src/gromacs/mdlib/sim_util.c

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  *          GROningen MAchine for Chemical Simulations
   9  *
  10  *                        VERSION 3.2.0
  11  * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
  12  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
  13  * Copyright (c) 2001-2004, The GROMACS development team,
  14  * check out http://www.gromacs.org for more information.
  15
  16  * This program is free software; you can redistribute it and/or
  17  * modify it under the terms of the GNU General Public License
  18  * as published by the Free Software Foundation; either version 2
  19  * of the License, or (at your option) any later version.
  20  *
  21  * If you want to redistribute modifications, please consider that
  22  * scientific software is very special. Version control is crucial -
  23  * bugs must be traceable. We will be happy to consider code for
  24  * inclusion in the official distribution, but derived work must not
  25  * be called official GROMACS. Details are found in the README & COPYING
  26  * files - if they are missing, get the official version at www.gromacs.org.
  27  *
  28  * To help us fund GROMACS development, we humbly ask that you cite
  29  * the papers on the package - you can find them in the top README file.
  30  *
  31  * For more info, check our website at http://www.gromacs.org
  32  *
  33  * And Hey:
  34  * GROwing Monsters And Cloning Shrimps
  35  */
  36 #ifdef HAVE_CONFIG_H
  37 #include <config.h>
  38 #endif
  39
  40 #ifdef GMX_CRAY_XT3
  41 #include <catamount/dclock.h>
  42 #endif
  43
  44
  45 #include <stdio.h>
  46 #include <time.h>
  47 #ifdef HAVE_SYS_TIME_H
  48 #include <sys/time.h>
  49 #endif
  50 #include <math.h>
  51 #include "typedefs.h"
  52 #include "string2.h"
  53 #include "gmxfio.h"
  54 #include "smalloc.h"
  55 #include "names.h"
  56 #include "confio.h"
  57 #include "mvdata.h"
  58 #include "txtdump.h"
  59 #include "pbc.h"
  60 #include "chargegroup.h"
  61 #include "vec.h"
  62 #include <time.h>
  63 #include "nrnb.h"
  64 #include "mshift.h"
  65 #include "mdrun.h"
  66 #include "sim_util.h"
  67 #include "update.h"
  68 #include "physics.h"
  69 #include "main.h"
  70 #include "mdatoms.h"
  71 #include "force.h"
  72 #include "bondf.h"
  73 #include "pme.h"
  74 #include "disre.h"
  75 #include "orires.h"
  76 #include "network.h"
  77 #include "calcmu.h"
  78 #include "constr.h"
  79 #include "xvgr.h"
  80 #include "trnio.h"
  81 #include "xtcio.h"
  82 #include "copyrite.h"
  83 #include "pull_rotation.h"
  84 #include "gmx_random.h"
  85 #include "domdec.h"
  86 #include "partdec.h"
  87 #include "gmx_wallcycle.h"
  88 #include "genborn.h"
  89 #include "nbnxn_atomdata.h"
  90 #include "nbnxn_search.h"
  91 #include "nbnxn_kernels/nbnxn_kernel_ref.h"
  92 #include "nbnxn_kernels/nbnxn_kernel_simd_4xn.h"
  93 #include "nbnxn_kernels/nbnxn_kernel_simd_2xnn.h"
  94 #include "nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  95
  96 #include "gromacs/utility/gmxmpi.h"
  97
  98 #include "adress.h"
  99 #include "qmmm.h"
 100
 101 #include "nbnxn_cuda_data_mgmt.h"
 102 #include "nbnxn_cuda/nbnxn_cuda.h"
 103
 104 double
 105 gmx_gettime()
 106 {
 107 #ifdef HAVE_GETTIMEOFDAY
 108     struct timeval t;
 109     double         seconds;
 110
 111     gettimeofday(&t, NULL);
 112
 113     seconds = (double) t.tv_sec + 1e-6*(double)t.tv_usec;
 114
 115     return seconds;
 116 #else
 117     double  seconds;
 118
 119     seconds = time(NULL);
 120
 121     return seconds;
 122 #endif
 123 }
 124
 125
 126 #define difftime(end, start) ((double)(end)-(double)(start))
 127
 128 void print_time(FILE *out, gmx_runtime_t *runtime, gmx_large_int_t step,
 129                 t_inputrec *ir, t_commrec gmx_unused *cr)
 130 {
 131     time_t finish;
 132     char   timebuf[STRLEN];
 133     double dt;
 134     char   buf[48];
 135
 136 #ifndef GMX_THREAD_MPI
 137     if (!PAR(cr))
 138 #endif
 139     {
 140         fprintf(out, "\r");
 141     }
 142     fprintf(out, "step %s", gmx_step_str(step, buf));
 143     if ((step >= ir->nstlist))
 144     {
 145         runtime->last          = gmx_gettime();
 146         dt                     = difftime(runtime->last, runtime->real);
 147         runtime->time_per_step = dt/(step - ir->init_step + 1);
 148
 149         dt = (ir->nsteps + ir->init_step - step)*runtime->time_per_step;
 150
 151         if (ir->nsteps >= 0)
 152         {
 153             if (dt >= 300)
 154             {
 155                 finish = (time_t) (runtime->last + dt);
 156                 gmx_ctime_r(&finish, timebuf, STRLEN);
 157                 sprintf(buf, "%s", timebuf);
 158                 buf[strlen(buf)-1] = '\0';
 159                 fprintf(out, ", will finish %s", buf);
 160             }
 161             else
 162             {
 163                 fprintf(out, ", remaining runtime: %5d s          ", (int)dt);
 164             }
 165         }
 166         else
 167         {
 168             fprintf(out, " performance: %.1f ns/day    ",
 169                     ir->delta_t/1000*24*60*60/runtime->time_per_step);
 170         }
 171     }
 172 #ifndef GMX_THREAD_MPI
 173     if (PAR(cr))
 174     {
 175         fprintf(out, "\n");
 176     }
 177 #endif
 178
 179     fflush(out);
 180 }
 181
 182 #ifdef NO_CLOCK
 183 #define clock() -1
 184 #endif
 185
 186 static double set_proctime(gmx_runtime_t *runtime)
 187 {
 188     double diff;
 189 #ifdef GMX_CRAY_XT3
 190     double prev;
 191
 192     prev          = runtime->proc;
 193     runtime->proc = dclock();
 194
 195     diff = runtime->proc - prev;
 196 #else
 197     clock_t prev;
 198
 199     prev          = runtime->proc;
 200     runtime->proc = clock();
 201
 202     diff = (double)(runtime->proc - prev)/(double)CLOCKS_PER_SEC;
 203 #endif
 204     if (diff < 0)
 205     {
 206         /* The counter has probably looped, ignore this data */
 207         diff = 0;
 208     }
 209
 210     return diff;
 211 }
 212
 213 void runtime_start(gmx_runtime_t *runtime)
 214 {
 215     runtime->real          = gmx_gettime();
 216     runtime->proc          = 0;
 217     set_proctime(runtime);
 218     runtime->realtime      = 0;
 219     runtime->proctime      = 0;
 220     runtime->last          = 0;
 221     runtime->time_per_step = 0;
 222 }
 223
 224 void runtime_end(gmx_runtime_t *runtime)
 225 {
 226     double now;
 227
 228     now = gmx_gettime();
 229
 230     runtime->proctime += set_proctime(runtime);
 231     runtime->realtime  = now - runtime->real;
 232     runtime->real      = now;
 233 }
 234
 235 void runtime_upd_proc(gmx_runtime_t *runtime)
 236 {
 237     runtime->proctime += set_proctime(runtime);
 238 }
 239
 240 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
 241                          const gmx_runtime_t *runtime)
 242 {
 243     int    i;
 244     char   timebuf[STRLEN];
 245     char   time_string[STRLEN];
 246     time_t tmptime;
 247
 248     if (fplog)
 249     {
 250         if (runtime != NULL)
 251         {
 252             tmptime = (time_t) runtime->real;
 253             gmx_ctime_r(&tmptime, timebuf, STRLEN);
 254         }
 255         else
 256         {
 257             tmptime = (time_t) gmx_gettime();
 258             gmx_ctime_r(&tmptime, timebuf, STRLEN);
 259         }
 260         for (i = 0; timebuf[i] >= ' '; i++)
 261         {
 262             time_string[i] = timebuf[i];
 263         }
 264         time_string[i] = '\0';
 265
 266         fprintf(fplog, "%s on node %d %s\n", title, nodeid, time_string);
 267     }
 268 }
 269
 270 static void sum_forces(int start, int end, rvec f[], rvec flr[])
 271 {
 272     int i;
 273
 274     if (gmx_debug_at)
 275     {
 276         pr_rvecs(debug, 0, "fsr", f+start, end-start);
 277         pr_rvecs(debug, 0, "flr", flr+start, end-start);
 278     }
 279     for (i = start; (i < end); i++)
 280     {
 281         rvec_inc(f[i], flr[i]);
 282     }
 283 }
 284
 285 /*
 286  * calc_f_el calculates forces due to an electric field.
 287  *
 288  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 289  *
 290  * Et[] contains the parameters for the time dependent
 291  * part of the field (not yet used).
 292  * Ex[] contains the parameters for
 293  * the spatial dependent part of the field. You can have cool periodic
 294  * fields in principle, but only a constant field is supported
 295  * now.
 296  * The function should return the energy due to the electric field
 297  * (if any) but for now returns 0.
 298  *
 299  * WARNING:
 300  * There can be problems with the virial.
 301  * Since the field is not self-consistent this is unavoidable.
 302  * For neutral molecules the virial is correct within this approximation.
 303  * For neutral systems with many charged molecules the error is small.
 304  * But for systems with a net charge or a few charged molecules
 305  * the error can be significant when the field is high.
 306  * Solution: implement a self-consitent electric field into PME.
 307  */
 308 static void calc_f_el(FILE *fp, int  start, int homenr,
 309                       real charge[], rvec f[],
 310                       t_cosines Ex[], t_cosines Et[], double t)
 311 {
 312     rvec Ext;
 313     real t0;
 314     int  i, m;
 315
 316     for (m = 0; (m < DIM); m++)
 317     {
 318         if (Et[m].n > 0)
 319         {
 320             if (Et[m].n == 3)
 321             {
 322                 t0     = Et[m].a[1];
 323                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 324             }
 325             else
 326             {
 327                 Ext[m] = cos(Et[m].a[0]*t);
 328             }
 329         }
 330         else
 331         {
 332             Ext[m] = 1.0;
 333         }
 334         if (Ex[m].n > 0)
 335         {
 336             /* Convert the field strength from V/nm to MD-units */
 337             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 338             for (i = start; (i < start+homenr); i++)
 339             {
 340                 f[i][m] += charge[i]*Ext[m];
 341             }
 342         }
 343         else
 344         {
 345             Ext[m] = 0;
 346         }
 347     }
 348     if (fp != NULL)
 349     {
 350         fprintf(fp, "%10g  %10g  %10g  %10g #FIELD\n", t,
 351                 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
 352     }
 353 }
 354
 355 static void calc_virial(int start, int homenr, rvec x[], rvec f[],
 356                         tensor vir_part, t_graph *graph, matrix box,
 357                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 358 {
 359     int    i, j;
 360     tensor virtest;
 361
 362     /* The short-range virial from surrounding boxes */
 363     clear_mat(vir_part);
 364     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 365     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 366
 367     /* Calculate partial virial, for local atoms only, based on short range.
 368      * Total virial is computed in global_stat, called from do_md
 369      */
 370     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 371     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 372
 373     /* Add position restraint contribution */
 374     for (i = 0; i < DIM; i++)
 375     {
 376         vir_part[i][i] += fr->vir_diag_posres[i];
 377     }
 378
 379     /* Add wall contribution */
 380     for (i = 0; i < DIM; i++)
 381     {
 382         vir_part[i][ZZ] += fr->vir_wall_z[i];
 383     }
 384
 385     if (debug)
 386     {
 387         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 388     }
 389 }
 390
 391 static void posres_wrapper(FILE *fplog,
 392                            int flags,
 393                            gmx_bool bSepDVDL,
 394                            t_inputrec *ir,
 395                            t_nrnb *nrnb,
 396                            gmx_localtop_t *top,
 397                            matrix box, rvec x[],
 398                            gmx_enerdata_t *enerd,
 399                            real *lambda,
 400                            t_forcerec *fr)
 401 {
 402     t_pbc pbc;
 403     real  v, dvdl;
 404     int   i;
 405
 406     /* Position restraints always require full pbc */
 407     set_pbc(&pbc, ir->ePBC, box);
 408     dvdl = 0;
 409     v    = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 410                   top->idef.iparams_posres,
 411                   (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 412                   ir->ePBC == epbcNONE ? NULL : &pbc,
 413                   lambda[efptRESTRAINT], &dvdl,
 414                   fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 415     if (bSepDVDL)
 416     {
 417         gmx_print_sepdvdl(fplog, interaction_function[F_POSRES].longname, v, dvdl);
 418     }
 419     enerd->term[F_POSRES] += v;
 420     /* If just the force constant changes, the FEP term is linear,
 421      * but if k changes, it is not.
 422      */
 423     enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
 424     inc_nrnb(nrnb, eNR_POSRES, top->idef.il[F_POSRES].nr/2);
 425
 426     if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
 427     {
 428         for (i = 0; i < enerd->n_lambda; i++)
 429         {
 430             real dvdl_dum, lambda_dum;
 431
 432             lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
 433             v          = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 434                                 top->idef.iparams_posres,
 435                                 (const rvec*)x, NULL, NULL,
 436                                 ir->ePBC == epbcNONE ? NULL : &pbc, lambda_dum, &dvdl,
 437                                 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 438             enerd->enerpart_lambda[i] += v;
 439         }
 440     }
 441 }
 442
 443 static void pull_potential_wrapper(FILE *fplog,
 444                                    gmx_bool bSepDVDL,
 445                                    t_commrec *cr,
 446                                    t_inputrec *ir,
 447                                    matrix box, rvec x[],
 448                                    rvec f[],
 449                                    tensor vir_force,
 450                                    t_mdatoms *mdatoms,
 451                                    gmx_enerdata_t *enerd,
 452                                    real *lambda,
 453                                    double t)
 454 {
 455     t_pbc  pbc;
 456     real   dvdl;
 457
 458     /* Calculate the center of mass forces, this requires communication,
 459      * which is why pull_potential is called close to other communication.
 460      * The virial contribution is calculated directly,
 461      * which is why we call pull_potential after calc_virial.
 462      */
 463     set_pbc(&pbc, ir->ePBC, box);
 464     dvdl                     = 0;
 465     enerd->term[F_COM_PULL] +=
 466         pull_potential(ir->ePull, ir->pull, mdatoms, &pbc,
 467                        cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
 468     if (bSepDVDL)
 469     {
 470         gmx_print_sepdvdl(fplog, "Com pull", enerd->term[F_COM_PULL], dvdl);
 471     }
 472     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 473 }
 474
 475 static void pme_receive_force_ener(FILE           *fplog,
 476                                    gmx_bool        bSepDVDL,
 477                                    t_commrec      *cr,
 478                                    gmx_wallcycle_t wcycle,
 479                                    gmx_enerdata_t *enerd,
 480                                    t_forcerec     *fr)
 481 {
 482     real   e, v, dvdl;
 483     float  cycles_ppdpme, cycles_seppme;
 484
 485     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 486     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 487
 488     /* In case of node-splitting, the PP nodes receive the long-range
 489      * forces, virial and energy from the PME nodes here.
 490      */
 491     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 492     dvdl = 0;
 493     gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e, &dvdl,
 494                       &cycles_seppme);
 495     if (bSepDVDL)
 496     {
 497         gmx_print_sepdvdl(fplog, "PME mesh", e, dvdl);
 498     }
 499     enerd->term[F_COUL_RECIP] += e;
 500     enerd->dvdl_lin[efptCOUL] += dvdl;
 501     if (wcycle)
 502     {
 503         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 504     }
 505     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 506 }
 507
 508 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
 509                                gmx_large_int_t step, real pforce, rvec *x, rvec *f)
 510 {
 511     int  i;
 512     real pf2, fn2;
 513     char buf[STEPSTRSIZE];
 514
 515     pf2 = sqr(pforce);
 516     for (i = md->start; i < md->start+md->homenr; i++)
 517     {
 518         fn2 = norm2(f[i]);
 519         /* We also catch NAN, if the compiler does not optimize this away. */
 520         if (fn2 >= pf2 || fn2 != fn2)
 521         {
 522             fprintf(fp, "step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 523                     gmx_step_str(step, buf),
 524                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
 525         }
 526     }
 527 }
 528
 529 static void post_process_forces(t_commrec *cr,
 530                                 gmx_large_int_t step,
 531                                 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 532                                 gmx_localtop_t *top,
 533                                 matrix box, rvec x[],
 534                                 rvec f[],
 535                                 tensor vir_force,
 536                                 t_mdatoms *mdatoms,
 537                                 t_graph *graph,
 538                                 t_forcerec *fr, gmx_vsite_t *vsite,
 539                                 int flags)
 540 {
 541     if (fr->bF_NoVirSum)
 542     {
 543         if (vsite)
 544         {
 545             /* Spread the mesh force on virtual sites to the other particles...
 546              * This is parallellized. MPI communication is performed
 547              * if the constructing atoms aren't local.
 548              */
 549             wallcycle_start(wcycle, ewcVSITESPREAD);
 550             spread_vsite_f(vsite, x, fr->f_novirsum, NULL,
 551                            (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
 552                            nrnb,
 553                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
 554             wallcycle_stop(wcycle, ewcVSITESPREAD);
 555         }
 556         if (flags & GMX_FORCE_VIRIAL)
 557         {
 558             /* Now add the forces, this is local */
 559             if (fr->bDomDec)
 560             {
 561                 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
 562             }
 563             else
 564             {
 565                 sum_forces(mdatoms->start, mdatoms->start+mdatoms->homenr,
 566                            f, fr->f_novirsum);
 567             }
 568             if (EEL_FULL(fr->eeltype))
 569             {
 570                 /* Add the mesh contribution to the virial */
 571                 m_add(vir_force, fr->vir_el_recip, vir_force);
 572             }
 573             if (debug)
 574             {
 575                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 576             }
 577         }
 578     }
 579
 580     if (fr->print_force >= 0)
 581     {
 582         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 583     }
 584 }
 585
 586 static void do_nb_verlet(t_forcerec *fr,
 587                          interaction_const_t *ic,
 588                          gmx_enerdata_t *enerd,
 589                          int flags, int ilocality,
 590                          int clearF,
 591                          t_nrnb *nrnb)
 592 {
 593     int                        nnbl, kernel_type, enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
 594     char                      *env;
 595     nonbonded_verlet_group_t  *nbvg;
 596     gmx_bool                  bCUDA;
 597
 598     if (!(flags & GMX_FORCE_NONBONDED))
 599     {
 600         /* skip non-bonded calculation */
 601         return;
 602     }
 603
 604     nbvg = &fr->nbv->grp[ilocality];
 605
 606     /* CUDA kernel launch overhead is already timed separately */
 607     if (fr->cutoff_scheme != ecutsVERLET)
 608     {
 609         gmx_incons("Invalid cut-off scheme passed!");
 610     }
 611
 612     bCUDA = (nbvg->kernel_type == nbnxnk8x8x8_CUDA);
 613
 614     if (!bCUDA)
 615     {
 616         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 617     }
 618     switch (nbvg->kernel_type)
 619     {
 620         case nbnxnk4x4_PlainC:
 621             nbnxn_kernel_ref(&nbvg->nbl_lists,
 622                              nbvg->nbat, ic,
 623                              fr->shift_vec,
 624                              flags,
 625                              clearF,
 626                              fr->fshift[0],
 627                              enerd->grpp.ener[egCOULSR],
 628                              fr->bBHAM ?
 629                              enerd->grpp.ener[egBHAMSR] :
 630                              enerd->grpp.ener[egLJSR]);
 631             break;
 632
 633         case nbnxnk4xN_SIMD_4xN:
 634             nbnxn_kernel_simd_4xn(&nbvg->nbl_lists,
 635                                   nbvg->nbat, ic,
 636                                   nbvg->ewald_excl,
 637                                   fr->shift_vec,
 638                                   flags,
 639                                   clearF,
 640                                   fr->fshift[0],
 641                                   enerd->grpp.ener[egCOULSR],
 642                                   fr->bBHAM ?
 643                                   enerd->grpp.ener[egBHAMSR] :
 644                                   enerd->grpp.ener[egLJSR]);
 645             break;
 646         case nbnxnk4xN_SIMD_2xNN:
 647             nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
 648                                    nbvg->nbat, ic,
 649                                    nbvg->ewald_excl,
 650                                    fr->shift_vec,
 651                                    flags,
 652                                    clearF,
 653                                    fr->fshift[0],
 654                                    enerd->grpp.ener[egCOULSR],
 655                                    fr->bBHAM ?
 656                                    enerd->grpp.ener[egBHAMSR] :
 657                                    enerd->grpp.ener[egLJSR]);
 658             break;
 659
 660         case nbnxnk8x8x8_CUDA:
 661             nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
 662             break;
 663
 664         case nbnxnk8x8x8_PlainC:
 665             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 666                                  nbvg->nbat, ic,
 667                                  fr->shift_vec,
 668                                  flags,
 669                                  clearF,
 670                                  nbvg->nbat->out[0].f,
 671                                  fr->fshift[0],
 672                                  enerd->grpp.ener[egCOULSR],
 673                                  fr->bBHAM ?
 674                                  enerd->grpp.ener[egBHAMSR] :
 675                                  enerd->grpp.ener[egLJSR]);
 676             break;
 677
 678         default:
 679             gmx_incons("Invalid nonbonded kernel type passed!");
 680
 681     }
 682     if (!bCUDA)
 683     {
 684         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 685     }
 686
 687     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 688     {
 689         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 690     }
 691     else if ((!bCUDA && nbvg->ewald_excl == ewaldexclAnalytical) ||
 692              (bCUDA && nbnxn_cuda_is_kernel_ewald_analytical(fr->nbv->cu_nbv)))
 693     {
 694         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 695     }
 696     else
 697     {
 698         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 699     }
 700     enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 701     if (flags & GMX_FORCE_ENERGY)
 702     {
 703         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 704         enr_nbnxn_kernel_ljc += 1;
 705         enr_nbnxn_kernel_lj  += 1;
 706     }
 707
 708     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 709              nbvg->nbl_lists.natpair_ljq);
 710     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 711              nbvg->nbl_lists.natpair_lj);
 712     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 713              nbvg->nbl_lists.natpair_q);
 714 }
 715
 716 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
 717                          t_inputrec *inputrec,
 718                          gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 719                          gmx_localtop_t *top,
 720                          gmx_groups_t gmx_unused *groups,
 721                          matrix box, rvec x[], history_t *hist,
 722                          rvec f[],
 723                          tensor vir_force,
 724                          t_mdatoms *mdatoms,
 725                          gmx_enerdata_t *enerd, t_fcdata *fcd,
 726                          real *lambda, t_graph *graph,
 727                          t_forcerec *fr, interaction_const_t *ic,
 728                          gmx_vsite_t *vsite, rvec mu_tot,
 729                          double t, FILE *field, gmx_edsam_t ed,
 730                          gmx_bool bBornRadii,
 731                          int flags)
 732 {
 733     int                 cg0, cg1, i, j;
 734     int                 start, homenr;
 735     int                 nb_kernel_type;
 736     double              mu[2*DIM];
 737     gmx_bool            bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
 738     gmx_bool            bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
 739     gmx_bool            bDiffKernels = FALSE;
 740     matrix              boxs;
 741     rvec                vzero, box_diag;
 742     real                e, v, dvdl;
 743     float               cycles_pme, cycles_force;
 744     nonbonded_verlet_t *nbv;
 745
 746     cycles_force   = 0;
 747     nbv            = fr->nbv;
 748     nb_kernel_type = fr->nbv->grp[0].kernel_type;
 749
 750     start  = mdatoms->start;
 751     homenr = mdatoms->homenr;
 752
 753     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
 754
 755     clear_mat(vir_force);
 756
 757     cg0 = 0;
 758     if (DOMAINDECOMP(cr))
 759     {
 760         cg1 = cr->dd->ncg_tot;
 761     }
 762     else
 763     {
 764         cg1 = top->cgs.nr;
 765     }
 766     if (fr->n_tpi > 0)
 767     {
 768         cg1--;
 769     }
 770
 771     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 772     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
 773     bFillGrid     = (bNS && bStateChanged);
 774     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 775     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 776     bDoForces     = (flags & GMX_FORCE_FORCES);
 777     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 778     bUseGPU       = fr->nbv->bUseGPU;
 779     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
 780
 781     if (bStateChanged)
 782     {
 783         update_forcerec(fr, box);
 784
 785         if (NEED_MUTOT(*inputrec))
 786         {
 787             /* Calculate total (local) dipole moment in a temporary common array.
 788              * This makes it possible to sum them over nodes faster.
 789              */
 790             calc_mu(start, homenr,
 791                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 792                     mu, mu+DIM);
 793         }
 794     }
 795
 796     if (fr->ePBC != epbcNONE)
 797     {
 798         /* Compute shift vectors every step,
 799          * because of pressure coupling or box deformation!
 800          */
 801         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 802         {
 803             calc_shifts(box, fr->shift_vec);
 804         }
 805
 806         if (bCalcCGCM)
 807         {
 808             put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
 809             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 810         }
 811         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 812         {
 813             unshift_self(graph, box, x);
 814         }
 815     }
 816
 817     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 818                                  fr->shift_vec, nbv->grp[0].nbat);
 819
 820 #ifdef GMX_MPI
 821     if (!(cr->duty & DUTY_PME))
 822     {
 823         /* Send particle coordinates to the pme nodes.
 824          * Since this is only implemented for domain decomposition
 825          * and domain decomposition does not use the graph,
 826          * we do not need to worry about shifting.
 827          */
 828
 829         wallcycle_start(wcycle, ewcPP_PMESENDX);
 830
 831         bBS = (inputrec->nwall == 2);
 832         if (bBS)
 833         {
 834             copy_mat(box, boxs);
 835             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
 836         }
 837
 838         gmx_pme_send_x(cr, bBS ? boxs : box, x,
 839                        mdatoms->nChargePerturbed, lambda[efptCOUL],
 840                        (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)), step);
 841
 842         wallcycle_stop(wcycle, ewcPP_PMESENDX);
 843     }
 844 #endif /* GMX_MPI */
 845
 846     /* do gridding for pair search */
 847     if (bNS)
 848     {
 849         if (graph && bStateChanged)
 850         {
 851             /* Calculate intramolecular shift vectors to make molecules whole */
 852             mk_mshift(fplog, graph, fr->ePBC, box, x);
 853         }
 854
 855         clear_rvec(vzero);
 856         box_diag[XX] = box[XX][XX];
 857         box_diag[YY] = box[YY][YY];
 858         box_diag[ZZ] = box[ZZ][ZZ];
 859
 860         wallcycle_start(wcycle, ewcNS);
 861         if (!fr->bDomDec)
 862         {
 863             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 864             nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
 865                               0, vzero, box_diag,
 866                               0, mdatoms->homenr, -1, fr->cginfo, x,
 867                               0, NULL,
 868                               nbv->grp[eintLocal].kernel_type,
 869                               nbv->grp[eintLocal].nbat);
 870             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 871         }
 872         else
 873         {
 874             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 875             nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
 876                                        fr->cginfo, x,
 877                                        nbv->grp[eintNonlocal].kernel_type,
 878                                        nbv->grp[eintNonlocal].nbat);
 879             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 880         }
 881
 882         if (nbv->ngrp == 1 ||
 883             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 884         {
 885             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
 886                                nbv->nbs, mdatoms, fr->cginfo);
 887         }
 888         else
 889         {
 890             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
 891                                nbv->nbs, mdatoms, fr->cginfo);
 892             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
 893                                nbv->nbs, mdatoms, fr->cginfo);
 894         }
 895         wallcycle_stop(wcycle, ewcNS);
 896     }
 897
 898     /* initialize the GPU atom data and copy shift vector */
 899     if (bUseGPU)
 900     {
 901         if (bNS)
 902         {
 903             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 904             nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 905             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 906         }
 907
 908         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 909         nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 910         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 911     }
 912
 913     /* do local pair search */
 914     if (bNS)
 915     {
 916         wallcycle_start_nocount(wcycle, ewcNS);
 917         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
 918         nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
 919                             &top->excls,
 920                             ic->rlist,
 921                             nbv->min_ci_balanced,
 922                             &nbv->grp[eintLocal].nbl_lists,
 923                             eintLocal,
 924                             nbv->grp[eintLocal].kernel_type,
 925                             nrnb);
 926         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
 927
 928         if (bUseGPU)
 929         {
 930             /* initialize local pair-list on the GPU */
 931             nbnxn_cuda_init_pairlist(nbv->cu_nbv,
 932                                      nbv->grp[eintLocal].nbl_lists.nbl[0],
 933                                      eintLocal);
 934         }
 935         wallcycle_stop(wcycle, ewcNS);
 936     }
 937     else
 938     {
 939         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 940         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
 941         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
 942                                         nbv->grp[eintLocal].nbat);
 943         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
 944         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
 945     }
 946
 947     if (bUseGPU)
 948     {
 949         wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
 950         /* launch local nonbonded F on GPU */
 951         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
 952                      nrnb);
 953         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 954     }
 955
 956     /* Communicate coordinates and sum dipole if necessary +
 957        do non-local pair search */
 958     if (DOMAINDECOMP(cr))
 959     {
 960         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
 961                         nbv->grp[eintLocal].kernel_type);
 962
 963         if (bDiffKernels)
 964         {
 965             /* With GPU+CPU non-bonded calculations we need to copy
 966              * the local coordinates to the non-local nbat struct
 967              * (in CPU format) as the non-local kernel call also
 968              * calculates the local - non-local interactions.
 969              */
 970             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 971             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
 972             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
 973                                             nbv->grp[eintNonlocal].nbat);
 974             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
 975             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
 976         }
 977
 978         if (bNS)
 979         {
 980             wallcycle_start_nocount(wcycle, ewcNS);
 981             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
 982
 983             if (bDiffKernels)
 984             {
 985                 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
 986             }
 987
 988             nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
 989                                 &top->excls,
 990                                 ic->rlist,
 991                                 nbv->min_ci_balanced,
 992                                 &nbv->grp[eintNonlocal].nbl_lists,
 993                                 eintNonlocal,
 994                                 nbv->grp[eintNonlocal].kernel_type,
 995                                 nrnb);
 996
 997             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
 998
 999             if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_CUDA)
1000             {
1001                 /* initialize non-local pair-list on the GPU */
1002                 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1003                                          nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1004                                          eintNonlocal);
1005             }
1006             wallcycle_stop(wcycle, ewcNS);
1007         }
1008         else
1009         {
1010             wallcycle_start(wcycle, ewcMOVEX);
1011             dd_move_x(cr->dd, box, x);
1012
1013             /* When we don't need the total dipole we sum it in global_stat */
1014             if (bStateChanged && NEED_MUTOT(*inputrec))
1015             {
1016                 gmx_sumd(2*DIM, mu, cr);
1017             }
1018             wallcycle_stop(wcycle, ewcMOVEX);
1019
1020             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1021             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1022             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1023                                             nbv->grp[eintNonlocal].nbat);
1024             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1025             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1026         }
1027
1028         if (bUseGPU && !bDiffKernels)
1029         {
1030             wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1031             /* launch non-local nonbonded F on GPU */
1032             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1033                          nrnb);
1034             cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1035         }
1036     }
1037
1038     if (bUseGPU)
1039     {
1040         /* launch D2H copy-back F */
1041         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1042         if (DOMAINDECOMP(cr) && !bDiffKernels)
1043         {
1044             nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1045                                       flags, eatNonlocal);
1046         }
1047         nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1048                                   flags, eatLocal);
1049         cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1050     }
1051
1052     if (bStateChanged && NEED_MUTOT(*inputrec))
1053     {
1054         if (PAR(cr))
1055         {
1056             gmx_sumd(2*DIM, mu, cr);
1057         }
1058
1059         for (i = 0; i < 2; i++)
1060         {
1061             for (j = 0; j < DIM; j++)
1062             {
1063                 fr->mu_tot[i][j] = mu[i*DIM + j];
1064             }
1065         }
1066     }
1067     if (fr->efep == efepNO)
1068     {
1069         copy_rvec(fr->mu_tot[0], mu_tot);
1070     }
1071     else
1072     {
1073         for (j = 0; j < DIM; j++)
1074         {
1075             mu_tot[j] =
1076                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1077                 lambda[efptCOUL]*fr->mu_tot[1][j];
1078         }
1079     }
1080
1081     /* Reset energies */
1082     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1083     clear_rvecs(SHIFTS, fr->fshift);
1084
1085     if (DOMAINDECOMP(cr))
1086     {
1087         if (!(cr->duty & DUTY_PME))
1088         {
1089             wallcycle_start(wcycle, ewcPPDURINGPME);
1090             dd_force_flop_start(cr->dd, nrnb);
1091         }
1092     }
1093
1094     if (inputrec->bRot)
1095     {
1096         /* Enforced rotation has its own cycle counter that starts after the collective
1097          * coordinates have been communicated. It is added to ddCyclF to allow
1098          * for proper load-balancing */
1099         wallcycle_start(wcycle, ewcROT);
1100         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1101         wallcycle_stop(wcycle, ewcROT);
1102     }
1103
1104     /* Start the force cycle counter.
1105      * This counter is stopped in do_forcelow_level.
1106      * No parallel communication should occur while this counter is running,
1107      * since that will interfere with the dynamic load balancing.
1108      */
1109     wallcycle_start(wcycle, ewcFORCE);
1110     if (bDoForces)
1111     {
1112         /* Reset forces for which the virial is calculated separately:
1113          * PME/Ewald forces if necessary */
1114         if (fr->bF_NoVirSum)
1115         {
1116             if (flags & GMX_FORCE_VIRIAL)
1117             {
1118                 fr->f_novirsum = fr->f_novirsum_alloc;
1119                 if (fr->bDomDec)
1120                 {
1121                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1122                 }
1123                 else
1124                 {
1125                     clear_rvecs(homenr, fr->f_novirsum+start);
1126                 }
1127             }
1128             else
1129             {
1130                 /* We are not calculating the pressure so we do not need
1131                  * a separate array for forces that do not contribute
1132                  * to the pressure.
1133                  */
1134                 fr->f_novirsum = f;
1135             }
1136         }
1137
1138         /* Clear the short- and long-range forces */
1139         clear_rvecs(fr->natoms_force_constr, f);
1140         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1141         {
1142             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1143         }
1144
1145         clear_rvec(fr->vir_diag_posres);
1146     }
1147
1148     if (inputrec->ePull == epullCONSTRAINT)
1149     {
1150         clear_pull_forces(inputrec->pull);
1151     }
1152
1153     /* We calculate the non-bonded forces, when done on the CPU, here.
1154      * We do this before calling do_force_lowlevel, as in there bondeds
1155      * forces are calculated before PME, which does communication.
1156      * With this order, non-bonded and bonded force calculation imbalance
1157      * can be balanced out by the domain decomposition load balancing.
1158      */
1159
1160     if (!bUseOrEmulGPU)
1161     {
1162         /* Maybe we should move this into do_force_lowlevel */
1163         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1164                      nrnb);
1165     }
1166
1167     if (!bUseOrEmulGPU || bDiffKernels)
1168     {
1169         int aloc;
1170
1171         if (DOMAINDECOMP(cr))
1172         {
1173             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1174                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1175                          nrnb);
1176         }
1177
1178         if (!bUseOrEmulGPU)
1179         {
1180             aloc = eintLocal;
1181         }
1182         else
1183         {
1184             aloc = eintNonlocal;
1185         }
1186
1187         /* Add all the non-bonded force to the normal force array.
1188          * This can be split into a local a non-local part when overlapping
1189          * communication with calculation with domain decomposition.
1190          */
1191         cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1192         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1193         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1194         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1195         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1196         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1197         wallcycle_start_nocount(wcycle, ewcFORCE);
1198
1199         /* if there are multiple fshift output buffers reduce them */
1200         if ((flags & GMX_FORCE_VIRIAL) &&
1201             nbv->grp[aloc].nbl_lists.nnbl > 1)
1202         {
1203             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1204                                                      fr->fshift);
1205         }
1206     }
1207
1208     /* update QMMMrec, if necessary */
1209     if (fr->bQMMM)
1210     {
1211         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1212     }
1213
1214     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1215     {
1216         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1217                        enerd, lambda, fr);
1218     }
1219
1220     /* Compute the bonded and non-bonded energies and optionally forces */
1221     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1222                       cr, nrnb, wcycle, mdatoms,
1223                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1224                       &(top->atomtypes), bBornRadii, box,
1225                       inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1226                       flags, &cycles_pme);
1227
1228     if (bSepLRF)
1229     {
1230         if (do_per_step(step, inputrec->nstcalclr))
1231         {
1232             /* Add the long range forces to the short range forces */
1233             for (i = 0; i < fr->natoms_force_constr; i++)
1234             {
1235                 rvec_add(fr->f_twin[i], f[i], f[i]);
1236             }
1237         }
1238     }
1239
1240     cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1241
1242     if (ed)
1243     {
1244         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1245     }
1246
1247     if (bUseOrEmulGPU && !bDiffKernels)
1248     {
1249         /* wait for non-local forces (or calculate in emulation mode) */
1250         if (DOMAINDECOMP(cr))
1251         {
1252             if (bUseGPU)
1253             {
1254                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1255                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1256                                     nbv->grp[eintNonlocal].nbat,
1257                                     flags, eatNonlocal,
1258                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1259                                     fr->fshift);
1260                 cycles_force += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1261             }
1262             else
1263             {
1264                 wallcycle_start_nocount(wcycle, ewcFORCE);
1265                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1266                              nrnb);
1267                 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1268             }
1269             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1270             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1271             /* skip the reduction if there was no non-local work to do */
1272             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1273             {
1274                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1275                                                nbv->grp[eintNonlocal].nbat, f);
1276             }
1277             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1278             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1279         }
1280     }
1281
1282     if (bDoForces)
1283     {
1284         /* Communicate the forces */
1285         if (PAR(cr))
1286         {
1287             wallcycle_start(wcycle, ewcMOVEF);
1288             if (DOMAINDECOMP(cr))
1289             {
1290                 dd_move_f(cr->dd, f, fr->fshift);
1291                 /* Do we need to communicate the separate force array
1292                  * for terms that do not contribute to the single sum virial?
1293                  * Position restraints and electric fields do not introduce
1294                  * inter-cg forces, only full electrostatics methods do.
1295                  * When we do not calculate the virial, fr->f_novirsum = f,
1296                  * so we have already communicated these forces.
1297                  */
1298                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1299                     (flags & GMX_FORCE_VIRIAL))
1300                 {
1301                     dd_move_f(cr->dd, fr->f_novirsum, NULL);
1302                 }
1303                 if (bSepLRF)
1304                 {
1305                     /* We should not update the shift forces here,
1306                      * since f_twin is already included in f.
1307                      */
1308                     dd_move_f(cr->dd, fr->f_twin, NULL);
1309                 }
1310             }
1311             wallcycle_stop(wcycle, ewcMOVEF);
1312         }
1313     }
1314
1315     if (bUseOrEmulGPU)
1316     {
1317         /* wait for local forces (or calculate in emulation mode) */
1318         if (bUseGPU)
1319         {
1320             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1321             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1322                                 nbv->grp[eintLocal].nbat,
1323                                 flags, eatLocal,
1324                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1325                                 fr->fshift);
1326             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1327
1328             /* now clear the GPU outputs while we finish the step on the CPU */
1329
1330             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1331             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1332             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1333         }
1334         else
1335         {
1336             wallcycle_start_nocount(wcycle, ewcFORCE);
1337             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1338                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1339                          nrnb);
1340             wallcycle_stop(wcycle, ewcFORCE);
1341         }
1342         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1343         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1344         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1345         {
1346             /* skip the reduction if there was no non-local work to do */
1347             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1348                                            nbv->grp[eintLocal].nbat, f);
1349         }
1350         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1351         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1352     }
1353
1354     if (DOMAINDECOMP(cr))
1355     {
1356         dd_force_flop_stop(cr->dd, nrnb);
1357         if (wcycle)
1358         {
1359             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1360         }
1361     }
1362
1363     if (bDoForces)
1364     {
1365         if (IR_ELEC_FIELD(*inputrec))
1366         {
1367             /* Compute forces due to electric field */
1368             calc_f_el(MASTER(cr) ? field : NULL,
1369                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1370                       inputrec->ex, inputrec->et, t);
1371         }
1372
1373         /* If we have NoVirSum forces, but we do not calculate the virial,
1374          * we sum fr->f_novirum=f later.
1375          */
1376         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1377         {
1378             wallcycle_start(wcycle, ewcVSITESPREAD);
1379             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1380                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1381             wallcycle_stop(wcycle, ewcVSITESPREAD);
1382
1383             if (bSepLRF)
1384             {
1385                 wallcycle_start(wcycle, ewcVSITESPREAD);
1386                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1387                                nrnb,
1388                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1389                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1390             }
1391         }
1392
1393         if (flags & GMX_FORCE_VIRIAL)
1394         {
1395             /* Calculation of the virial must be done after vsites! */
1396             calc_virial(mdatoms->start, mdatoms->homenr, x, f,
1397                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1398         }
1399     }
1400
1401     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1402     {
1403         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1404                                f, vir_force, mdatoms, enerd, lambda, t);
1405     }
1406
1407     /* Add the forces from enforced rotation potentials (if any) */
1408     if (inputrec->bRot)
1409     {
1410         wallcycle_start(wcycle, ewcROTadd);
1411         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1412         wallcycle_stop(wcycle, ewcROTadd);
1413     }
1414
1415     if (PAR(cr) && !(cr->duty & DUTY_PME))
1416     {
1417         /* In case of node-splitting, the PP nodes receive the long-range
1418          * forces, virial and energy from the PME nodes here.
1419          */
1420         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1421     }
1422
1423     if (bDoForces)
1424     {
1425         post_process_forces(cr, step, nrnb, wcycle,
1426                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1427                             flags);
1428     }
1429
1430     /* Sum the potential energy terms from group contributions */
1431     sum_epot(&(enerd->grpp), enerd->term);
1432 }
1433
1434 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1435                         t_inputrec *inputrec,
1436                         gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1437                         gmx_localtop_t *top,
1438                         gmx_groups_t *groups,
1439                         matrix box, rvec x[], history_t *hist,
1440                         rvec f[],
1441                         tensor vir_force,
1442                         t_mdatoms *mdatoms,
1443                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1444                         real *lambda, t_graph *graph,
1445                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1446                         double t, FILE *field, gmx_edsam_t ed,
1447                         gmx_bool bBornRadii,
1448                         int flags)
1449 {
1450     int        cg0, cg1, i, j;
1451     int        start, homenr;
1452     double     mu[2*DIM];
1453     gmx_bool   bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1454     gmx_bool   bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1455     gmx_bool   bDoAdressWF;
1456     matrix     boxs;
1457     rvec       vzero, box_diag;
1458     real       e, v, dvdlambda[efptNR];
1459     t_pbc      pbc;
1460     float      cycles_pme, cycles_force;
1461
1462     start  = mdatoms->start;
1463     homenr = mdatoms->homenr;
1464
1465     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
1466
1467     clear_mat(vir_force);
1468
1469     if (PARTDECOMP(cr))
1470     {
1471         pd_cg_range(cr, &cg0, &cg1);
1472     }
1473     else
1474     {
1475         cg0 = 0;
1476         if (DOMAINDECOMP(cr))
1477         {
1478             cg1 = cr->dd->ncg_tot;
1479         }
1480         else
1481         {
1482             cg1 = top->cgs.nr;
1483         }
1484         if (fr->n_tpi > 0)
1485         {
1486             cg1--;
1487         }
1488     }
1489
1490     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1491     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1492     /* Should we update the long-range neighborlists at this step? */
1493     bDoLongRangeNS = fr->bTwinRange && bNS;
1494     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1495     bFillGrid      = (bNS && bStateChanged);
1496     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1497     bDoForces      = (flags & GMX_FORCE_FORCES);
1498     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1499     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1500                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1501
1502     /* should probably move this to the forcerec since it doesn't change */
1503     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1504
1505     if (bStateChanged)
1506     {
1507         update_forcerec(fr, box);
1508
1509         if (NEED_MUTOT(*inputrec))
1510         {
1511             /* Calculate total (local) dipole moment in a temporary common array.
1512              * This makes it possible to sum them over nodes faster.
1513              */
1514             calc_mu(start, homenr,
1515                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1516                     mu, mu+DIM);
1517         }
1518     }
1519
1520     if (fr->ePBC != epbcNONE)
1521     {
1522         /* Compute shift vectors every step,
1523          * because of pressure coupling or box deformation!
1524          */
1525         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1526         {
1527             calc_shifts(box, fr->shift_vec);
1528         }
1529
1530         if (bCalcCGCM)
1531         {
1532             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1533                                      &(top->cgs), x, fr->cg_cm);
1534             inc_nrnb(nrnb, eNR_CGCM, homenr);
1535             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1536         }
1537         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1538         {
1539             unshift_self(graph, box, x);
1540         }
1541     }
1542     else if (bCalcCGCM)
1543     {
1544         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1545         inc_nrnb(nrnb, eNR_CGCM, homenr);
1546     }
1547
1548     if (bCalcCGCM)
1549     {
1550         if (PAR(cr))
1551         {
1552             move_cgcm(fplog, cr, fr->cg_cm);
1553         }
1554         if (gmx_debug_at)
1555         {
1556             pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1557         }
1558     }
1559
1560 #ifdef GMX_MPI
1561     if (!(cr->duty & DUTY_PME))
1562     {
1563         /* Send particle coordinates to the pme nodes.
1564          * Since this is only implemented for domain decomposition
1565          * and domain decomposition does not use the graph,
1566          * we do not need to worry about shifting.
1567          */
1568
1569         wallcycle_start(wcycle, ewcPP_PMESENDX);
1570
1571         bBS = (inputrec->nwall == 2);
1572         if (bBS)
1573         {
1574             copy_mat(box, boxs);
1575             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1576         }
1577
1578         gmx_pme_send_x(cr, bBS ? boxs : box, x,
1579                        mdatoms->nChargePerturbed, lambda[efptCOUL],
1580                        (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)), step);
1581
1582         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1583     }
1584 #endif /* GMX_MPI */
1585
1586     /* Communicate coordinates and sum dipole if necessary */
1587     if (PAR(cr))
1588     {
1589         wallcycle_start(wcycle, ewcMOVEX);
1590         if (DOMAINDECOMP(cr))
1591         {
1592             dd_move_x(cr->dd, box, x);
1593         }
1594         else
1595         {
1596             move_x(cr, x, nrnb);
1597         }
1598         wallcycle_stop(wcycle, ewcMOVEX);
1599     }
1600
1601     /* update adress weight beforehand */
1602     if (bStateChanged && bDoAdressWF)
1603     {
1604         /* need pbc for adress weight calculation with pbc_dx */
1605         set_pbc(&pbc, inputrec->ePBC, box);
1606         if (fr->adress_site == eAdressSITEcog)
1607         {
1608             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1609                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1610         }
1611         else if (fr->adress_site == eAdressSITEcom)
1612         {
1613             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1614                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1615         }
1616         else if (fr->adress_site == eAdressSITEatomatom)
1617         {
1618             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1619                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1620         }
1621         else
1622         {
1623             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1624                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1625         }
1626     }
1627
1628     if (NEED_MUTOT(*inputrec))
1629     {
1630
1631         if (bStateChanged)
1632         {
1633             if (PAR(cr))
1634             {
1635                 gmx_sumd(2*DIM, mu, cr);
1636             }
1637             for (i = 0; i < 2; i++)
1638             {
1639                 for (j = 0; j < DIM; j++)
1640                 {
1641                     fr->mu_tot[i][j] = mu[i*DIM + j];
1642                 }
1643             }
1644         }
1645         if (fr->efep == efepNO)
1646         {
1647             copy_rvec(fr->mu_tot[0], mu_tot);
1648         }
1649         else
1650         {
1651             for (j = 0; j < DIM; j++)
1652             {
1653                 mu_tot[j] =
1654                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1655             }
1656         }
1657     }
1658
1659     /* Reset energies */
1660     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1661     clear_rvecs(SHIFTS, fr->fshift);
1662
1663     if (bNS)
1664     {
1665         wallcycle_start(wcycle, ewcNS);
1666
1667         if (graph && bStateChanged)
1668         {
1669             /* Calculate intramolecular shift vectors to make molecules whole */
1670             mk_mshift(fplog, graph, fr->ePBC, box, x);
1671         }
1672
1673         /* Do the actual neighbour searching and if twin range electrostatics
1674          * also do the calculation of long range forces and energies.
1675          */
1676         for (i = 0; i < efptNR; i++)
1677         {
1678             dvdlambda[i] = 0;
1679         }
1680         ns(fplog, fr, box,
1681            groups, top, mdatoms,
1682            cr, nrnb, bFillGrid,
1683            bDoLongRangeNS);
1684
1685         wallcycle_stop(wcycle, ewcNS);
1686     }
1687
1688     if (inputrec->implicit_solvent && bNS)
1689     {
1690         make_gb_nblist(cr, inputrec->gb_algorithm,
1691                        x, box, fr, &top->idef, graph, fr->born);
1692     }
1693
1694     if (DOMAINDECOMP(cr))
1695     {
1696         if (!(cr->duty & DUTY_PME))
1697         {
1698             wallcycle_start(wcycle, ewcPPDURINGPME);
1699             dd_force_flop_start(cr->dd, nrnb);
1700         }
1701     }
1702
1703     if (inputrec->bRot)
1704     {
1705         /* Enforced rotation has its own cycle counter that starts after the collective
1706          * coordinates have been communicated. It is added to ddCyclF to allow
1707          * for proper load-balancing */
1708         wallcycle_start(wcycle, ewcROT);
1709         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1710         wallcycle_stop(wcycle, ewcROT);
1711     }
1712
1713     /* Start the force cycle counter.
1714      * This counter is stopped in do_forcelow_level.
1715      * No parallel communication should occur while this counter is running,
1716      * since that will interfere with the dynamic load balancing.
1717      */
1718     wallcycle_start(wcycle, ewcFORCE);
1719
1720     if (bDoForces)
1721     {
1722         /* Reset forces for which the virial is calculated separately:
1723          * PME/Ewald forces if necessary */
1724         if (fr->bF_NoVirSum)
1725         {
1726             if (flags & GMX_FORCE_VIRIAL)
1727             {
1728                 fr->f_novirsum = fr->f_novirsum_alloc;
1729                 if (fr->bDomDec)
1730                 {
1731                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1732                 }
1733                 else
1734                 {
1735                     clear_rvecs(homenr, fr->f_novirsum+start);
1736                 }
1737             }
1738             else
1739             {
1740                 /* We are not calculating the pressure so we do not need
1741                  * a separate array for forces that do not contribute
1742                  * to the pressure.
1743                  */
1744                 fr->f_novirsum = f;
1745             }
1746         }
1747
1748         /* Clear the short- and long-range forces */
1749         clear_rvecs(fr->natoms_force_constr, f);
1750         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1751         {
1752             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1753         }
1754
1755         clear_rvec(fr->vir_diag_posres);
1756     }
1757     if (inputrec->ePull == epullCONSTRAINT)
1758     {
1759         clear_pull_forces(inputrec->pull);
1760     }
1761
1762     /* update QMMMrec, if necessary */
1763     if (fr->bQMMM)
1764     {
1765         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1766     }
1767
1768     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1769     {
1770         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1771                        enerd, lambda, fr);
1772     }
1773
1774     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1775     {
1776         /* Flat-bottomed position restraints always require full pbc */
1777         if (!(bStateChanged && bDoAdressWF))
1778         {
1779             set_pbc(&pbc, inputrec->ePBC, box);
1780         }
1781         v = fbposres(top->idef.il[F_FBPOSRES].nr, top->idef.il[F_FBPOSRES].iatoms,
1782                      top->idef.iparams_fbposres,
1783                      (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
1784                      inputrec->ePBC == epbcNONE ? NULL : &pbc,
1785                      fr->rc_scaling, fr->ePBC, fr->posres_com);
1786         enerd->term[F_FBPOSRES] += v;
1787         inc_nrnb(nrnb, eNR_FBPOSRES, top->idef.il[F_FBPOSRES].nr/2);
1788     }
1789
1790     /* Compute the bonded and non-bonded energies and optionally forces */
1791     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1792                       cr, nrnb, wcycle, mdatoms,
1793                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1794                       &(top->atomtypes), bBornRadii, box,
1795                       inputrec->fepvals, lambda,
1796                       graph, &(top->excls), fr->mu_tot,
1797                       flags,
1798                       &cycles_pme);
1799
1800     if (bSepLRF)
1801     {
1802         if (do_per_step(step, inputrec->nstcalclr))
1803         {
1804             /* Add the long range forces to the short range forces */
1805             for (i = 0; i < fr->natoms_force_constr; i++)
1806             {
1807                 rvec_add(fr->f_twin[i], f[i], f[i]);
1808             }
1809         }
1810     }
1811
1812     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1813
1814     if (ed)
1815     {
1816         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1817     }
1818
1819     if (DOMAINDECOMP(cr))
1820     {
1821         dd_force_flop_stop(cr->dd, nrnb);
1822         if (wcycle)
1823         {
1824             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1825         }
1826     }
1827
1828     if (bDoForces)
1829     {
1830         if (IR_ELEC_FIELD(*inputrec))
1831         {
1832             /* Compute forces due to electric field */
1833             calc_f_el(MASTER(cr) ? field : NULL,
1834                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1835                       inputrec->ex, inputrec->et, t);
1836         }
1837
1838         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1839         {
1840             /* Compute thermodynamic force in hybrid AdResS region */
1841             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1842                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1843         }
1844
1845         /* Communicate the forces */
1846         if (PAR(cr))
1847         {
1848             wallcycle_start(wcycle, ewcMOVEF);
1849             if (DOMAINDECOMP(cr))
1850             {
1851                 dd_move_f(cr->dd, f, fr->fshift);
1852                 /* Do we need to communicate the separate force array
1853                  * for terms that do not contribute to the single sum virial?
1854                  * Position restraints and electric fields do not introduce
1855                  * inter-cg forces, only full electrostatics methods do.
1856                  * When we do not calculate the virial, fr->f_novirsum = f,
1857                  * so we have already communicated these forces.
1858                  */
1859                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1860                     (flags & GMX_FORCE_VIRIAL))
1861                 {
1862                     dd_move_f(cr->dd, fr->f_novirsum, NULL);
1863                 }
1864                 if (bSepLRF)
1865                 {
1866                     /* We should not update the shift forces here,
1867                      * since f_twin is already included in f.
1868                      */
1869                     dd_move_f(cr->dd, fr->f_twin, NULL);
1870                 }
1871             }
1872             else
1873             {
1874                 pd_move_f(cr, f, nrnb);
1875                 if (bSepLRF)
1876                 {
1877                     pd_move_f(cr, fr->f_twin, nrnb);
1878                 }
1879             }
1880             wallcycle_stop(wcycle, ewcMOVEF);
1881         }
1882
1883         /* If we have NoVirSum forces, but we do not calculate the virial,
1884          * we sum fr->f_novirum=f later.
1885          */
1886         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1887         {
1888             wallcycle_start(wcycle, ewcVSITESPREAD);
1889             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1890                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1891             wallcycle_stop(wcycle, ewcVSITESPREAD);
1892
1893             if (bSepLRF)
1894             {
1895                 wallcycle_start(wcycle, ewcVSITESPREAD);
1896                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1897                                nrnb,
1898                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1899                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1900             }
1901         }
1902
1903         if (flags & GMX_FORCE_VIRIAL)
1904         {
1905             /* Calculation of the virial must be done after vsites! */
1906             calc_virial(mdatoms->start, mdatoms->homenr, x, f,
1907                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1908         }
1909     }
1910
1911     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1912     {
1913         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1914                                f, vir_force, mdatoms, enerd, lambda, t);
1915     }
1916
1917     /* Add the forces from enforced rotation potentials (if any) */
1918     if (inputrec->bRot)
1919     {
1920         wallcycle_start(wcycle, ewcROTadd);
1921         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1922         wallcycle_stop(wcycle, ewcROTadd);
1923     }
1924
1925     if (PAR(cr) && !(cr->duty & DUTY_PME))
1926     {
1927         /* In case of node-splitting, the PP nodes receive the long-range
1928          * forces, virial and energy from the PME nodes here.
1929          */
1930         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1931     }
1932
1933     if (bDoForces)
1934     {
1935         post_process_forces(cr, step, nrnb, wcycle,
1936                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1937                             flags);
1938     }
1939
1940     /* Sum the potential energy terms from group contributions */
1941     sum_epot(&(enerd->grpp), enerd->term);
1942 }
1943
1944 void do_force(FILE *fplog, t_commrec *cr,
1945               t_inputrec *inputrec,
1946               gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1947               gmx_localtop_t *top,
1948               gmx_groups_t *groups,
1949               matrix box, rvec x[], history_t *hist,
1950               rvec f[],
1951               tensor vir_force,
1952               t_mdatoms *mdatoms,
1953               gmx_enerdata_t *enerd, t_fcdata *fcd,
1954               real *lambda, t_graph *graph,
1955               t_forcerec *fr,
1956               gmx_vsite_t *vsite, rvec mu_tot,
1957               double t, FILE *field, gmx_edsam_t ed,
1958               gmx_bool bBornRadii,
1959               int flags)
1960 {
1961     /* modify force flag if not doing nonbonded */
1962     if (!fr->bNonbonded)
1963     {
1964         flags &= ~GMX_FORCE_NONBONDED;
1965     }
1966
1967     switch (inputrec->cutoff_scheme)
1968     {
1969         case ecutsVERLET:
1970             do_force_cutsVERLET(fplog, cr, inputrec,
1971                                 step, nrnb, wcycle,
1972                                 top,
1973                                 groups,
1974                                 box, x, hist,
1975                                 f, vir_force,
1976                                 mdatoms,
1977                                 enerd, fcd,
1978                                 lambda, graph,
1979                                 fr, fr->ic,
1980                                 vsite, mu_tot,
1981                                 t, field, ed,
1982                                 bBornRadii,
1983                                 flags);
1984             break;
1985         case ecutsGROUP:
1986             do_force_cutsGROUP(fplog, cr, inputrec,
1987                                step, nrnb, wcycle,
1988                                top,
1989                                groups,
1990                                box, x, hist,
1991                                f, vir_force,
1992                                mdatoms,
1993                                enerd, fcd,
1994                                lambda, graph,
1995                                fr, vsite, mu_tot,
1996                                t, field, ed,
1997                                bBornRadii,
1998                                flags);
1999             break;
2000         default:
2001             gmx_incons("Invalid cut-off scheme passed!");
2002     }
2003 }
2004
2005
2006 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2007                         t_inputrec *ir, t_mdatoms *md,
2008                         t_state *state, t_commrec *cr, t_nrnb *nrnb,
2009                         t_forcerec *fr, gmx_localtop_t *top)
2010 {
2011     int             i, m, start, end;
2012     gmx_large_int_t step;
2013     real            dt = ir->delta_t;
2014     real            dvdl_dum;
2015     rvec           *savex;
2016
2017     snew(savex, state->natoms);
2018
2019     start = md->start;
2020     end   = md->homenr + start;
2021
2022     if (debug)
2023     {
2024         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2025                 start, md->homenr, end);
2026     }
2027     /* Do a first constrain to reset particles... */
2028     step = ir->init_step;
2029     if (fplog)
2030     {
2031         char buf[STEPSTRSIZE];
2032         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2033                 gmx_step_str(step, buf));
2034     }
2035     dvdl_dum = 0;
2036
2037     /* constrain the current position */
2038     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2039               ir, NULL, cr, step, 0, md,
2040               state->x, state->x, NULL,
2041               fr->bMolPBC, state->box,
2042               state->lambda[efptBONDED], &dvdl_dum,
2043               NULL, NULL, nrnb, econqCoord,
2044               ir->epc == epcMTTK, state->veta, state->veta);
2045     if (EI_VV(ir->eI))
2046     {
2047         /* constrain the inital velocity, and save it */
2048         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2049         /* might not yet treat veta correctly */
2050         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2051                   ir, NULL, cr, step, 0, md,
2052                   state->x, state->v, state->v,
2053                   fr->bMolPBC, state->box,
2054                   state->lambda[efptBONDED], &dvdl_dum,
2055                   NULL, NULL, nrnb, econqVeloc,
2056                   ir->epc == epcMTTK, state->veta, state->veta);
2057     }
2058     /* constrain the inital velocities at t-dt/2 */
2059     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2060     {
2061         for (i = start; (i < end); i++)
2062         {
2063             for (m = 0; (m < DIM); m++)
2064             {
2065                 /* Reverse the velocity */
2066                 state->v[i][m] = -state->v[i][m];
2067                 /* Store the position at t-dt in buf */
2068                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2069             }
2070         }
2071         /* Shake the positions at t=-dt with the positions at t=0
2072          * as reference coordinates.
2073          */
2074         if (fplog)
2075         {
2076             char buf[STEPSTRSIZE];
2077             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2078                     gmx_step_str(step, buf));
2079         }
2080         dvdl_dum = 0;
2081         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2082                   ir, NULL, cr, step, -1, md,
2083                   state->x, savex, NULL,
2084                   fr->bMolPBC, state->box,
2085                   state->lambda[efptBONDED], &dvdl_dum,
2086                   state->v, NULL, nrnb, econqCoord,
2087                   ir->epc == epcMTTK, state->veta, state->veta);
2088
2089         for (i = start; i < end; i++)
2090         {
2091             for (m = 0; m < DIM; m++)
2092             {
2093                 /* Re-reverse the velocities */
2094                 state->v[i][m] = -state->v[i][m];
2095             }
2096         }
2097     }
2098     sfree(savex);
2099 }
2100
2101 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2102 {
2103     double eners[2], virs[2], enersum, virsum, y0, f, g, h;
2104     double r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2105     double invscale, invscale2, invscale3;
2106     int    ri0, ri1, ri, i, offstart, offset;
2107     real   scale, *vdwtab, tabfactor, tmp;
2108
2109     fr->enershiftsix    = 0;
2110     fr->enershifttwelve = 0;
2111     fr->enerdiffsix     = 0;
2112     fr->enerdifftwelve  = 0;
2113     fr->virdiffsix      = 0;
2114     fr->virdifftwelve   = 0;
2115
2116     if (eDispCorr != edispcNO)
2117     {
2118         for (i = 0; i < 2; i++)
2119         {
2120             eners[i] = 0;
2121             virs[i]  = 0;
2122         }
2123         if ((fr->vdwtype == evdwSWITCH) || (fr->vdwtype == evdwSHIFT))
2124         {
2125             if (fr->rvdw_switch == 0)
2126             {
2127                 gmx_fatal(FARGS,
2128                           "With dispersion correction rvdw-switch can not be zero "
2129                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2130             }
2131
2132             scale  = fr->nblists[0].table_elec_vdw.scale;
2133             vdwtab = fr->nblists[0].table_vdw.data;
2134
2135             /* Round the cut-offs to exact table values for precision */
2136             ri0  = floor(fr->rvdw_switch*scale);
2137             ri1  = ceil(fr->rvdw*scale);
2138             r0   = ri0/scale;
2139             r1   = ri1/scale;
2140             rc3  = r0*r0*r0;
2141             rc9  = rc3*rc3*rc3;
2142
2143             if (fr->vdwtype == evdwSHIFT)
2144             {
2145                 /* Determine the constant energy shift below rvdw_switch.
2146                  * Table has a scale factor since we have scaled it down to compensate
2147                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2148                  */
2149                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2150                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2151             }
2152             /* Add the constant part from 0 to rvdw_switch.
2153              * This integration from 0 to rvdw_switch overcounts the number
2154              * of interactions by 1, as it also counts the self interaction.
2155              * We will correct for this later.
2156              */
2157             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2158             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2159
2160             invscale  = 1.0/(scale);
2161             invscale2 = invscale*invscale;
2162             invscale3 = invscale*invscale2;
2163
2164             /* following summation derived from cubic spline definition,
2165                Numerical Recipies in C, second edition, p. 113-116.  Exact
2166                for the cubic spline.  We first calculate the negative of
2167                the energy from rvdw to rvdw_switch, assuming that g(r)=1,
2168                and then add the more standard, abrupt cutoff correction to
2169                that result, yielding the long-range correction for a
2170                switched function.  We perform both the pressure and energy
2171                loops at the same time for simplicity, as the computational
2172                cost is low. */
2173
2174             for (i = 0; i < 2; i++)
2175             {
2176                 enersum = 0.0; virsum = 0.0;
2177                 if (i == 0)
2178                 {
2179                     offstart = 0;
2180                     /* Since the dispersion table has been scaled down a factor 6.0 and the repulsion
2181                      * a factor 12.0 to compensate for the c6/c12 parameters inside nbfp[] being scaled
2182                      * up (to save flops in kernels), we need to correct for this.
2183                      */
2184                     tabfactor = 6.0;
2185                 }
2186                 else
2187                 {
2188                     offstart  = 4;
2189                     tabfactor = 12.0;
2190                 }
2191                 for (ri = ri0; ri < ri1; ri++)
2192                 {
2193                     r  = ri*invscale;
2194                     ea = invscale3;
2195                     eb = 2.0*invscale2*r;
2196                     ec = invscale*r*r;
2197
2198                     pa = invscale3;
2199                     pb = 3.0*invscale2*r;
2200                     pc = 3.0*invscale*r*r;
2201                     pd = r*r*r;
2202
2203                     /* this "8" is from the packing in the vdwtab array - perhaps should be #define'ed? */
2204                     offset = 8*ri + offstart;
2205                     y0     = vdwtab[offset];
2206                     f      = vdwtab[offset+1];
2207                     g      = vdwtab[offset+2];
2208                     h      = vdwtab[offset+3];
2209
2210                     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);
2211                     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);
2212                 }
2213
2214                 enersum  *= 4.0*M_PI*tabfactor;
2215                 virsum   *= 4.0*M_PI*tabfactor;
2216                 eners[i] -= enersum;
2217                 virs[i]  -= virsum;
2218             }
2219
2220             /* now add the correction for rvdw_switch to infinity */
2221             eners[0] += -4.0*M_PI/(3.0*rc3);
2222             eners[1] +=  4.0*M_PI/(9.0*rc9);
2223             virs[0]  +=  8.0*M_PI/rc3;
2224             virs[1]  += -16.0*M_PI/(3.0*rc9);
2225         }
2226         else if ((fr->vdwtype == evdwCUT) || (fr->vdwtype == evdwUSER))
2227         {
2228             if (fr->vdwtype == evdwUSER && fplog)
2229             {
2230                 fprintf(fplog,
2231                         "WARNING: using dispersion correction with user tables\n");
2232             }
2233             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2234             rc9  = rc3*rc3*rc3;
2235             /* Contribution beyond the cut-off */
2236             eners[0] += -4.0*M_PI/(3.0*rc3);
2237             eners[1] +=  4.0*M_PI/(9.0*rc9);
2238             if (fr->vdw_modifier == eintmodPOTSHIFT)
2239             {
2240                 /* Contribution within the cut-off */
2241                 eners[0] += -4.0*M_PI/(3.0*rc3);
2242                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2243             }
2244             /* Contribution beyond the cut-off */
2245             virs[0]  +=  8.0*M_PI/rc3;
2246             virs[1]  += -16.0*M_PI/(3.0*rc9);
2247         }
2248         else
2249         {
2250             gmx_fatal(FARGS,
2251                       "Dispersion correction is not implemented for vdw-type = %s",
2252                       evdw_names[fr->vdwtype]);
2253         }
2254         fr->enerdiffsix    = eners[0];
2255         fr->enerdifftwelve = eners[1];
2256         /* The 0.5 is due to the Gromacs definition of the virial */
2257         fr->virdiffsix     = 0.5*virs[0];
2258         fr->virdifftwelve  = 0.5*virs[1];
2259     }
2260 }
2261
2262 void calc_dispcorr(FILE *fplog, t_inputrec *ir, t_forcerec *fr,
2263                    gmx_large_int_t step, int natoms,
2264                    matrix box, real lambda, tensor pres, tensor virial,
2265                    real *prescorr, real *enercorr, real *dvdlcorr)
2266 {
2267     gmx_bool bCorrAll, bCorrPres;
2268     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2269     int      m;
2270
2271     *prescorr = 0;
2272     *enercorr = 0;
2273     *dvdlcorr = 0;
2274
2275     clear_mat(virial);
2276     clear_mat(pres);
2277
2278     if (ir->eDispCorr != edispcNO)
2279     {
2280         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2281                      ir->eDispCorr == edispcAllEnerPres);
2282         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2283                      ir->eDispCorr == edispcAllEnerPres);
2284
2285         invvol = 1/det(box);
2286         if (fr->n_tpi)
2287         {
2288             /* Only correct for the interactions with the inserted molecule */
2289             dens   = (natoms - fr->n_tpi)*invvol;
2290             ninter = fr->n_tpi;
2291         }
2292         else
2293         {
2294             dens   = natoms*invvol;
2295             ninter = 0.5*natoms;
2296         }
2297
2298         if (ir->efep == efepNO)
2299         {
2300             avcsix    = fr->avcsix[0];
2301             avctwelve = fr->avctwelve[0];
2302         }
2303         else
2304         {
2305             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2306             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2307         }
2308
2309         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2310         *enercorr += avcsix*enerdiff;
2311         dvdlambda  = 0.0;
2312         if (ir->efep != efepNO)
2313         {
2314             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2315         }
2316         if (bCorrAll)
2317         {
2318             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2319             *enercorr += avctwelve*enerdiff;
2320             if (fr->efep != efepNO)
2321             {
2322                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2323             }
2324         }
2325
2326         if (bCorrPres)
2327         {
2328             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2329             if (ir->eDispCorr == edispcAllEnerPres)
2330             {
2331                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2332             }
2333             /* The factor 2 is because of the Gromacs virial definition */
2334             spres = -2.0*invvol*svir*PRESFAC;
2335
2336             for (m = 0; m < DIM; m++)
2337             {
2338                 virial[m][m] += svir;
2339                 pres[m][m]   += spres;
2340             }
2341             *prescorr += spres;
2342         }
2343
2344         /* Can't currently control when it prints, for now, just print when degugging */
2345         if (debug)
2346         {
2347             if (bCorrAll)
2348             {
2349                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2350                         avcsix, avctwelve);
2351             }
2352             if (bCorrPres)
2353             {
2354                 fprintf(debug,
2355                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2356                         *enercorr, spres, svir);
2357             }
2358             else
2359             {
2360                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2361             }
2362         }
2363
2364         if (fr->bSepDVDL && do_per_step(step, ir->nstlog))
2365         {
2366             gmx_print_sepdvdl(fplog, "Dispersion correction", *enercorr, dvdlambda);
2367         }
2368         if (fr->efep != efepNO)
2369         {
2370             *dvdlcorr += dvdlambda;
2371         }
2372     }
2373 }
2374
2375 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2376                   t_graph *graph, rvec x[])
2377 {
2378     if (fplog)
2379     {
2380         fprintf(fplog, "Removing pbc first time\n");
2381     }
2382     calc_shifts(box, fr->shift_vec);
2383     if (graph)
2384     {
2385         mk_mshift(fplog, graph, fr->ePBC, box, x);
2386         if (gmx_debug_at)
2387         {
2388             p_graph(debug, "do_pbc_first 1", graph);
2389         }
2390         shift_self(graph, box, x);
2391         /* By doing an extra mk_mshift the molecules that are broken
2392          * because they were e.g. imported from another software
2393          * will be made whole again. Such are the healing powers
2394          * of GROMACS.
2395          */
2396         mk_mshift(fplog, graph, fr->ePBC, box, x);
2397         if (gmx_debug_at)
2398         {
2399             p_graph(debug, "do_pbc_first 2", graph);
2400         }
2401     }
2402     if (fplog)
2403     {
2404         fprintf(fplog, "Done rmpbc\n");
2405     }
2406 }
2407
2408 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2409                             gmx_mtop_t *mtop, rvec x[],
2410                             gmx_bool bFirst)
2411 {
2412     t_graph        *graph;
2413     int             mb, as, mol;
2414     gmx_molblock_t *molb;
2415
2416     if (bFirst && fplog)
2417     {
2418         fprintf(fplog, "Removing pbc first time\n");
2419     }
2420
2421     snew(graph, 1);
2422     as = 0;
2423     for (mb = 0; mb < mtop->nmolblock; mb++)
2424     {
2425         molb = &mtop->molblock[mb];
2426         if (molb->natoms_mol == 1 ||
2427             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2428         {
2429             /* Just one atom or charge group in the molecule, no PBC required */
2430             as += molb->nmol*molb->natoms_mol;
2431         }
2432         else
2433         {
2434             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2435             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2436                            0, molb->natoms_mol, FALSE, FALSE, graph);
2437
2438             for (mol = 0; mol < molb->nmol; mol++)
2439             {
2440                 mk_mshift(fplog, graph, ePBC, box, x+as);
2441
2442                 shift_self(graph, box, x+as);
2443                 /* The molecule is whole now.
2444                  * We don't need the second mk_mshift call as in do_pbc_first,
2445                  * since we no longer need this graph.
2446                  */
2447
2448                 as += molb->natoms_mol;
2449             }
2450             done_graph(graph);
2451         }
2452     }
2453     sfree(graph);
2454 }
2455
2456 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2457                        gmx_mtop_t *mtop, rvec x[])
2458 {
2459     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2460 }
2461
2462 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2463                  gmx_mtop_t *mtop, rvec x[])
2464 {
2465     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2466 }
2467
2468 void finish_run(FILE *fplog, t_commrec *cr,
2469                 t_inputrec *inputrec,
2470                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2471                 gmx_runtime_t *runtime,
2472                 wallclock_gpu_t *gputimes,
2473                 gmx_bool bWriteStat)
2474 {
2475     int     i, j;
2476     t_nrnb *nrnb_tot = NULL;
2477     real    delta_t;
2478     double  nbfs, mflop;
2479
2480     wallcycle_sum(cr, wcycle);
2481
2482     if (cr->nnodes > 1)
2483     {
2484         snew(nrnb_tot, 1);
2485 #ifdef GMX_MPI
2486         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2487                       cr->mpi_comm_mysim);
2488 #endif
2489     }
2490     else
2491     {
2492         nrnb_tot = nrnb;
2493     }
2494
2495 #if defined(GMX_MPI) && !defined(GMX_THREAD_MPI)
2496     if (cr->nnodes > 1)
2497     {
2498         /* reduce nodetime over all MPI processes in the current simulation */
2499         double sum;
2500         MPI_Allreduce(&runtime->proctime, &sum, 1, MPI_DOUBLE, MPI_SUM,
2501                       cr->mpi_comm_mysim);
2502         runtime->proctime = sum;
2503     }
2504 #endif
2505
2506     if (SIMMASTER(cr))
2507     {
2508         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2509     }
2510     if (cr->nnodes > 1)
2511     {
2512         sfree(nrnb_tot);
2513     }
2514
2515     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2516     {
2517         print_dd_statistics(cr, inputrec, fplog);
2518     }
2519
2520 #ifdef GMX_MPI
2521     if (PARTDECOMP(cr))
2522     {
2523         if (MASTER(cr))
2524         {
2525             t_nrnb     *nrnb_all;
2526             int         s;
2527             MPI_Status  stat;
2528
2529             snew(nrnb_all, cr->nnodes);
2530             nrnb_all[0] = *nrnb;
2531             for (s = 1; s < cr->nnodes; s++)
2532             {
2533                 MPI_Recv(nrnb_all[s].n, eNRNB, MPI_DOUBLE, s, 0,
2534                          cr->mpi_comm_mysim, &stat);
2535             }
2536             pr_load(fplog, cr, nrnb_all);
2537             sfree(nrnb_all);
2538         }
2539         else
2540         {
2541             MPI_Send(nrnb->n, eNRNB, MPI_DOUBLE, MASTERRANK(cr), 0,
2542                      cr->mpi_comm_mysim);
2543         }
2544     }
2545 #endif
2546
2547     if (SIMMASTER(cr))
2548     {
2549         wallcycle_print(fplog, cr->nnodes, cr->npmenodes, runtime->realtime,
2550                         wcycle, gputimes);
2551
2552         if (EI_DYNAMICS(inputrec->eI))
2553         {
2554             delta_t = inputrec->delta_t;
2555         }
2556         else
2557         {
2558             delta_t = 0;
2559         }
2560
2561         if (fplog)
2562         {
2563             print_perf(fplog, runtime->proctime, runtime->realtime,
2564                        runtime->nsteps_done, delta_t, nbfs, mflop);
2565         }
2566         if (bWriteStat)
2567         {
2568             print_perf(stderr, runtime->proctime, runtime->realtime,
2569                        runtime->nsteps_done, delta_t, nbfs, mflop);
2570         }
2571     }
2572 }
2573
2574 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2575 {
2576     /* this function works, but could probably use a logic rewrite to keep all the different
2577        types of efep straight. */
2578
2579     int       i;
2580     t_lambda *fep = ir->fepvals;
2581
2582     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2583     {
2584         for (i = 0; i < efptNR; i++)
2585         {
2586             lambda[i] = 0.0;
2587             if (lam0)
2588             {
2589                 lam0[i] = 0.0;
2590             }
2591         }
2592         return;
2593     }
2594     else
2595     {
2596         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2597                                              if checkpoint is set -- a kludge is in for now
2598                                              to prevent this.*/
2599         for (i = 0; i < efptNR; i++)
2600         {
2601             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2602             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2603             {
2604                 lambda[i] = fep->init_lambda;
2605                 if (lam0)
2606                 {
2607                     lam0[i] = lambda[i];
2608                 }
2609             }
2610             else
2611             {
2612                 lambda[i] = fep->all_lambda[i][*fep_state];
2613                 if (lam0)
2614                 {
2615                     lam0[i] = lambda[i];
2616                 }
2617             }
2618         }
2619         if (ir->bSimTemp)
2620         {
2621             /* need to rescale control temperatures to match current state */
2622             for (i = 0; i < ir->opts.ngtc; i++)
2623             {
2624                 if (ir->opts.ref_t[i] > 0)
2625                 {
2626                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2627                 }
2628             }
2629         }
2630     }
2631
2632     /* Send to the log the information on the current lambdas */
2633     if (fplog != NULL)
2634     {
2635         fprintf(fplog, "Initial vector of lambda components:[ ");
2636         for (i = 0; i < efptNR; i++)
2637         {
2638             fprintf(fplog, "%10.4f ", lambda[i]);
2639         }
2640         fprintf(fplog, "]\n");
2641     }
2642     return;
2643 }
2644
2645
2646 void init_md(FILE *fplog,
2647              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2648              double *t, double *t0,
2649              real *lambda, int *fep_state, double *lam0,
2650              t_nrnb *nrnb, gmx_mtop_t *mtop,
2651              gmx_update_t *upd,
2652              int nfile, const t_filenm fnm[],
2653              gmx_mdoutf_t **outf, t_mdebin **mdebin,
2654              tensor force_vir, tensor shake_vir, rvec mu_tot,
2655              gmx_bool *bSimAnn, t_vcm **vcm, unsigned long Flags)
2656 {
2657     int  i, j, n;
2658     real tmpt, mod;
2659
2660     /* Initial values */
2661     *t = *t0       = ir->init_t;
2662
2663     *bSimAnn = FALSE;
2664     for (i = 0; i < ir->opts.ngtc; i++)
2665     {
2666         /* set bSimAnn if any group is being annealed */
2667         if (ir->opts.annealing[i] != eannNO)
2668         {
2669             *bSimAnn = TRUE;
2670         }
2671     }
2672     if (*bSimAnn)
2673     {
2674         update_annealing_target_temp(&(ir->opts), ir->init_t);
2675     }
2676
2677     /* Initialize lambda variables */
2678     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2679
2680     if (upd)
2681     {
2682         *upd = init_update(ir);
2683     }
2684
2685
2686     if (vcm != NULL)
2687     {
2688         *vcm = init_vcm(fplog, &mtop->groups, ir);
2689     }
2690
2691     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2692     {
2693         if (ir->etc == etcBERENDSEN)
2694         {
2695             please_cite(fplog, "Berendsen84a");
2696         }
2697         if (ir->etc == etcVRESCALE)
2698         {
2699             please_cite(fplog, "Bussi2007a");
2700         }
2701     }
2702
2703     init_nrnb(nrnb);
2704
2705     if (nfile != -1)
2706     {
2707         *outf = init_mdoutf(nfile, fnm, Flags, cr, ir, oenv);
2708
2709         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : (*outf)->fp_ene,
2710                               mtop, ir, (*outf)->fp_dhdl);
2711     }
2712
2713     if (ir->bAdress)
2714     {
2715         please_cite(fplog, "Fritsch12");
2716         please_cite(fplog, "Junghans10");
2717     }
2718     /* Initiate variables */
2719     clear_mat(force_vir);
2720     clear_mat(shake_vir);
2721     clear_rvec(mu_tot);
2722
2723     debug_gmx();
2724 }