(&ewaldOutput);
 358
 359         if (EEL_PME_EWALD(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
 360         {
 361             /* With the Verlet scheme exclusion forces are calculated
 362              * in the non-bonded kernel.
 363              */
 364             /* The TPI molecule does not have exclusions with the rest
 365              * of the system and no intra-molecular PME grid
 366              * contributions will be calculated in
 367              * gmx_pme_calc_energy.
 368              */
 369             if ((ir->cutoff_scheme == ecutsGROUP && fr->n_tpi == 0) ||
 370                 ir->ewald_geometry != eewg3D ||
 371                 ir->epsilon_surface != 0)
 372             {
 373                 int nthreads, t;
 374
 375                 wallcycle_sub_start(wcycle, ewcsEWALD_CORRECTION);
 376
 377                 if (fr->n_tpi > 0)
 378                 {
 379                     gmx_fatal(FARGS, "TPI with PME currently only works in a 3D geometry with tin-foil boundary conditions");
 380                 }
 381
 382                 nthreads = fr->nthread_ewc;
 383 #pragma omp parallel for num_threads(nthreads) schedule(static)
 384                 for (t = 0; t < nthreads; t++)
 385                 {
 386                     try
 387                     {
 388                         ewald_corr_thread_t &ewc_t = fr->ewc_t[t];
 389                         if (t > 0)
 390                         {
 391                             clearEwaldThreadOutput(&ewc_t);
 392                         }
 393
 394                         /* Threading is only supported with the Verlet cut-off
 395                          * scheme and then only single particle forces (no
 396                          * exclusion forces) are calculated, so we can store
 397                          * the forces in the normal, single forceWithVirial->force_ array.
 398                          */
 399                         ewald_LRcorrection(md->homenr, cr, nthreads, t, fr, ir,
 400                                            md->chargeA, md->chargeB,
 401                                            md->sqrt_c6A, md->sqrt_c6B,
 402                                            md->sigmaA, md->sigmaB,
 403                                            md->sigma3A, md->sigma3B,
 404                                            (md->nChargePerturbed != 0) || (md->nTypePerturbed != 0),
 405                                            ir->cutoff_scheme != ecutsVERLET,
 406                                            excl, x, box, mu_tot,
 407                                            ir->ewald_geometry,
 408                                            ir->epsilon_surface,
 409                                            as_rvec_array(forceWithVirial->force_.data()),
 410                                            ewc_t.vir_q, ewc_t.vir_lj,
 411                                            &ewc_t.Vcorr_q, &ewc_t.Vcorr_lj,
 412                                            lambda[efptCOUL], lambda[efptVDW],
 413                                            &ewc_t.dvdl[efptCOUL], &ewc_t.dvdl[efptVDW]);
 414                     }
 415                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 416                 }
 417                 if (nthreads > 1)
 418                 {
 419                     reduceEwaldThreadOuput(nthreads, fr->ewc_t);
 420                 }
 421                 wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
 422             }
 423
 424             if (EEL_PME_EWALD(fr->ic->eeltype) && fr->n_tpi == 0)
 425             {
 426                 /* This is not in a subcounter because it takes a
 427                    negligible and constant-sized amount of time */
 428                 ewaldOutput.Vcorr_q +=
 429                     ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
 430                                             &ewaldOutput.dvdl[efptCOUL],
 431                                             ewaldOutput.vir_q);
 432             }
 433
 434             if ((EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype)) &&
 435                 thisRankHasDuty(cr, DUTY_PME) && (pme_run_mode(fr->pmedata) == PmeRunMode::CPU))
 436             {
 437                 /* Do reciprocal PME for Coulomb and/or LJ. */
 438                 assert(fr->n_tpi >= 0);
 439                 if (fr->n_tpi == 0 || (flags & GMX_FORCE_STATECHANGED))
 440                 {
 441                     pme_flags = GMX_PME_SPREAD | GMX_PME_SOLVE;
 442
 443                     if (flags & GMX_FORCE_FORCES)
 444                     {
 445                         pme_flags |= GMX_PME_CALC_F;
 446                     }
 447                     if (flags & GMX_FORCE_VIRIAL)
 448                     {
 449                         pme_flags |= GMX_PME_CALC_ENER_VIR;
 450                     }
 451                     if (fr->n_tpi > 0)
 452                     {
 453                         /* We don't calculate f, but we do want the potential */
 454                         pme_flags |= GMX_PME_CALC_POT;
 455                     }
 456
 457                     /* With domain decomposition we close the CPU side load
 458                      * balancing region here, because PME does global
 459                      * communication that acts as a global barrier.
 460                      */
 461                     ddBalanceRegionHandler.closeAfterForceComputationCpu();
 462
 463                     wallcycle_start(wcycle, ewcPMEMESH);
 464                     status = gmx_pme_do(fr->pmedata,
 465                                         0, md->homenr - fr->n_tpi,
 466                                         x,
 467                                         as_rvec_array(forceWithVirial->force_.data()),
 468                                         md->chargeA, md->chargeB,
 469                                         md->sqrt_c6A, md->sqrt_c6B,
 470                                         md->sigmaA, md->sigmaB,
 471                                         box, cr,
 472                                         DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
 473                                         DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
 474                                         nrnb, wcycle,
 475                                         ewaldOutput.vir_q, ewaldOutput.vir_lj,
 476                                         &Vlr_q, &Vlr_lj,
 477                                         lambda[efptCOUL], lambda[efptVDW],
 478                                         &ewaldOutput.dvdl[efptCOUL],
 479                                         &ewaldOutput.dvdl[efptVDW],
 480                                         pme_flags);
 481                     wallcycle_stop(wcycle, ewcPMEMESH);
 482                     if (status != 0)
 483                     {
 484                         gmx_fatal(FARGS, "Error %d in reciprocal PME routine", status);
 485                     }
 486
 487                     /* We should try to do as little computation after
 488                      * this as possible, because parallel PME synchronizes
 489                      * the nodes, so we want all load imbalance of the
 490                      * rest of the force calculation to be before the PME
 491                      * call.  DD load balancing is done on the whole time
 492                      * of the force call (without PME).
 493                      */
 494                 }
 495                 if (fr->n_tpi > 0)
 496                 {
 497                     if (EVDW_PME(ir->vdwtype))
 498                     {
 499
 500                         gmx_fatal(FARGS, "Test particle insertion not implemented with LJ-PME");
 501                     }
 502                     /* Determine the PME grid energy of the test molecule
 503                      * with the PME grid potential of the other charges.
 504                      */
 505                     gmx_pme_calc_energy(fr->pmedata, fr->n_tpi,
 506                                         x + md->homenr - fr->n_tpi,
 507                                         md->chargeA + md->homenr - fr->n_tpi,
 508                                         &Vlr_q);
 509                 }
 510             }
 511         }
 512
 513         if (!EEL_PME(fr->ic->eeltype) && EEL_PME_EWALD(fr->ic->eeltype))
 514         {
 515             Vlr_q = do_ewald(ir, x, as_rvec_array(forceWithVirial->force_.data()),
 516                              md->chargeA, md->chargeB,
 517                              box, cr, md->homenr,
 518                              ewaldOutput.vir_q, fr->ic->ewaldcoeff_q,
 519                              lambda[efptCOUL], &ewaldOutput.dvdl[efptCOUL],
 520                              fr->ewald_table);
 521         }
 522
 523         /* Note that with separate PME nodes we get the real energies later */
 524         // TODO it would be simpler if we just accumulated a single
 525         // long-range virial contribution.
 526         forceWithVirial->addVirialContribution(ewaldOutput.vir_q);
 527         forceWithVirial->addVirialContribution(ewaldOutput.vir_lj);
 528         enerd->dvdl_lin[efptCOUL] += ewaldOutput.dvdl[efptCOUL];
 529         enerd->dvdl_lin[efptVDW]  += ewaldOutput.dvdl[efptVDW];
 530         enerd->term[F_COUL_RECIP]  = Vlr_q + ewaldOutput.Vcorr_q;
 531         enerd->term[F_LJ_RECIP]    = Vlr_lj + ewaldOutput.Vcorr_lj;
 532
 533         if (debug)
 534         {
 535             fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
 536                     Vlr_q, ewaldOutput.Vcorr_q, enerd->term[F_COUL_RECIP]);
 537             pr_rvecs(debug, 0, "vir_el_recip after corr", ewaldOutput.vir_q, DIM);
 538             pr_rvecs(debug, 0, "fshift after LR Corrections", fr->fshift, SHIFTS);
 539             fprintf(debug, "Vlr_lj: %g, Vcorr_lj = %g, Vlr_corr_lj = %g\n",
 540                     Vlr_lj, ewaldOutput.Vcorr_lj, enerd->term[F_LJ_RECIP]);
 541             pr_rvecs(debug, 0, "vir_lj_recip after corr", ewaldOutput.vir_lj, DIM);
 542         }
 543     }
 544     else
 545     {
 546         /* Is there a reaction-field exclusion correction needed?
 547          * With the Verlet scheme, exclusion forces are calculated
 548          * in the non-bonded kernel.
 549          */
 550         if (useRfWithGroupScheme)
 551         {
 552             real dvdl_rf_excl      = 0;
 553             enerd->term[F_RF_EXCL] =
 554                 RF_excl_correction(fr, graph, md, excl, DOMAINDECOMP(cr),
 555                                    x, forceForUseWithShiftForces,
 556                                    fr->fshift, &pbc, lambda[efptCOUL], &dvdl_rf_excl);
 557
 558             enerd->dvdl_lin[efptCOUL] += dvdl_rf_excl;
 559         }
 560     }
 561
 562     if (debug)
 563     {
 564         print_nrnb(debug, nrnb);
 565     }
 566
 567 #if GMX_MPI
 568     if (TAKETIME)
 569     {
 570         t2 = MPI_Wtime();
 571         MPI_Barrier(cr->mpi_comm_mygroup);
 572         t3          = MPI_Wtime();
 573         fr->t_wait += t3-t2;
 574         if (fr->timesteps == 11)
 575         {
 576             char buf[22];
 577             fprintf(stderr, "* PP load balancing info: rank %d, step %s, rel wait time=%3.0f%% , load string value: %7.2f\n",
 578                     cr->nodeid, gmx_step_str(fr->timesteps, buf),
 579                     100*fr->t_wait/(fr->t_wait+fr->t_fnbf),
 580                     (fr->t_fnbf+fr->t_wait)/fr->t_fnbf);
 581         }
 582         fr->timesteps++;
 583     }
 584 #endif
 585
 586     if (debug)
 587     {
 588         pr_rvecs(debug, 0, "fshift after bondeds", fr->fshift, SHIFTS);
 589     }
 590
 591 }
 592
 593 void init_enerdata(int ngener, int n_lambda, gmx_enerdata_t *enerd)
 594 {
 595     int i, n2;
 596
 597     for (i = 0; i < F_NRE; i++)
 598     {
 599         enerd->term[i]         = 0;
 600         enerd->foreign_term[i] = 0;
 601     }
 602
 603
 604     for (i = 0; i < efptNR; i++)
 605     {
 606         enerd->dvdl_lin[i]     = 0;
 607         enerd->dvdl_nonlin[i]  = 0;
 608     }
 609
 610     n2 = ngener*ngener;
 611     if (debug)
 612     {
 613         fprintf(debug, "Creating %d sized group matrix for energies\n", n2);
 614     }
 615     enerd->grpp.nener         = n2;
 616     enerd->foreign_grpp.nener = n2;
 617     for (i = 0; (i < egNR); i++)
 618     {
 619         snew(enerd->grpp.ener[i], n2);
 620         snew(enerd->foreign_grpp.ener[i], n2);
 621     }
 622
 623     if (n_lambda)
 624     {
 625         enerd->n_lambda = 1 + n_lambda;
 626         snew(enerd->enerpart_lambda, enerd->n_lambda);
 627     }
 628     else
 629     {
 630         enerd->n_lambda = 0;
 631     }
 632 }
 633
 634 void destroy_enerdata(gmx_enerdata_t *enerd)
 635 {
 636     int i;
 637
 638     for (i = 0; (i < egNR); i++)
 639     {
 640         sfree(enerd->grpp.ener[i]);
 641     }
 642
 643     for (i = 0; (i < egNR); i++)
 644     {
 645         sfree(enerd->foreign_grpp.ener[i]);
 646     }
 647
 648     if (enerd->n_lambda)
 649     {
 650         sfree(enerd->enerpart_lambda);
 651     }
 652 }
 653
 654 static real sum_v(int n, const real v[])
 655 {
 656     real t;
 657     int  i;
 658
 659     t = 0.0;
 660     for (i = 0; (i < n); i++)
 661     {
 662         t = t + v[i];
 663     }
 664
 665     return t;
 666 }
 667
 668 void sum_epot(gmx_grppairener_t *grpp, real *epot)
 669 {
 670     int i;
 671
 672     /* Accumulate energies */
 673     epot[F_COUL_SR]  = sum_v(grpp->nener, grpp->ener[egCOULSR]);
 674     epot[F_LJ]       = sum_v(grpp->nener, grpp->ener[egLJSR]);
 675     epot[F_LJ14]     = sum_v(grpp->nener, grpp->ener[egLJ14]);
 676     epot[F_COUL14]   = sum_v(grpp->nener, grpp->ener[egCOUL14]);
 677
 678 /* lattice part of LR doesnt belong to any group
 679  * and has been added earlier
 680  */
 681     epot[F_BHAM]     = sum_v(grpp->nener, grpp->ener[egBHAMSR]);
 682
 683     epot[F_EPOT] = 0;
 684     for (i = 0; (i < F_EPOT); i++)
 685     {
 686         if (i != F_DISRESVIOL && i != F_ORIRESDEV)
 687         {
 688             epot[F_EPOT] += epot[i];
 689         }
 690     }
 691 }
 692
 693 void sum_dhdl(gmx_enerdata_t *enerd, gmx::ArrayRef<const real> lambda, t_lambda *fepvals)
 694 {
 695     int    index;
 696
 697     enerd->dvdl_lin[efptVDW] += enerd->term[F_DVDL_VDW];  /* include dispersion correction */
 698     enerd->term[F_DVDL]       = 0.0;
 699     for (int i = 0; i < efptNR; i++)
 700     {
 701         if (fepvals->separate_dvdl[i])
 702         {
 703             /* could this be done more readably/compactly? */
 704             switch (i)
 705             {
 706                 case (efptMASS):
 707                     index = F_DKDL;
 708                     break;
 709                 case (efptCOUL):
 710                     index = F_DVDL_COUL;
 711                     break;
 712                 case (efptVDW):
 713                     index = F_DVDL_VDW;
 714                     break;
 715                 case (efptBONDED):
 716                     index = F_DVDL_BONDED;
 717                     break;
 718                 case (efptRESTRAINT):
 719                     index = F_DVDL_RESTRAINT;
 720                     break;
 721                 default:
 722                     index = F_DVDL;
 723                     break;
 724             }
 725             enerd->term[index] = enerd->dvdl_lin[i] + enerd->dvdl_nonlin[i];
 726             if (debug)
 727             {
 728                 fprintf(debug, "dvdl-%s[%2d]: %f: non-linear %f + linear %f\n",
 729                         efpt_names[i], i, enerd->term[index], enerd->dvdl_nonlin[i], enerd->dvdl_lin[i]);
 730             }
 731         }
 732         else
 733         {
 734             enerd->term[F_DVDL] += enerd->dvdl_lin[i] + enerd->dvdl_nonlin[i];
 735             if (debug)
 736             {
 737                 fprintf(debug, "dvd-%sl[%2d]: %f: non-linear %f + linear %f\n",
 738                         efpt_names[0], i, enerd->term[F_DVDL], enerd->dvdl_nonlin[i], enerd->dvdl_lin[i]);
 739             }
 740         }
 741     }
 742
 743     if (fepvals->separate_dvdl[efptBONDED])
 744     {
 745         enerd->term[F_DVDL_BONDED] += enerd->term[F_DVDL_CONSTR];
 746     }
 747     else
 748     {
 749         enerd->term[F_DVDL] += enerd->term[F_DVDL_CONSTR];
 750     }
 751
 752     for (int i = 0; i < fepvals->n_lambda; i++)
 753     {
 754         /* note we are iterating over fepvals here!
 755            For the current lam, dlam = 0 automatically,
 756            so we don't need to add anything to the
 757            enerd->enerpart_lambda[0] */
 758
 759         /* we don't need to worry about dvdl_lin contributions to dE at
 760            current lambda, because the contributions to the current
 761            lambda are automatically zeroed */
 762
 763         double &enerpart_lambda = enerd->enerpart_lambda[i + 1];
 764
 765         for (gmx::index j = 0; j < lambda.ssize(); j++)
 766         {
 767             /* Note that this loop is over all dhdl components, not just the separated ones */
 768             const double dlam  = fepvals->all_lambda[j][i] - lambda[j];
 769
 770             enerpart_lambda   += dlam*enerd->dvdl_lin[j];
 771
 772             /* Constraints can not be evaluated at foreign lambdas, so we add
 773              * a linear extrapolation. This is an approximation, but usually
 774              * quite accurate since constraints change little between lambdas.
 775              */
 776             if ((j == efptBONDED && fepvals->separate_dvdl[efptBONDED]) ||
 777                 (j == efptFEP && !fepvals->separate_dvdl[efptBONDED]))
 778             {
 779                 enerpart_lambda += dlam*enerd->term[F_DVDL_CONSTR];
 780             }
 781
 782             if (j == efptMASS && !fepvals->separate_dvdl[j])
 783             {
 784                 enerpart_lambda += dlam*enerd->term[F_DKDL];
 785             }
 786
 787             if (debug)
 788             {
 789                 fprintf(debug, "enerdiff lam %g: (%15s), non-linear %f linear %f*%f\n",
 790                         fepvals->all_lambda[j][i], efpt_names[j],
 791                         enerpart_lambda - enerd->enerpart_lambda[0],
 792                         dlam, enerd->dvdl_lin[j]);
 793             }
 794         }
 795     }
 796
 797     /* The constrain contribution is now included in other terms, so clear it */
 798     enerd->term[F_DVDL_CONSTR] = 0;
 799 }
 800
 801
 802 void reset_foreign_enerdata(gmx_enerdata_t *enerd)
 803 {
 804     int  i, j;
 805
 806     /* First reset all foreign energy components.  Foreign energies always called on
 807        neighbor search steps */
 808     for (i = 0; (i < egNR); i++)
 809     {
 810         for (j = 0; (j < enerd->grpp.nener); j++)
 811         {
 812             enerd->foreign_grpp.ener[i][j] = 0.0;
 813         }
 814     }
 815
 816     /* potential energy components */
 817     for (i = 0; (i <= F_EPOT); i++)
 818     {
 819         enerd->foreign_term[i] = 0.0;
 820     }
 821 }
 822
 823 void reset_enerdata(gmx_enerdata_t *enerd)
 824 {
 825     int      i, j;
 826
 827     /* First reset all energy components. */
 828     for (i = 0; (i < egNR); i++)
 829     {
 830         for (j = 0; (j < enerd->grpp.nener); j++)
 831         {
 832             enerd->grpp.ener[i][j] = 0.0;
 833         }
 834     }
 835     for (i = 0; i < efptNR; i++)
 836     {
 837         enerd->dvdl_lin[i]    = 0.0;
 838         enerd->dvdl_nonlin[i] = 0.0;
 839     }
 840
 841     /* Normal potential energy components */
 842     for (i = 0; (i <= F_EPOT); i++)
 843     {
 844         enerd->term[i] = 0.0;
 845     }
 846     enerd->term[F_DVDL]            = 0.0;
 847     enerd->term[F_DVDL_COUL]       = 0.0;
 848     enerd->term[F_DVDL_VDW]        = 0.0;
 849     enerd->term[F_DVDL_BONDED]     = 0.0;
 850     enerd->term[F_DVDL_RESTRAINT]  = 0.0;
 851     enerd->term[F_DKDL]            = 0.0;
 852     if (enerd->n_lambda > 0)
 853     {
 854         for (i = 0; i < enerd->n_lambda; i++)
 855         {
 856             enerd->enerpart_lambda[i] = 0.0;
 857         }
 858     }
 859     /* reset foreign energy data - separate function since we also call it elsewhere */
 860     reset_foreign_enerdata(enerd);
 861 }