src/gromacs/mdlib/sim_util.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "gromacs/legacyheaders/sim_util.h"
  40
  41 #include "config.h"
  42
  43 #include <assert.h>
  44 #include <math.h>
  45 #include <stdio.h>
  46 #include <string.h>
  47
  48 #ifdef HAVE_SYS_TIME_H
  49 #include <sys/time.h>
  50 #endif
  51
  52 #include "gromacs/essentialdynamics/edsam.h"
  53 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  54 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  55 #include "gromacs/imd/imd.h"
  56 #include "gromacs/legacyheaders/calcmu.h"
  57 #include "gromacs/legacyheaders/chargegroup.h"
  58 #include "gromacs/legacyheaders/constr.h"
  59 #include "gromacs/legacyheaders/copyrite.h"
  60 #include "gromacs/legacyheaders/disre.h"
  61 #include "gromacs/legacyheaders/domdec.h"
  62 #include "gromacs/legacyheaders/force.h"
  63 #include "gromacs/legacyheaders/genborn.h"
  64 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
  65 #include "gromacs/legacyheaders/mdatoms.h"
  66 #include "gromacs/legacyheaders/mdrun.h"
  67 #include "gromacs/legacyheaders/names.h"
  68 #include "gromacs/legacyheaders/network.h"
  69 #include "gromacs/legacyheaders/nonbonded.h"
  70 #include "gromacs/legacyheaders/nrnb.h"
  71 #include "gromacs/legacyheaders/orires.h"
  72 #include "gromacs/legacyheaders/pme.h"
  73 #include "gromacs/legacyheaders/qmmm.h"
  74 #include "gromacs/legacyheaders/txtdump.h"
  75 #include "gromacs/legacyheaders/typedefs.h"
  76 #include "gromacs/legacyheaders/update.h"
  77 #include "gromacs/legacyheaders/types/commrec.h"
  78 #include "gromacs/listed-forces/bonded.h"
  79 #include "gromacs/math/units.h"
  80 #include "gromacs/math/vec.h"
  81 #include "gromacs/mdlib/nb_verlet.h"
  82 #include "gromacs/mdlib/nbnxn_atomdata.h"
  83 #include "gromacs/mdlib/nbnxn_search.h"
  84 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda.h"
  85 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.h"
  86 #include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  87 #include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_ref.h"
  88 #include "gromacs/mdlib/nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn.h"
  89 #include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.h"
  90 #include "gromacs/pbcutil/ishift.h"
  91 #include "gromacs/pbcutil/mshift.h"
  92 #include "gromacs/pbcutil/pbc.h"
  93 #include "gromacs/pulling/pull.h"
  94 #include "gromacs/pulling/pull_rotation.h"
  95 #include "gromacs/timing/wallcycle.h"
  96 #include "gromacs/timing/walltime_accounting.h"
  97 #include "gromacs/utility/cstringutil.h"
  98 #include "gromacs/utility/gmxmpi.h"
  99 #include "gromacs/utility/smalloc.h"
 100
 101 #include "adress.h"
 102
 103 void print_time(FILE                     *out,
 104                 gmx_walltime_accounting_t walltime_accounting,
 105                 gmx_int64_t               step,
 106                 t_inputrec               *ir,
 107                 t_commrec gmx_unused     *cr)
 108 {
 109     time_t finish;
 110     char   timebuf[STRLEN];
 111     double dt, elapsed_seconds, time_per_step;
 112     char   buf[48];
 113
 114 #ifndef GMX_THREAD_MPI
 115     if (!PAR(cr))
 116 #endif
 117     {
 118         fprintf(out, "\r");
 119     }
 120     fprintf(out, "step %s", gmx_step_str(step, buf));
 121     if ((step >= ir->nstlist))
 122     {
 123         double seconds_since_epoch = gmx_gettime();
 124         elapsed_seconds = seconds_since_epoch - walltime_accounting_get_start_time_stamp(walltime_accounting);
 125         time_per_step   = elapsed_seconds/(step - ir->init_step + 1);
 126         dt              = (ir->nsteps + ir->init_step - step) * time_per_step;
 127
 128         if (ir->nsteps >= 0)
 129         {
 130             if (dt >= 300)
 131             {
 132                 finish = (time_t) (seconds_since_epoch + dt);
 133                 gmx_ctime_r(&finish, timebuf, STRLEN);
 134                 sprintf(buf, "%s", timebuf);
 135                 buf[strlen(buf)-1] = '\0';
 136                 fprintf(out, ", will finish %s", buf);
 137             }
 138             else
 139             {
 140                 fprintf(out, ", remaining wall clock time: %5d s          ", (int)dt);
 141             }
 142         }
 143         else
 144         {
 145             fprintf(out, " performance: %.1f ns/day    ",
 146                     ir->delta_t/1000*24*60*60/time_per_step);
 147         }
 148     }
 149 #ifndef GMX_THREAD_MPI
 150     if (PAR(cr))
 151     {
 152         fprintf(out, "\n");
 153     }
 154 #endif
 155
 156     fflush(out);
 157 }
 158
 159 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
 160                          double the_time)
 161 {
 162     char   time_string[STRLEN];
 163
 164     if (!fplog)
 165     {
 166         return;
 167     }
 168
 169     {
 170         int    i;
 171         char   timebuf[STRLEN];
 172         time_t temp_time = (time_t) the_time;
 173
 174         gmx_ctime_r(&temp_time, timebuf, STRLEN);
 175         for (i = 0; timebuf[i] >= ' '; i++)
 176         {
 177             time_string[i] = timebuf[i];
 178         }
 179         time_string[i] = '\0';
 180     }
 181
 182     fprintf(fplog, "%s on rank %d %s\n", title, nodeid, time_string);
 183 }
 184
 185 void print_start(FILE *fplog, t_commrec *cr,
 186                  gmx_walltime_accounting_t walltime_accounting,
 187                  const char *name)
 188 {
 189     char buf[STRLEN];
 190
 191     sprintf(buf, "Started %s", name);
 192     print_date_and_time(fplog, cr->nodeid, buf,
 193                         walltime_accounting_get_start_time_stamp(walltime_accounting));
 194 }
 195
 196 static void sum_forces(int start, int end, rvec f[], rvec flr[])
 197 {
 198     int i;
 199
 200     if (gmx_debug_at)
 201     {
 202         pr_rvecs(debug, 0, "fsr", f+start, end-start);
 203         pr_rvecs(debug, 0, "flr", flr+start, end-start);
 204     }
 205     for (i = start; (i < end); i++)
 206     {
 207         rvec_inc(f[i], flr[i]);
 208     }
 209 }
 210
 211 /*
 212  * calc_f_el calculates forces due to an electric field.
 213  *
 214  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 215  *
 216  * Et[] contains the parameters for the time dependent
 217  * part of the field (not yet used).
 218  * Ex[] contains the parameters for
 219  * the spatial dependent part of the field. You can have cool periodic
 220  * fields in principle, but only a constant field is supported
 221  * now.
 222  * The function should return the energy due to the electric field
 223  * (if any) but for now returns 0.
 224  *
 225  * WARNING:
 226  * There can be problems with the virial.
 227  * Since the field is not self-consistent this is unavoidable.
 228  * For neutral molecules the virial is correct within this approximation.
 229  * For neutral systems with many charged molecules the error is small.
 230  * But for systems with a net charge or a few charged molecules
 231  * the error can be significant when the field is high.
 232  * Solution: implement a self-consitent electric field into PME.
 233  */
 234 static void calc_f_el(FILE *fp, int  start, int homenr,
 235                       real charge[], rvec f[],
 236                       t_cosines Ex[], t_cosines Et[], double t)
 237 {
 238     rvec Ext;
 239     real t0;
 240     int  i, m;
 241
 242     for (m = 0; (m < DIM); m++)
 243     {
 244         if (Et[m].n > 0)
 245         {
 246             if (Et[m].n == 3)
 247             {
 248                 t0     = Et[m].a[1];
 249                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 250             }
 251             else
 252             {
 253                 Ext[m] = cos(Et[m].a[0]*t);
 254             }
 255         }
 256         else
 257         {
 258             Ext[m] = 1.0;
 259         }
 260         if (Ex[m].n > 0)
 261         {
 262             /* Convert the field strength from V/nm to MD-units */
 263             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 264             for (i = start; (i < start+homenr); i++)
 265             {
 266                 f[i][m] += charge[i]*Ext[m];
 267             }
 268         }
 269         else
 270         {
 271             Ext[m] = 0;
 272         }
 273     }
 274     if (fp != NULL)
 275     {
 276         fprintf(fp, "%10g  %10g  %10g  %10g #FIELD\n", t,
 277                 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
 278     }
 279 }
 280
 281 static void calc_virial(int start, int homenr, rvec x[], rvec f[],
 282                         tensor vir_part, t_graph *graph, matrix box,
 283                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 284 {
 285     int    i, j;
 286     tensor virtest;
 287
 288     /* The short-range virial from surrounding boxes */
 289     clear_mat(vir_part);
 290     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 291     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 292
 293     /* Calculate partial virial, for local atoms only, based on short range.
 294      * Total virial is computed in global_stat, called from do_md
 295      */
 296     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 297     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 298
 299     /* Add position restraint contribution */
 300     for (i = 0; i < DIM; i++)
 301     {
 302         vir_part[i][i] += fr->vir_diag_posres[i];
 303     }
 304
 305     /* Add wall contribution */
 306     for (i = 0; i < DIM; i++)
 307     {
 308         vir_part[i][ZZ] += fr->vir_wall_z[i];
 309     }
 310
 311     if (debug)
 312     {
 313         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 314     }
 315 }
 316
 317 static void posres_wrapper(int flags,
 318                            t_inputrec *ir,
 319                            t_nrnb *nrnb,
 320                            gmx_localtop_t *top,
 321                            matrix box, rvec x[],
 322                            gmx_enerdata_t *enerd,
 323                            real *lambda,
 324                            t_forcerec *fr)
 325 {
 326     t_pbc pbc;
 327     real  v, dvdl;
 328     int   i;
 329
 330     /* Position restraints always require full pbc */
 331     set_pbc(&pbc, ir->ePBC, box);
 332     dvdl = 0;
 333     v    = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 334                   top->idef.iparams_posres,
 335                   (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 336                   ir->ePBC == epbcNONE ? NULL : &pbc,
 337                   lambda[efptRESTRAINT], &dvdl,
 338                   fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 339     enerd->term[F_POSRES] += v;
 340     /* If just the force constant changes, the FEP term is linear,
 341      * but if k changes, it is not.
 342      */
 343     enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
 344     inc_nrnb(nrnb, eNR_POSRES, top->idef.il[F_POSRES].nr/2);
 345
 346     if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
 347     {
 348         for (i = 0; i < enerd->n_lambda; i++)
 349         {
 350             real dvdl_dum, lambda_dum;
 351
 352             lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
 353             v          = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 354                                 top->idef.iparams_posres,
 355                                 (const rvec*)x, NULL, NULL,
 356                                 ir->ePBC == epbcNONE ? NULL : &pbc, lambda_dum, &dvdl,
 357                                 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 358             enerd->enerpart_lambda[i] += v;
 359         }
 360     }
 361 }
 362
 363 static void fbposres_wrapper(t_inputrec *ir,
 364                              t_nrnb *nrnb,
 365                              gmx_localtop_t *top,
 366                              matrix box, rvec x[],
 367                              gmx_enerdata_t *enerd,
 368                              t_forcerec *fr)
 369 {
 370     t_pbc pbc;
 371     real  v;
 372
 373     /* Flat-bottomed position restraints always require full pbc */
 374     set_pbc(&pbc, ir->ePBC, box);
 375     v = fbposres(top->idef.il[F_FBPOSRES].nr, top->idef.il[F_FBPOSRES].iatoms,
 376                  top->idef.iparams_fbposres,
 377                  (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 378                  ir->ePBC == epbcNONE ? NULL : &pbc,
 379                  fr->rc_scaling, fr->ePBC, fr->posres_com);
 380     enerd->term[F_FBPOSRES] += v;
 381     inc_nrnb(nrnb, eNR_FBPOSRES, top->idef.il[F_FBPOSRES].nr/2);
 382 }
 383
 384 static void pull_potential_wrapper(t_commrec *cr,
 385                                    t_inputrec *ir,
 386                                    matrix box, rvec x[],
 387                                    rvec f[],
 388                                    tensor vir_force,
 389                                    t_mdatoms *mdatoms,
 390                                    gmx_enerdata_t *enerd,
 391                                    real *lambda,
 392                                    double t,
 393                                    gmx_wallcycle_t wcycle)
 394 {
 395     t_pbc  pbc;
 396     real   dvdl;
 397
 398     /* Calculate the center of mass forces, this requires communication,
 399      * which is why pull_potential is called close to other communication.
 400      * The virial contribution is calculated directly,
 401      * which is why we call pull_potential after calc_virial.
 402      */
 403     wallcycle_start(wcycle, ewcPULLPOT);
 404     set_pbc(&pbc, ir->ePBC, box);
 405     dvdl                     = 0;
 406     enerd->term[F_COM_PULL] +=
 407         pull_potential(ir->ePull, ir->pull, mdatoms, &pbc,
 408                        cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
 409     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 410     wallcycle_stop(wcycle, ewcPULLPOT);
 411 }
 412
 413 static void pme_receive_force_ener(t_commrec      *cr,
 414                                    gmx_wallcycle_t wcycle,
 415                                    gmx_enerdata_t *enerd,
 416                                    t_forcerec     *fr)
 417 {
 418     real   e_q, e_lj, v, dvdl_q, dvdl_lj;
 419     float  cycles_ppdpme, cycles_seppme;
 420
 421     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 422     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 423
 424     /* In case of node-splitting, the PP nodes receive the long-range
 425      * forces, virial and energy from the PME nodes here.
 426      */
 427     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 428     dvdl_q  = 0;
 429     dvdl_lj = 0;
 430     gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e_q,
 431                       fr->vir_lj_recip, &e_lj, &dvdl_q, &dvdl_lj,
 432                       &cycles_seppme);
 433     enerd->term[F_COUL_RECIP] += e_q;
 434     enerd->term[F_LJ_RECIP]   += e_lj;
 435     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 436     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 437
 438     if (wcycle)
 439     {
 440         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 441     }
 442     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 443 }
 444
 445 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
 446                                gmx_int64_t step, real pforce, rvec *x, rvec *f)
 447 {
 448     int  i;
 449     real pf2, fn2;
 450     char buf[STEPSTRSIZE];
 451
 452     pf2 = sqr(pforce);
 453     for (i = 0; i < md->homenr; i++)
 454     {
 455         fn2 = norm2(f[i]);
 456         /* We also catch NAN, if the compiler does not optimize this away. */
 457         if (fn2 >= pf2 || fn2 != fn2)
 458         {
 459             fprintf(fp, "step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 460                     gmx_step_str(step, buf),
 461                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
 462         }
 463     }
 464 }
 465
 466 static void post_process_forces(t_commrec *cr,
 467                                 gmx_int64_t step,
 468                                 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 469                                 gmx_localtop_t *top,
 470                                 matrix box, rvec x[],
 471                                 rvec f[],
 472                                 tensor vir_force,
 473                                 t_mdatoms *mdatoms,
 474                                 t_graph *graph,
 475                                 t_forcerec *fr, gmx_vsite_t *vsite,
 476                                 int flags)
 477 {
 478     if (fr->bF_NoVirSum)
 479     {
 480         if (vsite)
 481         {
 482             /* Spread the mesh force on virtual sites to the other particles...
 483              * This is parallellized. MPI communication is performed
 484              * if the constructing atoms aren't local.
 485              */
 486             wallcycle_start(wcycle, ewcVSITESPREAD);
 487             spread_vsite_f(vsite, x, fr->f_novirsum, NULL,
 488                            (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
 489                            nrnb,
 490                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
 491             wallcycle_stop(wcycle, ewcVSITESPREAD);
 492         }
 493         if (flags & GMX_FORCE_VIRIAL)
 494         {
 495             /* Now add the forces, this is local */
 496             if (fr->bDomDec)
 497             {
 498                 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
 499             }
 500             else
 501             {
 502                 sum_forces(0, mdatoms->homenr,
 503                            f, fr->f_novirsum);
 504             }
 505             if (EEL_FULL(fr->eeltype))
 506             {
 507                 /* Add the mesh contribution to the virial */
 508                 m_add(vir_force, fr->vir_el_recip, vir_force);
 509             }
 510             if (EVDW_PME(fr->vdwtype))
 511             {
 512                 /* Add the mesh contribution to the virial */
 513                 m_add(vir_force, fr->vir_lj_recip, vir_force);
 514             }
 515             if (debug)
 516             {
 517                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 518             }
 519         }
 520     }
 521
 522     if (fr->print_force >= 0)
 523     {
 524         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 525     }
 526 }
 527
 528 static void do_nb_verlet(t_forcerec *fr,
 529                          interaction_const_t *ic,
 530                          gmx_enerdata_t *enerd,
 531                          int flags, int ilocality,
 532                          int clearF,
 533                          t_nrnb *nrnb,
 534                          gmx_wallcycle_t wcycle)
 535 {
 536     int                        nnbl, kernel_type, enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
 537     char                      *env;
 538     nonbonded_verlet_group_t  *nbvg;
 539     gmx_bool                   bCUDA;
 540
 541     if (!(flags & GMX_FORCE_NONBONDED))
 542     {
 543         /* skip non-bonded calculation */
 544         return;
 545     }
 546
 547     nbvg = &fr->nbv->grp[ilocality];
 548
 549     /* CUDA kernel launch overhead is already timed separately */
 550     if (fr->cutoff_scheme != ecutsVERLET)
 551     {
 552         gmx_incons("Invalid cut-off scheme passed!");
 553     }
 554
 555     bCUDA = (nbvg->kernel_type == nbnxnk8x8x8_CUDA);
 556
 557     if (!bCUDA)
 558     {
 559         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 560     }
 561     switch (nbvg->kernel_type)
 562     {
 563         case nbnxnk4x4_PlainC:
 564             nbnxn_kernel_ref(&nbvg->nbl_lists,
 565                              nbvg->nbat, ic,
 566                              fr->shift_vec,
 567                              flags,
 568                              clearF,
 569                              fr->fshift[0],
 570                              enerd->grpp.ener[egCOULSR],
 571                              fr->bBHAM ?
 572                              enerd->grpp.ener[egBHAMSR] :
 573                              enerd->grpp.ener[egLJSR]);
 574             break;
 575
 576         case nbnxnk4xN_SIMD_4xN:
 577             nbnxn_kernel_simd_4xn(&nbvg->nbl_lists,
 578                                   nbvg->nbat, ic,
 579                                   nbvg->ewald_excl,
 580                                   fr->shift_vec,
 581                                   flags,
 582                                   clearF,
 583                                   fr->fshift[0],
 584                                   enerd->grpp.ener[egCOULSR],
 585                                   fr->bBHAM ?
 586                                   enerd->grpp.ener[egBHAMSR] :
 587                                   enerd->grpp.ener[egLJSR]);
 588             break;
 589         case nbnxnk4xN_SIMD_2xNN:
 590             nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
 591                                    nbvg->nbat, ic,
 592                                    nbvg->ewald_excl,
 593                                    fr->shift_vec,
 594                                    flags,
 595                                    clearF,
 596                                    fr->fshift[0],
 597                                    enerd->grpp.ener[egCOULSR],
 598                                    fr->bBHAM ?
 599                                    enerd->grpp.ener[egBHAMSR] :
 600                                    enerd->grpp.ener[egLJSR]);
 601             break;
 602
 603         case nbnxnk8x8x8_CUDA:
 604             nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
 605             break;
 606
 607         case nbnxnk8x8x8_PlainC:
 608             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 609                                  nbvg->nbat, ic,
 610                                  fr->shift_vec,
 611                                  flags,
 612                                  clearF,
 613                                  nbvg->nbat->out[0].f,
 614                                  fr->fshift[0],
 615                                  enerd->grpp.ener[egCOULSR],
 616                                  fr->bBHAM ?
 617                                  enerd->grpp.ener[egBHAMSR] :
 618                                  enerd->grpp.ener[egLJSR]);
 619             break;
 620
 621         default:
 622             gmx_incons("Invalid nonbonded kernel type passed!");
 623
 624     }
 625     if (!bCUDA)
 626     {
 627         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 628     }
 629
 630     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 631     {
 632         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 633     }
 634     else if ((!bCUDA && nbvg->ewald_excl == ewaldexclAnalytical) ||
 635              (bCUDA && nbnxn_cuda_is_kernel_ewald_analytical(fr->nbv->cu_nbv)))
 636     {
 637         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 638     }
 639     else
 640     {
 641         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 642     }
 643     enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 644     if (flags & GMX_FORCE_ENERGY)
 645     {
 646         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 647         enr_nbnxn_kernel_ljc += 1;
 648         enr_nbnxn_kernel_lj  += 1;
 649     }
 650
 651     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 652              nbvg->nbl_lists.natpair_ljq);
 653     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 654              nbvg->nbl_lists.natpair_lj);
 655     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 656     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 657              nbvg->nbl_lists.natpair_q);
 658
 659     if (ic->vdw_modifier == eintmodFORCESWITCH)
 660     {
 661         /* We add up the switch cost separately */
 662         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_FSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 663                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 664     }
 665     if (ic->vdw_modifier == eintmodPOTSWITCH)
 666     {
 667         /* We add up the switch cost separately */
 668         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_PSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 669                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 670     }
 671     if (ic->vdwtype == evdwPME)
 672     {
 673         /* We add up the LJ Ewald cost separately */
 674         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_EWALD+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 675                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 676     }
 677 }
 678
 679 static void do_nb_verlet_fep(nbnxn_pairlist_set_t *nbl_lists,
 680                              t_forcerec           *fr,
 681                              rvec                  x[],
 682                              rvec                  f[],
 683                              t_mdatoms            *mdatoms,
 684                              t_lambda             *fepvals,
 685                              real                 *lambda,
 686                              gmx_enerdata_t       *enerd,
 687                              int                   flags,
 688                              t_nrnb               *nrnb,
 689                              gmx_wallcycle_t       wcycle)
 690 {
 691     int              donb_flags;
 692     nb_kernel_data_t kernel_data;
 693     real             lam_i[efptNR];
 694     real             dvdl_nb[efptNR];
 695     int              th;
 696     int              i, j;
 697
 698     donb_flags = 0;
 699     /* Add short-range interactions */
 700     donb_flags |= GMX_NONBONDED_DO_SR;
 701
 702     /* Currently all group scheme kernels always calculate (shift-)forces */
 703     if (flags & GMX_FORCE_FORCES)
 704     {
 705         donb_flags |= GMX_NONBONDED_DO_FORCE;
 706     }
 707     if (flags & GMX_FORCE_VIRIAL)
 708     {
 709         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 710     }
 711     if (flags & GMX_FORCE_ENERGY)
 712     {
 713         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 714     }
 715     if (flags & GMX_FORCE_DO_LR)
 716     {
 717         donb_flags |= GMX_NONBONDED_DO_LR;
 718     }
 719
 720     kernel_data.flags  = donb_flags;
 721     kernel_data.lambda = lambda;
 722     kernel_data.dvdl   = dvdl_nb;
 723
 724     kernel_data.energygrp_elec = enerd->grpp.ener[egCOULSR];
 725     kernel_data.energygrp_vdw  = enerd->grpp.ener[egLJSR];
 726
 727     /* reset free energy components */
 728     for (i = 0; i < efptNR; i++)
 729     {
 730         dvdl_nb[i]  = 0;
 731     }
 732
 733     assert(gmx_omp_nthreads_get(emntNonbonded) == nbl_lists->nnbl);
 734
 735     wallcycle_sub_start(wcycle, ewcsNONBONDED);
 736 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 737     for (th = 0; th < nbl_lists->nnbl; th++)
 738     {
 739         gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 740                                   x, f, fr, mdatoms, &kernel_data, nrnb);
 741     }
 742
 743     if (fepvals->sc_alpha != 0)
 744     {
 745         enerd->dvdl_nonlin[efptVDW]  += dvdl_nb[efptVDW];
 746         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 747     }
 748     else
 749     {
 750         enerd->dvdl_lin[efptVDW]  += dvdl_nb[efptVDW];
 751         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 752     }
 753
 754     /* If we do foreign lambda and we have soft-core interactions
 755      * we have to recalculate the (non-linear) energies contributions.
 756      */
 757     if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
 758     {
 759         kernel_data.flags          = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE)) | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 760         kernel_data.lambda         = lam_i;
 761         kernel_data.energygrp_elec = enerd->foreign_grpp.ener[egCOULSR];
 762         kernel_data.energygrp_vdw  = enerd->foreign_grpp.ener[egLJSR];
 763         /* Note that we add to kernel_data.dvdl, but ignore the result */
 764
 765         for (i = 0; i < enerd->n_lambda; i++)
 766         {
 767             for (j = 0; j < efptNR; j++)
 768             {
 769                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
 770             }
 771             reset_foreign_enerdata(enerd);
 772 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 773             for (th = 0; th < nbl_lists->nnbl; th++)
 774             {
 775                 gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 776                                           x, f, fr, mdatoms, &kernel_data, nrnb);
 777             }
 778
 779             sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
 780             enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
 781         }
 782     }
 783
 784     wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 785 }
 786
 787 gmx_bool use_GPU(const nonbonded_verlet_t *nbv)
 788 {
 789     return nbv != NULL && nbv->bUseGPU;
 790 }
 791
 792 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
 793                          t_inputrec *inputrec,
 794                          gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 795                          gmx_localtop_t *top,
 796                          gmx_groups_t gmx_unused *groups,
 797                          matrix box, rvec x[], history_t *hist,
 798                          rvec f[],
 799                          tensor vir_force,
 800                          t_mdatoms *mdatoms,
 801                          gmx_enerdata_t *enerd, t_fcdata *fcd,
 802                          real *lambda, t_graph *graph,
 803                          t_forcerec *fr, interaction_const_t *ic,
 804                          gmx_vsite_t *vsite, rvec mu_tot,
 805                          double t, FILE *field, gmx_edsam_t ed,
 806                          gmx_bool bBornRadii,
 807                          int flags)
 808 {
 809     int                 cg0, cg1, i, j;
 810     int                 start, homenr;
 811     int                 nb_kernel_type;
 812     double              mu[2*DIM];
 813     gmx_bool            bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
 814     gmx_bool            bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
 815     gmx_bool            bDiffKernels = FALSE;
 816     matrix              boxs;
 817     rvec                vzero, box_diag;
 818     real                e, v, dvdl;
 819     float               cycles_pme, cycles_force, cycles_wait_gpu;
 820     nonbonded_verlet_t *nbv;
 821
 822     cycles_force    = 0;
 823     cycles_wait_gpu = 0;
 824     nbv             = fr->nbv;
 825     nb_kernel_type  = fr->nbv->grp[0].kernel_type;
 826
 827     start  = 0;
 828     homenr = mdatoms->homenr;
 829
 830     clear_mat(vir_force);
 831
 832     cg0 = 0;
 833     if (DOMAINDECOMP(cr))
 834     {
 835         cg1 = cr->dd->ncg_tot;
 836     }
 837     else
 838     {
 839         cg1 = top->cgs.nr;
 840     }
 841     if (fr->n_tpi > 0)
 842     {
 843         cg1--;
 844     }
 845
 846     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 847     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
 848     bFillGrid     = (bNS && bStateChanged);
 849     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 850     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 851     bDoForces     = (flags & GMX_FORCE_FORCES);
 852     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 853     bUseGPU       = fr->nbv->bUseGPU;
 854     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
 855
 856     if (bStateChanged)
 857     {
 858         update_forcerec(fr, box);
 859
 860         if (NEED_MUTOT(*inputrec))
 861         {
 862             /* Calculate total (local) dipole moment in a temporary common array.
 863              * This makes it possible to sum them over nodes faster.
 864              */
 865             calc_mu(start, homenr,
 866                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 867                     mu, mu+DIM);
 868         }
 869     }
 870
 871     if (fr->ePBC != epbcNONE)
 872     {
 873         /* Compute shift vectors every step,
 874          * because of pressure coupling or box deformation!
 875          */
 876         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 877         {
 878             calc_shifts(box, fr->shift_vec);
 879         }
 880
 881         if (bCalcCGCM)
 882         {
 883             put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
 884             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 885         }
 886         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 887         {
 888             unshift_self(graph, box, x);
 889         }
 890     }
 891
 892     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 893                                  fr->shift_vec, nbv->grp[0].nbat);
 894
 895 #ifdef GMX_MPI
 896     if (!(cr->duty & DUTY_PME))
 897     {
 898         /* Send particle coordinates to the pme nodes.
 899          * Since this is only implemented for domain decomposition
 900          * and domain decomposition does not use the graph,
 901          * we do not need to worry about shifting.
 902          */
 903
 904         int pme_flags = 0;
 905
 906         wallcycle_start(wcycle, ewcPP_PMESENDX);
 907
 908         bBS = (inputrec->nwall == 2);
 909         if (bBS)
 910         {
 911             copy_mat(box, boxs);
 912             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
 913         }
 914
 915         if (EEL_PME(fr->eeltype))
 916         {
 917             pme_flags |= GMX_PME_DO_COULOMB;
 918         }
 919
 920         if (EVDW_PME(fr->vdwtype))
 921         {
 922             pme_flags |= GMX_PME_DO_LJ;
 923         }
 924
 925         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
 926                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
 927                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
 928                                  pme_flags, step);
 929
 930         wallcycle_stop(wcycle, ewcPP_PMESENDX);
 931     }
 932 #endif /* GMX_MPI */
 933
 934     /* do gridding for pair search */
 935     if (bNS)
 936     {
 937         if (graph && bStateChanged)
 938         {
 939             /* Calculate intramolecular shift vectors to make molecules whole */
 940             mk_mshift(fplog, graph, fr->ePBC, box, x);
 941         }
 942
 943         clear_rvec(vzero);
 944         box_diag[XX] = box[XX][XX];
 945         box_diag[YY] = box[YY][YY];
 946         box_diag[ZZ] = box[ZZ][ZZ];
 947
 948         wallcycle_start(wcycle, ewcNS);
 949         if (!fr->bDomDec)
 950         {
 951             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 952             nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
 953                               0, vzero, box_diag,
 954                               0, mdatoms->homenr, -1, fr->cginfo, x,
 955                               0, NULL,
 956                               nbv->grp[eintLocal].kernel_type,
 957                               nbv->grp[eintLocal].nbat);
 958             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 959         }
 960         else
 961         {
 962             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 963             nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
 964                                        fr->cginfo, x,
 965                                        nbv->grp[eintNonlocal].kernel_type,
 966                                        nbv->grp[eintNonlocal].nbat);
 967             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 968         }
 969
 970         if (nbv->ngrp == 1 ||
 971             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 972         {
 973             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
 974                                nbv->nbs, mdatoms, fr->cginfo);
 975         }
 976         else
 977         {
 978             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
 979                                nbv->nbs, mdatoms, fr->cginfo);
 980             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
 981                                nbv->nbs, mdatoms, fr->cginfo);
 982         }
 983         wallcycle_stop(wcycle, ewcNS);
 984     }
 985
 986     /* initialize the GPU atom data and copy shift vector */
 987     if (bUseGPU)
 988     {
 989         if (bNS)
 990         {
 991             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 992             nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 993             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 994         }
 995
 996         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 997         nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 998         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 999     }
1000
1001     /* do local pair search */
1002     if (bNS)
1003     {
1004         wallcycle_start_nocount(wcycle, ewcNS);
1005         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1006         nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
1007                             &top->excls,
1008                             ic->rlist,
1009                             nbv->min_ci_balanced,
1010                             &nbv->grp[eintLocal].nbl_lists,
1011                             eintLocal,
1012                             nbv->grp[eintLocal].kernel_type,
1013                             nrnb);
1014         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1015
1016         if (bUseGPU)
1017         {
1018             /* initialize local pair-list on the GPU */
1019             nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1020                                      nbv->grp[eintLocal].nbl_lists.nbl[0],
1021                                      eintLocal);
1022         }
1023         wallcycle_stop(wcycle, ewcNS);
1024     }
1025     else
1026     {
1027         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1028         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1029         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
1030                                         nbv->grp[eintLocal].nbat);
1031         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1032         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1033     }
1034
1035     if (bUseGPU)
1036     {
1037         wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1038         /* launch local nonbonded F on GPU */
1039         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
1040                      nrnb, wcycle);
1041         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1042     }
1043
1044     /* Communicate coordinates and sum dipole if necessary +
1045        do non-local pair search */
1046     if (DOMAINDECOMP(cr))
1047     {
1048         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
1049                         nbv->grp[eintLocal].kernel_type);
1050
1051         if (bDiffKernels)
1052         {
1053             /* With GPU+CPU non-bonded calculations we need to copy
1054              * the local coordinates to the non-local nbat struct
1055              * (in CPU format) as the non-local kernel call also
1056              * calculates the local - non-local interactions.
1057              */
1058             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1059             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1060             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
1061                                             nbv->grp[eintNonlocal].nbat);
1062             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1063             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1064         }
1065
1066         if (bNS)
1067         {
1068             wallcycle_start_nocount(wcycle, ewcNS);
1069             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1070
1071             if (bDiffKernels)
1072             {
1073                 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
1074             }
1075
1076             nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
1077                                 &top->excls,
1078                                 ic->rlist,
1079                                 nbv->min_ci_balanced,
1080                                 &nbv->grp[eintNonlocal].nbl_lists,
1081                                 eintNonlocal,
1082                                 nbv->grp[eintNonlocal].kernel_type,
1083                                 nrnb);
1084
1085             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1086
1087             if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_CUDA)
1088             {
1089                 /* initialize non-local pair-list on the GPU */
1090                 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1091                                          nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1092                                          eintNonlocal);
1093             }
1094             wallcycle_stop(wcycle, ewcNS);
1095         }
1096         else
1097         {
1098             wallcycle_start(wcycle, ewcMOVEX);
1099             dd_move_x(cr->dd, box, x);
1100
1101             /* When we don't need the total dipole we sum it in global_stat */
1102             if (bStateChanged && NEED_MUTOT(*inputrec))
1103             {
1104                 gmx_sumd(2*DIM, mu, cr);
1105             }
1106             wallcycle_stop(wcycle, ewcMOVEX);
1107
1108             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1109             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1110             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1111                                             nbv->grp[eintNonlocal].nbat);
1112             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1113             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1114         }
1115
1116         if (bUseGPU && !bDiffKernels)
1117         {
1118             wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1119             /* launch non-local nonbonded F on GPU */
1120             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1121                          nrnb, wcycle);
1122             cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1123         }
1124     }
1125
1126     if (bUseGPU)
1127     {
1128         /* launch D2H copy-back F */
1129         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1130         if (DOMAINDECOMP(cr) && !bDiffKernels)
1131         {
1132             nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1133                                       flags, eatNonlocal);
1134         }
1135         nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1136                                   flags, eatLocal);
1137         cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1138     }
1139
1140     if (bStateChanged && NEED_MUTOT(*inputrec))
1141     {
1142         if (PAR(cr))
1143         {
1144             gmx_sumd(2*DIM, mu, cr);
1145         }
1146
1147         for (i = 0; i < 2; i++)
1148         {
1149             for (j = 0; j < DIM; j++)
1150             {
1151                 fr->mu_tot[i][j] = mu[i*DIM + j];
1152             }
1153         }
1154     }
1155     if (fr->efep == efepNO)
1156     {
1157         copy_rvec(fr->mu_tot[0], mu_tot);
1158     }
1159     else
1160     {
1161         for (j = 0; j < DIM; j++)
1162         {
1163             mu_tot[j] =
1164                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1165                 lambda[efptCOUL]*fr->mu_tot[1][j];
1166         }
1167     }
1168
1169     /* Reset energies */
1170     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1171     clear_rvecs(SHIFTS, fr->fshift);
1172
1173     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1174     {
1175         wallcycle_start(wcycle, ewcPPDURINGPME);
1176         dd_force_flop_start(cr->dd, nrnb);
1177     }
1178
1179     if (inputrec->bRot)
1180     {
1181         /* Enforced rotation has its own cycle counter that starts after the collective
1182          * coordinates have been communicated. It is added to ddCyclF to allow
1183          * for proper load-balancing */
1184         wallcycle_start(wcycle, ewcROT);
1185         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1186         wallcycle_stop(wcycle, ewcROT);
1187     }
1188
1189     /* Start the force cycle counter.
1190      * This counter is stopped in do_forcelow_level.
1191      * No parallel communication should occur while this counter is running,
1192      * since that will interfere with the dynamic load balancing.
1193      */
1194     wallcycle_start(wcycle, ewcFORCE);
1195     if (bDoForces)
1196     {
1197         /* Reset forces for which the virial is calculated separately:
1198          * PME/Ewald forces if necessary */
1199         if (fr->bF_NoVirSum)
1200         {
1201             if (flags & GMX_FORCE_VIRIAL)
1202             {
1203                 fr->f_novirsum = fr->f_novirsum_alloc;
1204                 if (fr->bDomDec)
1205                 {
1206                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1207                 }
1208                 else
1209                 {
1210                     clear_rvecs(homenr, fr->f_novirsum+start);
1211                 }
1212             }
1213             else
1214             {
1215                 /* We are not calculating the pressure so we do not need
1216                  * a separate array for forces that do not contribute
1217                  * to the pressure.
1218                  */
1219                 fr->f_novirsum = f;
1220             }
1221         }
1222
1223         /* Clear the short- and long-range forces */
1224         clear_rvecs(fr->natoms_force_constr, f);
1225         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1226         {
1227             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1228         }
1229
1230         clear_rvec(fr->vir_diag_posres);
1231     }
1232
1233     if (inputrec->ePull == epullCONSTRAINT)
1234     {
1235         clear_pull_forces(inputrec->pull);
1236     }
1237
1238     /* We calculate the non-bonded forces, when done on the CPU, here.
1239      * We do this before calling do_force_lowlevel, because in that
1240      * function, the listed forces are calculated before PME, which
1241      * does communication.  With this order, non-bonded and listed
1242      * force calculation imbalance can be balanced out by the domain
1243      * decomposition load balancing.
1244      */
1245
1246     if (!bUseOrEmulGPU)
1247     {
1248         /* Maybe we should move this into do_force_lowlevel */
1249         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1250                      nrnb, wcycle);
1251     }
1252
1253     if (fr->efep != efepNO)
1254     {
1255         /* Calculate the local and non-local free energy interactions here.
1256          * Happens here on the CPU both with and without GPU.
1257          */
1258         if (fr->nbv->grp[eintLocal].nbl_lists.nbl_fep[0]->nrj > 0)
1259         {
1260             do_nb_verlet_fep(&fr->nbv->grp[eintLocal].nbl_lists,
1261                              fr, x, f, mdatoms,
1262                              inputrec->fepvals, lambda,
1263                              enerd, flags, nrnb, wcycle);
1264         }
1265
1266         if (DOMAINDECOMP(cr) &&
1267             fr->nbv->grp[eintNonlocal].nbl_lists.nbl_fep[0]->nrj > 0)
1268         {
1269             do_nb_verlet_fep(&fr->nbv->grp[eintNonlocal].nbl_lists,
1270                              fr, x, f, mdatoms,
1271                              inputrec->fepvals, lambda,
1272                              enerd, flags, nrnb, wcycle);
1273         }
1274     }
1275
1276     if (!bUseOrEmulGPU || bDiffKernels)
1277     {
1278         int aloc;
1279
1280         if (DOMAINDECOMP(cr))
1281         {
1282             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1283                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1284                          nrnb, wcycle);
1285         }
1286
1287         if (!bUseOrEmulGPU)
1288         {
1289             aloc = eintLocal;
1290         }
1291         else
1292         {
1293             aloc = eintNonlocal;
1294         }
1295
1296         /* Add all the non-bonded force to the normal force array.
1297          * This can be split into a local a non-local part when overlapping
1298          * communication with calculation with domain decomposition.
1299          */
1300         cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1301         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1302         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1303         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1304         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1305         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1306         wallcycle_start_nocount(wcycle, ewcFORCE);
1307
1308         /* if there are multiple fshift output buffers reduce them */
1309         if ((flags & GMX_FORCE_VIRIAL) &&
1310             nbv->grp[aloc].nbl_lists.nnbl > 1)
1311         {
1312             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1313                                                      fr->fshift);
1314         }
1315     }
1316
1317     /* update QMMMrec, if necessary */
1318     if (fr->bQMMM)
1319     {
1320         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1321     }
1322
1323     if ((flags & GMX_FORCE_LISTED) && top->idef.il[F_POSRES].nr > 0)
1324     {
1325         posres_wrapper(flags, inputrec, nrnb, top, box, x,
1326                        enerd, lambda, fr);
1327     }
1328
1329     if ((flags & GMX_FORCE_LISTED) && top->idef.il[F_FBPOSRES].nr > 0)
1330     {
1331         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1332     }
1333
1334     /* Compute the bonded and non-bonded energies and optionally forces */
1335     do_force_lowlevel(fr, inputrec, &(top->idef),
1336                       cr, nrnb, wcycle, mdatoms,
1337                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1338                       bBornRadii, box,
1339                       inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1340                       flags, &cycles_pme);
1341
1342     if (bSepLRF)
1343     {
1344         if (do_per_step(step, inputrec->nstcalclr))
1345         {
1346             /* Add the long range forces to the short range forces */
1347             for (i = 0; i < fr->natoms_force_constr; i++)
1348             {
1349                 rvec_add(fr->f_twin[i], f[i], f[i]);
1350             }
1351         }
1352     }
1353
1354     cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1355
1356     if (ed)
1357     {
1358         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1359     }
1360
1361     if (bUseOrEmulGPU && !bDiffKernels)
1362     {
1363         /* wait for non-local forces (or calculate in emulation mode) */
1364         if (DOMAINDECOMP(cr))
1365         {
1366             if (bUseGPU)
1367             {
1368                 float cycles_tmp;
1369
1370                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1371                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1372                                     nbv->grp[eintNonlocal].nbat,
1373                                     flags, eatNonlocal,
1374                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1375                                     fr->fshift);
1376                 cycles_tmp       = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1377                 cycles_wait_gpu += cycles_tmp;
1378                 cycles_force    += cycles_tmp;
1379             }
1380             else
1381             {
1382                 wallcycle_start_nocount(wcycle, ewcFORCE);
1383                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1384                              nrnb, wcycle);
1385                 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1386             }
1387             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1388             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1389             /* skip the reduction if there was no non-local work to do */
1390             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1391             {
1392                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1393                                                nbv->grp[eintNonlocal].nbat, f);
1394             }
1395             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1396             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1397         }
1398     }
1399
1400     if (bDoForces && DOMAINDECOMP(cr))
1401     {
1402         /* Communicate the forces */
1403         wallcycle_start(wcycle, ewcMOVEF);
1404         dd_move_f(cr->dd, f, fr->fshift);
1405         /* Do we need to communicate the separate force array
1406          * for terms that do not contribute to the single sum virial?
1407          * Position restraints and electric fields do not introduce
1408          * inter-cg forces, only full electrostatics methods do.
1409          * When we do not calculate the virial, fr->f_novirsum = f,
1410          * so we have already communicated these forces.
1411          */
1412         if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1413             (flags & GMX_FORCE_VIRIAL))
1414         {
1415             dd_move_f(cr->dd, fr->f_novirsum, NULL);
1416         }
1417         if (bSepLRF)
1418         {
1419             /* We should not update the shift forces here,
1420              * since f_twin is already included in f.
1421              */
1422             dd_move_f(cr->dd, fr->f_twin, NULL);
1423         }
1424         wallcycle_stop(wcycle, ewcMOVEF);
1425     }
1426
1427     if (bUseOrEmulGPU)
1428     {
1429         /* wait for local forces (or calculate in emulation mode) */
1430         if (bUseGPU)
1431         {
1432             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1433             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1434                                 nbv->grp[eintLocal].nbat,
1435                                 flags, eatLocal,
1436                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1437                                 fr->fshift);
1438             cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1439
1440             /* now clear the GPU outputs while we finish the step on the CPU */
1441
1442             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1443             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1444             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1445         }
1446         else
1447         {
1448             wallcycle_start_nocount(wcycle, ewcFORCE);
1449             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1450                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1451                          nrnb, wcycle);
1452             wallcycle_stop(wcycle, ewcFORCE);
1453         }
1454         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1455         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1456         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1457         {
1458             /* skip the reduction if there was no non-local work to do */
1459             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1460                                            nbv->grp[eintLocal].nbat, f);
1461         }
1462         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1463         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1464     }
1465
1466     if (DOMAINDECOMP(cr))
1467     {
1468         dd_force_flop_stop(cr->dd, nrnb);
1469         if (wcycle)
1470         {
1471             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1472             if (bUseGPU)
1473             {
1474                 dd_cycles_add(cr->dd, cycles_wait_gpu, ddCyclWaitGPU);
1475             }
1476         }
1477     }
1478
1479     if (bDoForces)
1480     {
1481         if (IR_ELEC_FIELD(*inputrec))
1482         {
1483             /* Compute forces due to electric field */
1484             calc_f_el(MASTER(cr) ? field : NULL,
1485                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1486                       inputrec->ex, inputrec->et, t);
1487         }
1488
1489         /* If we have NoVirSum forces, but we do not calculate the virial,
1490          * we sum fr->f_novirum=f later.
1491          */
1492         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1493         {
1494             wallcycle_start(wcycle, ewcVSITESPREAD);
1495             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1496                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1497             wallcycle_stop(wcycle, ewcVSITESPREAD);
1498
1499             if (bSepLRF)
1500             {
1501                 wallcycle_start(wcycle, ewcVSITESPREAD);
1502                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1503                                nrnb,
1504                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1505                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1506             }
1507         }
1508
1509         if (flags & GMX_FORCE_VIRIAL)
1510         {
1511             /* Calculation of the virial must be done after vsites! */
1512             calc_virial(0, mdatoms->homenr, x, f,
1513                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1514         }
1515     }
1516
1517     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1518     {
1519         /* Since the COM pulling is always done mass-weighted, no forces are
1520          * applied to vsites and this call can be done after vsite spreading.
1521          */
1522         pull_potential_wrapper(cr, inputrec, box, x,
1523                                f, vir_force, mdatoms, enerd, lambda, t,
1524                                wcycle);
1525     }
1526
1527     /* Add the forces from enforced rotation potentials (if any) */
1528     if (inputrec->bRot)
1529     {
1530         wallcycle_start(wcycle, ewcROTadd);
1531         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1532         wallcycle_stop(wcycle, ewcROTadd);
1533     }
1534
1535     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
1536     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
1537
1538     if (PAR(cr) && !(cr->duty & DUTY_PME))
1539     {
1540         /* In case of node-splitting, the PP nodes receive the long-range
1541          * forces, virial and energy from the PME nodes here.
1542          */
1543         pme_receive_force_ener(cr, wcycle, enerd, fr);
1544     }
1545
1546     if (bDoForces)
1547     {
1548         post_process_forces(cr, step, nrnb, wcycle,
1549                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1550                             flags);
1551     }
1552
1553     /* Sum the potential energy terms from group contributions */
1554     sum_epot(&(enerd->grpp), enerd->term);
1555 }
1556
1557 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1558                         t_inputrec *inputrec,
1559                         gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1560                         gmx_localtop_t *top,
1561                         gmx_groups_t *groups,
1562                         matrix box, rvec x[], history_t *hist,
1563                         rvec f[],
1564                         tensor vir_force,
1565                         t_mdatoms *mdatoms,
1566                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1567                         real *lambda, t_graph *graph,
1568                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1569                         double t, FILE *field, gmx_edsam_t ed,
1570                         gmx_bool bBornRadii,
1571                         int flags)
1572 {
1573     int        cg0, cg1, i, j;
1574     int        start, homenr;
1575     double     mu[2*DIM];
1576     gmx_bool   bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1577     gmx_bool   bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1578     gmx_bool   bDoAdressWF;
1579     matrix     boxs;
1580     rvec       vzero, box_diag;
1581     real       e, v, dvdlambda[efptNR];
1582     t_pbc      pbc;
1583     float      cycles_pme, cycles_force;
1584
1585     start  = 0;
1586     homenr = mdatoms->homenr;
1587
1588     clear_mat(vir_force);
1589
1590     cg0 = 0;
1591     if (DOMAINDECOMP(cr))
1592     {
1593         cg1 = cr->dd->ncg_tot;
1594     }
1595     else
1596     {
1597         cg1 = top->cgs.nr;
1598     }
1599     if (fr->n_tpi > 0)
1600     {
1601         cg1--;
1602     }
1603
1604     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1605     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1606     /* Should we update the long-range neighborlists at this step? */
1607     bDoLongRangeNS = fr->bTwinRange && bNS;
1608     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1609     bFillGrid      = (bNS && bStateChanged);
1610     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1611     bDoForces      = (flags & GMX_FORCE_FORCES);
1612     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1613     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1614                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1615
1616     /* should probably move this to the forcerec since it doesn't change */
1617     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1618
1619     if (bStateChanged)
1620     {
1621         update_forcerec(fr, box);
1622
1623         if (NEED_MUTOT(*inputrec))
1624         {
1625             /* Calculate total (local) dipole moment in a temporary common array.
1626              * This makes it possible to sum them over nodes faster.
1627              */
1628             calc_mu(start, homenr,
1629                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1630                     mu, mu+DIM);
1631         }
1632     }
1633
1634     if (fr->ePBC != epbcNONE)
1635     {
1636         /* Compute shift vectors every step,
1637          * because of pressure coupling or box deformation!
1638          */
1639         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1640         {
1641             calc_shifts(box, fr->shift_vec);
1642         }
1643
1644         if (bCalcCGCM)
1645         {
1646             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1647                                      &(top->cgs), x, fr->cg_cm);
1648             inc_nrnb(nrnb, eNR_CGCM, homenr);
1649             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1650         }
1651         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1652         {
1653             unshift_self(graph, box, x);
1654         }
1655     }
1656     else if (bCalcCGCM)
1657     {
1658         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1659         inc_nrnb(nrnb, eNR_CGCM, homenr);
1660     }
1661
1662     if (bCalcCGCM && gmx_debug_at)
1663     {
1664         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1665     }
1666
1667 #ifdef GMX_MPI
1668     if (!(cr->duty & DUTY_PME))
1669     {
1670         /* Send particle coordinates to the pme nodes.
1671          * Since this is only implemented for domain decomposition
1672          * and domain decomposition does not use the graph,
1673          * we do not need to worry about shifting.
1674          */
1675
1676         int pme_flags = 0;
1677
1678         wallcycle_start(wcycle, ewcPP_PMESENDX);
1679
1680         bBS = (inputrec->nwall == 2);
1681         if (bBS)
1682         {
1683             copy_mat(box, boxs);
1684             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1685         }
1686
1687         if (EEL_PME(fr->eeltype))
1688         {
1689             pme_flags |= GMX_PME_DO_COULOMB;
1690         }
1691
1692         if (EVDW_PME(fr->vdwtype))
1693         {
1694             pme_flags |= GMX_PME_DO_LJ;
1695         }
1696
1697         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
1698                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
1699                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
1700                                  pme_flags, step);
1701
1702         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1703     }
1704 #endif /* GMX_MPI */
1705
1706     /* Communicate coordinates and sum dipole if necessary */
1707     if (DOMAINDECOMP(cr))
1708     {
1709         wallcycle_start(wcycle, ewcMOVEX);
1710         dd_move_x(cr->dd, box, x);
1711         wallcycle_stop(wcycle, ewcMOVEX);
1712     }
1713
1714     /* update adress weight beforehand */
1715     if (bStateChanged && bDoAdressWF)
1716     {
1717         /* need pbc for adress weight calculation with pbc_dx */
1718         set_pbc(&pbc, inputrec->ePBC, box);
1719         if (fr->adress_site == eAdressSITEcog)
1720         {
1721             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1722                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1723         }
1724         else if (fr->adress_site == eAdressSITEcom)
1725         {
1726             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1727                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1728         }
1729         else if (fr->adress_site == eAdressSITEatomatom)
1730         {
1731             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1732                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1733         }
1734         else
1735         {
1736             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1737                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1738         }
1739     }
1740
1741     if (NEED_MUTOT(*inputrec))
1742     {
1743
1744         if (bStateChanged)
1745         {
1746             if (PAR(cr))
1747             {
1748                 gmx_sumd(2*DIM, mu, cr);
1749             }
1750             for (i = 0; i < 2; i++)
1751             {
1752                 for (j = 0; j < DIM; j++)
1753                 {
1754                     fr->mu_tot[i][j] = mu[i*DIM + j];
1755                 }
1756             }
1757         }
1758         if (fr->efep == efepNO)
1759         {
1760             copy_rvec(fr->mu_tot[0], mu_tot);
1761         }
1762         else
1763         {
1764             for (j = 0; j < DIM; j++)
1765             {
1766                 mu_tot[j] =
1767                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1768             }
1769         }
1770     }
1771
1772     /* Reset energies */
1773     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1774     clear_rvecs(SHIFTS, fr->fshift);
1775
1776     if (bNS)
1777     {
1778         wallcycle_start(wcycle, ewcNS);
1779
1780         if (graph && bStateChanged)
1781         {
1782             /* Calculate intramolecular shift vectors to make molecules whole */
1783             mk_mshift(fplog, graph, fr->ePBC, box, x);
1784         }
1785
1786         /* Do the actual neighbour searching */
1787         ns(fplog, fr, box,
1788            groups, top, mdatoms,
1789            cr, nrnb, bFillGrid,
1790            bDoLongRangeNS);
1791
1792         wallcycle_stop(wcycle, ewcNS);
1793     }
1794
1795     if (inputrec->implicit_solvent && bNS)
1796     {
1797         make_gb_nblist(cr, inputrec->gb_algorithm,
1798                        x, box, fr, &top->idef, graph, fr->born);
1799     }
1800
1801     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1802     {
1803         wallcycle_start(wcycle, ewcPPDURINGPME);
1804         dd_force_flop_start(cr->dd, nrnb);
1805     }
1806
1807     if (inputrec->bRot)
1808     {
1809         /* Enforced rotation has its own cycle counter that starts after the collective
1810          * coordinates have been communicated. It is added to ddCyclF to allow
1811          * for proper load-balancing */
1812         wallcycle_start(wcycle, ewcROT);
1813         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1814         wallcycle_stop(wcycle, ewcROT);
1815     }
1816
1817     /* Start the force cycle counter.
1818      * This counter is stopped in do_forcelow_level.
1819      * No parallel communication should occur while this counter is running,
1820      * since that will interfere with the dynamic load balancing.
1821      */
1822     wallcycle_start(wcycle, ewcFORCE);
1823
1824     if (bDoForces)
1825     {
1826         /* Reset forces for which the virial is calculated separately:
1827          * PME/Ewald forces if necessary */
1828         if (fr->bF_NoVirSum)
1829         {
1830             if (flags & GMX_FORCE_VIRIAL)
1831             {
1832                 fr->f_novirsum = fr->f_novirsum_alloc;
1833                 if (fr->bDomDec)
1834                 {
1835                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1836                 }
1837                 else
1838                 {
1839                     clear_rvecs(homenr, fr->f_novirsum+start);
1840                 }
1841             }
1842             else
1843             {
1844                 /* We are not calculating the pressure so we do not need
1845                  * a separate array for forces that do not contribute
1846                  * to the pressure.
1847                  */
1848                 fr->f_novirsum = f;
1849             }
1850         }
1851
1852         /* Clear the short- and long-range forces */
1853         clear_rvecs(fr->natoms_force_constr, f);
1854         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1855         {
1856             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1857         }
1858
1859         clear_rvec(fr->vir_diag_posres);
1860     }
1861     if (inputrec->ePull == epullCONSTRAINT)
1862     {
1863         clear_pull_forces(inputrec->pull);
1864     }
1865
1866     /* update QMMMrec, if necessary */
1867     if (fr->bQMMM)
1868     {
1869         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1870     }
1871
1872     if ((flags & GMX_FORCE_LISTED) && top->idef.il[F_POSRES].nr > 0)
1873     {
1874         posres_wrapper(flags, inputrec, nrnb, top, box, x,
1875                        enerd, lambda, fr);
1876     }
1877
1878     if ((flags & GMX_FORCE_LISTED) && top->idef.il[F_FBPOSRES].nr > 0)
1879     {
1880         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1881     }
1882
1883     /* Compute the bonded and non-bonded energies and optionally forces */
1884     do_force_lowlevel(fr, inputrec, &(top->idef),
1885                       cr, nrnb, wcycle, mdatoms,
1886                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1887                       bBornRadii, box,
1888                       inputrec->fepvals, lambda,
1889                       graph, &(top->excls), fr->mu_tot,
1890                       flags,
1891                       &cycles_pme);
1892
1893     if (bSepLRF)
1894     {
1895         if (do_per_step(step, inputrec->nstcalclr))
1896         {
1897             /* Add the long range forces to the short range forces */
1898             for (i = 0; i < fr->natoms_force_constr; i++)
1899             {
1900                 rvec_add(fr->f_twin[i], f[i], f[i]);
1901             }
1902         }
1903     }
1904
1905     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1906
1907     if (ed)
1908     {
1909         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1910     }
1911
1912     if (DOMAINDECOMP(cr))
1913     {
1914         dd_force_flop_stop(cr->dd, nrnb);
1915         if (wcycle)
1916         {
1917             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1918         }
1919     }
1920
1921     if (bDoForces)
1922     {
1923         if (IR_ELEC_FIELD(*inputrec))
1924         {
1925             /* Compute forces due to electric field */
1926             calc_f_el(MASTER(cr) ? field : NULL,
1927                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1928                       inputrec->ex, inputrec->et, t);
1929         }
1930
1931         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1932         {
1933             /* Compute thermodynamic force in hybrid AdResS region */
1934             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1935                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1936         }
1937
1938         /* Communicate the forces */
1939         if (DOMAINDECOMP(cr))
1940         {
1941             wallcycle_start(wcycle, ewcMOVEF);
1942             dd_move_f(cr->dd, f, fr->fshift);
1943             /* Do we need to communicate the separate force array
1944              * for terms that do not contribute to the single sum virial?
1945              * Position restraints and electric fields do not introduce
1946              * inter-cg forces, only full electrostatics methods do.
1947              * When we do not calculate the virial, fr->f_novirsum = f,
1948              * so we have already communicated these forces.
1949              */
1950             if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1951                 (flags & GMX_FORCE_VIRIAL))
1952             {
1953                 dd_move_f(cr->dd, fr->f_novirsum, NULL);
1954             }
1955             if (bSepLRF)
1956             {
1957                 /* We should not update the shift forces here,
1958                  * since f_twin is already included in f.
1959                  */
1960                 dd_move_f(cr->dd, fr->f_twin, NULL);
1961             }
1962             wallcycle_stop(wcycle, ewcMOVEF);
1963         }
1964
1965         /* If we have NoVirSum forces, but we do not calculate the virial,
1966          * we sum fr->f_novirum=f later.
1967          */
1968         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1969         {
1970             wallcycle_start(wcycle, ewcVSITESPREAD);
1971             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1972                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1973             wallcycle_stop(wcycle, ewcVSITESPREAD);
1974
1975             if (bSepLRF)
1976             {
1977                 wallcycle_start(wcycle, ewcVSITESPREAD);
1978                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1979                                nrnb,
1980                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1981                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1982             }
1983         }
1984
1985         if (flags & GMX_FORCE_VIRIAL)
1986         {
1987             /* Calculation of the virial must be done after vsites! */
1988             calc_virial(0, mdatoms->homenr, x, f,
1989                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1990         }
1991     }
1992
1993     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1994     {
1995         pull_potential_wrapper(cr, inputrec, box, x,
1996                                f, vir_force, mdatoms, enerd, lambda, t,
1997                                wcycle);
1998     }
1999
2000     /* Add the forces from enforced rotation potentials (if any) */
2001     if (inputrec->bRot)
2002     {
2003         wallcycle_start(wcycle, ewcROTadd);
2004         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
2005         wallcycle_stop(wcycle, ewcROTadd);
2006     }
2007
2008     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
2009     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
2010
2011     if (PAR(cr) && !(cr->duty & DUTY_PME))
2012     {
2013         /* In case of node-splitting, the PP nodes receive the long-range
2014          * forces, virial and energy from the PME nodes here.
2015          */
2016         pme_receive_force_ener(cr, wcycle, enerd, fr);
2017     }
2018
2019     if (bDoForces)
2020     {
2021         post_process_forces(cr, step, nrnb, wcycle,
2022                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
2023                             flags);
2024     }
2025
2026     /* Sum the potential energy terms from group contributions */
2027     sum_epot(&(enerd->grpp), enerd->term);
2028 }
2029
2030 void do_force(FILE *fplog, t_commrec *cr,
2031               t_inputrec *inputrec,
2032               gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
2033               gmx_localtop_t *top,
2034               gmx_groups_t *groups,
2035               matrix box, rvec x[], history_t *hist,
2036               rvec f[],
2037               tensor vir_force,
2038               t_mdatoms *mdatoms,
2039               gmx_enerdata_t *enerd, t_fcdata *fcd,
2040               real *lambda, t_graph *graph,
2041               t_forcerec *fr,
2042               gmx_vsite_t *vsite, rvec mu_tot,
2043               double t, FILE *field, gmx_edsam_t ed,
2044               gmx_bool bBornRadii,
2045               int flags)
2046 {
2047     /* modify force flag if not doing nonbonded */
2048     if (!fr->bNonbonded)
2049     {
2050         flags &= ~GMX_FORCE_NONBONDED;
2051     }
2052
2053     switch (inputrec->cutoff_scheme)
2054     {
2055         case ecutsVERLET:
2056             do_force_cutsVERLET(fplog, cr, inputrec,
2057                                 step, nrnb, wcycle,
2058                                 top,
2059                                 groups,
2060                                 box, x, hist,
2061                                 f, vir_force,
2062                                 mdatoms,
2063                                 enerd, fcd,
2064                                 lambda, graph,
2065                                 fr, fr->ic,
2066                                 vsite, mu_tot,
2067                                 t, field, ed,
2068                                 bBornRadii,
2069                                 flags);
2070             break;
2071         case ecutsGROUP:
2072             do_force_cutsGROUP(fplog, cr, inputrec,
2073                                step, nrnb, wcycle,
2074                                top,
2075                                groups,
2076                                box, x, hist,
2077                                f, vir_force,
2078                                mdatoms,
2079                                enerd, fcd,
2080                                lambda, graph,
2081                                fr, vsite, mu_tot,
2082                                t, field, ed,
2083                                bBornRadii,
2084                                flags);
2085             break;
2086         default:
2087             gmx_incons("Invalid cut-off scheme passed!");
2088     }
2089 }
2090
2091
2092 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2093                         t_inputrec *ir, t_mdatoms *md,
2094                         t_state *state, t_commrec *cr, t_nrnb *nrnb,
2095                         t_forcerec *fr, gmx_localtop_t *top)
2096 {
2097     int             i, m, start, end;
2098     gmx_int64_t     step;
2099     real            dt = ir->delta_t;
2100     real            dvdl_dum;
2101     rvec           *savex;
2102
2103     snew(savex, state->natoms);
2104
2105     start = 0;
2106     end   = md->homenr;
2107
2108     if (debug)
2109     {
2110         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2111                 start, md->homenr, end);
2112     }
2113     /* Do a first constrain to reset particles... */
2114     step = ir->init_step;
2115     if (fplog)
2116     {
2117         char buf[STEPSTRSIZE];
2118         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2119                 gmx_step_str(step, buf));
2120     }
2121     dvdl_dum = 0;
2122
2123     /* constrain the current position */
2124     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2125               ir, NULL, cr, step, 0, 1.0, md,
2126               state->x, state->x, NULL,
2127               fr->bMolPBC, state->box,
2128               state->lambda[efptBONDED], &dvdl_dum,
2129               NULL, NULL, nrnb, econqCoord,
2130               ir->epc == epcMTTK, state->veta, state->veta);
2131     if (EI_VV(ir->eI))
2132     {
2133         /* constrain the inital velocity, and save it */
2134         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2135         /* might not yet treat veta correctly */
2136         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2137                   ir, NULL, cr, step, 0, 1.0, md,
2138                   state->x, state->v, state->v,
2139                   fr->bMolPBC, state->box,
2140                   state->lambda[efptBONDED], &dvdl_dum,
2141                   NULL, NULL, nrnb, econqVeloc,
2142                   ir->epc == epcMTTK, state->veta, state->veta);
2143     }
2144     /* constrain the inital velocities at t-dt/2 */
2145     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2146     {
2147         for (i = start; (i < end); i++)
2148         {
2149             for (m = 0; (m < DIM); m++)
2150             {
2151                 /* Reverse the velocity */
2152                 state->v[i][m] = -state->v[i][m];
2153                 /* Store the position at t-dt in buf */
2154                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2155             }
2156         }
2157         /* Shake the positions at t=-dt with the positions at t=0
2158          * as reference coordinates.
2159          */
2160         if (fplog)
2161         {
2162             char buf[STEPSTRSIZE];
2163             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2164                     gmx_step_str(step, buf));
2165         }
2166         dvdl_dum = 0;
2167         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2168                   ir, NULL, cr, step, -1, 1.0, md,
2169                   state->x, savex, NULL,
2170                   fr->bMolPBC, state->box,
2171                   state->lambda[efptBONDED], &dvdl_dum,
2172                   state->v, NULL, nrnb, econqCoord,
2173                   ir->epc == epcMTTK, state->veta, state->veta);
2174
2175         for (i = start; i < end; i++)
2176         {
2177             for (m = 0; m < DIM; m++)
2178             {
2179                 /* Re-reverse the velocities */
2180                 state->v[i][m] = -state->v[i][m];
2181             }
2182         }
2183     }
2184     sfree(savex);
2185 }
2186
2187
2188 static void
2189 integrate_table(real vdwtab[], real scale, int offstart, int rstart, int rend,
2190                 double *enerout, double *virout)
2191 {
2192     double enersum, virsum;
2193     double invscale, invscale2, invscale3;
2194     double r, ea, eb, ec, pa, pb, pc, pd;
2195     double y0, f, g, h;
2196     int    ri, offset, tabfactor;
2197
2198     invscale  = 1.0/scale;
2199     invscale2 = invscale*invscale;
2200     invscale3 = invscale*invscale2;
2201
2202     /* Following summation derived from cubic spline definition,
2203      * Numerical Recipies in C, second edition, p. 113-116.  Exact for
2204      * the cubic spline.  We first calculate the negative of the
2205      * energy from rvdw to rvdw_switch, assuming that g(r)=1, and then
2206      * add the more standard, abrupt cutoff correction to that result,
2207      * yielding the long-range correction for a switched function.  We
2208      * perform both the pressure and energy loops at the same time for
2209      * simplicity, as the computational cost is low. */
2210
2211     if (offstart == 0)
2212     {
2213         /* Since the dispersion table has been scaled down a factor
2214          * 6.0 and the repulsion a factor 12.0 to compensate for the
2215          * c6/c12 parameters inside nbfp[] being scaled up (to save
2216          * flops in kernels), we need to correct for this.
2217          */
2218         tabfactor = 6.0;
2219     }
2220     else
2221     {
2222         tabfactor = 12.0;
2223     }
2224
2225     enersum = 0.0;
2226     virsum  = 0.0;
2227     for (ri = rstart; ri < rend; ++ri)
2228     {
2229         r  = ri*invscale;
2230         ea = invscale3;
2231         eb = 2.0*invscale2*r;
2232         ec = invscale*r*r;
2233
2234         pa = invscale3;
2235         pb = 3.0*invscale2*r;
2236         pc = 3.0*invscale*r*r;
2237         pd = r*r*r;
2238
2239         /* this "8" is from the packing in the vdwtab array - perhaps
2240            should be defined? */
2241
2242         offset = 8*ri + offstart;
2243         y0     = vdwtab[offset];
2244         f      = vdwtab[offset+1];
2245         g      = vdwtab[offset+2];
2246         h      = vdwtab[offset+3];
2247
2248         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);
2249         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);
2250     }
2251     *enerout = 4.0*M_PI*enersum*tabfactor;
2252     *virout  = 4.0*M_PI*virsum*tabfactor;
2253 }
2254
2255 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2256 {
2257     double   eners[2], virs[2], enersum, virsum, y0, f, g, h;
2258     double   r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2259     double   invscale, invscale2, invscale3;
2260     int      ri0, ri1, ri, i, offstart, offset;
2261     real     scale, *vdwtab, tabfactor, tmp;
2262
2263     fr->enershiftsix    = 0;
2264     fr->enershifttwelve = 0;
2265     fr->enerdiffsix     = 0;
2266     fr->enerdifftwelve  = 0;
2267     fr->virdiffsix      = 0;
2268     fr->virdifftwelve   = 0;
2269
2270     if (eDispCorr != edispcNO)
2271     {
2272         for (i = 0; i < 2; i++)
2273         {
2274             eners[i] = 0;
2275             virs[i]  = 0;
2276         }
2277         if ((fr->vdw_modifier == eintmodPOTSHIFT) ||
2278             (fr->vdw_modifier == eintmodPOTSWITCH) ||
2279             (fr->vdw_modifier == eintmodFORCESWITCH) ||
2280             (fr->vdwtype == evdwSHIFT) ||
2281             (fr->vdwtype == evdwSWITCH))
2282         {
2283             if (((fr->vdw_modifier == eintmodPOTSWITCH) ||
2284                  (fr->vdw_modifier == eintmodFORCESWITCH) ||
2285                  (fr->vdwtype == evdwSWITCH)) && fr->rvdw_switch == 0)
2286             {
2287                 gmx_fatal(FARGS,
2288                           "With dispersion correction rvdw-switch can not be zero "
2289                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2290             }
2291
2292             scale  = fr->nblists[0].table_vdw.scale;
2293             vdwtab = fr->nblists[0].table_vdw.data;
2294
2295             /* Round the cut-offs to exact table values for precision */
2296             ri0  = floor(fr->rvdw_switch*scale);
2297             ri1  = ceil(fr->rvdw*scale);
2298
2299             /* The code below has some support for handling force-switching, i.e.
2300              * when the force (instead of potential) is switched over a limited
2301              * region. This leads to a constant shift in the potential inside the
2302              * switching region, which we can handle by adding a constant energy
2303              * term in the force-switch case just like when we do potential-shift.
2304              *
2305              * For now this is not enabled, but to keep the functionality in the
2306              * code we check separately for switch and shift. When we do force-switch
2307              * the shifting point is rvdw_switch, while it is the cutoff when we
2308              * have a classical potential-shift.
2309              *
2310              * For a pure potential-shift the potential has a constant shift
2311              * all the way out to the cutoff, and that is it. For other forms
2312              * we need to calculate the constant shift up to the point where we
2313              * start modifying the potential.
2314              */
2315             ri0  = (fr->vdw_modifier == eintmodPOTSHIFT) ? ri1 : ri0;
2316
2317             r0   = ri0/scale;
2318             r1   = ri1/scale;
2319             rc3  = r0*r0*r0;
2320             rc9  = rc3*rc3*rc3;
2321
2322             if ((fr->vdw_modifier == eintmodFORCESWITCH) ||
2323                 (fr->vdwtype == evdwSHIFT))
2324             {
2325                 /* Determine the constant energy shift below rvdw_switch.
2326                  * Table has a scale factor since we have scaled it down to compensate
2327                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2328                  */
2329                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2330                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2331             }
2332             else if (fr->vdw_modifier == eintmodPOTSHIFT)
2333             {
2334                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3));
2335                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3));
2336             }
2337
2338             /* Add the constant part from 0 to rvdw_switch.
2339              * This integration from 0 to rvdw_switch overcounts the number
2340              * of interactions by 1, as it also counts the self interaction.
2341              * We will correct for this later.
2342              */
2343             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2344             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2345
2346             /* Calculate the contribution in the range [r0,r1] where we
2347              * modify the potential. For a pure potential-shift modifier we will
2348              * have ri0==ri1, and there will not be any contribution here.
2349              */
2350             for (i = 0; i < 2; i++)
2351             {
2352                 enersum = 0;
2353                 virsum  = 0;
2354                 integrate_table(vdwtab, scale, (i == 0 ? 0 : 4), ri0, ri1, &enersum, &virsum);
2355                 eners[i] -= enersum;
2356                 virs[i]  -= virsum;
2357             }
2358
2359             /* Alright: Above we compensated by REMOVING the parts outside r0
2360              * corresponding to the ideal VdW 1/r6 and /r12 potentials.
2361              *
2362              * Regardless of whether r0 is the point where we start switching,
2363              * or the cutoff where we calculated the constant shift, we include
2364              * all the parts we are missing out to infinity from r0 by
2365              * calculating the analytical dispersion correction.
2366              */
2367             eners[0] += -4.0*M_PI/(3.0*rc3);
2368             eners[1] +=  4.0*M_PI/(9.0*rc9);
2369             virs[0]  +=  8.0*M_PI/rc3;
2370             virs[1]  += -16.0*M_PI/(3.0*rc9);
2371         }
2372         else if (fr->vdwtype == evdwCUT ||
2373                  EVDW_PME(fr->vdwtype) ||
2374                  fr->vdwtype == evdwUSER)
2375         {
2376             if (fr->vdwtype == evdwUSER && fplog)
2377             {
2378                 fprintf(fplog,
2379                         "WARNING: using dispersion correction with user tables\n");
2380             }
2381
2382             /* Note that with LJ-PME, the dispersion correction is multiplied
2383              * by the difference between the actual C6 and the value of C6
2384              * that would produce the combination rule.
2385              * This means the normal energy and virial difference formulas
2386              * can be used here.
2387              */
2388
2389             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2390             rc9  = rc3*rc3*rc3;
2391             /* Contribution beyond the cut-off */
2392             eners[0] += -4.0*M_PI/(3.0*rc3);
2393             eners[1] +=  4.0*M_PI/(9.0*rc9);
2394             if (fr->vdw_modifier == eintmodPOTSHIFT)
2395             {
2396                 /* Contribution within the cut-off */
2397                 eners[0] += -4.0*M_PI/(3.0*rc3);
2398                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2399             }
2400             /* Contribution beyond the cut-off */
2401             virs[0]  +=  8.0*M_PI/rc3;
2402             virs[1]  += -16.0*M_PI/(3.0*rc9);
2403         }
2404         else
2405         {
2406             gmx_fatal(FARGS,
2407                       "Dispersion correction is not implemented for vdw-type = %s",
2408                       evdw_names[fr->vdwtype]);
2409         }
2410
2411         /* When we deprecate the group kernels the code below can go too */
2412         if (fr->vdwtype == evdwPME && fr->cutoff_scheme == ecutsGROUP)
2413         {
2414             /* Calculate self-interaction coefficient (assuming that
2415              * the reciprocal-space contribution is constant in the
2416              * region that contributes to the self-interaction).
2417              */
2418             fr->enershiftsix = pow(fr->ewaldcoeff_lj, 6) / 6.0;
2419
2420             eners[0] += -pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3)/3.0;
2421             virs[0]  +=  pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3);
2422         }
2423
2424         fr->enerdiffsix    = eners[0];
2425         fr->enerdifftwelve = eners[1];
2426         /* The 0.5 is due to the Gromacs definition of the virial */
2427         fr->virdiffsix     = 0.5*virs[0];
2428         fr->virdifftwelve  = 0.5*virs[1];
2429     }
2430 }
2431
2432 void calc_dispcorr(t_inputrec *ir, t_forcerec *fr,
2433                    int natoms,
2434                    matrix box, real lambda, tensor pres, tensor virial,
2435                    real *prescorr, real *enercorr, real *dvdlcorr)
2436 {
2437     gmx_bool bCorrAll, bCorrPres;
2438     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2439     int      m;
2440
2441     *prescorr = 0;
2442     *enercorr = 0;
2443     *dvdlcorr = 0;
2444
2445     clear_mat(virial);
2446     clear_mat(pres);
2447
2448     if (ir->eDispCorr != edispcNO)
2449     {
2450         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2451                      ir->eDispCorr == edispcAllEnerPres);
2452         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2453                      ir->eDispCorr == edispcAllEnerPres);
2454
2455         invvol = 1/det(box);
2456         if (fr->n_tpi)
2457         {
2458             /* Only correct for the interactions with the inserted molecule */
2459             dens   = (natoms - fr->n_tpi)*invvol;
2460             ninter = fr->n_tpi;
2461         }
2462         else
2463         {
2464             dens   = natoms*invvol;
2465             ninter = 0.5*natoms;
2466         }
2467
2468         if (ir->efep == efepNO)
2469         {
2470             avcsix    = fr->avcsix[0];
2471             avctwelve = fr->avctwelve[0];
2472         }
2473         else
2474         {
2475             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2476             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2477         }
2478
2479         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2480         *enercorr += avcsix*enerdiff;
2481         dvdlambda  = 0.0;
2482         if (ir->efep != efepNO)
2483         {
2484             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2485         }
2486         if (bCorrAll)
2487         {
2488             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2489             *enercorr += avctwelve*enerdiff;
2490             if (fr->efep != efepNO)
2491             {
2492                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2493             }
2494         }
2495
2496         if (bCorrPres)
2497         {
2498             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2499             if (ir->eDispCorr == edispcAllEnerPres)
2500             {
2501                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2502             }
2503             /* The factor 2 is because of the Gromacs virial definition */
2504             spres = -2.0*invvol*svir*PRESFAC;
2505
2506             for (m = 0; m < DIM; m++)
2507             {
2508                 virial[m][m] += svir;
2509                 pres[m][m]   += spres;
2510             }
2511             *prescorr += spres;
2512         }
2513
2514         /* Can't currently control when it prints, for now, just print when degugging */
2515         if (debug)
2516         {
2517             if (bCorrAll)
2518             {
2519                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2520                         avcsix, avctwelve);
2521             }
2522             if (bCorrPres)
2523             {
2524                 fprintf(debug,
2525                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2526                         *enercorr, spres, svir);
2527             }
2528             else
2529             {
2530                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2531             }
2532         }
2533
2534         if (fr->efep != efepNO)
2535         {
2536             *dvdlcorr += dvdlambda;
2537         }
2538     }
2539 }
2540
2541 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2542                   t_graph *graph, rvec x[])
2543 {
2544     if (fplog)
2545     {
2546         fprintf(fplog, "Removing pbc first time\n");
2547     }
2548     calc_shifts(box, fr->shift_vec);
2549     if (graph)
2550     {
2551         mk_mshift(fplog, graph, fr->ePBC, box, x);
2552         if (gmx_debug_at)
2553         {
2554             p_graph(debug, "do_pbc_first 1", graph);
2555         }
2556         shift_self(graph, box, x);
2557         /* By doing an extra mk_mshift the molecules that are broken
2558          * because they were e.g. imported from another software
2559          * will be made whole again. Such are the healing powers
2560          * of GROMACS.
2561          */
2562         mk_mshift(fplog, graph, fr->ePBC, box, x);
2563         if (gmx_debug_at)
2564         {
2565             p_graph(debug, "do_pbc_first 2", graph);
2566         }
2567     }
2568     if (fplog)
2569     {
2570         fprintf(fplog, "Done rmpbc\n");
2571     }
2572 }
2573
2574 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2575                             gmx_mtop_t *mtop, rvec x[],
2576                             gmx_bool bFirst)
2577 {
2578     t_graph        *graph;
2579     int             mb, as, mol;
2580     gmx_molblock_t *molb;
2581
2582     if (bFirst && fplog)
2583     {
2584         fprintf(fplog, "Removing pbc first time\n");
2585     }
2586
2587     snew(graph, 1);
2588     as = 0;
2589     for (mb = 0; mb < mtop->nmolblock; mb++)
2590     {
2591         molb = &mtop->molblock[mb];
2592         if (molb->natoms_mol == 1 ||
2593             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2594         {
2595             /* Just one atom or charge group in the molecule, no PBC required */
2596             as += molb->nmol*molb->natoms_mol;
2597         }
2598         else
2599         {
2600             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2601             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2602                            0, molb->natoms_mol, FALSE, FALSE, graph);
2603
2604             for (mol = 0; mol < molb->nmol; mol++)
2605             {
2606                 mk_mshift(fplog, graph, ePBC, box, x+as);
2607
2608                 shift_self(graph, box, x+as);
2609                 /* The molecule is whole now.
2610                  * We don't need the second mk_mshift call as in do_pbc_first,
2611                  * since we no longer need this graph.
2612                  */
2613
2614                 as += molb->natoms_mol;
2615             }
2616             done_graph(graph);
2617         }
2618     }
2619     sfree(graph);
2620 }
2621
2622 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2623                        gmx_mtop_t *mtop, rvec x[])
2624 {
2625     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2626 }
2627
2628 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2629                  gmx_mtop_t *mtop, rvec x[])
2630 {
2631     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2632 }
2633
2634 void finish_run(FILE *fplog, t_commrec *cr,
2635                 t_inputrec *inputrec,
2636                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2637                 gmx_walltime_accounting_t walltime_accounting,
2638                 nonbonded_verlet_t *nbv,
2639                 gmx_bool bWriteStat)
2640 {
2641     int     i, j;
2642     t_nrnb *nrnb_tot = NULL;
2643     real    delta_t;
2644     double  nbfs, mflop;
2645     double  elapsed_time,
2646             elapsed_time_over_all_ranks,
2647             elapsed_time_over_all_threads,
2648             elapsed_time_over_all_threads_over_all_ranks;
2649     wallcycle_sum(cr, wcycle);
2650
2651     if (cr->nnodes > 1)
2652     {
2653         snew(nrnb_tot, 1);
2654 #ifdef GMX_MPI
2655         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2656                       cr->mpi_comm_mysim);
2657 #endif
2658     }
2659     else
2660     {
2661         nrnb_tot = nrnb;
2662     }
2663
2664     elapsed_time                                 = walltime_accounting_get_elapsed_time(walltime_accounting);
2665     elapsed_time_over_all_ranks                  = elapsed_time;
2666     elapsed_time_over_all_threads                = walltime_accounting_get_elapsed_time_over_all_threads(walltime_accounting);
2667     elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
2668 #ifdef GMX_MPI
2669     if (cr->nnodes > 1)
2670     {
2671         /* reduce elapsed_time over all MPI ranks in the current simulation */
2672         MPI_Allreduce(&elapsed_time,
2673                       &elapsed_time_over_all_ranks,
2674                       1, MPI_DOUBLE, MPI_SUM,
2675                       cr->mpi_comm_mysim);
2676         elapsed_time_over_all_ranks /= cr->nnodes;
2677         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
2678          * current simulation. */
2679         MPI_Allreduce(&elapsed_time_over_all_threads,
2680                       &elapsed_time_over_all_threads_over_all_ranks,
2681                       1, MPI_DOUBLE, MPI_SUM,
2682                       cr->mpi_comm_mysim);
2683     }
2684 #endif
2685
2686     if (SIMMASTER(cr))
2687     {
2688         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2689     }
2690     if (cr->nnodes > 1)
2691     {
2692         sfree(nrnb_tot);
2693     }
2694
2695     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2696     {
2697         print_dd_statistics(cr, inputrec, fplog);
2698     }
2699
2700     if (SIMMASTER(cr))
2701     {
2702         wallclock_gpu_t* gputimes = use_GPU(nbv) ?
2703             nbnxn_cuda_get_timings(nbv->cu_nbv) : NULL;
2704         wallcycle_print(fplog, cr->nnodes, cr->npmenodes,
2705                         elapsed_time_over_all_ranks,
2706                         wcycle, gputimes);
2707
2708         if (EI_DYNAMICS(inputrec->eI))
2709         {
2710             delta_t = inputrec->delta_t;
2711         }
2712         else
2713         {
2714             delta_t = 0;
2715         }
2716
2717         if (fplog)
2718         {
2719             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
2720                        elapsed_time_over_all_ranks,
2721                        walltime_accounting_get_nsteps_done(walltime_accounting),
2722                        delta_t, nbfs, mflop);
2723         }
2724         if (bWriteStat)
2725         {
2726             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
2727                        elapsed_time_over_all_ranks,
2728                        walltime_accounting_get_nsteps_done(walltime_accounting),
2729                        delta_t, nbfs, mflop);
2730         }
2731     }
2732 }
2733
2734 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2735 {
2736     /* this function works, but could probably use a logic rewrite to keep all the different
2737        types of efep straight. */
2738
2739     int       i;
2740     t_lambda *fep = ir->fepvals;
2741
2742     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2743     {
2744         for (i = 0; i < efptNR; i++)
2745         {
2746             lambda[i] = 0.0;
2747             if (lam0)
2748             {
2749                 lam0[i] = 0.0;
2750             }
2751         }
2752         return;
2753     }
2754     else
2755     {
2756         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2757                                              if checkpoint is set -- a kludge is in for now
2758                                              to prevent this.*/
2759         for (i = 0; i < efptNR; i++)
2760         {
2761             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2762             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2763             {
2764                 lambda[i] = fep->init_lambda;
2765                 if (lam0)
2766                 {
2767                     lam0[i] = lambda[i];
2768                 }
2769             }
2770             else
2771             {
2772                 lambda[i] = fep->all_lambda[i][*fep_state];
2773                 if (lam0)
2774                 {
2775                     lam0[i] = lambda[i];
2776                 }
2777             }
2778         }
2779         if (ir->bSimTemp)
2780         {
2781             /* need to rescale control temperatures to match current state */
2782             for (i = 0; i < ir->opts.ngtc; i++)
2783             {
2784                 if (ir->opts.ref_t[i] > 0)
2785                 {
2786                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2787                 }
2788             }
2789         }
2790     }
2791
2792     /* Send to the log the information on the current lambdas */
2793     if (fplog != NULL)
2794     {
2795         fprintf(fplog, "Initial vector of lambda components:[ ");
2796         for (i = 0; i < efptNR; i++)
2797         {
2798             fprintf(fplog, "%10.4f ", lambda[i]);
2799         }
2800         fprintf(fplog, "]\n");
2801     }
2802     return;
2803 }
2804
2805
2806 void init_md(FILE *fplog,
2807              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2808              double *t, double *t0,
2809              real *lambda, int *fep_state, double *lam0,
2810              t_nrnb *nrnb, gmx_mtop_t *mtop,
2811              gmx_update_t *upd,
2812              int nfile, const t_filenm fnm[],
2813              gmx_mdoutf_t *outf, t_mdebin **mdebin,
2814              tensor force_vir, tensor shake_vir, rvec mu_tot,
2815              gmx_bool *bSimAnn, t_vcm **vcm, unsigned long Flags,
2816              gmx_wallcycle_t wcycle)
2817 {
2818     int  i, j, n;
2819     real tmpt, mod;
2820
2821     /* Initial values */
2822     *t = *t0       = ir->init_t;
2823
2824     *bSimAnn = FALSE;
2825     for (i = 0; i < ir->opts.ngtc; i++)
2826     {
2827         /* set bSimAnn if any group is being annealed */
2828         if (ir->opts.annealing[i] != eannNO)
2829         {
2830             *bSimAnn = TRUE;
2831         }
2832     }
2833     if (*bSimAnn)
2834     {
2835         update_annealing_target_temp(&(ir->opts), ir->init_t);
2836     }
2837
2838     /* Initialize lambda variables */
2839     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2840
2841     if (upd)
2842     {
2843         *upd = init_update(ir);
2844     }
2845
2846
2847     if (vcm != NULL)
2848     {
2849         *vcm = init_vcm(fplog, &mtop->groups, ir);
2850     }
2851
2852     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2853     {
2854         if (ir->etc == etcBERENDSEN)
2855         {
2856             please_cite(fplog, "Berendsen84a");
2857         }
2858         if (ir->etc == etcVRESCALE)
2859         {
2860             please_cite(fplog, "Bussi2007a");
2861         }
2862         if (ir->eI == eiSD1)
2863         {
2864             please_cite(fplog, "Goga2012");
2865         }
2866     }
2867
2868     init_nrnb(nrnb);
2869
2870     if (nfile != -1)
2871     {
2872         *outf = init_mdoutf(fplog, nfile, fnm, Flags, cr, ir, mtop, oenv, wcycle);
2873
2874         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : mdoutf_get_fp_ene(*outf),
2875                               mtop, ir, mdoutf_get_fp_dhdl(*outf));
2876     }
2877
2878     if (ir->bAdress)
2879     {
2880         please_cite(fplog, "Fritsch12");
2881         please_cite(fplog, "Junghans10");
2882     }
2883     /* Initiate variables */
2884     clear_mat(force_vir);
2885     clear_mat(shake_vir);
2886     clear_rvec(mu_tot);
2887
2888     debug_gmx();
2889 }