src/mdlib/sim_util.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team,
   6  * check out http://www.gromacs.org for more information.
   7  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   8  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   9  * others, as listed in the AUTHORS file in the top-level source
  10  * directory and at http://www.gromacs.org.
  11  *
  12  * GROMACS is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public License
  14  * as published by the Free Software Foundation; either version 2.1
  15  * of the License, or (at your option) any later version.
  16  *
  17  * GROMACS is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with GROMACS; if not, see
  24  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  25  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  26  *
  27  * If you want to redistribute modifications to GROMACS, please
  28  * consider that scientific software is very special. Version
  29  * control is crucial - bugs must be traceable. We will be happy to
  30  * consider code for inclusion in the official distribution, but
  31  * derived work must not be called official GROMACS. Details are found
  32  * in the README & COPYING files - if they are missing, get the
  33  * official version at http://www.gromacs.org.
  34  *
  35  * To help us fund GROMACS development, we humbly ask that you cite
  36  * the research papers on the package. Check out http://www.gromacs.org.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #ifdef GMX_CRAY_XT3
  43 #include <catamount/dclock.h>
  44 #endif
  45
  46
  47 #include <stdio.h>
  48 #include <time.h>
  49 #ifdef HAVE_SYS_TIME_H
  50 #include <sys/time.h>
  51 #endif
  52 #include <math.h>
  53 #include "visibility.h"
  54 #include "typedefs.h"
  55 #include "string2.h"
  56 #include "gmxfio.h"
  57 #include "smalloc.h"
  58 #include "names.h"
  59 #include "confio.h"
  60 #include "mvdata.h"
  61 #include "txtdump.h"
  62 #include "pbc.h"
  63 #include "chargegroup.h"
  64 #include "vec.h"
  65 #include <time.h>
  66 #include "nrnb.h"
  67 #include "mshift.h"
  68 #include "mdrun.h"
  69 #include "sim_util.h"
  70 #include "update.h"
  71 #include "physics.h"
  72 #include "main.h"
  73 #include "mdatoms.h"
  74 #include "force.h"
  75 #include "bondf.h"
  76 #include "pme.h"
  77 #include "disre.h"
  78 #include "orires.h"
  79 #include "network.h"
  80 #include "calcmu.h"
  81 #include "constr.h"
  82 #include "xvgr.h"
  83 #include "trnio.h"
  84 #include "xtcio.h"
  85 #include "copyrite.h"
  86 #include "pull_rotation.h"
  87 #include "gmx_random.h"
  88 #include "mpelogging.h"
  89 #include "domdec.h"
  90 #include "partdec.h"
  91 #include "gmx_wallcycle.h"
  92 #include "genborn.h"
  93 #include "nbnxn_atomdata.h"
  94 #include "nbnxn_search.h"
  95 #include "nbnxn_kernels/nbnxn_kernel_ref.h"
  96 #include "nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.h"
  97 #include "nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn.h"
  98 #include "nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  99
 100 #ifdef GMX_LIB_MPI
 101 #include <mpi.h>
 102 #endif
 103 #ifdef GMX_THREAD_MPI
 104 #include "tmpi.h"
 105 #endif
 106
 107 #include "adress.h"
 108 #include "qmmm.h"
 109
 110 #include "nbnxn_cuda_data_mgmt.h"
 111 #include "nbnxn_cuda/nbnxn_cuda.h"
 112
 113 #if 0
 114 typedef struct gmx_timeprint {
 115
 116 } t_gmx_timeprint;
 117 #endif
 118
 119 /* Portable version of ctime_r implemented in src/gmxlib/string2.c, but we do not want it declared in public installed headers */
 120 GMX_LIBGMX_EXPORT
 121 char *
 122 gmx_ctime_r(const time_t *clock, char *buf, int n);
 123
 124
 125 double
 126 gmx_gettime()
 127 {
 128 #ifdef HAVE_GETTIMEOFDAY
 129     struct timeval t;
 130     double         seconds;
 131
 132     gettimeofday(&t, NULL);
 133
 134     seconds = (double) t.tv_sec + 1e-6*(double)t.tv_usec;
 135
 136     return seconds;
 137 #else
 138     double  seconds;
 139
 140     seconds = time(NULL);
 141
 142     return seconds;
 143 #endif
 144 }
 145
 146
 147 #define difftime(end, start) ((double)(end)-(double)(start))
 148
 149 void print_time(FILE *out, gmx_runtime_t *runtime, gmx_large_int_t step,
 150                 t_inputrec *ir, t_commrec *cr)
 151 {
 152     time_t finish;
 153     char   timebuf[STRLEN];
 154     double dt;
 155     char   buf[48];
 156
 157 #ifndef GMX_THREAD_MPI
 158     if (!PAR(cr))
 159 #endif
 160     {
 161         fprintf(out, "\r");
 162     }
 163     fprintf(out, "step %s", gmx_step_str(step, buf));
 164     if ((step >= ir->nstlist))
 165     {
 166         runtime->last          = gmx_gettime();
 167         dt                     = difftime(runtime->last, runtime->real);
 168         runtime->time_per_step = dt/(step - ir->init_step + 1);
 169
 170         dt = (ir->nsteps + ir->init_step - step)*runtime->time_per_step;
 171
 172         if (ir->nsteps >= 0)
 173         {
 174             if (dt >= 300)
 175             {
 176                 finish = (time_t) (runtime->last + dt);
 177                 gmx_ctime_r(&finish, timebuf, STRLEN);
 178                 sprintf(buf, "%s", timebuf);
 179                 buf[strlen(buf)-1] = '\0';
 180                 fprintf(out, ", will finish %s", buf);
 181             }
 182             else
 183             {
 184                 fprintf(out, ", remaining runtime: %5d s          ", (int)dt);
 185             }
 186         }
 187         else
 188         {
 189             fprintf(out, " performance: %.1f ns/day    ",
 190                     ir->delta_t/1000*24*60*60/runtime->time_per_step);
 191         }
 192     }
 193 #ifndef GMX_THREAD_MPI
 194     if (PAR(cr))
 195     {
 196         fprintf(out, "\n");
 197     }
 198 #endif
 199
 200     fflush(out);
 201 }
 202
 203 #ifdef NO_CLOCK
 204 #define clock() -1
 205 #endif
 206
 207 static double set_proctime(gmx_runtime_t *runtime)
 208 {
 209     double diff;
 210 #ifdef GMX_CRAY_XT3
 211     double prev;
 212
 213     prev          = runtime->proc;
 214     runtime->proc = dclock();
 215
 216     diff = runtime->proc - prev;
 217 #else
 218     clock_t prev;
 219
 220     prev          = runtime->proc;
 221     runtime->proc = clock();
 222
 223     diff = (double)(runtime->proc - prev)/(double)CLOCKS_PER_SEC;
 224 #endif
 225     if (diff < 0)
 226     {
 227         /* The counter has probably looped, ignore this data */
 228         diff = 0;
 229     }
 230
 231     return diff;
 232 }
 233
 234 void runtime_start(gmx_runtime_t *runtime)
 235 {
 236     runtime->real          = gmx_gettime();
 237     runtime->proc          = 0;
 238     set_proctime(runtime);
 239     runtime->realtime      = 0;
 240     runtime->proctime      = 0;
 241     runtime->last          = 0;
 242     runtime->time_per_step = 0;
 243 }
 244
 245 void runtime_end(gmx_runtime_t *runtime)
 246 {
 247     double now;
 248
 249     now = gmx_gettime();
 250
 251     runtime->proctime += set_proctime(runtime);
 252     runtime->realtime  = now - runtime->real;
 253     runtime->real      = now;
 254 }
 255
 256 void runtime_upd_proc(gmx_runtime_t *runtime)
 257 {
 258     runtime->proctime += set_proctime(runtime);
 259 }
 260
 261 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
 262                          const gmx_runtime_t *runtime)
 263 {
 264     int    i;
 265     char   timebuf[STRLEN];
 266     char   time_string[STRLEN];
 267     time_t tmptime;
 268
 269     if (fplog)
 270     {
 271         if (runtime != NULL)
 272         {
 273             tmptime = (time_t) runtime->real;
 274             gmx_ctime_r(&tmptime, timebuf, STRLEN);
 275         }
 276         else
 277         {
 278             tmptime = (time_t) gmx_gettime();
 279             gmx_ctime_r(&tmptime, timebuf, STRLEN);
 280         }
 281         for (i = 0; timebuf[i] >= ' '; i++)
 282         {
 283             time_string[i] = timebuf[i];
 284         }
 285         time_string[i] = '\0';
 286
 287         fprintf(fplog, "%s on node %d %s\n", title, nodeid, time_string);
 288     }
 289 }
 290
 291 static void sum_forces(int start, int end, rvec f[], rvec flr[])
 292 {
 293     int i;
 294
 295     if (gmx_debug_at)
 296     {
 297         pr_rvecs(debug, 0, "fsr", f+start, end-start);
 298         pr_rvecs(debug, 0, "flr", flr+start, end-start);
 299     }
 300     for (i = start; (i < end); i++)
 301     {
 302         rvec_inc(f[i], flr[i]);
 303     }
 304 }
 305
 306 /*
 307  * calc_f_el calculates forces due to an electric field.
 308  *
 309  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 310  *
 311  * Et[] contains the parameters for the time dependent
 312  * part of the field (not yet used).
 313  * Ex[] contains the parameters for
 314  * the spatial dependent part of the field. You can have cool periodic
 315  * fields in principle, but only a constant field is supported
 316  * now.
 317  * The function should return the energy due to the electric field
 318  * (if any) but for now returns 0.
 319  *
 320  * WARNING:
 321  * There can be problems with the virial.
 322  * Since the field is not self-consistent this is unavoidable.
 323  * For neutral molecules the virial is correct within this approximation.
 324  * For neutral systems with many charged molecules the error is small.
 325  * But for systems with a net charge or a few charged molecules
 326  * the error can be significant when the field is high.
 327  * Solution: implement a self-consitent electric field into PME.
 328  */
 329 static void calc_f_el(FILE *fp, int  start, int homenr,
 330                       real charge[], rvec x[], rvec f[],
 331                       t_cosines Ex[], t_cosines Et[], double t)
 332 {
 333     rvec Ext;
 334     real t0;
 335     int  i, m;
 336
 337     for (m = 0; (m < DIM); m++)
 338     {
 339         if (Et[m].n > 0)
 340         {
 341             if (Et[m].n == 3)
 342             {
 343                 t0     = Et[m].a[1];
 344                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 345             }
 346             else
 347             {
 348                 Ext[m] = cos(Et[m].a[0]*t);
 349             }
 350         }
 351         else
 352         {
 353             Ext[m] = 1.0;
 354         }
 355         if (Ex[m].n > 0)
 356         {
 357             /* Convert the field strength from V/nm to MD-units */
 358             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 359             for (i = start; (i < start+homenr); i++)
 360             {
 361                 f[i][m] += charge[i]*Ext[m];
 362             }
 363         }
 364         else
 365         {
 366             Ext[m] = 0;
 367         }
 368     }
 369     if (fp != NULL)
 370     {
 371         fprintf(fp, "%10g  %10g  %10g  %10g #FIELD\n", t,
 372                 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
 373     }
 374 }
 375
 376 static void calc_virial(FILE *fplog, int start, int homenr, rvec x[], rvec f[],
 377                         tensor vir_part, t_graph *graph, matrix box,
 378                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 379 {
 380     int    i, j;
 381     tensor virtest;
 382
 383     /* The short-range virial from surrounding boxes */
 384     clear_mat(vir_part);
 385     calc_vir(fplog, SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 386     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 387
 388     /* Calculate partial virial, for local atoms only, based on short range.
 389      * Total virial is computed in global_stat, called from do_md
 390      */
 391     f_calc_vir(fplog, start, start+homenr, x, f, vir_part, graph, box);
 392     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 393
 394     /* Add position restraint contribution */
 395     for (i = 0; i < DIM; i++)
 396     {
 397         vir_part[i][i] += fr->vir_diag_posres[i];
 398     }
 399
 400     /* Add wall contribution */
 401     for (i = 0; i < DIM; i++)
 402     {
 403         vir_part[i][ZZ] += fr->vir_wall_z[i];
 404     }
 405
 406     if (debug)
 407     {
 408         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 409     }
 410 }
 411
 412 static void posres_wrapper(FILE *fplog,
 413                            int flags,
 414                            gmx_bool bSepDVDL,
 415                            t_inputrec *ir,
 416                            t_nrnb *nrnb,
 417                            gmx_localtop_t *top,
 418                            matrix box, rvec x[],
 419                            rvec f[],
 420                            gmx_enerdata_t *enerd,
 421                            real *lambda,
 422                            t_forcerec *fr)
 423 {
 424     t_pbc pbc;
 425     real  v, dvdl;
 426     int   i;
 427
 428     /* Position restraints always require full pbc */
 429     set_pbc(&pbc, ir->ePBC, box);
 430     dvdl = 0;
 431     v    = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 432                   top->idef.iparams_posres,
 433                   (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 434                   ir->ePBC == epbcNONE ? NULL : &pbc,
 435                   lambda[efptRESTRAINT], &dvdl,
 436                   fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 437     if (bSepDVDL)
 438     {
 439         fprintf(fplog, sepdvdlformat,
 440                 interaction_function[F_POSRES].longname, v, dvdl);
 441     }
 442     enerd->term[F_POSRES] += v;
 443     /* If just the force constant changes, the FEP term is linear,
 444      * but if k changes, it is not.
 445      */
 446     enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
 447     inc_nrnb(nrnb, eNR_POSRES, top->idef.il[F_POSRES].nr/2);
 448
 449     if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
 450     {
 451         for (i = 0; i < enerd->n_lambda; i++)
 452         {
 453             real dvdl_dum, lambda_dum;
 454
 455             lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
 456             v          = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 457                                 top->idef.iparams_posres,
 458                                 (const rvec*)x, NULL, NULL,
 459                                 ir->ePBC == epbcNONE ? NULL : &pbc, lambda_dum, &dvdl,
 460                                 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 461             enerd->enerpart_lambda[i] += v;
 462         }
 463     }
 464 }
 465
 466 static void pull_potential_wrapper(FILE *fplog,
 467                                    gmx_bool bSepDVDL,
 468                                    t_commrec *cr,
 469                                    t_inputrec *ir,
 470                                    matrix box, rvec x[],
 471                                    rvec f[],
 472                                    tensor vir_force,
 473                                    t_mdatoms *mdatoms,
 474                                    gmx_enerdata_t *enerd,
 475                                    real *lambda,
 476                                    double t)
 477 {
 478     t_pbc  pbc;
 479     real   dvdl;
 480
 481     /* Calculate the center of mass forces, this requires communication,
 482      * which is why pull_potential is called close to other communication.
 483      * The virial contribution is calculated directly,
 484      * which is why we call pull_potential after calc_virial.
 485      */
 486     set_pbc(&pbc, ir->ePBC, box);
 487     dvdl                     = 0;
 488     enerd->term[F_COM_PULL] +=
 489         pull_potential(ir->ePull, ir->pull, mdatoms, &pbc,
 490                        cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
 491     if (bSepDVDL)
 492     {
 493         fprintf(fplog, sepdvdlformat, "Com pull", enerd->term[F_COM_PULL], dvdl);
 494     }
 495     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 496 }
 497
 498 static void pme_receive_force_ener(FILE           *fplog,
 499                                    gmx_bool        bSepDVDL,
 500                                    t_commrec      *cr,
 501                                    gmx_wallcycle_t wcycle,
 502                                    gmx_enerdata_t *enerd,
 503                                    t_forcerec     *fr)
 504 {
 505     real   e, v, dvdl;
 506     float  cycles_ppdpme, cycles_seppme;
 507
 508     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 509     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 510
 511     /* In case of node-splitting, the PP nodes receive the long-range
 512      * forces, virial and energy from the PME nodes here.
 513      */
 514     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 515     dvdl = 0;
 516     gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e, &dvdl,
 517                       &cycles_seppme);
 518     if (bSepDVDL)
 519     {
 520         fprintf(fplog, sepdvdlformat, "PME mesh", e, dvdl);
 521     }
 522     enerd->term[F_COUL_RECIP] += e;
 523     enerd->dvdl_lin[efptCOUL] += dvdl;
 524     if (wcycle)
 525     {
 526         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 527     }
 528     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 529 }
 530
 531 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
 532                                gmx_large_int_t step, real pforce, rvec *x, rvec *f)
 533 {
 534     int  i;
 535     real pf2, fn2;
 536     char buf[STEPSTRSIZE];
 537
 538     pf2 = sqr(pforce);
 539     for (i = md->start; i < md->start+md->homenr; i++)
 540     {
 541         fn2 = norm2(f[i]);
 542         /* We also catch NAN, if the compiler does not optimize this away. */
 543         if (fn2 >= pf2 || fn2 != fn2)
 544         {
 545             fprintf(fp, "step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 546                     gmx_step_str(step, buf),
 547                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
 548         }
 549     }
 550 }
 551
 552 static void post_process_forces(FILE *fplog,
 553                                 t_commrec *cr,
 554                                 gmx_large_int_t step,
 555                                 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 556                                 gmx_localtop_t *top,
 557                                 matrix box, rvec x[],
 558                                 rvec f[],
 559                                 tensor vir_force,
 560                                 t_mdatoms *mdatoms,
 561                                 t_graph *graph,
 562                                 t_forcerec *fr, gmx_vsite_t *vsite,
 563                                 int flags)
 564 {
 565     if (fr->bF_NoVirSum)
 566     {
 567         if (vsite)
 568         {
 569             /* Spread the mesh force on virtual sites to the other particles...
 570              * This is parallellized. MPI communication is performed
 571              * if the constructing atoms aren't local.
 572              */
 573             wallcycle_start(wcycle, ewcVSITESPREAD);
 574             spread_vsite_f(fplog, vsite, x, fr->f_novirsum, NULL,
 575                            (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
 576                            nrnb,
 577                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
 578             wallcycle_stop(wcycle, ewcVSITESPREAD);
 579         }
 580         if (flags & GMX_FORCE_VIRIAL)
 581         {
 582             /* Now add the forces, this is local */
 583             if (fr->bDomDec)
 584             {
 585                 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
 586             }
 587             else
 588             {
 589                 sum_forces(mdatoms->start, mdatoms->start+mdatoms->homenr,
 590                            f, fr->f_novirsum);
 591             }
 592             if (EEL_FULL(fr->eeltype))
 593             {
 594                 /* Add the mesh contribution to the virial */
 595                 m_add(vir_force, fr->vir_el_recip, vir_force);
 596             }
 597             if (debug)
 598             {
 599                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 600             }
 601         }
 602     }
 603
 604     if (fr->print_force >= 0)
 605     {
 606         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 607     }
 608 }
 609
 610 static void do_nb_verlet(t_forcerec *fr,
 611                          interaction_const_t *ic,
 612                          gmx_enerdata_t *enerd,
 613                          int flags, int ilocality,
 614                          int clearF,
 615                          t_nrnb *nrnb,
 616                          gmx_wallcycle_t wcycle)
 617 {
 618     int                        nnbl, kernel_type, enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
 619     char                      *env;
 620     nonbonded_verlet_group_t  *nbvg;
 621     gmx_bool                  bCUDA;
 622
 623     if (!(flags & GMX_FORCE_NONBONDED))
 624     {
 625         /* skip non-bonded calculation */
 626         return;
 627     }
 628
 629     nbvg = &fr->nbv->grp[ilocality];
 630
 631     /* CUDA kernel launch overhead is already timed separately */
 632     if (fr->cutoff_scheme != ecutsVERLET)
 633     {
 634         gmx_incons("Invalid cut-off scheme passed!");
 635     }
 636
 637     bCUDA = (nbvg->kernel_type == nbnxnk8x8x8_CUDA);
 638
 639     if (!bCUDA)
 640     {
 641         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 642     }
 643     switch (nbvg->kernel_type)
 644     {
 645         case nbnxnk4x4_PlainC:
 646             nbnxn_kernel_ref(&nbvg->nbl_lists,
 647                              nbvg->nbat, ic,
 648                              fr->shift_vec,
 649                              flags,
 650                              clearF,
 651                              fr->fshift[0],
 652                              enerd->grpp.ener[egCOULSR],
 653                              fr->bBHAM ?
 654                              enerd->grpp.ener[egBHAMSR] :
 655                              enerd->grpp.ener[egLJSR]);
 656             break;
 657
 658         case nbnxnk4xN_SIMD_4xN:
 659             nbnxn_kernel_simd_4xn(&nbvg->nbl_lists,
 660                                   nbvg->nbat, ic,
 661                                   nbvg->ewald_excl,
 662                                   fr->shift_vec,
 663                                   flags,
 664                                   clearF,
 665                                   fr->fshift[0],
 666                                   enerd->grpp.ener[egCOULSR],
 667                                   fr->bBHAM ?
 668                                   enerd->grpp.ener[egBHAMSR] :
 669                                   enerd->grpp.ener[egLJSR]);
 670             break;
 671         case nbnxnk4xN_SIMD_2xNN:
 672             nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
 673                                    nbvg->nbat, ic,
 674                                    nbvg->ewald_excl,
 675                                    fr->shift_vec,
 676                                    flags,
 677                                    clearF,
 678                                    fr->fshift[0],
 679                                    enerd->grpp.ener[egCOULSR],
 680                                    fr->bBHAM ?
 681                                    enerd->grpp.ener[egBHAMSR] :
 682                                    enerd->grpp.ener[egLJSR]);
 683             break;
 684
 685         case nbnxnk8x8x8_CUDA:
 686             nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
 687             break;
 688
 689         case nbnxnk8x8x8_PlainC:
 690             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 691                                  nbvg->nbat, ic,
 692                                  fr->shift_vec,
 693                                  flags,
 694                                  clearF,
 695                                  nbvg->nbat->out[0].f,
 696                                  fr->fshift[0],
 697                                  enerd->grpp.ener[egCOULSR],
 698                                  fr->bBHAM ?
 699                                  enerd->grpp.ener[egBHAMSR] :
 700                                  enerd->grpp.ener[egLJSR]);
 701             break;
 702
 703         default:
 704             gmx_incons("Invalid nonbonded kernel type passed!");
 705
 706     }
 707     if (!bCUDA)
 708     {
 709         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 710     }
 711
 712     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 713     {
 714         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 715     }
 716     else if ((!bCUDA && nbvg->ewald_excl == ewaldexclAnalytical) ||
 717              (bCUDA && nbnxn_cuda_is_kernel_ewald_analytical(fr->nbv->cu_nbv)))
 718     {
 719         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 720     }
 721     else
 722     {
 723         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 724     }
 725     enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 726     if (flags & GMX_FORCE_ENERGY)
 727     {
 728         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 729         enr_nbnxn_kernel_ljc += 1;
 730         enr_nbnxn_kernel_lj  += 1;
 731     }
 732
 733     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 734              nbvg->nbl_lists.natpair_ljq);
 735     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 736              nbvg->nbl_lists.natpair_lj);
 737     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 738              nbvg->nbl_lists.natpair_q);
 739 }
 740
 741 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
 742                          t_inputrec *inputrec,
 743                          gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 744                          gmx_localtop_t *top,
 745                          gmx_mtop_t *mtop,
 746                          gmx_groups_t *groups,
 747                          matrix box, rvec x[], history_t *hist,
 748                          rvec f[],
 749                          tensor vir_force,
 750                          t_mdatoms *mdatoms,
 751                          gmx_enerdata_t *enerd, t_fcdata *fcd,
 752                          real *lambda, t_graph *graph,
 753                          t_forcerec *fr, interaction_const_t *ic,
 754                          gmx_vsite_t *vsite, rvec mu_tot,
 755                          double t, FILE *field, gmx_edsam_t ed,
 756                          gmx_bool bBornRadii,
 757                          int flags)
 758 {
 759     int                 cg0, cg1, i, j;
 760     int                 start, homenr;
 761     int                 nb_kernel_type;
 762     double              mu[2*DIM];
 763     gmx_bool            bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
 764     gmx_bool            bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
 765     gmx_bool            bDiffKernels = FALSE;
 766     matrix              boxs;
 767     rvec                vzero, box_diag;
 768     real                e, v, dvdl;
 769     float               cycles_pme, cycles_force;
 770     nonbonded_verlet_t *nbv;
 771
 772     cycles_force   = 0;
 773     nbv            = fr->nbv;
 774     nb_kernel_type = fr->nbv->grp[0].kernel_type;
 775
 776     start  = mdatoms->start;
 777     homenr = mdatoms->homenr;
 778
 779     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
 780
 781     clear_mat(vir_force);
 782
 783     cg0 = 0;
 784     if (DOMAINDECOMP(cr))
 785     {
 786         cg1 = cr->dd->ncg_tot;
 787     }
 788     else
 789     {
 790         cg1 = top->cgs.nr;
 791     }
 792     if (fr->n_tpi > 0)
 793     {
 794         cg1--;
 795     }
 796
 797     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 798     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
 799     bFillGrid     = (bNS && bStateChanged);
 800     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 801     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 802     bDoForces     = (flags & GMX_FORCE_FORCES);
 803     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 804     bUseGPU       = fr->nbv->bUseGPU;
 805     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
 806
 807     if (bStateChanged)
 808     {
 809         update_forcerec(fplog, fr, box);
 810
 811         if (NEED_MUTOT(*inputrec))
 812         {
 813             /* Calculate total (local) dipole moment in a temporary common array.
 814              * This makes it possible to sum them over nodes faster.
 815              */
 816             calc_mu(start, homenr,
 817                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 818                     mu, mu+DIM);
 819         }
 820     }
 821
 822     if (fr->ePBC != epbcNONE)
 823     {
 824         /* Compute shift vectors every step,
 825          * because of pressure coupling or box deformation!
 826          */
 827         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 828         {
 829             calc_shifts(box, fr->shift_vec);
 830         }
 831
 832         if (bCalcCGCM)
 833         {
 834             put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
 835             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 836         }
 837         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 838         {
 839             unshift_self(graph, box, x);
 840         }
 841     }
 842
 843     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 844                                  fr->shift_vec, nbv->grp[0].nbat);
 845
 846 #ifdef GMX_MPI
 847     if (!(cr->duty & DUTY_PME))
 848     {
 849         /* Send particle coordinates to the pme nodes.
 850          * Since this is only implemented for domain decomposition
 851          * and domain decomposition does not use the graph,
 852          * we do not need to worry about shifting.
 853          */
 854
 855         wallcycle_start(wcycle, ewcPP_PMESENDX);
 856         GMX_MPE_LOG(ev_send_coordinates_start);
 857
 858         bBS = (inputrec->nwall == 2);
 859         if (bBS)
 860         {
 861             copy_mat(box, boxs);
 862             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
 863         }
 864
 865         gmx_pme_send_x(cr, bBS ? boxs : box, x,
 866                        mdatoms->nChargePerturbed, lambda[efptCOUL],
 867                        (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)), step);
 868
 869         GMX_MPE_LOG(ev_send_coordinates_finish);
 870         wallcycle_stop(wcycle, ewcPP_PMESENDX);
 871     }
 872 #endif /* GMX_MPI */
 873
 874     /* do gridding for pair search */
 875     if (bNS)
 876     {
 877         if (graph && bStateChanged)
 878         {
 879             /* Calculate intramolecular shift vectors to make molecules whole */
 880             mk_mshift(fplog, graph, fr->ePBC, box, x);
 881         }
 882
 883         clear_rvec(vzero);
 884         box_diag[XX] = box[XX][XX];
 885         box_diag[YY] = box[YY][YY];
 886         box_diag[ZZ] = box[ZZ][ZZ];
 887
 888         wallcycle_start(wcycle, ewcNS);
 889         if (!fr->bDomDec)
 890         {
 891             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 892             nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
 893                               0, vzero, box_diag,
 894                               0, mdatoms->homenr, -1, fr->cginfo, x,
 895                               0, NULL,
 896                               nbv->grp[eintLocal].kernel_type,
 897                               nbv->grp[eintLocal].nbat);
 898             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 899         }
 900         else
 901         {
 902             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 903             nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
 904                                        fr->cginfo, x,
 905                                        nbv->grp[eintNonlocal].kernel_type,
 906                                        nbv->grp[eintNonlocal].nbat);
 907             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 908         }
 909
 910         if (nbv->ngrp == 1 ||
 911             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 912         {
 913             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
 914                                nbv->nbs, mdatoms, fr->cginfo);
 915         }
 916         else
 917         {
 918             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
 919                                nbv->nbs, mdatoms, fr->cginfo);
 920             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
 921                                nbv->nbs, mdatoms, fr->cginfo);
 922         }
 923         wallcycle_stop(wcycle, ewcNS);
 924     }
 925
 926     /* initialize the GPU atom data and copy shift vector */
 927     if (bUseGPU)
 928     {
 929         if (bNS)
 930         {
 931             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 932             nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 933             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 934         }
 935
 936         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 937         nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 938         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 939     }
 940
 941     /* do local pair search */
 942     if (bNS)
 943     {
 944         wallcycle_start_nocount(wcycle, ewcNS);
 945         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
 946         nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
 947                             &top->excls,
 948                             ic->rlist,
 949                             nbv->min_ci_balanced,
 950                             &nbv->grp[eintLocal].nbl_lists,
 951                             eintLocal,
 952                             nbv->grp[eintLocal].kernel_type,
 953                             nrnb);
 954         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
 955
 956         if (bUseGPU)
 957         {
 958             /* initialize local pair-list on the GPU */
 959             nbnxn_cuda_init_pairlist(nbv->cu_nbv,
 960                                      nbv->grp[eintLocal].nbl_lists.nbl[0],
 961                                      eintLocal);
 962         }
 963         wallcycle_stop(wcycle, ewcNS);
 964     }
 965     else
 966     {
 967         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 968         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
 969         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
 970                                         nbv->grp[eintLocal].nbat);
 971         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
 972         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
 973     }
 974
 975     if (bUseGPU)
 976     {
 977         wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
 978         /* launch local nonbonded F on GPU */
 979         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
 980                      nrnb, wcycle);
 981         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 982     }
 983
 984     /* Communicate coordinates and sum dipole if necessary +
 985        do non-local pair search */
 986     if (DOMAINDECOMP(cr))
 987     {
 988         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
 989                         nbv->grp[eintLocal].kernel_type);
 990
 991         if (bDiffKernels)
 992         {
 993             /* With GPU+CPU non-bonded calculations we need to copy
 994              * the local coordinates to the non-local nbat struct
 995              * (in CPU format) as the non-local kernel call also
 996              * calculates the local - non-local interactions.
 997              */
 998             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 999             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1000             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
1001                                             nbv->grp[eintNonlocal].nbat);
1002             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1003             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1004         }
1005
1006         if (bNS)
1007         {
1008             wallcycle_start_nocount(wcycle, ewcNS);
1009             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1010
1011             if (bDiffKernels)
1012             {
1013                 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
1014             }
1015
1016             nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
1017                                 &top->excls,
1018                                 ic->rlist,
1019                                 nbv->min_ci_balanced,
1020                                 &nbv->grp[eintNonlocal].nbl_lists,
1021                                 eintNonlocal,
1022                                 nbv->grp[eintNonlocal].kernel_type,
1023                                 nrnb);
1024
1025             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1026
1027             if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_CUDA)
1028             {
1029                 /* initialize non-local pair-list on the GPU */
1030                 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1031                                          nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1032                                          eintNonlocal);
1033             }
1034             wallcycle_stop(wcycle, ewcNS);
1035         }
1036         else
1037         {
1038             wallcycle_start(wcycle, ewcMOVEX);
1039             dd_move_x(cr->dd, box, x);
1040
1041             /* When we don't need the total dipole we sum it in global_stat */
1042             if (bStateChanged && NEED_MUTOT(*inputrec))
1043             {
1044                 gmx_sumd(2*DIM, mu, cr);
1045             }
1046             wallcycle_stop(wcycle, ewcMOVEX);
1047
1048             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1049             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1050             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1051                                             nbv->grp[eintNonlocal].nbat);
1052             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1053             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1054         }
1055
1056         if (bUseGPU && !bDiffKernels)
1057         {
1058             wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1059             /* launch non-local nonbonded F on GPU */
1060             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1061                          nrnb, wcycle);
1062             cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1063         }
1064     }
1065
1066     if (bUseGPU)
1067     {
1068         /* launch D2H copy-back F */
1069         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1070         if (DOMAINDECOMP(cr) && !bDiffKernels)
1071         {
1072             nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1073                                       flags, eatNonlocal);
1074         }
1075         nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1076                                   flags, eatLocal);
1077         cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1078     }
1079
1080     if (bStateChanged && NEED_MUTOT(*inputrec))
1081     {
1082         if (PAR(cr))
1083         {
1084             gmx_sumd(2*DIM, mu, cr);
1085         }
1086
1087         for (i = 0; i < 2; i++)
1088         {
1089             for (j = 0; j < DIM; j++)
1090             {
1091                 fr->mu_tot[i][j] = mu[i*DIM + j];
1092             }
1093         }
1094     }
1095     if (fr->efep == efepNO)
1096     {
1097         copy_rvec(fr->mu_tot[0], mu_tot);
1098     }
1099     else
1100     {
1101         for (j = 0; j < DIM; j++)
1102         {
1103             mu_tot[j] =
1104                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1105                 lambda[efptCOUL]*fr->mu_tot[1][j];
1106         }
1107     }
1108
1109     /* Reset energies */
1110     reset_enerdata(&(inputrec->opts), fr, bNS, enerd, MASTER(cr));
1111     clear_rvecs(SHIFTS, fr->fshift);
1112
1113     if (DOMAINDECOMP(cr))
1114     {
1115         if (!(cr->duty & DUTY_PME))
1116         {
1117             wallcycle_start(wcycle, ewcPPDURINGPME);
1118             dd_force_flop_start(cr->dd, nrnb);
1119         }
1120     }
1121
1122     if (inputrec->bRot)
1123     {
1124         /* Enforced rotation has its own cycle counter that starts after the collective
1125          * coordinates have been communicated. It is added to ddCyclF to allow
1126          * for proper load-balancing */
1127         wallcycle_start(wcycle, ewcROT);
1128         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1129         wallcycle_stop(wcycle, ewcROT);
1130     }
1131
1132     /* Start the force cycle counter.
1133      * This counter is stopped in do_forcelow_level.
1134      * No parallel communication should occur while this counter is running,
1135      * since that will interfere with the dynamic load balancing.
1136      */
1137     wallcycle_start(wcycle, ewcFORCE);
1138     if (bDoForces)
1139     {
1140         /* Reset forces for which the virial is calculated separately:
1141          * PME/Ewald forces if necessary */
1142         if (fr->bF_NoVirSum)
1143         {
1144             if (flags & GMX_FORCE_VIRIAL)
1145             {
1146                 fr->f_novirsum = fr->f_novirsum_alloc;
1147                 GMX_BARRIER(cr->mpi_comm_mygroup);
1148                 if (fr->bDomDec)
1149                 {
1150                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1151                 }
1152                 else
1153                 {
1154                     clear_rvecs(homenr, fr->f_novirsum+start);
1155                 }
1156                 GMX_BARRIER(cr->mpi_comm_mygroup);
1157             }
1158             else
1159             {
1160                 /* We are not calculating the pressure so we do not need
1161                  * a separate array for forces that do not contribute
1162                  * to the pressure.
1163                  */
1164                 fr->f_novirsum = f;
1165             }
1166         }
1167
1168         /* Clear the short- and long-range forces */
1169         clear_rvecs(fr->natoms_force_constr, f);
1170         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1171         {
1172             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1173         }
1174
1175         clear_rvec(fr->vir_diag_posres);
1176
1177         GMX_BARRIER(cr->mpi_comm_mygroup);
1178     }
1179
1180     if (inputrec->ePull == epullCONSTRAINT)
1181     {
1182         clear_pull_forces(inputrec->pull);
1183     }
1184
1185     /* We calculate the non-bonded forces, when done on the CPU, here.
1186      * We do this before calling do_force_lowlevel, as in there bondeds
1187      * forces are calculated before PME, which does communication.
1188      * With this order, non-bonded and bonded force calculation imbalance
1189      * can be balanced out by the domain decomposition load balancing.
1190      */
1191
1192     if (!bUseOrEmulGPU)
1193     {
1194         /* Maybe we should move this into do_force_lowlevel */
1195         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1196                      nrnb, wcycle);
1197     }
1198
1199     if (!bUseOrEmulGPU || bDiffKernels)
1200     {
1201         int aloc;
1202
1203         if (DOMAINDECOMP(cr))
1204         {
1205             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1206                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1207                          nrnb, wcycle);
1208         }
1209
1210         if (!bUseOrEmulGPU)
1211         {
1212             aloc = eintLocal;
1213         }
1214         else
1215         {
1216             aloc = eintNonlocal;
1217         }
1218
1219         /* Add all the non-bonded force to the normal force array.
1220          * This can be split into a local a non-local part when overlapping
1221          * communication with calculation with domain decomposition.
1222          */
1223         cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1224         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1225         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1226         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1227         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1228         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1229         wallcycle_start_nocount(wcycle, ewcFORCE);
1230
1231         /* if there are multiple fshift output buffers reduce them */
1232         if ((flags & GMX_FORCE_VIRIAL) &&
1233             nbv->grp[aloc].nbl_lists.nnbl > 1)
1234         {
1235             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1236                                                      fr->fshift);
1237         }
1238     }
1239
1240     /* update QMMMrec, if necessary */
1241     if (fr->bQMMM)
1242     {
1243         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1244     }
1245
1246     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1247     {
1248         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1249                        f, enerd, lambda, fr);
1250     }
1251
1252     /* Compute the bonded and non-bonded energies and optionally forces */
1253     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1254                       cr, nrnb, wcycle, mdatoms, &(inputrec->opts),
1255                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, mtop, top, fr->born,
1256                       &(top->atomtypes), bBornRadii, box,
1257                       inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1258                       flags, &cycles_pme);
1259
1260     if (bSepLRF)
1261     {
1262         if (do_per_step(step, inputrec->nstcalclr))
1263         {
1264             /* Add the long range forces to the short range forces */
1265             for (i = 0; i < fr->natoms_force_constr; i++)
1266             {
1267                 rvec_add(fr->f_twin[i], f[i], f[i]);
1268             }
1269         }
1270     }
1271
1272     cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1273     GMX_BARRIER(cr->mpi_comm_mygroup);
1274
1275     if (ed)
1276     {
1277         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1278     }
1279
1280     if (bUseOrEmulGPU && !bDiffKernels)
1281     {
1282         /* wait for non-local forces (or calculate in emulation mode) */
1283         if (DOMAINDECOMP(cr))
1284         {
1285             if (bUseGPU)
1286             {
1287                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1288                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1289                                     nbv->grp[eintNonlocal].nbat,
1290                                     flags, eatNonlocal,
1291                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1292                                     fr->fshift);
1293                 cycles_force += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1294             }
1295             else
1296             {
1297                 wallcycle_start_nocount(wcycle, ewcFORCE);
1298                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1299                              nrnb, wcycle);
1300                 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1301             }
1302             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1303             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1304             /* skip the reduction if there was no non-local work to do */
1305             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1306             {
1307                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1308                                                nbv->grp[eintNonlocal].nbat, f);
1309             }
1310             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1311             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1312         }
1313     }
1314
1315     if (bDoForces)
1316     {
1317         /* Communicate the forces */
1318         if (PAR(cr))
1319         {
1320             wallcycle_start(wcycle, ewcMOVEF);
1321             if (DOMAINDECOMP(cr))
1322             {
1323                 dd_move_f(cr->dd, f, fr->fshift);
1324                 /* Do we need to communicate the separate force array
1325                  * for terms that do not contribute to the single sum virial?
1326                  * Position restraints and electric fields do not introduce
1327                  * inter-cg forces, only full electrostatics methods do.
1328                  * When we do not calculate the virial, fr->f_novirsum = f,
1329                  * so we have already communicated these forces.
1330                  */
1331                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1332                     (flags & GMX_FORCE_VIRIAL))
1333                 {
1334                     dd_move_f(cr->dd, fr->f_novirsum, NULL);
1335                 }
1336                 if (bSepLRF)
1337                 {
1338                     /* We should not update the shift forces here,
1339                      * since f_twin is already included in f.
1340                      */
1341                     dd_move_f(cr->dd, fr->f_twin, NULL);
1342                 }
1343             }
1344             wallcycle_stop(wcycle, ewcMOVEF);
1345         }
1346     }
1347
1348     if (bUseOrEmulGPU)
1349     {
1350         /* wait for local forces (or calculate in emulation mode) */
1351         if (bUseGPU)
1352         {
1353             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1354             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1355                                 nbv->grp[eintLocal].nbat,
1356                                 flags, eatLocal,
1357                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1358                                 fr->fshift);
1359             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1360
1361             /* now clear the GPU outputs while we finish the step on the CPU */
1362
1363             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1364             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1365             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1366         }
1367         else
1368         {
1369             wallcycle_start_nocount(wcycle, ewcFORCE);
1370             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1371                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1372                          nrnb, wcycle);
1373             wallcycle_stop(wcycle, ewcFORCE);
1374         }
1375         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1376         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1377         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1378         {
1379             /* skip the reduction if there was no non-local work to do */
1380             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1381                                            nbv->grp[eintLocal].nbat, f);
1382         }
1383         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1384         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1385     }
1386
1387     if (DOMAINDECOMP(cr))
1388     {
1389         dd_force_flop_stop(cr->dd, nrnb);
1390         if (wcycle)
1391         {
1392             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1393         }
1394     }
1395
1396     if (bDoForces)
1397     {
1398         if (IR_ELEC_FIELD(*inputrec))
1399         {
1400             /* Compute forces due to electric field */
1401             calc_f_el(MASTER(cr) ? field : NULL,
1402                       start, homenr, mdatoms->chargeA, x, fr->f_novirsum,
1403                       inputrec->ex, inputrec->et, t);
1404         }
1405
1406         /* If we have NoVirSum forces, but we do not calculate the virial,
1407          * we sum fr->f_novirum=f later.
1408          */
1409         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1410         {
1411             wallcycle_start(wcycle, ewcVSITESPREAD);
1412             spread_vsite_f(fplog, vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1413                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1414             wallcycle_stop(wcycle, ewcVSITESPREAD);
1415
1416             if (bSepLRF)
1417             {
1418                 wallcycle_start(wcycle, ewcVSITESPREAD);
1419                 spread_vsite_f(fplog, vsite, x, fr->f_twin, NULL, FALSE, NULL,
1420                                nrnb,
1421                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1422                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1423             }
1424         }
1425
1426         if (flags & GMX_FORCE_VIRIAL)
1427         {
1428             /* Calculation of the virial must be done after vsites! */
1429             calc_virial(fplog, mdatoms->start, mdatoms->homenr, x, f,
1430                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1431         }
1432     }
1433
1434     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1435     {
1436         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1437                                f, vir_force, mdatoms, enerd, lambda, t);
1438     }
1439
1440     /* Add the forces from enforced rotation potentials (if any) */
1441     if (inputrec->bRot)
1442     {
1443         wallcycle_start(wcycle, ewcROTadd);
1444         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1445         wallcycle_stop(wcycle, ewcROTadd);
1446     }
1447
1448     if (PAR(cr) && !(cr->duty & DUTY_PME))
1449     {
1450         /* In case of node-splitting, the PP nodes receive the long-range
1451          * forces, virial and energy from the PME nodes here.
1452          */
1453         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1454     }
1455
1456     if (bDoForces)
1457     {
1458         post_process_forces(fplog, cr, step, nrnb, wcycle,
1459                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1460                             flags);
1461     }
1462
1463     /* Sum the potential energy terms from group contributions */
1464     sum_epot(&(inputrec->opts), &(enerd->grpp), enerd->term);
1465 }
1466
1467 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1468                         t_inputrec *inputrec,
1469                         gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1470                         gmx_localtop_t *top,
1471                         gmx_mtop_t *mtop,
1472                         gmx_groups_t *groups,
1473                         matrix box, rvec x[], history_t *hist,
1474                         rvec f[],
1475                         tensor vir_force,
1476                         t_mdatoms *mdatoms,
1477                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1478                         real *lambda, t_graph *graph,
1479                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1480                         double t, FILE *field, gmx_edsam_t ed,
1481                         gmx_bool bBornRadii,
1482                         int flags)
1483 {
1484     int        cg0, cg1, i, j;
1485     int        start, homenr;
1486     double     mu[2*DIM];
1487     gmx_bool   bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1488     gmx_bool   bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1489     gmx_bool   bDoAdressWF;
1490     matrix     boxs;
1491     rvec       vzero, box_diag;
1492     real       e, v, dvdlambda[efptNR];
1493     t_pbc      pbc;
1494     float      cycles_pme, cycles_force;
1495
1496     start  = mdatoms->start;
1497     homenr = mdatoms->homenr;
1498
1499     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
1500
1501     clear_mat(vir_force);
1502
1503     if (PARTDECOMP(cr))
1504     {
1505         pd_cg_range(cr, &cg0, &cg1);
1506     }
1507     else
1508     {
1509         cg0 = 0;
1510         if (DOMAINDECOMP(cr))
1511         {
1512             cg1 = cr->dd->ncg_tot;
1513         }
1514         else
1515         {
1516             cg1 = top->cgs.nr;
1517         }
1518         if (fr->n_tpi > 0)
1519         {
1520             cg1--;
1521         }
1522     }
1523
1524     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1525     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1526     /* Should we update the long-range neighborlists at this step? */
1527     bDoLongRangeNS = fr->bTwinRange && bNS;
1528     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1529     bFillGrid      = (bNS && bStateChanged);
1530     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1531     bDoForces      = (flags & GMX_FORCE_FORCES);
1532     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1533     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1534                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1535
1536     /* should probably move this to the forcerec since it doesn't change */
1537     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1538
1539     if (bStateChanged)
1540     {
1541         update_forcerec(fplog, fr, box);
1542
1543         if (NEED_MUTOT(*inputrec))
1544         {
1545             /* Calculate total (local) dipole moment in a temporary common array.
1546              * This makes it possible to sum them over nodes faster.
1547              */
1548             calc_mu(start, homenr,
1549                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1550                     mu, mu+DIM);
1551         }
1552     }
1553
1554     if (fr->ePBC != epbcNONE)
1555     {
1556         /* Compute shift vectors every step,
1557          * because of pressure coupling or box deformation!
1558          */
1559         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1560         {
1561             calc_shifts(box, fr->shift_vec);
1562         }
1563
1564         if (bCalcCGCM)
1565         {
1566             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1567                                      &(top->cgs), x, fr->cg_cm);
1568             inc_nrnb(nrnb, eNR_CGCM, homenr);
1569             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1570         }
1571         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1572         {
1573             unshift_self(graph, box, x);
1574         }
1575     }
1576     else if (bCalcCGCM)
1577     {
1578         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1579         inc_nrnb(nrnb, eNR_CGCM, homenr);
1580     }
1581
1582     if (bCalcCGCM)
1583     {
1584         if (PAR(cr))
1585         {
1586             move_cgcm(fplog, cr, fr->cg_cm);
1587         }
1588         if (gmx_debug_at)
1589         {
1590             pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1591         }
1592     }
1593
1594 #ifdef GMX_MPI
1595     if (!(cr->duty & DUTY_PME))
1596     {
1597         /* Send particle coordinates to the pme nodes.
1598          * Since this is only implemented for domain decomposition
1599          * and domain decomposition does not use the graph,
1600          * we do not need to worry about shifting.
1601          */
1602
1603         wallcycle_start(wcycle, ewcPP_PMESENDX);
1604         GMX_MPE_LOG(ev_send_coordinates_start);
1605
1606         bBS = (inputrec->nwall == 2);
1607         if (bBS)
1608         {
1609             copy_mat(box, boxs);
1610             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1611         }
1612
1613         gmx_pme_send_x(cr, bBS ? boxs : box, x,
1614                        mdatoms->nChargePerturbed, lambda[efptCOUL],
1615                        (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)), step);
1616
1617         GMX_MPE_LOG(ev_send_coordinates_finish);
1618         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1619     }
1620 #endif /* GMX_MPI */
1621
1622     /* Communicate coordinates and sum dipole if necessary */
1623     if (PAR(cr))
1624     {
1625         wallcycle_start(wcycle, ewcMOVEX);
1626         if (DOMAINDECOMP(cr))
1627         {
1628             dd_move_x(cr->dd, box, x);
1629         }
1630         else
1631         {
1632             move_x(fplog, cr, GMX_LEFT, GMX_RIGHT, x, nrnb);
1633         }
1634         wallcycle_stop(wcycle, ewcMOVEX);
1635     }
1636
1637     /* update adress weight beforehand */
1638     if (bStateChanged && bDoAdressWF)
1639     {
1640         /* need pbc for adress weight calculation with pbc_dx */
1641         set_pbc(&pbc, inputrec->ePBC, box);
1642         if (fr->adress_site == eAdressSITEcog)
1643         {
1644             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1645                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1646         }
1647         else if (fr->adress_site == eAdressSITEcom)
1648         {
1649             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1650                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1651         }
1652         else if (fr->adress_site == eAdressSITEatomatom)
1653         {
1654             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1655                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1656         }
1657         else
1658         {
1659             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1660                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1661         }
1662     }
1663
1664     if (NEED_MUTOT(*inputrec))
1665     {
1666
1667         if (bStateChanged)
1668         {
1669             if (PAR(cr))
1670             {
1671                 gmx_sumd(2*DIM, mu, cr);
1672             }
1673             for (i = 0; i < 2; i++)
1674             {
1675                 for (j = 0; j < DIM; j++)
1676                 {
1677                     fr->mu_tot[i][j] = mu[i*DIM + j];
1678                 }
1679             }
1680         }
1681         if (fr->efep == efepNO)
1682         {
1683             copy_rvec(fr->mu_tot[0], mu_tot);
1684         }
1685         else
1686         {
1687             for (j = 0; j < DIM; j++)
1688             {
1689                 mu_tot[j] =
1690                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1691             }
1692         }
1693     }
1694
1695     /* Reset energies */
1696     reset_enerdata(&(inputrec->opts), fr, bNS, enerd, MASTER(cr));
1697     clear_rvecs(SHIFTS, fr->fshift);
1698
1699     if (bNS)
1700     {
1701         wallcycle_start(wcycle, ewcNS);
1702
1703         if (graph && bStateChanged)
1704         {
1705             /* Calculate intramolecular shift vectors to make molecules whole */
1706             mk_mshift(fplog, graph, fr->ePBC, box, x);
1707         }
1708
1709         /* Do the actual neighbour searching */
1710         ns(fplog, fr, x, box,
1711            groups, &(inputrec->opts), top, mdatoms,
1712            cr, nrnb, lambda, dvdlambda, &enerd->grpp, bFillGrid,
1713            bDoLongRangeNS);
1714
1715         wallcycle_stop(wcycle, ewcNS);
1716     }
1717
1718     if (inputrec->implicit_solvent && bNS)
1719     {
1720         make_gb_nblist(cr, inputrec->gb_algorithm, inputrec->rlist,
1721                        x, box, fr, &top->idef, graph, fr->born);
1722     }
1723
1724     if (DOMAINDECOMP(cr))
1725     {
1726         if (!(cr->duty & DUTY_PME))
1727         {
1728             wallcycle_start(wcycle, ewcPPDURINGPME);
1729             dd_force_flop_start(cr->dd, nrnb);
1730         }
1731     }
1732
1733     if (inputrec->bRot)
1734     {
1735         /* Enforced rotation has its own cycle counter that starts after the collective
1736          * coordinates have been communicated. It is added to ddCyclF to allow
1737          * for proper load-balancing */
1738         wallcycle_start(wcycle, ewcROT);
1739         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1740         wallcycle_stop(wcycle, ewcROT);
1741     }
1742
1743     /* Start the force cycle counter.
1744      * This counter is stopped in do_forcelow_level.
1745      * No parallel communication should occur while this counter is running,
1746      * since that will interfere with the dynamic load balancing.
1747      */
1748     wallcycle_start(wcycle, ewcFORCE);
1749
1750     if (bDoForces)
1751     {
1752         /* Reset forces for which the virial is calculated separately:
1753          * PME/Ewald forces if necessary */
1754         if (fr->bF_NoVirSum)
1755         {
1756             if (flags & GMX_FORCE_VIRIAL)
1757             {
1758                 fr->f_novirsum = fr->f_novirsum_alloc;
1759                 GMX_BARRIER(cr->mpi_comm_mygroup);
1760                 if (fr->bDomDec)
1761                 {
1762                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1763                 }
1764                 else
1765                 {
1766                     clear_rvecs(homenr, fr->f_novirsum+start);
1767                 }
1768                 GMX_BARRIER(cr->mpi_comm_mygroup);
1769             }
1770             else
1771             {
1772                 /* We are not calculating the pressure so we do not need
1773                  * a separate array for forces that do not contribute
1774                  * to the pressure.
1775                  */
1776                 fr->f_novirsum = f;
1777             }
1778         }
1779
1780         /* Clear the short- and long-range forces */
1781         clear_rvecs(fr->natoms_force_constr, f);
1782         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1783         {
1784             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1785         }
1786
1787         clear_rvec(fr->vir_diag_posres);
1788
1789         GMX_BARRIER(cr->mpi_comm_mygroup);
1790     }
1791     if (inputrec->ePull == epullCONSTRAINT)
1792     {
1793         clear_pull_forces(inputrec->pull);
1794     }
1795
1796     /* update QMMMrec, if necessary */
1797     if (fr->bQMMM)
1798     {
1799         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1800     }
1801
1802     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1803     {
1804         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1805                        f, enerd, lambda, fr);
1806     }
1807
1808     /* Compute the bonded and non-bonded energies and optionally forces */
1809     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1810                       cr, nrnb, wcycle, mdatoms, &(inputrec->opts),
1811                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, mtop, top, fr->born,
1812                       &(top->atomtypes), bBornRadii, box,
1813                       inputrec->fepvals, lambda,
1814                       graph, &(top->excls), fr->mu_tot,
1815                       flags,
1816                       &cycles_pme);
1817
1818     if (bSepLRF)
1819     {
1820         if (do_per_step(step, inputrec->nstcalclr))
1821         {
1822             /* Add the long range forces to the short range forces */
1823             for (i = 0; i < fr->natoms_force_constr; i++)
1824             {
1825                 rvec_add(fr->f_twin[i], f[i], f[i]);
1826             }
1827         }
1828     }
1829
1830     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1831     GMX_BARRIER(cr->mpi_comm_mygroup);
1832
1833     if (ed)
1834     {
1835         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1836     }
1837
1838     if (DOMAINDECOMP(cr))
1839     {
1840         dd_force_flop_stop(cr->dd, nrnb);
1841         if (wcycle)
1842         {
1843             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1844         }
1845     }
1846
1847     if (bDoForces)
1848     {
1849         if (IR_ELEC_FIELD(*inputrec))
1850         {
1851             /* Compute forces due to electric field */
1852             calc_f_el(MASTER(cr) ? field : NULL,
1853                       start, homenr, mdatoms->chargeA, x, fr->f_novirsum,
1854                       inputrec->ex, inputrec->et, t);
1855         }
1856
1857         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1858         {
1859             /* Compute thermodynamic force in hybrid AdResS region */
1860             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1861                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1862         }
1863
1864         /* Communicate the forces */
1865         if (PAR(cr))
1866         {
1867             wallcycle_start(wcycle, ewcMOVEF);
1868             if (DOMAINDECOMP(cr))
1869             {
1870                 dd_move_f(cr->dd, f, fr->fshift);
1871                 /* Do we need to communicate the separate force array
1872                  * for terms that do not contribute to the single sum virial?
1873                  * Position restraints and electric fields do not introduce
1874                  * inter-cg forces, only full electrostatics methods do.
1875                  * When we do not calculate the virial, fr->f_novirsum = f,
1876                  * so we have already communicated these forces.
1877                  */
1878                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1879                     (flags & GMX_FORCE_VIRIAL))
1880                 {
1881                     dd_move_f(cr->dd, fr->f_novirsum, NULL);
1882                 }
1883                 if (bSepLRF)
1884                 {
1885                     /* We should not update the shift forces here,
1886                      * since f_twin is already included in f.
1887                      */
1888                     dd_move_f(cr->dd, fr->f_twin, NULL);
1889                 }
1890             }
1891             else
1892             {
1893                 pd_move_f(cr, f, nrnb);
1894                 if (bSepLRF)
1895                 {
1896                     pd_move_f(cr, fr->f_twin, nrnb);
1897                 }
1898             }
1899             wallcycle_stop(wcycle, ewcMOVEF);
1900         }
1901
1902         /* If we have NoVirSum forces, but we do not calculate the virial,
1903          * we sum fr->f_novirum=f later.
1904          */
1905         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1906         {
1907             wallcycle_start(wcycle, ewcVSITESPREAD);
1908             spread_vsite_f(fplog, vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1909                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1910             wallcycle_stop(wcycle, ewcVSITESPREAD);
1911
1912             if (bSepLRF)
1913             {
1914                 wallcycle_start(wcycle, ewcVSITESPREAD);
1915                 spread_vsite_f(fplog, vsite, x, fr->f_twin, NULL, FALSE, NULL,
1916                                nrnb,
1917                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1918                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1919             }
1920         }
1921
1922         if (flags & GMX_FORCE_VIRIAL)
1923         {
1924             /* Calculation of the virial must be done after vsites! */
1925             calc_virial(fplog, mdatoms->start, mdatoms->homenr, x, f,
1926                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1927         }
1928     }
1929
1930     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1931     {
1932         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1933                                f, vir_force, mdatoms, enerd, lambda, t);
1934     }
1935
1936     /* Add the forces from enforced rotation potentials (if any) */
1937     if (inputrec->bRot)
1938     {
1939         wallcycle_start(wcycle, ewcROTadd);
1940         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1941         wallcycle_stop(wcycle, ewcROTadd);
1942     }
1943
1944     if (PAR(cr) && !(cr->duty & DUTY_PME))
1945     {
1946         /* In case of node-splitting, the PP nodes receive the long-range
1947          * forces, virial and energy from the PME nodes here.
1948          */
1949         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1950     }
1951
1952     if (bDoForces)
1953     {
1954         post_process_forces(fplog, cr, step, nrnb, wcycle,
1955                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1956                             flags);
1957     }
1958
1959     /* Sum the potential energy terms from group contributions */
1960     sum_epot(&(inputrec->opts), &(enerd->grpp), enerd->term);
1961 }
1962
1963 void do_force(FILE *fplog, t_commrec *cr,
1964               t_inputrec *inputrec,
1965               gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1966               gmx_localtop_t *top,
1967               gmx_mtop_t *mtop,
1968               gmx_groups_t *groups,
1969               matrix box, rvec x[], history_t *hist,
1970               rvec f[],
1971               tensor vir_force,
1972               t_mdatoms *mdatoms,
1973               gmx_enerdata_t *enerd, t_fcdata *fcd,
1974               real *lambda, t_graph *graph,
1975               t_forcerec *fr,
1976               gmx_vsite_t *vsite, rvec mu_tot,
1977               double t, FILE *field, gmx_edsam_t ed,
1978               gmx_bool bBornRadii,
1979               int flags)
1980 {
1981     /* modify force flag if not doing nonbonded */
1982     if (!fr->bNonbonded)
1983     {
1984         flags &= ~GMX_FORCE_NONBONDED;
1985     }
1986
1987     switch (inputrec->cutoff_scheme)
1988     {
1989         case ecutsVERLET:
1990             do_force_cutsVERLET(fplog, cr, inputrec,
1991                                 step, nrnb, wcycle,
1992                                 top, mtop,
1993                                 groups,
1994                                 box, x, hist,
1995                                 f, vir_force,
1996                                 mdatoms,
1997                                 enerd, fcd,
1998                                 lambda, graph,
1999                                 fr, fr->ic,
2000                                 vsite, mu_tot,
2001                                 t, field, ed,
2002                                 bBornRadii,
2003                                 flags);
2004             break;
2005         case ecutsGROUP:
2006             do_force_cutsGROUP(fplog, cr, inputrec,
2007                                step, nrnb, wcycle,
2008                                top, mtop,
2009                                groups,
2010                                box, x, hist,
2011                                f, vir_force,
2012                                mdatoms,
2013                                enerd, fcd,
2014                                lambda, graph,
2015                                fr, vsite, mu_tot,
2016                                t, field, ed,
2017                                bBornRadii,
2018                                flags);
2019             break;
2020         default:
2021             gmx_incons("Invalid cut-off scheme passed!");
2022     }
2023 }
2024
2025
2026 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2027                         t_inputrec *ir, t_mdatoms *md,
2028                         t_state *state, rvec *f,
2029                         t_graph *graph, t_commrec *cr, t_nrnb *nrnb,
2030                         t_forcerec *fr, gmx_localtop_t *top, tensor shake_vir)
2031 {
2032     int             i, m, start, end;
2033     gmx_large_int_t step;
2034     real            dt = ir->delta_t;
2035     real            dvdl_dum;
2036     rvec           *savex;
2037
2038     snew(savex, state->natoms);
2039
2040     start = md->start;
2041     end   = md->homenr + start;
2042
2043     if (debug)
2044     {
2045         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2046                 start, md->homenr, end);
2047     }
2048     /* Do a first constrain to reset particles... */
2049     step = ir->init_step;
2050     if (fplog)
2051     {
2052         char buf[STEPSTRSIZE];
2053         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2054                 gmx_step_str(step, buf));
2055     }
2056     dvdl_dum = 0;
2057
2058     /* constrain the current position */
2059     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2060               ir, NULL, cr, step, 0, md,
2061               state->x, state->x, NULL,
2062               fr->bMolPBC, state->box,
2063               state->lambda[efptBONDED], &dvdl_dum,
2064               NULL, NULL, nrnb, econqCoord,
2065               ir->epc == epcMTTK, state->veta, state->veta);
2066     if (EI_VV(ir->eI))
2067     {
2068         /* constrain the inital velocity, and save it */
2069         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2070         /* might not yet treat veta correctly */
2071         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2072                   ir, NULL, cr, step, 0, md,
2073                   state->x, state->v, state->v,
2074                   fr->bMolPBC, state->box,
2075                   state->lambda[efptBONDED], &dvdl_dum,
2076                   NULL, NULL, nrnb, econqVeloc,
2077                   ir->epc == epcMTTK, state->veta, state->veta);
2078     }
2079     /* constrain the inital velocities at t-dt/2 */
2080     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2081     {
2082         for (i = start; (i < end); i++)
2083         {
2084             for (m = 0; (m < DIM); m++)
2085             {
2086                 /* Reverse the velocity */
2087                 state->v[i][m] = -state->v[i][m];
2088                 /* Store the position at t-dt in buf */
2089                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2090             }
2091         }
2092         /* Shake the positions at t=-dt with the positions at t=0
2093          * as reference coordinates.
2094          */
2095         if (fplog)
2096         {
2097             char buf[STEPSTRSIZE];
2098             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2099                     gmx_step_str(step, buf));
2100         }
2101         dvdl_dum = 0;
2102         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2103                   ir, NULL, cr, step, -1, md,
2104                   state->x, savex, NULL,
2105                   fr->bMolPBC, state->box,
2106                   state->lambda[efptBONDED], &dvdl_dum,
2107                   state->v, NULL, nrnb, econqCoord,
2108                   ir->epc == epcMTTK, state->veta, state->veta);
2109
2110         for (i = start; i < end; i++)
2111         {
2112             for (m = 0; m < DIM; m++)
2113             {
2114                 /* Re-reverse the velocities */
2115                 state->v[i][m] = -state->v[i][m];
2116             }
2117         }
2118     }
2119     sfree(savex);
2120 }
2121
2122 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2123 {
2124     double eners[2], virs[2], enersum, virsum, y0, f, g, h;
2125     double r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2126     double invscale, invscale2, invscale3;
2127     int    ri0, ri1, ri, i, offstart, offset;
2128     real   scale, *vdwtab, tabfactor, tmp;
2129
2130     fr->enershiftsix    = 0;
2131     fr->enershifttwelve = 0;
2132     fr->enerdiffsix     = 0;
2133     fr->enerdifftwelve  = 0;
2134     fr->virdiffsix      = 0;
2135     fr->virdifftwelve   = 0;
2136
2137     if (eDispCorr != edispcNO)
2138     {
2139         for (i = 0; i < 2; i++)
2140         {
2141             eners[i] = 0;
2142             virs[i]  = 0;
2143         }
2144         if ((fr->vdwtype == evdwSWITCH) || (fr->vdwtype == evdwSHIFT))
2145         {
2146             if (fr->rvdw_switch == 0)
2147             {
2148                 gmx_fatal(FARGS,
2149                           "With dispersion correction rvdw-switch can not be zero "
2150                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2151             }
2152
2153             scale  = fr->nblists[0].table_elec_vdw.scale;
2154             vdwtab = fr->nblists[0].table_vdw.data;
2155
2156             /* Round the cut-offs to exact table values for precision */
2157             ri0  = floor(fr->rvdw_switch*scale);
2158             ri1  = ceil(fr->rvdw*scale);
2159             r0   = ri0/scale;
2160             r1   = ri1/scale;
2161             rc3  = r0*r0*r0;
2162             rc9  = rc3*rc3*rc3;
2163
2164             if (fr->vdwtype == evdwSHIFT)
2165             {
2166                 /* Determine the constant energy shift below rvdw_switch.
2167                  * Table has a scale factor since we have scaled it down to compensate
2168                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2169                  */
2170                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2171                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2172             }
2173             /* Add the constant part from 0 to rvdw_switch.
2174              * This integration from 0 to rvdw_switch overcounts the number
2175              * of interactions by 1, as it also counts the self interaction.
2176              * We will correct for this later.
2177              */
2178             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2179             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2180
2181             invscale  = 1.0/(scale);
2182             invscale2 = invscale*invscale;
2183             invscale3 = invscale*invscale2;
2184
2185             /* following summation derived from cubic spline definition,
2186                Numerical Recipies in C, second edition, p. 113-116.  Exact
2187                for the cubic spline.  We first calculate the negative of
2188                the energy from rvdw to rvdw_switch, assuming that g(r)=1,
2189                and then add the more standard, abrupt cutoff correction to
2190                that result, yielding the long-range correction for a
2191                switched function.  We perform both the pressure and energy
2192                loops at the same time for simplicity, as the computational
2193                cost is low. */
2194
2195             for (i = 0; i < 2; i++)
2196             {
2197                 enersum = 0.0; virsum = 0.0;
2198                 if (i == 0)
2199                 {
2200                     offstart = 0;
2201                     /* Since the dispersion table has been scaled down a factor 6.0 and the repulsion
2202                      * a factor 12.0 to compensate for the c6/c12 parameters inside nbfp[] being scaled
2203                      * up (to save flops in kernels), we need to correct for this.
2204                      */
2205                     tabfactor = 6.0;
2206                 }
2207                 else
2208                 {
2209                     offstart  = 4;
2210                     tabfactor = 12.0;
2211                 }
2212                 for (ri = ri0; ri < ri1; ri++)
2213                 {
2214                     r  = ri*invscale;
2215                     ea = invscale3;
2216                     eb = 2.0*invscale2*r;
2217                     ec = invscale*r*r;
2218
2219                     pa = invscale3;
2220                     pb = 3.0*invscale2*r;
2221                     pc = 3.0*invscale*r*r;
2222                     pd = r*r*r;
2223
2224                     /* this "8" is from the packing in the vdwtab array - perhaps should be #define'ed? */
2225                     offset = 8*ri + offstart;
2226                     y0     = vdwtab[offset];
2227                     f      = vdwtab[offset+1];
2228                     g      = vdwtab[offset+2];
2229                     h      = vdwtab[offset+3];
2230
2231                     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);
2232                     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);
2233                 }
2234
2235                 enersum  *= 4.0*M_PI*tabfactor;
2236                 virsum   *= 4.0*M_PI*tabfactor;
2237                 eners[i] -= enersum;
2238                 virs[i]  -= virsum;
2239             }
2240
2241             /* now add the correction for rvdw_switch to infinity */
2242             eners[0] += -4.0*M_PI/(3.0*rc3);
2243             eners[1] +=  4.0*M_PI/(9.0*rc9);
2244             virs[0]  +=  8.0*M_PI/rc3;
2245             virs[1]  += -16.0*M_PI/(3.0*rc9);
2246         }
2247         else if ((fr->vdwtype == evdwCUT) || (fr->vdwtype == evdwUSER))
2248         {
2249             if (fr->vdwtype == evdwUSER && fplog)
2250             {
2251                 fprintf(fplog,
2252                         "WARNING: using dispersion correction with user tables\n");
2253             }
2254             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2255             rc9  = rc3*rc3*rc3;
2256             /* Contribution beyond the cut-off */
2257             eners[0] += -4.0*M_PI/(3.0*rc3);
2258             eners[1] +=  4.0*M_PI/(9.0*rc9);
2259             if (fr->vdw_modifier == eintmodPOTSHIFT)
2260             {
2261                 /* Contribution within the cut-off */
2262                 eners[0] += -4.0*M_PI/(3.0*rc3);
2263                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2264             }
2265             /* Contribution beyond the cut-off */
2266             virs[0]  +=  8.0*M_PI/rc3;
2267             virs[1]  += -16.0*M_PI/(3.0*rc9);
2268         }
2269         else
2270         {
2271             gmx_fatal(FARGS,
2272                       "Dispersion correction is not implemented for vdw-type = %s",
2273                       evdw_names[fr->vdwtype]);
2274         }
2275         fr->enerdiffsix    = eners[0];
2276         fr->enerdifftwelve = eners[1];
2277         /* The 0.5 is due to the Gromacs definition of the virial */
2278         fr->virdiffsix     = 0.5*virs[0];
2279         fr->virdifftwelve  = 0.5*virs[1];
2280     }
2281 }
2282
2283 void calc_dispcorr(FILE *fplog, t_inputrec *ir, t_forcerec *fr,
2284                    gmx_large_int_t step, int natoms,
2285                    matrix box, real lambda, tensor pres, tensor virial,
2286                    real *prescorr, real *enercorr, real *dvdlcorr)
2287 {
2288     gmx_bool bCorrAll, bCorrPres;
2289     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2290     int      m;
2291
2292     *prescorr = 0;
2293     *enercorr = 0;
2294     *dvdlcorr = 0;
2295
2296     clear_mat(virial);
2297     clear_mat(pres);
2298
2299     if (ir->eDispCorr != edispcNO)
2300     {
2301         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2302                      ir->eDispCorr == edispcAllEnerPres);
2303         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2304                      ir->eDispCorr == edispcAllEnerPres);
2305
2306         invvol = 1/det(box);
2307         if (fr->n_tpi)
2308         {
2309             /* Only correct for the interactions with the inserted molecule */
2310             dens   = (natoms - fr->n_tpi)*invvol;
2311             ninter = fr->n_tpi;
2312         }
2313         else
2314         {
2315             dens   = natoms*invvol;
2316             ninter = 0.5*natoms;
2317         }
2318
2319         if (ir->efep == efepNO)
2320         {
2321             avcsix    = fr->avcsix[0];
2322             avctwelve = fr->avctwelve[0];
2323         }
2324         else
2325         {
2326             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2327             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2328         }
2329
2330         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2331         *enercorr += avcsix*enerdiff;
2332         dvdlambda  = 0.0;
2333         if (ir->efep != efepNO)
2334         {
2335             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2336         }
2337         if (bCorrAll)
2338         {
2339             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2340             *enercorr += avctwelve*enerdiff;
2341             if (fr->efep != efepNO)
2342             {
2343                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2344             }
2345         }
2346
2347         if (bCorrPres)
2348         {
2349             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2350             if (ir->eDispCorr == edispcAllEnerPres)
2351             {
2352                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2353             }
2354             /* The factor 2 is because of the Gromacs virial definition */
2355             spres = -2.0*invvol*svir*PRESFAC;
2356
2357             for (m = 0; m < DIM; m++)
2358             {
2359                 virial[m][m] += svir;
2360                 pres[m][m]   += spres;
2361             }
2362             *prescorr += spres;
2363         }
2364
2365         /* Can't currently control when it prints, for now, just print when degugging */
2366         if (debug)
2367         {
2368             if (bCorrAll)
2369             {
2370                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2371                         avcsix, avctwelve);
2372             }
2373             if (bCorrPres)
2374             {
2375                 fprintf(debug,
2376                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2377                         *enercorr, spres, svir);
2378             }
2379             else
2380             {
2381                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2382             }
2383         }
2384
2385         if (fr->bSepDVDL && do_per_step(step, ir->nstlog))
2386         {
2387             fprintf(fplog, sepdvdlformat, "Dispersion correction",
2388                     *enercorr, dvdlambda);
2389         }
2390         if (fr->efep != efepNO)
2391         {
2392             *dvdlcorr += dvdlambda;
2393         }
2394     }
2395 }
2396
2397 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2398                   t_graph *graph, rvec x[])
2399 {
2400     if (fplog)
2401     {
2402         fprintf(fplog, "Removing pbc first time\n");
2403     }
2404     calc_shifts(box, fr->shift_vec);
2405     if (graph)
2406     {
2407         mk_mshift(fplog, graph, fr->ePBC, box, x);
2408         if (gmx_debug_at)
2409         {
2410             p_graph(debug, "do_pbc_first 1", graph);
2411         }
2412         shift_self(graph, box, x);
2413         /* By doing an extra mk_mshift the molecules that are broken
2414          * because they were e.g. imported from another software
2415          * will be made whole again. Such are the healing powers
2416          * of GROMACS.
2417          */
2418         mk_mshift(fplog, graph, fr->ePBC, box, x);
2419         if (gmx_debug_at)
2420         {
2421             p_graph(debug, "do_pbc_first 2", graph);
2422         }
2423     }
2424     if (fplog)
2425     {
2426         fprintf(fplog, "Done rmpbc\n");
2427     }
2428 }
2429
2430 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2431                             gmx_mtop_t *mtop, rvec x[],
2432                             gmx_bool bFirst)
2433 {
2434     t_graph        *graph;
2435     int             mb, as, mol;
2436     gmx_molblock_t *molb;
2437
2438     if (bFirst && fplog)
2439     {
2440         fprintf(fplog, "Removing pbc first time\n");
2441     }
2442
2443     snew(graph, 1);
2444     as = 0;
2445     for (mb = 0; mb < mtop->nmolblock; mb++)
2446     {
2447         molb = &mtop->molblock[mb];
2448         if (molb->natoms_mol == 1 ||
2449             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2450         {
2451             /* Just one atom or charge group in the molecule, no PBC required */
2452             as += molb->nmol*molb->natoms_mol;
2453         }
2454         else
2455         {
2456             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2457             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2458                            0, molb->natoms_mol, FALSE, FALSE, graph);
2459
2460             for (mol = 0; mol < molb->nmol; mol++)
2461             {
2462                 mk_mshift(fplog, graph, ePBC, box, x+as);
2463
2464                 shift_self(graph, box, x+as);
2465                 /* The molecule is whole now.
2466                  * We don't need the second mk_mshift call as in do_pbc_first,
2467                  * since we no longer need this graph.
2468                  */
2469
2470                 as += molb->natoms_mol;
2471             }
2472             done_graph(graph);
2473         }
2474     }
2475     sfree(graph);
2476 }
2477
2478 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2479                        gmx_mtop_t *mtop, rvec x[])
2480 {
2481     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2482 }
2483
2484 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2485                  gmx_mtop_t *mtop, rvec x[])
2486 {
2487     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2488 }
2489
2490 void finish_run(FILE *fplog, t_commrec *cr, const char *confout,
2491                 t_inputrec *inputrec,
2492                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2493                 gmx_runtime_t *runtime,
2494                 wallclock_gpu_t *gputimes,
2495                 int omp_nth_pp,
2496                 gmx_bool bWriteStat)
2497 {
2498     int     i, j;
2499     t_nrnb *nrnb_tot = NULL;
2500     real    delta_t;
2501     double  nbfs, mflop;
2502
2503     wallcycle_sum(cr, wcycle);
2504
2505     if (cr->nnodes > 1)
2506     {
2507         snew(nrnb_tot, 1);
2508 #ifdef GMX_MPI
2509         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2510                       cr->mpi_comm_mysim);
2511 #endif
2512     }
2513     else
2514     {
2515         nrnb_tot = nrnb;
2516     }
2517
2518 #if defined(GMX_MPI) && !defined(GMX_THREAD_MPI)
2519     if (cr->nnodes > 1)
2520     {
2521         /* reduce nodetime over all MPI processes in the current simulation */
2522         double sum;
2523         MPI_Allreduce(&runtime->proctime, &sum, 1, MPI_DOUBLE, MPI_SUM,
2524                       cr->mpi_comm_mysim);
2525         runtime->proctime = sum;
2526     }
2527 #endif
2528
2529     if (SIMMASTER(cr))
2530     {
2531         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2532     }
2533     if (cr->nnodes > 1)
2534     {
2535         sfree(nrnb_tot);
2536     }
2537
2538     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2539     {
2540         print_dd_statistics(cr, inputrec, fplog);
2541     }
2542
2543 #ifdef GMX_MPI
2544     if (PARTDECOMP(cr))
2545     {
2546         if (MASTER(cr))
2547         {
2548             t_nrnb     *nrnb_all;
2549             int         s;
2550             MPI_Status  stat;
2551
2552             snew(nrnb_all, cr->nnodes);
2553             nrnb_all[0] = *nrnb;
2554             for (s = 1; s < cr->nnodes; s++)
2555             {
2556                 MPI_Recv(nrnb_all[s].n, eNRNB, MPI_DOUBLE, s, 0,
2557                          cr->mpi_comm_mysim, &stat);
2558             }
2559             pr_load(fplog, cr, nrnb_all);
2560             sfree(nrnb_all);
2561         }
2562         else
2563         {
2564             MPI_Send(nrnb->n, eNRNB, MPI_DOUBLE, MASTERRANK(cr), 0,
2565                      cr->mpi_comm_mysim);
2566         }
2567     }
2568 #endif
2569
2570     if (SIMMASTER(cr))
2571     {
2572         wallcycle_print(fplog, cr->nnodes, cr->npmenodes, runtime->realtime,
2573                         wcycle, gputimes);
2574
2575         if (EI_DYNAMICS(inputrec->eI))
2576         {
2577             delta_t = inputrec->delta_t;
2578         }
2579         else
2580         {
2581             delta_t = 0;
2582         }
2583
2584         if (fplog)
2585         {
2586             print_perf(fplog, runtime->proctime, runtime->realtime,
2587                        cr->nnodes-cr->npmenodes,
2588                        runtime->nsteps_done, delta_t, nbfs, mflop,
2589                        omp_nth_pp);
2590         }
2591         if (bWriteStat)
2592         {
2593             print_perf(stderr, runtime->proctime, runtime->realtime,
2594                        cr->nnodes-cr->npmenodes,
2595                        runtime->nsteps_done, delta_t, nbfs, mflop,
2596                        omp_nth_pp);
2597         }
2598     }
2599 }
2600
2601 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2602 {
2603     /* this function works, but could probably use a logic rewrite to keep all the different
2604        types of efep straight. */
2605
2606     int       i;
2607     t_lambda *fep = ir->fepvals;
2608
2609     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2610     {
2611         for (i = 0; i < efptNR; i++)
2612         {
2613             lambda[i] = 0.0;
2614             if (lam0)
2615             {
2616                 lam0[i] = 0.0;
2617             }
2618         }
2619         return;
2620     }
2621     else
2622     {
2623         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2624                                              if checkpoint is set -- a kludge is in for now
2625                                              to prevent this.*/
2626         for (i = 0; i < efptNR; i++)
2627         {
2628             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2629             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2630             {
2631                 lambda[i] = fep->init_lambda;
2632                 if (lam0)
2633                 {
2634                     lam0[i] = lambda[i];
2635                 }
2636             }
2637             else
2638             {
2639                 lambda[i] = fep->all_lambda[i][*fep_state];
2640                 if (lam0)
2641                 {
2642                     lam0[i] = lambda[i];
2643                 }
2644             }
2645         }
2646         if (ir->bSimTemp)
2647         {
2648             /* need to rescale control temperatures to match current state */
2649             for (i = 0; i < ir->opts.ngtc; i++)
2650             {
2651                 if (ir->opts.ref_t[i] > 0)
2652                 {
2653                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2654                 }
2655             }
2656         }
2657     }
2658
2659     /* Send to the log the information on the current lambdas */
2660     if (fplog != NULL)
2661     {
2662         fprintf(fplog, "Initial vector of lambda components:[ ");
2663         for (i = 0; i < efptNR; i++)
2664         {
2665             fprintf(fplog, "%10.4f ", lambda[i]);
2666         }
2667         fprintf(fplog, "]\n");
2668     }
2669     return;
2670 }
2671
2672
2673 void init_md(FILE *fplog,
2674              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2675              double *t, double *t0,
2676              real *lambda, int *fep_state, double *lam0,
2677              t_nrnb *nrnb, gmx_mtop_t *mtop,
2678              gmx_update_t *upd,
2679              int nfile, const t_filenm fnm[],
2680              gmx_mdoutf_t **outf, t_mdebin **mdebin,
2681              tensor force_vir, tensor shake_vir, rvec mu_tot,
2682              gmx_bool *bSimAnn, t_vcm **vcm, t_state *state, unsigned long Flags)
2683 {
2684     int  i, j, n;
2685     real tmpt, mod;
2686
2687     /* Initial values */
2688     *t = *t0       = ir->init_t;
2689
2690     *bSimAnn = FALSE;
2691     for (i = 0; i < ir->opts.ngtc; i++)
2692     {
2693         /* set bSimAnn if any group is being annealed */
2694         if (ir->opts.annealing[i] != eannNO)
2695         {
2696             *bSimAnn = TRUE;
2697         }
2698     }
2699     if (*bSimAnn)
2700     {
2701         update_annealing_target_temp(&(ir->opts), ir->init_t);
2702     }
2703
2704     /* Initialize lambda variables */
2705     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2706
2707     if (upd)
2708     {
2709         *upd = init_update(fplog, ir);
2710     }
2711
2712
2713     if (vcm != NULL)
2714     {
2715         *vcm = init_vcm(fplog, &mtop->groups, ir);
2716     }
2717
2718     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2719     {
2720         if (ir->etc == etcBERENDSEN)
2721         {
2722             please_cite(fplog, "Berendsen84a");
2723         }
2724         if (ir->etc == etcVRESCALE)
2725         {
2726             please_cite(fplog, "Bussi2007a");
2727         }
2728     }
2729
2730     init_nrnb(nrnb);
2731
2732     if (nfile != -1)
2733     {
2734         *outf = init_mdoutf(nfile, fnm, Flags, cr, ir, oenv);
2735
2736         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : (*outf)->fp_ene,
2737                               mtop, ir, (*outf)->fp_dhdl);
2738     }
2739
2740     if (ir->bAdress)
2741     {
2742         please_cite(fplog, "Fritsch12");
2743         please_cite(fplog, "Junghans10");
2744     }
2745     /* Initiate variables */
2746     clear_mat(force_vir);
2747     clear_mat(shake_vir);
2748     clear_rvec(mu_tot);
2749
2750     debug_gmx();
2751 }