src/gromacs/mdlib/sim_util.c

   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, 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 #ifdef HAVE_CONFIG_H
  38 #include <config.h>
  39 #endif
  40
  41 #include <stdio.h>
  42 #ifdef HAVE_SYS_TIME_H
  43 #include <sys/time.h>
  44 #endif
  45 #include <math.h>
  46 #include <assert.h>
  47
  48 #include "typedefs.h"
  49 #include "string2.h"
  50 #include "smalloc.h"
  51 #include "names.h"
  52 #include "txtdump.h"
  53 #include "pbc.h"
  54 #include "chargegroup.h"
  55 #include "vec.h"
  56 #include "nrnb.h"
  57 #include "mshift.h"
  58 #include "mdrun.h"
  59 #include "sim_util.h"
  60 #include "update.h"
  61 #include "physics.h"
  62 #include "main.h"
  63 #include "mdatoms.h"
  64 #include "force.h"
  65 #include "bondf.h"
  66 #include "pme.h"
  67 #include "disre.h"
  68 #include "orires.h"
  69 #include "network.h"
  70 #include "calcmu.h"
  71 #include "constr.h"
  72 #include "xvgr.h"
  73 #include "copyrite.h"
  74 #include "domdec.h"
  75 #include "genborn.h"
  76 #include "nbnxn_atomdata.h"
  77 #include "nbnxn_search.h"
  78 #include "nbnxn_kernels/nbnxn_kernel_ref.h"
  79 #include "nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.h"
  80 #include "nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn.h"
  81 #include "nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  82 #include "nonbonded.h"
  83 #include "../gmxlib/nonbonded/nb_kernel.h"
  84 #include "../gmxlib/nonbonded/nb_free_energy.h"
  85
  86 #include "gromacs/timing/wallcycle.h"
  87 #include "gromacs/timing/walltime_accounting.h"
  88 #include "gromacs/utility/gmxmpi.h"
  89 #include "gromacs/essentialdynamics/edsam.h"
  90 #include "gromacs/pulling/pull.h"
  91 #include "gromacs/pulling/pull_rotation.h"
  92
  93 #include "adress.h"
  94 #include "qmmm.h"
  95
  96 #include "gmx_omp_nthreads.h"
  97
  98 #include "nbnxn_cuda_data_mgmt.h"
  99 #include "nbnxn_cuda/nbnxn_cuda.h"
 100
 101 void print_time(FILE                     *out,
 102                 gmx_walltime_accounting_t walltime_accounting,
 103                 gmx_int64_t               step,
 104                 t_inputrec               *ir,
 105                 t_commrec gmx_unused     *cr)
 106 {
 107     time_t finish;
 108     char   timebuf[STRLEN];
 109     double dt, elapsed_seconds, time_per_step;
 110     char   buf[48];
 111
 112 #ifndef GMX_THREAD_MPI
 113     if (!PAR(cr))
 114 #endif
 115     {
 116         fprintf(out, "\r");
 117     }
 118     fprintf(out, "step %s", gmx_step_str(step, buf));
 119     if ((step >= ir->nstlist))
 120     {
 121         double seconds_since_epoch = gmx_gettime();
 122         elapsed_seconds = seconds_since_epoch - walltime_accounting_get_start_time_stamp(walltime_accounting);
 123         time_per_step   = elapsed_seconds/(step - ir->init_step + 1);
 124         dt              = (ir->nsteps + ir->init_step - step) * time_per_step;
 125
 126         if (ir->nsteps >= 0)
 127         {
 128             if (dt >= 300)
 129             {
 130                 finish = (time_t) (seconds_since_epoch + dt);
 131                 gmx_ctime_r(&finish, timebuf, STRLEN);
 132                 sprintf(buf, "%s", timebuf);
 133                 buf[strlen(buf)-1] = '\0';
 134                 fprintf(out, ", will finish %s", buf);
 135             }
 136             else
 137             {
 138                 fprintf(out, ", remaining wall clock time: %5d s          ", (int)dt);
 139             }
 140         }
 141         else
 142         {
 143             fprintf(out, " performance: %.1f ns/day    ",
 144                     ir->delta_t/1000*24*60*60/time_per_step);
 145         }
 146     }
 147 #ifndef GMX_THREAD_MPI
 148     if (PAR(cr))
 149     {
 150         fprintf(out, "\n");
 151     }
 152 #endif
 153
 154     fflush(out);
 155 }
 156
 157 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
 158                          const gmx_walltime_accounting_t walltime_accounting)
 159 {
 160     int    i;
 161     char   timebuf[STRLEN];
 162     char   time_string[STRLEN];
 163     time_t tmptime;
 164
 165     if (fplog)
 166     {
 167         if (walltime_accounting != NULL)
 168         {
 169             tmptime = (time_t) walltime_accounting_get_start_time_stamp(walltime_accounting);
 170             gmx_ctime_r(&tmptime, timebuf, STRLEN);
 171         }
 172         else
 173         {
 174             tmptime = (time_t) gmx_gettime();
 175             gmx_ctime_r(&tmptime, timebuf, STRLEN);
 176         }
 177         for (i = 0; timebuf[i] >= ' '; i++)
 178         {
 179             time_string[i] = timebuf[i];
 180         }
 181         time_string[i] = '\0';
 182
 183         fprintf(fplog, "%s on node %d %s\n", title, nodeid, time_string);
 184     }
 185 }
 186
 187 void print_start(FILE *fplog, t_commrec *cr,
 188                  gmx_walltime_accounting_t walltime_accounting,
 189                  const char *name)
 190 {
 191     char buf[STRLEN];
 192
 193     sprintf(buf, "Started %s", name);
 194     print_date_and_time(fplog, cr->nodeid, buf, walltime_accounting);
 195 }
 196
 197 static void sum_forces(int start, int end, rvec f[], rvec flr[])
 198 {
 199     int i;
 200
 201     if (gmx_debug_at)
 202     {
 203         pr_rvecs(debug, 0, "fsr", f+start, end-start);
 204         pr_rvecs(debug, 0, "flr", flr+start, end-start);
 205     }
 206     for (i = start; (i < end); i++)
 207     {
 208         rvec_inc(f[i], flr[i]);
 209     }
 210 }
 211
 212 /*
 213  * calc_f_el calculates forces due to an electric field.
 214  *
 215  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 216  *
 217  * Et[] contains the parameters for the time dependent
 218  * part of the field (not yet used).
 219  * Ex[] contains the parameters for
 220  * the spatial dependent part of the field. You can have cool periodic
 221  * fields in principle, but only a constant field is supported
 222  * now.
 223  * The function should return the energy due to the electric field
 224  * (if any) but for now returns 0.
 225  *
 226  * WARNING:
 227  * There can be problems with the virial.
 228  * Since the field is not self-consistent this is unavoidable.
 229  * For neutral molecules the virial is correct within this approximation.
 230  * For neutral systems with many charged molecules the error is small.
 231  * But for systems with a net charge or a few charged molecules
 232  * the error can be significant when the field is high.
 233  * Solution: implement a self-consitent electric field into PME.
 234  */
 235 static void calc_f_el(FILE *fp, int  start, int homenr,
 236                       real charge[], rvec f[],
 237                       t_cosines Ex[], t_cosines Et[], double t)
 238 {
 239     rvec Ext;
 240     real t0;
 241     int  i, m;
 242
 243     for (m = 0; (m < DIM); m++)
 244     {
 245         if (Et[m].n > 0)
 246         {
 247             if (Et[m].n == 3)
 248             {
 249                 t0     = Et[m].a[1];
 250                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 251             }
 252             else
 253             {
 254                 Ext[m] = cos(Et[m].a[0]*t);
 255             }
 256         }
 257         else
 258         {
 259             Ext[m] = 1.0;
 260         }
 261         if (Ex[m].n > 0)
 262         {
 263             /* Convert the field strength from V/nm to MD-units */
 264             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 265             for (i = start; (i < start+homenr); i++)
 266             {
 267                 f[i][m] += charge[i]*Ext[m];
 268             }
 269         }
 270         else
 271         {
 272             Ext[m] = 0;
 273         }
 274     }
 275     if (fp != NULL)
 276     {
 277         fprintf(fp, "%10g  %10g  %10g  %10g #FIELD\n", t,
 278                 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
 279     }
 280 }
 281
 282 static void calc_virial(int start, int homenr, rvec x[], rvec f[],
 283                         tensor vir_part, t_graph *graph, matrix box,
 284                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 285 {
 286     int    i, j;
 287     tensor virtest;
 288
 289     /* The short-range virial from surrounding boxes */
 290     clear_mat(vir_part);
 291     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 292     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 293
 294     /* Calculate partial virial, for local atoms only, based on short range.
 295      * Total virial is computed in global_stat, called from do_md
 296      */
 297     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 298     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 299
 300     /* Add position restraint contribution */
 301     for (i = 0; i < DIM; i++)
 302     {
 303         vir_part[i][i] += fr->vir_diag_posres[i];
 304     }
 305
 306     /* Add wall contribution */
 307     for (i = 0; i < DIM; i++)
 308     {
 309         vir_part[i][ZZ] += fr->vir_wall_z[i];
 310     }
 311
 312     if (debug)
 313     {
 314         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 315     }
 316 }
 317
 318 static void posres_wrapper(FILE *fplog,
 319                            int flags,
 320                            gmx_bool bSepDVDL,
 321                            t_inputrec *ir,
 322                            t_nrnb *nrnb,
 323                            gmx_localtop_t *top,
 324                            matrix box, rvec x[],
 325                            gmx_enerdata_t *enerd,
 326                            real *lambda,
 327                            t_forcerec *fr)
 328 {
 329     t_pbc pbc;
 330     real  v, dvdl;
 331     int   i;
 332
 333     /* Position restraints always require full pbc */
 334     set_pbc(&pbc, ir->ePBC, box);
 335     dvdl = 0;
 336     v    = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 337                   top->idef.iparams_posres,
 338                   (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 339                   ir->ePBC == epbcNONE ? NULL : &pbc,
 340                   lambda[efptRESTRAINT], &dvdl,
 341                   fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 342     if (bSepDVDL)
 343     {
 344         gmx_print_sepdvdl(fplog, interaction_function[F_POSRES].longname, v, dvdl);
 345     }
 346     enerd->term[F_POSRES] += v;
 347     /* If just the force constant changes, the FEP term is linear,
 348      * but if k changes, it is not.
 349      */
 350     enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
 351     inc_nrnb(nrnb, eNR_POSRES, top->idef.il[F_POSRES].nr/2);
 352
 353     if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
 354     {
 355         for (i = 0; i < enerd->n_lambda; i++)
 356         {
 357             real dvdl_dum, lambda_dum;
 358
 359             lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
 360             v          = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 361                                 top->idef.iparams_posres,
 362                                 (const rvec*)x, NULL, NULL,
 363                                 ir->ePBC == epbcNONE ? NULL : &pbc, lambda_dum, &dvdl,
 364                                 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 365             enerd->enerpart_lambda[i] += v;
 366         }
 367     }
 368 }
 369
 370 static void fbposres_wrapper(t_inputrec *ir,
 371                              t_nrnb *nrnb,
 372                              gmx_localtop_t *top,
 373                              matrix box, rvec x[],
 374                              gmx_enerdata_t *enerd,
 375                              t_forcerec *fr)
 376 {
 377     t_pbc pbc;
 378     real  v;
 379
 380     /* Flat-bottomed position restraints always require full pbc */
 381     set_pbc(&pbc, ir->ePBC, box);
 382     v = fbposres(top->idef.il[F_FBPOSRES].nr, top->idef.il[F_FBPOSRES].iatoms,
 383                  top->idef.iparams_fbposres,
 384                  (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 385                  ir->ePBC == epbcNONE ? NULL : &pbc,
 386                  fr->rc_scaling, fr->ePBC, fr->posres_com);
 387     enerd->term[F_FBPOSRES] += v;
 388     inc_nrnb(nrnb, eNR_FBPOSRES, top->idef.il[F_FBPOSRES].nr/2);
 389 }
 390
 391 static void pull_potential_wrapper(FILE *fplog,
 392                                    gmx_bool bSepDVDL,
 393                                    t_commrec *cr,
 394                                    t_inputrec *ir,
 395                                    matrix box, rvec x[],
 396                                    rvec f[],
 397                                    tensor vir_force,
 398                                    t_mdatoms *mdatoms,
 399                                    gmx_enerdata_t *enerd,
 400                                    real *lambda,
 401                                    double t)
 402 {
 403     t_pbc  pbc;
 404     real   dvdl;
 405
 406     /* Calculate the center of mass forces, this requires communication,
 407      * which is why pull_potential is called close to other communication.
 408      * The virial contribution is calculated directly,
 409      * which is why we call pull_potential after calc_virial.
 410      */
 411     set_pbc(&pbc, ir->ePBC, box);
 412     dvdl                     = 0;
 413     enerd->term[F_COM_PULL] +=
 414         pull_potential(ir->ePull, ir->pull, mdatoms, &pbc,
 415                        cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
 416     if (bSepDVDL)
 417     {
 418         gmx_print_sepdvdl(fplog, "Com pull", enerd->term[F_COM_PULL], dvdl);
 419     }
 420     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 421 }
 422
 423 static void pme_receive_force_ener(FILE           *fplog,
 424                                    gmx_bool        bSepDVDL,
 425                                    t_commrec      *cr,
 426                                    gmx_wallcycle_t wcycle,
 427                                    gmx_enerdata_t *enerd,
 428                                    t_forcerec     *fr)
 429 {
 430     real   e_q, e_lj, v, dvdl_q, dvdl_lj;
 431     float  cycles_ppdpme, cycles_seppme;
 432
 433     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 434     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 435
 436     /* In case of node-splitting, the PP nodes receive the long-range
 437      * forces, virial and energy from the PME nodes here.
 438      */
 439     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 440     dvdl_q  = 0;
 441     dvdl_lj = 0;
 442     gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e_q,
 443                       fr->vir_lj_recip, &e_lj, &dvdl_q, &dvdl_lj,
 444                       &cycles_seppme);
 445     if (bSepDVDL)
 446     {
 447         gmx_print_sepdvdl(fplog, "Electrostatic PME mesh", e_q, dvdl_q);
 448         gmx_print_sepdvdl(fplog, "Lennard-Jones PME mesh", e_lj, dvdl_lj);
 449     }
 450     enerd->term[F_COUL_RECIP] += e_q;
 451     enerd->term[F_LJ_RECIP]   += e_lj;
 452     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 453     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 454
 455     if (wcycle)
 456     {
 457         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 458     }
 459     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 460 }
 461
 462 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
 463                                gmx_int64_t step, real pforce, rvec *x, rvec *f)
 464 {
 465     int  i;
 466     real pf2, fn2;
 467     char buf[STEPSTRSIZE];
 468
 469     pf2 = sqr(pforce);
 470     for (i = 0; i < md->homenr; i++)
 471     {
 472         fn2 = norm2(f[i]);
 473         /* We also catch NAN, if the compiler does not optimize this away. */
 474         if (fn2 >= pf2 || fn2 != fn2)
 475         {
 476             fprintf(fp, "step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 477                     gmx_step_str(step, buf),
 478                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
 479         }
 480     }
 481 }
 482
 483 static void post_process_forces(t_commrec *cr,
 484                                 gmx_int64_t step,
 485                                 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 486                                 gmx_localtop_t *top,
 487                                 matrix box, rvec x[],
 488                                 rvec f[],
 489                                 tensor vir_force,
 490                                 t_mdatoms *mdatoms,
 491                                 t_graph *graph,
 492                                 t_forcerec *fr, gmx_vsite_t *vsite,
 493                                 int flags)
 494 {
 495     if (fr->bF_NoVirSum)
 496     {
 497         if (vsite)
 498         {
 499             /* Spread the mesh force on virtual sites to the other particles...
 500              * This is parallellized. MPI communication is performed
 501              * if the constructing atoms aren't local.
 502              */
 503             wallcycle_start(wcycle, ewcVSITESPREAD);
 504             spread_vsite_f(vsite, x, fr->f_novirsum, NULL,
 505                            (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
 506                            nrnb,
 507                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
 508             wallcycle_stop(wcycle, ewcVSITESPREAD);
 509         }
 510         if (flags & GMX_FORCE_VIRIAL)
 511         {
 512             /* Now add the forces, this is local */
 513             if (fr->bDomDec)
 514             {
 515                 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
 516             }
 517             else
 518             {
 519                 sum_forces(0, mdatoms->homenr,
 520                            f, fr->f_novirsum);
 521             }
 522             if (EEL_FULL(fr->eeltype))
 523             {
 524                 /* Add the mesh contribution to the virial */
 525                 m_add(vir_force, fr->vir_el_recip, vir_force);
 526             }
 527             if (EVDW_PME(fr->vdwtype))
 528             {
 529                 /* Add the mesh contribution to the virial */
 530                 m_add(vir_force, fr->vir_lj_recip, vir_force);
 531             }
 532             if (debug)
 533             {
 534                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 535             }
 536         }
 537     }
 538
 539     if (fr->print_force >= 0)
 540     {
 541         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 542     }
 543 }
 544
 545 static void do_nb_verlet(t_forcerec *fr,
 546                          interaction_const_t *ic,
 547                          gmx_enerdata_t *enerd,
 548                          int flags, int ilocality,
 549                          int clearF,
 550                          t_nrnb *nrnb,
 551                          gmx_wallcycle_t wcycle)
 552 {
 553     int                        nnbl, kernel_type, enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
 554     char                      *env;
 555     nonbonded_verlet_group_t  *nbvg;
 556     gmx_bool                   bCUDA;
 557
 558     if (!(flags & GMX_FORCE_NONBONDED))
 559     {
 560         /* skip non-bonded calculation */
 561         return;
 562     }
 563
 564     nbvg = &fr->nbv->grp[ilocality];
 565
 566     /* CUDA kernel launch overhead is already timed separately */
 567     if (fr->cutoff_scheme != ecutsVERLET)
 568     {
 569         gmx_incons("Invalid cut-off scheme passed!");
 570     }
 571
 572     bCUDA = (nbvg->kernel_type == nbnxnk8x8x8_CUDA);
 573
 574     if (!bCUDA)
 575     {
 576         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 577     }
 578     switch (nbvg->kernel_type)
 579     {
 580         case nbnxnk4x4_PlainC:
 581             nbnxn_kernel_ref(&nbvg->nbl_lists,
 582                              nbvg->nbat, ic,
 583                              fr->shift_vec,
 584                              flags,
 585                              clearF,
 586                              fr->fshift[0],
 587                              enerd->grpp.ener[egCOULSR],
 588                              fr->bBHAM ?
 589                              enerd->grpp.ener[egBHAMSR] :
 590                              enerd->grpp.ener[egLJSR]);
 591             break;
 592
 593         case nbnxnk4xN_SIMD_4xN:
 594             nbnxn_kernel_simd_4xn(&nbvg->nbl_lists,
 595                                   nbvg->nbat, ic,
 596                                   nbvg->ewald_excl,
 597                                   fr->shift_vec,
 598                                   flags,
 599                                   clearF,
 600                                   fr->fshift[0],
 601                                   enerd->grpp.ener[egCOULSR],
 602                                   fr->bBHAM ?
 603                                   enerd->grpp.ener[egBHAMSR] :
 604                                   enerd->grpp.ener[egLJSR]);
 605             break;
 606         case nbnxnk4xN_SIMD_2xNN:
 607             nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
 608                                    nbvg->nbat, ic,
 609                                    nbvg->ewald_excl,
 610                                    fr->shift_vec,
 611                                    flags,
 612                                    clearF,
 613                                    fr->fshift[0],
 614                                    enerd->grpp.ener[egCOULSR],
 615                                    fr->bBHAM ?
 616                                    enerd->grpp.ener[egBHAMSR] :
 617                                    enerd->grpp.ener[egLJSR]);
 618             break;
 619
 620         case nbnxnk8x8x8_CUDA:
 621             nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
 622             break;
 623
 624         case nbnxnk8x8x8_PlainC:
 625             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 626                                  nbvg->nbat, ic,
 627                                  fr->shift_vec,
 628                                  flags,
 629                                  clearF,
 630                                  nbvg->nbat->out[0].f,
 631                                  fr->fshift[0],
 632                                  enerd->grpp.ener[egCOULSR],
 633                                  fr->bBHAM ?
 634                                  enerd->grpp.ener[egBHAMSR] :
 635                                  enerd->grpp.ener[egLJSR]);
 636             break;
 637
 638         default:
 639             gmx_incons("Invalid nonbonded kernel type passed!");
 640
 641     }
 642     if (!bCUDA)
 643     {
 644         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 645     }
 646
 647     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 648     {
 649         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 650     }
 651     else if ((!bCUDA && nbvg->ewald_excl == ewaldexclAnalytical) ||
 652              (bCUDA && nbnxn_cuda_is_kernel_ewald_analytical(fr->nbv->cu_nbv)))
 653     {
 654         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 655     }
 656     else
 657     {
 658         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 659     }
 660     enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 661     if (flags & GMX_FORCE_ENERGY)
 662     {
 663         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 664         enr_nbnxn_kernel_ljc += 1;
 665         enr_nbnxn_kernel_lj  += 1;
 666     }
 667
 668     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 669              nbvg->nbl_lists.natpair_ljq);
 670     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 671              nbvg->nbl_lists.natpair_lj);
 672     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 673     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 674              nbvg->nbl_lists.natpair_q);
 675
 676     if (ic->vdw_modifier == eintmodFORCESWITCH)
 677     {
 678         /* We add up the switch cost separately */
 679         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_FSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 680                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 681     }
 682     if (ic->vdw_modifier == eintmodPOTSWITCH)
 683     {
 684         /* We add up the switch cost separately */
 685         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_PSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 686                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 687     }
 688     if (ic->vdwtype == evdwPME)
 689     {
 690         /* We add up the LJ Ewald cost separately */
 691         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_EWALD+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 692                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 693     }
 694 }
 695
 696 static void do_nb_verlet_fep(nbnxn_pairlist_set_t *nbl_lists,
 697                              t_forcerec           *fr,
 698                              rvec                  x[],
 699                              rvec                  f[],
 700                              t_mdatoms            *mdatoms,
 701                              t_lambda             *fepvals,
 702                              real                 *lambda,
 703                              gmx_enerdata_t       *enerd,
 704                              int                   flags,
 705                              t_nrnb               *nrnb,
 706                              gmx_wallcycle_t       wcycle)
 707 {
 708     int              donb_flags;
 709     nb_kernel_data_t kernel_data;
 710     real             lam_i[efptNR];
 711     real             dvdl_nb[efptNR];
 712     int              th;
 713     int              i, j;
 714
 715     donb_flags = 0;
 716     /* Add short-range interactions */
 717     donb_flags |= GMX_NONBONDED_DO_SR;
 718
 719     /* Currently all group scheme kernels always calculate (shift-)forces */
 720     if (flags & GMX_FORCE_FORCES)
 721     {
 722         donb_flags |= GMX_NONBONDED_DO_FORCE;
 723     }
 724     if (flags & GMX_FORCE_VIRIAL)
 725     {
 726         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 727     }
 728     if (flags & GMX_FORCE_ENERGY)
 729     {
 730         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 731     }
 732     if (flags & GMX_FORCE_DO_LR)
 733     {
 734         donb_flags |= GMX_NONBONDED_DO_LR;
 735     }
 736
 737     kernel_data.flags  = donb_flags;
 738     kernel_data.lambda = lambda;
 739     kernel_data.dvdl   = dvdl_nb;
 740
 741     kernel_data.energygrp_elec = enerd->grpp.ener[egCOULSR];
 742     kernel_data.energygrp_vdw  = enerd->grpp.ener[egLJSR];
 743
 744     /* reset free energy components */
 745     for (i = 0; i < efptNR; i++)
 746     {
 747         dvdl_nb[i]  = 0;
 748     }
 749
 750     assert(gmx_omp_nthreads_get(emntNonbonded) == nbl_lists->nnbl);
 751
 752     wallcycle_sub_start(wcycle, ewcsNONBONDED);
 753 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 754     for (th = 0; th < nbl_lists->nnbl; th++)
 755     {
 756         gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 757                                   x, f, fr, mdatoms, &kernel_data, nrnb);
 758     }
 759
 760     if (fepvals->sc_alpha != 0)
 761     {
 762         enerd->dvdl_nonlin[efptVDW]  += dvdl_nb[efptVDW];
 763         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 764     }
 765     else
 766     {
 767         enerd->dvdl_lin[efptVDW]  += dvdl_nb[efptVDW];
 768         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 769     }
 770
 771     /* If we do foreign lambda and we have soft-core interactions
 772      * we have to recalculate the (non-linear) energies contributions.
 773      */
 774     if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
 775     {
 776         kernel_data.flags          = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE)) | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 777         kernel_data.lambda         = lam_i;
 778         kernel_data.energygrp_elec = enerd->foreign_grpp.ener[egCOULSR];
 779         kernel_data.energygrp_vdw  = enerd->foreign_grpp.ener[egLJSR];
 780         /* Note that we add to kernel_data.dvdl, but ignore the result */
 781
 782         for (i = 0; i < enerd->n_lambda; i++)
 783         {
 784             for (j = 0; j < efptNR; j++)
 785             {
 786                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
 787             }
 788             reset_foreign_enerdata(enerd);
 789 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 790             for (th = 0; th < nbl_lists->nnbl; th++)
 791             {
 792                 gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 793                                           x, f, fr, mdatoms, &kernel_data, nrnb);
 794             }
 795
 796             sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
 797             enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
 798         }
 799     }
 800
 801     wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 802 }
 803
 804 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
 805                          t_inputrec *inputrec,
 806                          gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 807                          gmx_localtop_t *top,
 808                          gmx_groups_t gmx_unused *groups,
 809                          matrix box, rvec x[], history_t *hist,
 810                          rvec f[],
 811                          tensor vir_force,
 812                          t_mdatoms *mdatoms,
 813                          gmx_enerdata_t *enerd, t_fcdata *fcd,
 814                          real *lambda, t_graph *graph,
 815                          t_forcerec *fr, interaction_const_t *ic,
 816                          gmx_vsite_t *vsite, rvec mu_tot,
 817                          double t, FILE *field, gmx_edsam_t ed,
 818                          gmx_bool bBornRadii,
 819                          int flags)
 820 {
 821     int                 cg0, cg1, i, j;
 822     int                 start, homenr;
 823     int                 nb_kernel_type;
 824     double              mu[2*DIM];
 825     gmx_bool            bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
 826     gmx_bool            bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
 827     gmx_bool            bDiffKernels = FALSE;
 828     matrix              boxs;
 829     rvec                vzero, box_diag;
 830     real                e, v, dvdl;
 831     float               cycles_pme, cycles_force, cycles_wait_gpu;
 832     nonbonded_verlet_t *nbv;
 833
 834     cycles_force    = 0;
 835     cycles_wait_gpu = 0;
 836     nbv             = fr->nbv;
 837     nb_kernel_type  = fr->nbv->grp[0].kernel_type;
 838
 839     start  = 0;
 840     homenr = mdatoms->homenr;
 841
 842     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
 843
 844     clear_mat(vir_force);
 845
 846     cg0 = 0;
 847     if (DOMAINDECOMP(cr))
 848     {
 849         cg1 = cr->dd->ncg_tot;
 850     }
 851     else
 852     {
 853         cg1 = top->cgs.nr;
 854     }
 855     if (fr->n_tpi > 0)
 856     {
 857         cg1--;
 858     }
 859
 860     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 861     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
 862     bFillGrid     = (bNS && bStateChanged);
 863     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 864     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 865     bDoForces     = (flags & GMX_FORCE_FORCES);
 866     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 867     bUseGPU       = fr->nbv->bUseGPU;
 868     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
 869
 870     if (bStateChanged)
 871     {
 872         update_forcerec(fr, box);
 873
 874         if (NEED_MUTOT(*inputrec))
 875         {
 876             /* Calculate total (local) dipole moment in a temporary common array.
 877              * This makes it possible to sum them over nodes faster.
 878              */
 879             calc_mu(start, homenr,
 880                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 881                     mu, mu+DIM);
 882         }
 883     }
 884
 885     if (fr->ePBC != epbcNONE)
 886     {
 887         /* Compute shift vectors every step,
 888          * because of pressure coupling or box deformation!
 889          */
 890         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 891         {
 892             calc_shifts(box, fr->shift_vec);
 893         }
 894
 895         if (bCalcCGCM)
 896         {
 897             put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
 898             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 899         }
 900         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 901         {
 902             unshift_self(graph, box, x);
 903         }
 904     }
 905
 906     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 907                                  fr->shift_vec, nbv->grp[0].nbat);
 908
 909 #ifdef GMX_MPI
 910     if (!(cr->duty & DUTY_PME))
 911     {
 912         /* Send particle coordinates to the pme nodes.
 913          * Since this is only implemented for domain decomposition
 914          * and domain decomposition does not use the graph,
 915          * we do not need to worry about shifting.
 916          */
 917
 918         int pme_flags = 0;
 919
 920         wallcycle_start(wcycle, ewcPP_PMESENDX);
 921
 922         bBS = (inputrec->nwall == 2);
 923         if (bBS)
 924         {
 925             copy_mat(box, boxs);
 926             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
 927         }
 928
 929         if (EEL_PME(fr->eeltype))
 930         {
 931             pme_flags |= GMX_PME_DO_COULOMB;
 932         }
 933
 934         if (EVDW_PME(fr->vdwtype))
 935         {
 936             pme_flags |= GMX_PME_DO_LJ;
 937         }
 938
 939         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
 940                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
 941                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
 942                                  pme_flags, step);
 943
 944         wallcycle_stop(wcycle, ewcPP_PMESENDX);
 945     }
 946 #endif /* GMX_MPI */
 947
 948     /* do gridding for pair search */
 949     if (bNS)
 950     {
 951         if (graph && bStateChanged)
 952         {
 953             /* Calculate intramolecular shift vectors to make molecules whole */
 954             mk_mshift(fplog, graph, fr->ePBC, box, x);
 955         }
 956
 957         clear_rvec(vzero);
 958         box_diag[XX] = box[XX][XX];
 959         box_diag[YY] = box[YY][YY];
 960         box_diag[ZZ] = box[ZZ][ZZ];
 961
 962         wallcycle_start(wcycle, ewcNS);
 963         if (!fr->bDomDec)
 964         {
 965             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 966             nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
 967                               0, vzero, box_diag,
 968                               0, mdatoms->homenr, -1, fr->cginfo, x,
 969                               0, NULL,
 970                               nbv->grp[eintLocal].kernel_type,
 971                               nbv->grp[eintLocal].nbat);
 972             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 973         }
 974         else
 975         {
 976             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 977             nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
 978                                        fr->cginfo, x,
 979                                        nbv->grp[eintNonlocal].kernel_type,
 980                                        nbv->grp[eintNonlocal].nbat);
 981             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 982         }
 983
 984         if (nbv->ngrp == 1 ||
 985             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 986         {
 987             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
 988                                nbv->nbs, mdatoms, fr->cginfo);
 989         }
 990         else
 991         {
 992             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
 993                                nbv->nbs, mdatoms, fr->cginfo);
 994             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
 995                                nbv->nbs, mdatoms, fr->cginfo);
 996         }
 997         wallcycle_stop(wcycle, ewcNS);
 998     }
 999
1000     /* initialize the GPU atom data and copy shift vector */
1001     if (bUseGPU)
1002     {
1003         if (bNS)
1004         {
1005             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1006             nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
1007             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1008         }
1009
1010         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1011         nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
1012         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1013     }
1014
1015     /* do local pair search */
1016     if (bNS)
1017     {
1018         wallcycle_start_nocount(wcycle, ewcNS);
1019         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1020         nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
1021                             &top->excls,
1022                             ic->rlist,
1023                             nbv->min_ci_balanced,
1024                             &nbv->grp[eintLocal].nbl_lists,
1025                             eintLocal,
1026                             nbv->grp[eintLocal].kernel_type,
1027                             nrnb);
1028         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1029
1030         if (bUseGPU)
1031         {
1032             /* initialize local pair-list on the GPU */
1033             nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1034                                      nbv->grp[eintLocal].nbl_lists.nbl[0],
1035                                      eintLocal);
1036         }
1037         wallcycle_stop(wcycle, ewcNS);
1038     }
1039     else
1040     {
1041         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1042         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1043         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
1044                                         nbv->grp[eintLocal].nbat);
1045         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1046         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1047     }
1048
1049     if (bUseGPU)
1050     {
1051         wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1052         /* launch local nonbonded F on GPU */
1053         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
1054                      nrnb, wcycle);
1055         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1056     }
1057
1058     /* Communicate coordinates and sum dipole if necessary +
1059        do non-local pair search */
1060     if (DOMAINDECOMP(cr))
1061     {
1062         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
1063                         nbv->grp[eintLocal].kernel_type);
1064
1065         if (bDiffKernels)
1066         {
1067             /* With GPU+CPU non-bonded calculations we need to copy
1068              * the local coordinates to the non-local nbat struct
1069              * (in CPU format) as the non-local kernel call also
1070              * calculates the local - non-local interactions.
1071              */
1072             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1073             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1074             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
1075                                             nbv->grp[eintNonlocal].nbat);
1076             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1077             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1078         }
1079
1080         if (bNS)
1081         {
1082             wallcycle_start_nocount(wcycle, ewcNS);
1083             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1084
1085             if (bDiffKernels)
1086             {
1087                 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
1088             }
1089
1090             nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
1091                                 &top->excls,
1092                                 ic->rlist,
1093                                 nbv->min_ci_balanced,
1094                                 &nbv->grp[eintNonlocal].nbl_lists,
1095                                 eintNonlocal,
1096                                 nbv->grp[eintNonlocal].kernel_type,
1097                                 nrnb);
1098
1099             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1100
1101             if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_CUDA)
1102             {
1103                 /* initialize non-local pair-list on the GPU */
1104                 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1105                                          nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1106                                          eintNonlocal);
1107             }
1108             wallcycle_stop(wcycle, ewcNS);
1109         }
1110         else
1111         {
1112             wallcycle_start(wcycle, ewcMOVEX);
1113             dd_move_x(cr->dd, box, x);
1114
1115             /* When we don't need the total dipole we sum it in global_stat */
1116             if (bStateChanged && NEED_MUTOT(*inputrec))
1117             {
1118                 gmx_sumd(2*DIM, mu, cr);
1119             }
1120             wallcycle_stop(wcycle, ewcMOVEX);
1121
1122             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1123             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1124             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1125                                             nbv->grp[eintNonlocal].nbat);
1126             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1127             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1128         }
1129
1130         if (bUseGPU && !bDiffKernels)
1131         {
1132             wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1133             /* launch non-local nonbonded F on GPU */
1134             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1135                          nrnb, wcycle);
1136             cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1137         }
1138     }
1139
1140     if (bUseGPU)
1141     {
1142         /* launch D2H copy-back F */
1143         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1144         if (DOMAINDECOMP(cr) && !bDiffKernels)
1145         {
1146             nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1147                                       flags, eatNonlocal);
1148         }
1149         nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1150                                   flags, eatLocal);
1151         cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1152     }
1153
1154     if (bStateChanged && NEED_MUTOT(*inputrec))
1155     {
1156         if (PAR(cr))
1157         {
1158             gmx_sumd(2*DIM, mu, cr);
1159         }
1160
1161         for (i = 0; i < 2; i++)
1162         {
1163             for (j = 0; j < DIM; j++)
1164             {
1165                 fr->mu_tot[i][j] = mu[i*DIM + j];
1166             }
1167         }
1168     }
1169     if (fr->efep == efepNO)
1170     {
1171         copy_rvec(fr->mu_tot[0], mu_tot);
1172     }
1173     else
1174     {
1175         for (j = 0; j < DIM; j++)
1176         {
1177             mu_tot[j] =
1178                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1179                 lambda[efptCOUL]*fr->mu_tot[1][j];
1180         }
1181     }
1182
1183     /* Reset energies */
1184     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1185     clear_rvecs(SHIFTS, fr->fshift);
1186
1187     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1188     {
1189         wallcycle_start(wcycle, ewcPPDURINGPME);
1190         dd_force_flop_start(cr->dd, nrnb);
1191     }
1192
1193     if (inputrec->bRot)
1194     {
1195         /* Enforced rotation has its own cycle counter that starts after the collective
1196          * coordinates have been communicated. It is added to ddCyclF to allow
1197          * for proper load-balancing */
1198         wallcycle_start(wcycle, ewcROT);
1199         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1200         wallcycle_stop(wcycle, ewcROT);
1201     }
1202
1203     /* Start the force cycle counter.
1204      * This counter is stopped in do_forcelow_level.
1205      * No parallel communication should occur while this counter is running,
1206      * since that will interfere with the dynamic load balancing.
1207      */
1208     wallcycle_start(wcycle, ewcFORCE);
1209     if (bDoForces)
1210     {
1211         /* Reset forces for which the virial is calculated separately:
1212          * PME/Ewald forces if necessary */
1213         if (fr->bF_NoVirSum)
1214         {
1215             if (flags & GMX_FORCE_VIRIAL)
1216             {
1217                 fr->f_novirsum = fr->f_novirsum_alloc;
1218                 if (fr->bDomDec)
1219                 {
1220                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1221                 }
1222                 else
1223                 {
1224                     clear_rvecs(homenr, fr->f_novirsum+start);
1225                 }
1226             }
1227             else
1228             {
1229                 /* We are not calculating the pressure so we do not need
1230                  * a separate array for forces that do not contribute
1231                  * to the pressure.
1232                  */
1233                 fr->f_novirsum = f;
1234             }
1235         }
1236
1237         /* Clear the short- and long-range forces */
1238         clear_rvecs(fr->natoms_force_constr, f);
1239         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1240         {
1241             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1242         }
1243
1244         clear_rvec(fr->vir_diag_posres);
1245     }
1246
1247     if (inputrec->ePull == epullCONSTRAINT)
1248     {
1249         clear_pull_forces(inputrec->pull);
1250     }
1251
1252     /* We calculate the non-bonded forces, when done on the CPU, here.
1253      * We do this before calling do_force_lowlevel, as in there bondeds
1254      * forces are calculated before PME, which does communication.
1255      * With this order, non-bonded and bonded force calculation imbalance
1256      * can be balanced out by the domain decomposition load balancing.
1257      */
1258
1259     if (!bUseOrEmulGPU)
1260     {
1261         /* Maybe we should move this into do_force_lowlevel */
1262         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1263                      nrnb, wcycle);
1264     }
1265
1266     if (fr->efep != efepNO)
1267     {
1268         /* Calculate the local and non-local free energy interactions here.
1269          * Happens here on the CPU both with and without GPU.
1270          */
1271         if (fr->nbv->grp[eintLocal].nbl_lists.nbl_fep[0]->nrj > 0)
1272         {
1273             do_nb_verlet_fep(&fr->nbv->grp[eintLocal].nbl_lists,
1274                              fr, x, f, mdatoms,
1275                              inputrec->fepvals, lambda,
1276                              enerd, flags, nrnb, wcycle);
1277         }
1278
1279         if (DOMAINDECOMP(cr) &&
1280             fr->nbv->grp[eintNonlocal].nbl_lists.nbl_fep[0]->nrj > 0)
1281         {
1282             do_nb_verlet_fep(&fr->nbv->grp[eintNonlocal].nbl_lists,
1283                              fr, x, f, mdatoms,
1284                              inputrec->fepvals, lambda,
1285                              enerd, flags, nrnb, wcycle);
1286         }
1287     }
1288
1289     if (!bUseOrEmulGPU || bDiffKernels)
1290     {
1291         int aloc;
1292
1293         if (DOMAINDECOMP(cr))
1294         {
1295             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1296                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1297                          nrnb, wcycle);
1298         }
1299
1300         if (!bUseOrEmulGPU)
1301         {
1302             aloc = eintLocal;
1303         }
1304         else
1305         {
1306             aloc = eintNonlocal;
1307         }
1308
1309         /* Add all the non-bonded force to the normal force array.
1310          * This can be split into a local a non-local part when overlapping
1311          * communication with calculation with domain decomposition.
1312          */
1313         cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1314         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1315         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1316         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1317         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1318         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1319         wallcycle_start_nocount(wcycle, ewcFORCE);
1320
1321         /* if there are multiple fshift output buffers reduce them */
1322         if ((flags & GMX_FORCE_VIRIAL) &&
1323             nbv->grp[aloc].nbl_lists.nnbl > 1)
1324         {
1325             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1326                                                      fr->fshift);
1327         }
1328     }
1329
1330     /* update QMMMrec, if necessary */
1331     if (fr->bQMMM)
1332     {
1333         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1334     }
1335
1336     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1337     {
1338         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1339                        enerd, lambda, fr);
1340     }
1341
1342     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1343     {
1344         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1345     }
1346
1347     /* Compute the bonded and non-bonded energies and optionally forces */
1348     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1349                       cr, nrnb, wcycle, mdatoms,
1350                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1351                       &(top->atomtypes), bBornRadii, box,
1352                       inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1353                       flags, &cycles_pme);
1354
1355     if (bSepLRF)
1356     {
1357         if (do_per_step(step, inputrec->nstcalclr))
1358         {
1359             /* Add the long range forces to the short range forces */
1360             for (i = 0; i < fr->natoms_force_constr; i++)
1361             {
1362                 rvec_add(fr->f_twin[i], f[i], f[i]);
1363             }
1364         }
1365     }
1366
1367     cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1368
1369     if (ed)
1370     {
1371         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1372     }
1373
1374     if (bUseOrEmulGPU && !bDiffKernels)
1375     {
1376         /* wait for non-local forces (or calculate in emulation mode) */
1377         if (DOMAINDECOMP(cr))
1378         {
1379             if (bUseGPU)
1380             {
1381                 float cycles_tmp;
1382
1383                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1384                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1385                                     nbv->grp[eintNonlocal].nbat,
1386                                     flags, eatNonlocal,
1387                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1388                                     fr->fshift);
1389                 cycles_tmp       = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1390                 cycles_wait_gpu += cycles_tmp;
1391                 cycles_force    += cycles_tmp;
1392             }
1393             else
1394             {
1395                 wallcycle_start_nocount(wcycle, ewcFORCE);
1396                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1397                              nrnb, wcycle);
1398                 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1399             }
1400             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1401             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1402             /* skip the reduction if there was no non-local work to do */
1403             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1404             {
1405                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1406                                                nbv->grp[eintNonlocal].nbat, f);
1407             }
1408             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1409             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1410         }
1411     }
1412
1413     if (bDoForces && DOMAINDECOMP(cr))
1414     {
1415         /* Communicate the forces */
1416         wallcycle_start(wcycle, ewcMOVEF);
1417         dd_move_f(cr->dd, f, fr->fshift);
1418         /* Do we need to communicate the separate force array
1419          * for terms that do not contribute to the single sum virial?
1420          * Position restraints and electric fields do not introduce
1421          * inter-cg forces, only full electrostatics methods do.
1422          * When we do not calculate the virial, fr->f_novirsum = f,
1423          * so we have already communicated these forces.
1424          */
1425         if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1426             (flags & GMX_FORCE_VIRIAL))
1427         {
1428             dd_move_f(cr->dd, fr->f_novirsum, NULL);
1429         }
1430         if (bSepLRF)
1431         {
1432             /* We should not update the shift forces here,
1433              * since f_twin is already included in f.
1434              */
1435             dd_move_f(cr->dd, fr->f_twin, NULL);
1436         }
1437         wallcycle_stop(wcycle, ewcMOVEF);
1438     }
1439
1440     if (bUseOrEmulGPU)
1441     {
1442         /* wait for local forces (or calculate in emulation mode) */
1443         if (bUseGPU)
1444         {
1445             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1446             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1447                                 nbv->grp[eintLocal].nbat,
1448                                 flags, eatLocal,
1449                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1450                                 fr->fshift);
1451             cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1452
1453             /* now clear the GPU outputs while we finish the step on the CPU */
1454
1455             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1456             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1457             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1458         }
1459         else
1460         {
1461             wallcycle_start_nocount(wcycle, ewcFORCE);
1462             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1463                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1464                          nrnb, wcycle);
1465             wallcycle_stop(wcycle, ewcFORCE);
1466         }
1467         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1468         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1469         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1470         {
1471             /* skip the reduction if there was no non-local work to do */
1472             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1473                                            nbv->grp[eintLocal].nbat, f);
1474         }
1475         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1476         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1477     }
1478
1479     if (DOMAINDECOMP(cr))
1480     {
1481         dd_force_flop_stop(cr->dd, nrnb);
1482         if (wcycle)
1483         {
1484             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1485             if (bUseGPU)
1486             {
1487                 dd_cycles_add(cr->dd, cycles_wait_gpu, ddCyclWaitGPU);
1488             }
1489         }
1490     }
1491
1492     if (bDoForces)
1493     {
1494         if (IR_ELEC_FIELD(*inputrec))
1495         {
1496             /* Compute forces due to electric field */
1497             calc_f_el(MASTER(cr) ? field : NULL,
1498                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1499                       inputrec->ex, inputrec->et, t);
1500         }
1501
1502         /* If we have NoVirSum forces, but we do not calculate the virial,
1503          * we sum fr->f_novirum=f later.
1504          */
1505         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1506         {
1507             wallcycle_start(wcycle, ewcVSITESPREAD);
1508             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1509                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1510             wallcycle_stop(wcycle, ewcVSITESPREAD);
1511
1512             if (bSepLRF)
1513             {
1514                 wallcycle_start(wcycle, ewcVSITESPREAD);
1515                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1516                                nrnb,
1517                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1518                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1519             }
1520         }
1521
1522         if (flags & GMX_FORCE_VIRIAL)
1523         {
1524             /* Calculation of the virial must be done after vsites! */
1525             calc_virial(0, mdatoms->homenr, x, f,
1526                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1527         }
1528     }
1529
1530     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1531     {
1532         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1533                                f, vir_force, mdatoms, enerd, lambda, t);
1534     }
1535
1536     /* Add the forces from enforced rotation potentials (if any) */
1537     if (inputrec->bRot)
1538     {
1539         wallcycle_start(wcycle, ewcROTadd);
1540         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1541         wallcycle_stop(wcycle, ewcROTadd);
1542     }
1543
1544     if (PAR(cr) && !(cr->duty & DUTY_PME))
1545     {
1546         /* In case of node-splitting, the PP nodes receive the long-range
1547          * forces, virial and energy from the PME nodes here.
1548          */
1549         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1550     }
1551
1552     if (bDoForces)
1553     {
1554         post_process_forces(cr, step, nrnb, wcycle,
1555                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1556                             flags);
1557     }
1558
1559     /* Sum the potential energy terms from group contributions */
1560     sum_epot(&(enerd->grpp), enerd->term);
1561 }
1562
1563 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1564                         t_inputrec *inputrec,
1565                         gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1566                         gmx_localtop_t *top,
1567                         gmx_groups_t *groups,
1568                         matrix box, rvec x[], history_t *hist,
1569                         rvec f[],
1570                         tensor vir_force,
1571                         t_mdatoms *mdatoms,
1572                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1573                         real *lambda, t_graph *graph,
1574                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1575                         double t, FILE *field, gmx_edsam_t ed,
1576                         gmx_bool bBornRadii,
1577                         int flags)
1578 {
1579     int        cg0, cg1, i, j;
1580     int        start, homenr;
1581     double     mu[2*DIM];
1582     gmx_bool   bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1583     gmx_bool   bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1584     gmx_bool   bDoAdressWF;
1585     matrix     boxs;
1586     rvec       vzero, box_diag;
1587     real       e, v, dvdlambda[efptNR];
1588     t_pbc      pbc;
1589     float      cycles_pme, cycles_force;
1590
1591     start  = 0;
1592     homenr = mdatoms->homenr;
1593
1594     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
1595
1596     clear_mat(vir_force);
1597
1598     cg0 = 0;
1599     if (DOMAINDECOMP(cr))
1600     {
1601         cg1 = cr->dd->ncg_tot;
1602     }
1603     else
1604     {
1605         cg1 = top->cgs.nr;
1606     }
1607     if (fr->n_tpi > 0)
1608     {
1609         cg1--;
1610     }
1611
1612     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1613     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1614     /* Should we update the long-range neighborlists at this step? */
1615     bDoLongRangeNS = fr->bTwinRange && bNS;
1616     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1617     bFillGrid      = (bNS && bStateChanged);
1618     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1619     bDoForces      = (flags & GMX_FORCE_FORCES);
1620     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1621     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1622                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1623
1624     /* should probably move this to the forcerec since it doesn't change */
1625     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1626
1627     if (bStateChanged)
1628     {
1629         update_forcerec(fr, box);
1630
1631         if (NEED_MUTOT(*inputrec))
1632         {
1633             /* Calculate total (local) dipole moment in a temporary common array.
1634              * This makes it possible to sum them over nodes faster.
1635              */
1636             calc_mu(start, homenr,
1637                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1638                     mu, mu+DIM);
1639         }
1640     }
1641
1642     if (fr->ePBC != epbcNONE)
1643     {
1644         /* Compute shift vectors every step,
1645          * because of pressure coupling or box deformation!
1646          */
1647         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1648         {
1649             calc_shifts(box, fr->shift_vec);
1650         }
1651
1652         if (bCalcCGCM)
1653         {
1654             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1655                                      &(top->cgs), x, fr->cg_cm);
1656             inc_nrnb(nrnb, eNR_CGCM, homenr);
1657             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1658         }
1659         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1660         {
1661             unshift_self(graph, box, x);
1662         }
1663     }
1664     else if (bCalcCGCM)
1665     {
1666         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1667         inc_nrnb(nrnb, eNR_CGCM, homenr);
1668     }
1669
1670     if (bCalcCGCM && gmx_debug_at)
1671     {
1672         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1673     }
1674
1675 #ifdef GMX_MPI
1676     if (!(cr->duty & DUTY_PME))
1677     {
1678         /* Send particle coordinates to the pme nodes.
1679          * Since this is only implemented for domain decomposition
1680          * and domain decomposition does not use the graph,
1681          * we do not need to worry about shifting.
1682          */
1683
1684         int pme_flags = 0;
1685
1686         wallcycle_start(wcycle, ewcPP_PMESENDX);
1687
1688         bBS = (inputrec->nwall == 2);
1689         if (bBS)
1690         {
1691             copy_mat(box, boxs);
1692             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1693         }
1694
1695         if (EEL_PME(fr->eeltype))
1696         {
1697             pme_flags |= GMX_PME_DO_COULOMB;
1698         }
1699
1700         if (EVDW_PME(fr->vdwtype))
1701         {
1702             pme_flags |= GMX_PME_DO_LJ;
1703         }
1704
1705         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
1706                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
1707                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
1708                                  pme_flags, step);
1709
1710         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1711     }
1712 #endif /* GMX_MPI */
1713
1714     /* Communicate coordinates and sum dipole if necessary */
1715     if (DOMAINDECOMP(cr))
1716     {
1717         wallcycle_start(wcycle, ewcMOVEX);
1718         dd_move_x(cr->dd, box, x);
1719         wallcycle_stop(wcycle, ewcMOVEX);
1720     }
1721
1722     /* update adress weight beforehand */
1723     if (bStateChanged && bDoAdressWF)
1724     {
1725         /* need pbc for adress weight calculation with pbc_dx */
1726         set_pbc(&pbc, inputrec->ePBC, box);
1727         if (fr->adress_site == eAdressSITEcog)
1728         {
1729             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1730                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1731         }
1732         else if (fr->adress_site == eAdressSITEcom)
1733         {
1734             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1735                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1736         }
1737         else if (fr->adress_site == eAdressSITEatomatom)
1738         {
1739             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1740                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1741         }
1742         else
1743         {
1744             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1745                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1746         }
1747     }
1748
1749     if (NEED_MUTOT(*inputrec))
1750     {
1751
1752         if (bStateChanged)
1753         {
1754             if (PAR(cr))
1755             {
1756                 gmx_sumd(2*DIM, mu, cr);
1757             }
1758             for (i = 0; i < 2; i++)
1759             {
1760                 for (j = 0; j < DIM; j++)
1761                 {
1762                     fr->mu_tot[i][j] = mu[i*DIM + j];
1763                 }
1764             }
1765         }
1766         if (fr->efep == efepNO)
1767         {
1768             copy_rvec(fr->mu_tot[0], mu_tot);
1769         }
1770         else
1771         {
1772             for (j = 0; j < DIM; j++)
1773             {
1774                 mu_tot[j] =
1775                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1776             }
1777         }
1778     }
1779
1780     /* Reset energies */
1781     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1782     clear_rvecs(SHIFTS, fr->fshift);
1783
1784     if (bNS)
1785     {
1786         wallcycle_start(wcycle, ewcNS);
1787
1788         if (graph && bStateChanged)
1789         {
1790             /* Calculate intramolecular shift vectors to make molecules whole */
1791             mk_mshift(fplog, graph, fr->ePBC, box, x);
1792         }
1793
1794         /* Do the actual neighbour searching */
1795         ns(fplog, fr, box,
1796            groups, top, mdatoms,
1797            cr, nrnb, bFillGrid,
1798            bDoLongRangeNS);
1799
1800         wallcycle_stop(wcycle, ewcNS);
1801     }
1802
1803     if (inputrec->implicit_solvent && bNS)
1804     {
1805         make_gb_nblist(cr, inputrec->gb_algorithm,
1806                        x, box, fr, &top->idef, graph, fr->born);
1807     }
1808
1809     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1810     {
1811         wallcycle_start(wcycle, ewcPPDURINGPME);
1812         dd_force_flop_start(cr->dd, nrnb);
1813     }
1814
1815     if (inputrec->bRot)
1816     {
1817         /* Enforced rotation has its own cycle counter that starts after the collective
1818          * coordinates have been communicated. It is added to ddCyclF to allow
1819          * for proper load-balancing */
1820         wallcycle_start(wcycle, ewcROT);
1821         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1822         wallcycle_stop(wcycle, ewcROT);
1823     }
1824
1825     /* Start the force cycle counter.
1826      * This counter is stopped in do_forcelow_level.
1827      * No parallel communication should occur while this counter is running,
1828      * since that will interfere with the dynamic load balancing.
1829      */
1830     wallcycle_start(wcycle, ewcFORCE);
1831
1832     if (bDoForces)
1833     {
1834         /* Reset forces for which the virial is calculated separately:
1835          * PME/Ewald forces if necessary */
1836         if (fr->bF_NoVirSum)
1837         {
1838             if (flags & GMX_FORCE_VIRIAL)
1839             {
1840                 fr->f_novirsum = fr->f_novirsum_alloc;
1841                 if (fr->bDomDec)
1842                 {
1843                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1844                 }
1845                 else
1846                 {
1847                     clear_rvecs(homenr, fr->f_novirsum+start);
1848                 }
1849             }
1850             else
1851             {
1852                 /* We are not calculating the pressure so we do not need
1853                  * a separate array for forces that do not contribute
1854                  * to the pressure.
1855                  */
1856                 fr->f_novirsum = f;
1857             }
1858         }
1859
1860         /* Clear the short- and long-range forces */
1861         clear_rvecs(fr->natoms_force_constr, f);
1862         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1863         {
1864             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1865         }
1866
1867         clear_rvec(fr->vir_diag_posres);
1868     }
1869     if (inputrec->ePull == epullCONSTRAINT)
1870     {
1871         clear_pull_forces(inputrec->pull);
1872     }
1873
1874     /* update QMMMrec, if necessary */
1875     if (fr->bQMMM)
1876     {
1877         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1878     }
1879
1880     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1881     {
1882         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1883                        enerd, lambda, fr);
1884     }
1885
1886     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1887     {
1888         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1889     }
1890
1891     /* Compute the bonded and non-bonded energies and optionally forces */
1892     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1893                       cr, nrnb, wcycle, mdatoms,
1894                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1895                       &(top->atomtypes), bBornRadii, box,
1896                       inputrec->fepvals, lambda,
1897                       graph, &(top->excls), fr->mu_tot,
1898                       flags,
1899                       &cycles_pme);
1900
1901     if (bSepLRF)
1902     {
1903         if (do_per_step(step, inputrec->nstcalclr))
1904         {
1905             /* Add the long range forces to the short range forces */
1906             for (i = 0; i < fr->natoms_force_constr; i++)
1907             {
1908                 rvec_add(fr->f_twin[i], f[i], f[i]);
1909             }
1910         }
1911     }
1912
1913     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1914
1915     if (ed)
1916     {
1917         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1918     }
1919
1920     if (DOMAINDECOMP(cr))
1921     {
1922         dd_force_flop_stop(cr->dd, nrnb);
1923         if (wcycle)
1924         {
1925             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1926         }
1927     }
1928
1929     if (bDoForces)
1930     {
1931         if (IR_ELEC_FIELD(*inputrec))
1932         {
1933             /* Compute forces due to electric field */
1934             calc_f_el(MASTER(cr) ? field : NULL,
1935                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1936                       inputrec->ex, inputrec->et, t);
1937         }
1938
1939         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1940         {
1941             /* Compute thermodynamic force in hybrid AdResS region */
1942             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1943                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1944         }
1945
1946         /* Communicate the forces */
1947         if (DOMAINDECOMP(cr))
1948         {
1949             wallcycle_start(wcycle, ewcMOVEF);
1950             dd_move_f(cr->dd, f, fr->fshift);
1951             /* Do we need to communicate the separate force array
1952              * for terms that do not contribute to the single sum virial?
1953              * Position restraints and electric fields do not introduce
1954              * inter-cg forces, only full electrostatics methods do.
1955              * When we do not calculate the virial, fr->f_novirsum = f,
1956              * so we have already communicated these forces.
1957              */
1958             if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1959                 (flags & GMX_FORCE_VIRIAL))
1960             {
1961                 dd_move_f(cr->dd, fr->f_novirsum, NULL);
1962             }
1963             if (bSepLRF)
1964             {
1965                 /* We should not update the shift forces here,
1966                  * since f_twin is already included in f.
1967                  */
1968                 dd_move_f(cr->dd, fr->f_twin, NULL);
1969             }
1970             wallcycle_stop(wcycle, ewcMOVEF);
1971         }
1972
1973         /* If we have NoVirSum forces, but we do not calculate the virial,
1974          * we sum fr->f_novirum=f later.
1975          */
1976         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1977         {
1978             wallcycle_start(wcycle, ewcVSITESPREAD);
1979             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1980                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1981             wallcycle_stop(wcycle, ewcVSITESPREAD);
1982
1983             if (bSepLRF)
1984             {
1985                 wallcycle_start(wcycle, ewcVSITESPREAD);
1986                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1987                                nrnb,
1988                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1989                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1990             }
1991         }
1992
1993         if (flags & GMX_FORCE_VIRIAL)
1994         {
1995             /* Calculation of the virial must be done after vsites! */
1996             calc_virial(0, mdatoms->homenr, x, f,
1997                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1998         }
1999     }
2000
2001     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
2002     {
2003         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
2004                                f, vir_force, mdatoms, enerd, lambda, t);
2005     }
2006
2007     /* Add the forces from enforced rotation potentials (if any) */
2008     if (inputrec->bRot)
2009     {
2010         wallcycle_start(wcycle, ewcROTadd);
2011         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
2012         wallcycle_stop(wcycle, ewcROTadd);
2013     }
2014
2015     if (PAR(cr) && !(cr->duty & DUTY_PME))
2016     {
2017         /* In case of node-splitting, the PP nodes receive the long-range
2018          * forces, virial and energy from the PME nodes here.
2019          */
2020         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
2021     }
2022
2023     if (bDoForces)
2024     {
2025         post_process_forces(cr, step, nrnb, wcycle,
2026                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
2027                             flags);
2028     }
2029
2030     /* Sum the potential energy terms from group contributions */
2031     sum_epot(&(enerd->grpp), enerd->term);
2032 }
2033
2034 void do_force(FILE *fplog, t_commrec *cr,
2035               t_inputrec *inputrec,
2036               gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
2037               gmx_localtop_t *top,
2038               gmx_groups_t *groups,
2039               matrix box, rvec x[], history_t *hist,
2040               rvec f[],
2041               tensor vir_force,
2042               t_mdatoms *mdatoms,
2043               gmx_enerdata_t *enerd, t_fcdata *fcd,
2044               real *lambda, t_graph *graph,
2045               t_forcerec *fr,
2046               gmx_vsite_t *vsite, rvec mu_tot,
2047               double t, FILE *field, gmx_edsam_t ed,
2048               gmx_bool bBornRadii,
2049               int flags)
2050 {
2051     /* modify force flag if not doing nonbonded */
2052     if (!fr->bNonbonded)
2053     {
2054         flags &= ~GMX_FORCE_NONBONDED;
2055     }
2056
2057     switch (inputrec->cutoff_scheme)
2058     {
2059         case ecutsVERLET:
2060             do_force_cutsVERLET(fplog, cr, inputrec,
2061                                 step, nrnb, wcycle,
2062                                 top,
2063                                 groups,
2064                                 box, x, hist,
2065                                 f, vir_force,
2066                                 mdatoms,
2067                                 enerd, fcd,
2068                                 lambda, graph,
2069                                 fr, fr->ic,
2070                                 vsite, mu_tot,
2071                                 t, field, ed,
2072                                 bBornRadii,
2073                                 flags);
2074             break;
2075         case ecutsGROUP:
2076             do_force_cutsGROUP(fplog, cr, inputrec,
2077                                step, nrnb, wcycle,
2078                                top,
2079                                groups,
2080                                box, x, hist,
2081                                f, vir_force,
2082                                mdatoms,
2083                                enerd, fcd,
2084                                lambda, graph,
2085                                fr, vsite, mu_tot,
2086                                t, field, ed,
2087                                bBornRadii,
2088                                flags);
2089             break;
2090         default:
2091             gmx_incons("Invalid cut-off scheme passed!");
2092     }
2093 }
2094
2095
2096 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2097                         t_inputrec *ir, t_mdatoms *md,
2098                         t_state *state, t_commrec *cr, t_nrnb *nrnb,
2099                         t_forcerec *fr, gmx_localtop_t *top)
2100 {
2101     int             i, m, start, end;
2102     gmx_int64_t     step;
2103     real            dt = ir->delta_t;
2104     real            dvdl_dum;
2105     rvec           *savex;
2106
2107     snew(savex, state->natoms);
2108
2109     start = 0;
2110     end   = md->homenr;
2111
2112     if (debug)
2113     {
2114         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2115                 start, md->homenr, end);
2116     }
2117     /* Do a first constrain to reset particles... */
2118     step = ir->init_step;
2119     if (fplog)
2120     {
2121         char buf[STEPSTRSIZE];
2122         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2123                 gmx_step_str(step, buf));
2124     }
2125     dvdl_dum = 0;
2126
2127     /* constrain the current position */
2128     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2129               ir, NULL, cr, step, 0, md,
2130               state->x, state->x, NULL,
2131               fr->bMolPBC, state->box,
2132               state->lambda[efptBONDED], &dvdl_dum,
2133               NULL, NULL, nrnb, econqCoord,
2134               ir->epc == epcMTTK, state->veta, state->veta);
2135     if (EI_VV(ir->eI))
2136     {
2137         /* constrain the inital velocity, and save it */
2138         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2139         /* might not yet treat veta correctly */
2140         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2141                   ir, NULL, cr, step, 0, md,
2142                   state->x, state->v, state->v,
2143                   fr->bMolPBC, state->box,
2144                   state->lambda[efptBONDED], &dvdl_dum,
2145                   NULL, NULL, nrnb, econqVeloc,
2146                   ir->epc == epcMTTK, state->veta, state->veta);
2147     }
2148     /* constrain the inital velocities at t-dt/2 */
2149     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2150     {
2151         for (i = start; (i < end); i++)
2152         {
2153             for (m = 0; (m < DIM); m++)
2154             {
2155                 /* Reverse the velocity */
2156                 state->v[i][m] = -state->v[i][m];
2157                 /* Store the position at t-dt in buf */
2158                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2159             }
2160         }
2161         /* Shake the positions at t=-dt with the positions at t=0
2162          * as reference coordinates.
2163          */
2164         if (fplog)
2165         {
2166             char buf[STEPSTRSIZE];
2167             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2168                     gmx_step_str(step, buf));
2169         }
2170         dvdl_dum = 0;
2171         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2172                   ir, NULL, cr, step, -1, md,
2173                   state->x, savex, NULL,
2174                   fr->bMolPBC, state->box,
2175                   state->lambda[efptBONDED], &dvdl_dum,
2176                   state->v, NULL, nrnb, econqCoord,
2177                   ir->epc == epcMTTK, state->veta, state->veta);
2178
2179         for (i = start; i < end; i++)
2180         {
2181             for (m = 0; m < DIM; m++)
2182             {
2183                 /* Re-reverse the velocities */
2184                 state->v[i][m] = -state->v[i][m];
2185             }
2186         }
2187     }
2188     sfree(savex);
2189 }
2190
2191
2192 static void
2193 integrate_table(real vdwtab[], real scale, int offstart, int rstart, int rend,
2194                 double *enerout, double *virout)
2195 {
2196     double enersum, virsum;
2197     double invscale, invscale2, invscale3;
2198     double r, ea, eb, ec, pa, pb, pc, pd;
2199     double y0, f, g, h;
2200     int    ri, offset, tabfactor;
2201
2202     invscale  = 1.0/scale;
2203     invscale2 = invscale*invscale;
2204     invscale3 = invscale*invscale2;
2205
2206     /* Following summation derived from cubic spline definition,
2207      * Numerical Recipies in C, second edition, p. 113-116.  Exact for
2208      * the cubic spline.  We first calculate the negative of the
2209      * energy from rvdw to rvdw_switch, assuming that g(r)=1, and then
2210      * add the more standard, abrupt cutoff correction to that result,
2211      * yielding the long-range correction for a switched function.  We
2212      * perform both the pressure and energy loops at the same time for
2213      * simplicity, as the computational cost is low. */
2214
2215     if (offstart == 0)
2216     {
2217         /* Since the dispersion table has been scaled down a factor
2218          * 6.0 and the repulsion a factor 12.0 to compensate for the
2219          * c6/c12 parameters inside nbfp[] being scaled up (to save
2220          * flops in kernels), we need to correct for this.
2221          */
2222         tabfactor = 6.0;
2223     }
2224     else
2225     {
2226         tabfactor = 12.0;
2227     }
2228
2229     enersum = 0.0;
2230     virsum  = 0.0;
2231     for (ri = rstart; ri < rend; ++ri)
2232     {
2233         r  = ri*invscale;
2234         ea = invscale3;
2235         eb = 2.0*invscale2*r;
2236         ec = invscale*r*r;
2237
2238         pa = invscale3;
2239         pb = 3.0*invscale2*r;
2240         pc = 3.0*invscale*r*r;
2241         pd = r*r*r;
2242
2243         /* this "8" is from the packing in the vdwtab array - perhaps
2244            should be defined? */
2245
2246         offset = 8*ri + offstart;
2247         y0     = vdwtab[offset];
2248         f      = vdwtab[offset+1];
2249         g      = vdwtab[offset+2];
2250         h      = vdwtab[offset+3];
2251
2252         enersum += y0*(ea/3 + eb/2 + ec) + f*(ea/4 + eb/3 + ec/2) + g*(ea/5 + eb/4 + ec/3) + h*(ea/6 + eb/5 + ec/4);
2253         virsum  +=  f*(pa/4 + pb/3 + pc/2 + pd) + 2*g*(pa/5 + pb/4 + pc/3 + pd/2) + 3*h*(pa/6 + pb/5 + pc/4 + pd/3);
2254     }
2255     *enerout = 4.0*M_PI*enersum*tabfactor;
2256     *virout  = 4.0*M_PI*virsum*tabfactor;
2257 }
2258
2259 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2260 {
2261     double eners[2], virs[2], enersum, virsum, y0, f, g, h;
2262     double r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2263     double invscale, invscale2, invscale3;
2264     int    ri0, ri1, ri, i, offstart, offset;
2265     real   scale, *vdwtab, tabfactor, tmp;
2266
2267     fr->enershiftsix    = 0;
2268     fr->enershifttwelve = 0;
2269     fr->enerdiffsix     = 0;
2270     fr->enerdifftwelve  = 0;
2271     fr->virdiffsix      = 0;
2272     fr->virdifftwelve   = 0;
2273
2274     if (eDispCorr != edispcNO)
2275     {
2276         for (i = 0; i < 2; i++)
2277         {
2278             eners[i] = 0;
2279             virs[i]  = 0;
2280         }
2281         if (fr->vdwtype == evdwSWITCH || fr->vdwtype == evdwSHIFT ||
2282             fr->vdw_modifier == eintmodPOTSWITCH ||
2283             fr->vdw_modifier == eintmodFORCESWITCH)
2284         {
2285             if (fr->rvdw_switch == 0)
2286             {
2287                 gmx_fatal(FARGS,
2288                           "With dispersion correction rvdw-switch can not be zero "
2289                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2290             }
2291
2292             scale  = fr->nblists[0].table_elec_vdw.scale;
2293             vdwtab = fr->nblists[0].table_vdw.data;
2294
2295             /* Round the cut-offs to exact table values for precision */
2296             ri0  = floor(fr->rvdw_switch*scale);
2297             ri1  = ceil(fr->rvdw*scale);
2298             r0   = ri0/scale;
2299             r1   = ri1/scale;
2300             rc3  = r0*r0*r0;
2301             rc9  = rc3*rc3*rc3;
2302
2303             if (fr->vdwtype == evdwSHIFT ||
2304                 fr->vdw_modifier == eintmodFORCESWITCH)
2305             {
2306                 /* Determine the constant energy shift below rvdw_switch.
2307                  * Table has a scale factor since we have scaled it down to compensate
2308                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2309                  */
2310                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2311                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2312             }
2313             /* Add the constant part from 0 to rvdw_switch.
2314              * This integration from 0 to rvdw_switch overcounts the number
2315              * of interactions by 1, as it also counts the self interaction.
2316              * We will correct for this later.
2317              */
2318             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2319             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2320             for (i = 0; i < 2; i++)
2321             {
2322                 enersum = 0;
2323                 virsum  = 0;
2324                 integrate_table(vdwtab, scale, (i == 0 ? 0 : 4), ri0, ri1, &enersum, &virsum);
2325                 eners[i] -= enersum;
2326                 virs[i]  -= virsum;
2327             }
2328
2329             /* now add the correction for rvdw_switch to infinity */
2330             eners[0] += -4.0*M_PI/(3.0*rc3);
2331             eners[1] +=  4.0*M_PI/(9.0*rc9);
2332             virs[0]  +=  8.0*M_PI/rc3;
2333             virs[1]  += -16.0*M_PI/(3.0*rc9);
2334         }
2335         else if (fr->vdwtype == evdwCUT ||
2336                  EVDW_PME(fr->vdwtype) ||
2337                  fr->vdwtype == evdwUSER)
2338         {
2339             if (fr->vdwtype == evdwUSER && fplog)
2340             {
2341                 fprintf(fplog,
2342                         "WARNING: using dispersion correction with user tables\n");
2343             }
2344
2345             /* Note that with LJ-PME, the dispersion correction is multiplied
2346              * by the difference between the actual C6 and the value of C6
2347              * that would produce the combination rule.
2348              * This means the normal energy and virial difference formulas
2349              * can be used here.
2350              */
2351
2352             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2353             rc9  = rc3*rc3*rc3;
2354             /* Contribution beyond the cut-off */
2355             eners[0] += -4.0*M_PI/(3.0*rc3);
2356             eners[1] +=  4.0*M_PI/(9.0*rc9);
2357             if (fr->vdw_modifier == eintmodPOTSHIFT)
2358             {
2359                 /* Contribution within the cut-off */
2360                 eners[0] += -4.0*M_PI/(3.0*rc3);
2361                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2362             }
2363             /* Contribution beyond the cut-off */
2364             virs[0]  +=  8.0*M_PI/rc3;
2365             virs[1]  += -16.0*M_PI/(3.0*rc9);
2366         }
2367         else
2368         {
2369             gmx_fatal(FARGS,
2370                       "Dispersion correction is not implemented for vdw-type = %s",
2371                       evdw_names[fr->vdwtype]);
2372         }
2373
2374         /* TODO: remove this code once we have group LJ-PME kernels
2375          * that calculate the exact, full LJ param C6/r^6 within the cut-off,
2376          * as the current nbnxn kernels do.
2377          */
2378         if (fr->vdwtype == evdwPME && fr->cutoff_scheme == ecutsGROUP)
2379         {
2380             /* Calculate self-interaction coefficient (assuming that
2381              * the reciprocal-space contribution is constant in the
2382              * region that contributes to the self-interaction).
2383              */
2384             fr->enershiftsix = pow(fr->ewaldcoeff_lj, 6) / 6.0;
2385
2386             eners[0] += -pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3)/3.0;
2387             virs[0]  +=  pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3);
2388         }
2389
2390         fr->enerdiffsix    = eners[0];
2391         fr->enerdifftwelve = eners[1];
2392         /* The 0.5 is due to the Gromacs definition of the virial */
2393         fr->virdiffsix     = 0.5*virs[0];
2394         fr->virdifftwelve  = 0.5*virs[1];
2395     }
2396 }
2397
2398 void calc_dispcorr(FILE *fplog, t_inputrec *ir, t_forcerec *fr,
2399                    gmx_int64_t step, int natoms,
2400                    matrix box, real lambda, tensor pres, tensor virial,
2401                    real *prescorr, real *enercorr, real *dvdlcorr)
2402 {
2403     gmx_bool bCorrAll, bCorrPres;
2404     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2405     int      m;
2406
2407     *prescorr = 0;
2408     *enercorr = 0;
2409     *dvdlcorr = 0;
2410
2411     clear_mat(virial);
2412     clear_mat(pres);
2413
2414     if (ir->eDispCorr != edispcNO)
2415     {
2416         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2417                      ir->eDispCorr == edispcAllEnerPres);
2418         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2419                      ir->eDispCorr == edispcAllEnerPres);
2420
2421         invvol = 1/det(box);
2422         if (fr->n_tpi)
2423         {
2424             /* Only correct for the interactions with the inserted molecule */
2425             dens   = (natoms - fr->n_tpi)*invvol;
2426             ninter = fr->n_tpi;
2427         }
2428         else
2429         {
2430             dens   = natoms*invvol;
2431             ninter = 0.5*natoms;
2432         }
2433
2434         if (ir->efep == efepNO)
2435         {
2436             avcsix    = fr->avcsix[0];
2437             avctwelve = fr->avctwelve[0];
2438         }
2439         else
2440         {
2441             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2442             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2443         }
2444
2445         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2446         *enercorr += avcsix*enerdiff;
2447         dvdlambda  = 0.0;
2448         if (ir->efep != efepNO)
2449         {
2450             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2451         }
2452         if (bCorrAll)
2453         {
2454             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2455             *enercorr += avctwelve*enerdiff;
2456             if (fr->efep != efepNO)
2457             {
2458                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2459             }
2460         }
2461
2462         if (bCorrPres)
2463         {
2464             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2465             if (ir->eDispCorr == edispcAllEnerPres)
2466             {
2467                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2468             }
2469             /* The factor 2 is because of the Gromacs virial definition */
2470             spres = -2.0*invvol*svir*PRESFAC;
2471
2472             for (m = 0; m < DIM; m++)
2473             {
2474                 virial[m][m] += svir;
2475                 pres[m][m]   += spres;
2476             }
2477             *prescorr += spres;
2478         }
2479
2480         /* Can't currently control when it prints, for now, just print when degugging */
2481         if (debug)
2482         {
2483             if (bCorrAll)
2484             {
2485                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2486                         avcsix, avctwelve);
2487             }
2488             if (bCorrPres)
2489             {
2490                 fprintf(debug,
2491                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2492                         *enercorr, spres, svir);
2493             }
2494             else
2495             {
2496                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2497             }
2498         }
2499
2500         if (fr->bSepDVDL && do_per_step(step, ir->nstlog))
2501         {
2502             gmx_print_sepdvdl(fplog, "Dispersion correction", *enercorr, dvdlambda);
2503         }
2504         if (fr->efep != efepNO)
2505         {
2506             *dvdlcorr += dvdlambda;
2507         }
2508     }
2509 }
2510
2511 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2512                   t_graph *graph, rvec x[])
2513 {
2514     if (fplog)
2515     {
2516         fprintf(fplog, "Removing pbc first time\n");
2517     }
2518     calc_shifts(box, fr->shift_vec);
2519     if (graph)
2520     {
2521         mk_mshift(fplog, graph, fr->ePBC, box, x);
2522         if (gmx_debug_at)
2523         {
2524             p_graph(debug, "do_pbc_first 1", graph);
2525         }
2526         shift_self(graph, box, x);
2527         /* By doing an extra mk_mshift the molecules that are broken
2528          * because they were e.g. imported from another software
2529          * will be made whole again. Such are the healing powers
2530          * of GROMACS.
2531          */
2532         mk_mshift(fplog, graph, fr->ePBC, box, x);
2533         if (gmx_debug_at)
2534         {
2535             p_graph(debug, "do_pbc_first 2", graph);
2536         }
2537     }
2538     if (fplog)
2539     {
2540         fprintf(fplog, "Done rmpbc\n");
2541     }
2542 }
2543
2544 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2545                             gmx_mtop_t *mtop, rvec x[],
2546                             gmx_bool bFirst)
2547 {
2548     t_graph        *graph;
2549     int             mb, as, mol;
2550     gmx_molblock_t *molb;
2551
2552     if (bFirst && fplog)
2553     {
2554         fprintf(fplog, "Removing pbc first time\n");
2555     }
2556
2557     snew(graph, 1);
2558     as = 0;
2559     for (mb = 0; mb < mtop->nmolblock; mb++)
2560     {
2561         molb = &mtop->molblock[mb];
2562         if (molb->natoms_mol == 1 ||
2563             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2564         {
2565             /* Just one atom or charge group in the molecule, no PBC required */
2566             as += molb->nmol*molb->natoms_mol;
2567         }
2568         else
2569         {
2570             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2571             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2572                            0, molb->natoms_mol, FALSE, FALSE, graph);
2573
2574             for (mol = 0; mol < molb->nmol; mol++)
2575             {
2576                 mk_mshift(fplog, graph, ePBC, box, x+as);
2577
2578                 shift_self(graph, box, x+as);
2579                 /* The molecule is whole now.
2580                  * We don't need the second mk_mshift call as in do_pbc_first,
2581                  * since we no longer need this graph.
2582                  */
2583
2584                 as += molb->natoms_mol;
2585             }
2586             done_graph(graph);
2587         }
2588     }
2589     sfree(graph);
2590 }
2591
2592 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2593                        gmx_mtop_t *mtop, rvec x[])
2594 {
2595     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2596 }
2597
2598 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2599                  gmx_mtop_t *mtop, rvec x[])
2600 {
2601     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2602 }
2603
2604 void finish_run(FILE *fplog, t_commrec *cr,
2605                 t_inputrec *inputrec,
2606                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2607                 gmx_walltime_accounting_t walltime_accounting,
2608                 wallclock_gpu_t *gputimes,
2609                 gmx_bool bWriteStat)
2610 {
2611     int     i, j;
2612     t_nrnb *nrnb_tot = NULL;
2613     real    delta_t;
2614     double  nbfs, mflop;
2615     double  elapsed_time,
2616             elapsed_time_over_all_ranks,
2617             elapsed_time_over_all_threads,
2618             elapsed_time_over_all_threads_over_all_ranks;
2619     wallcycle_sum(cr, wcycle);
2620
2621     if (cr->nnodes > 1)
2622     {
2623         snew(nrnb_tot, 1);
2624 #ifdef GMX_MPI
2625         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2626                       cr->mpi_comm_mysim);
2627 #endif
2628     }
2629     else
2630     {
2631         nrnb_tot = nrnb;
2632     }
2633
2634     elapsed_time                                 = walltime_accounting_get_elapsed_time(walltime_accounting);
2635     elapsed_time_over_all_ranks                  = elapsed_time;
2636     elapsed_time_over_all_threads                = walltime_accounting_get_elapsed_time_over_all_threads(walltime_accounting);
2637     elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
2638 #ifdef GMX_MPI
2639     if (cr->nnodes > 1)
2640     {
2641         /* reduce elapsed_time over all MPI ranks in the current simulation */
2642         MPI_Allreduce(&elapsed_time,
2643                       &elapsed_time_over_all_ranks,
2644                       1, MPI_DOUBLE, MPI_SUM,
2645                       cr->mpi_comm_mysim);
2646         elapsed_time_over_all_ranks /= cr->nnodes;
2647         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
2648          * current simulation. */
2649         MPI_Allreduce(&elapsed_time_over_all_threads,
2650                       &elapsed_time_over_all_threads_over_all_ranks,
2651                       1, MPI_DOUBLE, MPI_SUM,
2652                       cr->mpi_comm_mysim);
2653     }
2654 #endif
2655
2656     if (SIMMASTER(cr))
2657     {
2658         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2659     }
2660     if (cr->nnodes > 1)
2661     {
2662         sfree(nrnb_tot);
2663     }
2664
2665     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2666     {
2667         print_dd_statistics(cr, inputrec, fplog);
2668     }
2669
2670     if (SIMMASTER(cr))
2671     {
2672         wallcycle_print(fplog, cr->nnodes, cr->npmenodes,
2673                         elapsed_time_over_all_ranks,
2674                         wcycle, gputimes);
2675
2676         if (EI_DYNAMICS(inputrec->eI))
2677         {
2678             delta_t = inputrec->delta_t;
2679         }
2680         else
2681         {
2682             delta_t = 0;
2683         }
2684
2685         if (fplog)
2686         {
2687             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
2688                        elapsed_time_over_all_ranks,
2689                        walltime_accounting_get_nsteps_done(walltime_accounting),
2690                        delta_t, nbfs, mflop);
2691         }
2692         if (bWriteStat)
2693         {
2694             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
2695                        elapsed_time_over_all_ranks,
2696                        walltime_accounting_get_nsteps_done(walltime_accounting),
2697                        delta_t, nbfs, mflop);
2698         }
2699     }
2700 }
2701
2702 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2703 {
2704     /* this function works, but could probably use a logic rewrite to keep all the different
2705        types of efep straight. */
2706
2707     int       i;
2708     t_lambda *fep = ir->fepvals;
2709
2710     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2711     {
2712         for (i = 0; i < efptNR; i++)
2713         {
2714             lambda[i] = 0.0;
2715             if (lam0)
2716             {
2717                 lam0[i] = 0.0;
2718             }
2719         }
2720         return;
2721     }
2722     else
2723     {
2724         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2725                                              if checkpoint is set -- a kludge is in for now
2726                                              to prevent this.*/
2727         for (i = 0; i < efptNR; i++)
2728         {
2729             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2730             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2731             {
2732                 lambda[i] = fep->init_lambda;
2733                 if (lam0)
2734                 {
2735                     lam0[i] = lambda[i];
2736                 }
2737             }
2738             else
2739             {
2740                 lambda[i] = fep->all_lambda[i][*fep_state];
2741                 if (lam0)
2742                 {
2743                     lam0[i] = lambda[i];
2744                 }
2745             }
2746         }
2747         if (ir->bSimTemp)
2748         {
2749             /* need to rescale control temperatures to match current state */
2750             for (i = 0; i < ir->opts.ngtc; i++)
2751             {
2752                 if (ir->opts.ref_t[i] > 0)
2753                 {
2754                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2755                 }
2756             }
2757         }
2758     }
2759
2760     /* Send to the log the information on the current lambdas */
2761     if (fplog != NULL)
2762     {
2763         fprintf(fplog, "Initial vector of lambda components:[ ");
2764         for (i = 0; i < efptNR; i++)
2765         {
2766             fprintf(fplog, "%10.4f ", lambda[i]);
2767         }
2768         fprintf(fplog, "]\n");
2769     }
2770     return;
2771 }
2772
2773
2774 void init_md(FILE *fplog,
2775              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2776              double *t, double *t0,
2777              real *lambda, int *fep_state, double *lam0,
2778              t_nrnb *nrnb, gmx_mtop_t *mtop,
2779              gmx_update_t *upd,
2780              int nfile, const t_filenm fnm[],
2781              gmx_mdoutf_t *outf, t_mdebin **mdebin,
2782              tensor force_vir, tensor shake_vir, rvec mu_tot,
2783              gmx_bool *bSimAnn, t_vcm **vcm, unsigned long Flags)
2784 {
2785     int  i, j, n;
2786     real tmpt, mod;
2787
2788     /* Initial values */
2789     *t = *t0       = ir->init_t;
2790
2791     *bSimAnn = FALSE;
2792     for (i = 0; i < ir->opts.ngtc; i++)
2793     {
2794         /* set bSimAnn if any group is being annealed */
2795         if (ir->opts.annealing[i] != eannNO)
2796         {
2797             *bSimAnn = TRUE;
2798         }
2799     }
2800     if (*bSimAnn)
2801     {
2802         update_annealing_target_temp(&(ir->opts), ir->init_t);
2803     }
2804
2805     /* Initialize lambda variables */
2806     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2807
2808     if (upd)
2809     {
2810         *upd = init_update(ir);
2811     }
2812
2813
2814     if (vcm != NULL)
2815     {
2816         *vcm = init_vcm(fplog, &mtop->groups, ir);
2817     }
2818
2819     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2820     {
2821         if (ir->etc == etcBERENDSEN)
2822         {
2823             please_cite(fplog, "Berendsen84a");
2824         }
2825         if (ir->etc == etcVRESCALE)
2826         {
2827             please_cite(fplog, "Bussi2007a");
2828         }
2829     }
2830
2831     init_nrnb(nrnb);
2832
2833     if (nfile != -1)
2834     {
2835         *outf = init_mdoutf(fplog, nfile, fnm, Flags, cr, ir, mtop, oenv);
2836
2837         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : mdoutf_get_fp_ene(*outf),
2838                               mtop, ir, mdoutf_get_fp_dhdl(*outf));
2839     }
2840
2841     if (ir->bAdress)
2842     {
2843         please_cite(fplog, "Fritsch12");
2844         please_cite(fplog, "Junghans10");
2845     }
2846     /* Initiate variables */
2847     clear_mat(force_vir);
2848     clear_mat(shake_vir);
2849     clear_rvec(mu_tot);
2850
2851     debug_gmx();
2852 }