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 "config.h"
  38
  39 #include <assert.h>
  40 #include <math.h>
  41 #include <stdio.h>
  42 #include <string.h>
  43 #ifdef HAVE_SYS_TIME_H
  44 #include <sys/time.h>
  45 #endif
  46
  47 #include "typedefs.h"
  48 #include "gromacs/utility/cstringutil.h"
  49 #include "names.h"
  50 #include "txtdump.h"
  51 #include "gromacs/pbcutil/pbc.h"
  52 #include "chargegroup.h"
  53 #include "gromacs/math/vec.h"
  54 #include "nrnb.h"
  55 #include "mdrun.h"
  56 #include "sim_util.h"
  57 #include "update.h"
  58 #include "gromacs/math/units.h"
  59 #include "mdatoms.h"
  60 #include "force.h"
  61 #include "bondf.h"
  62 #include "pme.h"
  63 #include "disre.h"
  64 #include "orires.h"
  65 #include "network.h"
  66 #include "calcmu.h"
  67 #include "constr.h"
  68 #include "copyrite.h"
  69 #include "domdec.h"
  70 #include "genborn.h"
  71 #include "nbnxn_atomdata.h"
  72 #include "nbnxn_search.h"
  73 #include "nbnxn_kernels/nbnxn_kernel_ref.h"
  74 #include "nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.h"
  75 #include "nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn.h"
  76 #include "nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  77 #include "nonbonded.h"
  78 #include "../gmxlib/nonbonded/nb_kernel.h"
  79 #include "../gmxlib/nonbonded/nb_free_energy.h"
  80
  81 #include "gromacs/legacyheaders/types/commrec.h"
  82 #include "gromacs/pbcutil/ishift.h"
  83 #include "gromacs/pbcutil/mshift.h"
  84 #include "gromacs/timing/wallcycle.h"
  85 #include "gromacs/timing/walltime_accounting.h"
  86 #include "gromacs/utility/gmxmpi.h"
  87 #include "gromacs/utility/smalloc.h"
  88 #include "gromacs/essentialdynamics/edsam.h"
  89 #include "gromacs/pulling/pull.h"
  90 #include "gromacs/pulling/pull_rotation.h"
  91 #include "gromacs/imd/imd.h"
  92 #include "adress.h"
  93 #include "qmmm.h"
  94
  95 #include "gmx_omp_nthreads.h"
  96
  97 #include "nbnxn_cuda_data_mgmt.h"
  98 #include "nbnxn_cuda/nbnxn_cuda.h"
  99
 100 void print_time(FILE                     *out,
 101                 gmx_walltime_accounting_t walltime_accounting,
 102                 gmx_int64_t               step,
 103                 t_inputrec               *ir,
 104                 t_commrec gmx_unused     *cr)
 105 {
 106     time_t finish;
 107     char   timebuf[STRLEN];
 108     double dt, elapsed_seconds, time_per_step;
 109     char   buf[48];
 110
 111 #ifndef GMX_THREAD_MPI
 112     if (!PAR(cr))
 113 #endif
 114     {
 115         fprintf(out, "\r");
 116     }
 117     fprintf(out, "step %s", gmx_step_str(step, buf));
 118     if ((step >= ir->nstlist))
 119     {
 120         double seconds_since_epoch = gmx_gettime();
 121         elapsed_seconds = seconds_since_epoch - walltime_accounting_get_start_time_stamp(walltime_accounting);
 122         time_per_step   = elapsed_seconds/(step - ir->init_step + 1);
 123         dt              = (ir->nsteps + ir->init_step - step) * time_per_step;
 124
 125         if (ir->nsteps >= 0)
 126         {
 127             if (dt >= 300)
 128             {
 129                 finish = (time_t) (seconds_since_epoch + dt);
 130                 gmx_ctime_r(&finish, timebuf, STRLEN);
 131                 sprintf(buf, "%s", timebuf);
 132                 buf[strlen(buf)-1] = '\0';
 133                 fprintf(out, ", will finish %s", buf);
 134             }
 135             else
 136             {
 137                 fprintf(out, ", remaining wall clock time: %5d s          ", (int)dt);
 138             }
 139         }
 140         else
 141         {
 142             fprintf(out, " performance: %.1f ns/day    ",
 143                     ir->delta_t/1000*24*60*60/time_per_step);
 144         }
 145     }
 146 #ifndef GMX_THREAD_MPI
 147     if (PAR(cr))
 148     {
 149         fprintf(out, "\n");
 150     }
 151 #endif
 152
 153     fflush(out);
 154 }
 155
 156 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
 157                          double the_time)
 158 {
 159     char   time_string[STRLEN];
 160
 161     if (!fplog)
 162     {
 163         return;
 164     }
 165
 166     {
 167         int    i;
 168         char   timebuf[STRLEN];
 169         time_t temp_time = (time_t) the_time;
 170
 171         gmx_ctime_r(&temp_time, timebuf, STRLEN);
 172         for (i = 0; timebuf[i] >= ' '; i++)
 173         {
 174             time_string[i] = timebuf[i];
 175         }
 176         time_string[i] = '\0';
 177     }
 178
 179     fprintf(fplog, "%s on rank %d %s\n", title, nodeid, time_string);
 180 }
 181
 182 void print_start(FILE *fplog, t_commrec *cr,
 183                  gmx_walltime_accounting_t walltime_accounting,
 184                  const char *name)
 185 {
 186     char buf[STRLEN];
 187
 188     sprintf(buf, "Started %s", name);
 189     print_date_and_time(fplog, cr->nodeid, buf,
 190                         walltime_accounting_get_start_time_stamp(walltime_accounting));
 191 }
 192
 193 static void sum_forces(int start, int end, rvec f[], rvec flr[])
 194 {
 195     int i;
 196
 197     if (gmx_debug_at)
 198     {
 199         pr_rvecs(debug, 0, "fsr", f+start, end-start);
 200         pr_rvecs(debug, 0, "flr", flr+start, end-start);
 201     }
 202     for (i = start; (i < end); i++)
 203     {
 204         rvec_inc(f[i], flr[i]);
 205     }
 206 }
 207
 208 /*
 209  * calc_f_el calculates forces due to an electric field.
 210  *
 211  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 212  *
 213  * Et[] contains the parameters for the time dependent
 214  * part of the field (not yet used).
 215  * Ex[] contains the parameters for
 216  * the spatial dependent part of the field. You can have cool periodic
 217  * fields in principle, but only a constant field is supported
 218  * now.
 219  * The function should return the energy due to the electric field
 220  * (if any) but for now returns 0.
 221  *
 222  * WARNING:
 223  * There can be problems with the virial.
 224  * Since the field is not self-consistent this is unavoidable.
 225  * For neutral molecules the virial is correct within this approximation.
 226  * For neutral systems with many charged molecules the error is small.
 227  * But for systems with a net charge or a few charged molecules
 228  * the error can be significant when the field is high.
 229  * Solution: implement a self-consitent electric field into PME.
 230  */
 231 static void calc_f_el(FILE *fp, int  start, int homenr,
 232                       real charge[], rvec f[],
 233                       t_cosines Ex[], t_cosines Et[], double t)
 234 {
 235     rvec Ext;
 236     real t0;
 237     int  i, m;
 238
 239     for (m = 0; (m < DIM); m++)
 240     {
 241         if (Et[m].n > 0)
 242         {
 243             if (Et[m].n == 3)
 244             {
 245                 t0     = Et[m].a[1];
 246                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 247             }
 248             else
 249             {
 250                 Ext[m] = cos(Et[m].a[0]*t);
 251             }
 252         }
 253         else
 254         {
 255             Ext[m] = 1.0;
 256         }
 257         if (Ex[m].n > 0)
 258         {
 259             /* Convert the field strength from V/nm to MD-units */
 260             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 261             for (i = start; (i < start+homenr); i++)
 262             {
 263                 f[i][m] += charge[i]*Ext[m];
 264             }
 265         }
 266         else
 267         {
 268             Ext[m] = 0;
 269         }
 270     }
 271     if (fp != NULL)
 272     {
 273         fprintf(fp, "%10g  %10g  %10g  %10g #FIELD\n", t,
 274                 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
 275     }
 276 }
 277
 278 static void calc_virial(int start, int homenr, rvec x[], rvec f[],
 279                         tensor vir_part, t_graph *graph, matrix box,
 280                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 281 {
 282     int    i, j;
 283     tensor virtest;
 284
 285     /* The short-range virial from surrounding boxes */
 286     clear_mat(vir_part);
 287     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 288     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 289
 290     /* Calculate partial virial, for local atoms only, based on short range.
 291      * Total virial is computed in global_stat, called from do_md
 292      */
 293     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 294     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 295
 296     /* Add position restraint contribution */
 297     for (i = 0; i < DIM; i++)
 298     {
 299         vir_part[i][i] += fr->vir_diag_posres[i];
 300     }
 301
 302     /* Add wall contribution */
 303     for (i = 0; i < DIM; i++)
 304     {
 305         vir_part[i][ZZ] += fr->vir_wall_z[i];
 306     }
 307
 308     if (debug)
 309     {
 310         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 311     }
 312 }
 313
 314 static void posres_wrapper(FILE *fplog,
 315                            int flags,
 316                            gmx_bool bSepDVDL,
 317                            t_inputrec *ir,
 318                            t_nrnb *nrnb,
 319                            gmx_localtop_t *top,
 320                            matrix box, rvec x[],
 321                            gmx_enerdata_t *enerd,
 322                            real *lambda,
 323                            t_forcerec *fr)
 324 {
 325     t_pbc pbc;
 326     real  v, dvdl;
 327     int   i;
 328
 329     /* Position restraints always require full pbc */
 330     set_pbc(&pbc, ir->ePBC, box);
 331     dvdl = 0;
 332     v    = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 333                   top->idef.iparams_posres,
 334                   (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 335                   ir->ePBC == epbcNONE ? NULL : &pbc,
 336                   lambda[efptRESTRAINT], &dvdl,
 337                   fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 338     if (bSepDVDL)
 339     {
 340         gmx_print_sepdvdl(fplog, interaction_function[F_POSRES].longname, v, dvdl);
 341     }
 342     enerd->term[F_POSRES] += v;
 343     /* If just the force constant changes, the FEP term is linear,
 344      * but if k changes, it is not.
 345      */
 346     enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
 347     inc_nrnb(nrnb, eNR_POSRES, top->idef.il[F_POSRES].nr/2);
 348
 349     if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
 350     {
 351         for (i = 0; i < enerd->n_lambda; i++)
 352         {
 353             real dvdl_dum, lambda_dum;
 354
 355             lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
 356             v          = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 357                                 top->idef.iparams_posres,
 358                                 (const rvec*)x, NULL, NULL,
 359                                 ir->ePBC == epbcNONE ? NULL : &pbc, lambda_dum, &dvdl,
 360                                 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 361             enerd->enerpart_lambda[i] += v;
 362         }
 363     }
 364 }
 365
 366 static void fbposres_wrapper(t_inputrec *ir,
 367                              t_nrnb *nrnb,
 368                              gmx_localtop_t *top,
 369                              matrix box, rvec x[],
 370                              gmx_enerdata_t *enerd,
 371                              t_forcerec *fr)
 372 {
 373     t_pbc pbc;
 374     real  v;
 375
 376     /* Flat-bottomed position restraints always require full pbc */
 377     set_pbc(&pbc, ir->ePBC, box);
 378     v = fbposres(top->idef.il[F_FBPOSRES].nr, top->idef.il[F_FBPOSRES].iatoms,
 379                  top->idef.iparams_fbposres,
 380                  (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 381                  ir->ePBC == epbcNONE ? NULL : &pbc,
 382                  fr->rc_scaling, fr->ePBC, fr->posres_com);
 383     enerd->term[F_FBPOSRES] += v;
 384     inc_nrnb(nrnb, eNR_FBPOSRES, top->idef.il[F_FBPOSRES].nr/2);
 385 }
 386
 387 static void pull_potential_wrapper(FILE *fplog,
 388                                    gmx_bool bSepDVDL,
 389                                    t_commrec *cr,
 390                                    t_inputrec *ir,
 391                                    matrix box, rvec x[],
 392                                    rvec f[],
 393                                    tensor vir_force,
 394                                    t_mdatoms *mdatoms,
 395                                    gmx_enerdata_t *enerd,
 396                                    real *lambda,
 397                                    double t)
 398 {
 399     t_pbc  pbc;
 400     real   dvdl;
 401
 402     /* Calculate the center of mass forces, this requires communication,
 403      * which is why pull_potential is called close to other communication.
 404      * The virial contribution is calculated directly,
 405      * which is why we call pull_potential after calc_virial.
 406      */
 407     set_pbc(&pbc, ir->ePBC, box);
 408     dvdl                     = 0;
 409     enerd->term[F_COM_PULL] +=
 410         pull_potential(ir->ePull, ir->pull, mdatoms, &pbc,
 411                        cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
 412     if (bSepDVDL)
 413     {
 414         gmx_print_sepdvdl(fplog, "Com pull", enerd->term[F_COM_PULL], dvdl);
 415     }
 416     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 417 }
 418
 419 static void pme_receive_force_ener(FILE           *fplog,
 420                                    gmx_bool        bSepDVDL,
 421                                    t_commrec      *cr,
 422                                    gmx_wallcycle_t wcycle,
 423                                    gmx_enerdata_t *enerd,
 424                                    t_forcerec     *fr)
 425 {
 426     real   e_q, e_lj, v, dvdl_q, dvdl_lj;
 427     float  cycles_ppdpme, cycles_seppme;
 428
 429     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 430     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 431
 432     /* In case of node-splitting, the PP nodes receive the long-range
 433      * forces, virial and energy from the PME nodes here.
 434      */
 435     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 436     dvdl_q  = 0;
 437     dvdl_lj = 0;
 438     gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e_q,
 439                       fr->vir_lj_recip, &e_lj, &dvdl_q, &dvdl_lj,
 440                       &cycles_seppme);
 441     if (bSepDVDL)
 442     {
 443         gmx_print_sepdvdl(fplog, "Electrostatic PME mesh", e_q, dvdl_q);
 444         gmx_print_sepdvdl(fplog, "Lennard-Jones PME mesh", e_lj, dvdl_lj);
 445     }
 446     enerd->term[F_COUL_RECIP] += e_q;
 447     enerd->term[F_LJ_RECIP]   += e_lj;
 448     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 449     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 450
 451     if (wcycle)
 452     {
 453         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 454     }
 455     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 456 }
 457
 458 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
 459                                gmx_int64_t step, real pforce, rvec *x, rvec *f)
 460 {
 461     int  i;
 462     real pf2, fn2;
 463     char buf[STEPSTRSIZE];
 464
 465     pf2 = sqr(pforce);
 466     for (i = 0; i < md->homenr; i++)
 467     {
 468         fn2 = norm2(f[i]);
 469         /* We also catch NAN, if the compiler does not optimize this away. */
 470         if (fn2 >= pf2 || fn2 != fn2)
 471         {
 472             fprintf(fp, "step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 473                     gmx_step_str(step, buf),
 474                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
 475         }
 476     }
 477 }
 478
 479 static void post_process_forces(t_commrec *cr,
 480                                 gmx_int64_t step,
 481                                 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 482                                 gmx_localtop_t *top,
 483                                 matrix box, rvec x[],
 484                                 rvec f[],
 485                                 tensor vir_force,
 486                                 t_mdatoms *mdatoms,
 487                                 t_graph *graph,
 488                                 t_forcerec *fr, gmx_vsite_t *vsite,
 489                                 int flags)
 490 {
 491     if (fr->bF_NoVirSum)
 492     {
 493         if (vsite)
 494         {
 495             /* Spread the mesh force on virtual sites to the other particles...
 496              * This is parallellized. MPI communication is performed
 497              * if the constructing atoms aren't local.
 498              */
 499             wallcycle_start(wcycle, ewcVSITESPREAD);
 500             spread_vsite_f(vsite, x, fr->f_novirsum, NULL,
 501                            (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
 502                            nrnb,
 503                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
 504             wallcycle_stop(wcycle, ewcVSITESPREAD);
 505         }
 506         if (flags & GMX_FORCE_VIRIAL)
 507         {
 508             /* Now add the forces, this is local */
 509             if (fr->bDomDec)
 510             {
 511                 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
 512             }
 513             else
 514             {
 515                 sum_forces(0, mdatoms->homenr,
 516                            f, fr->f_novirsum);
 517             }
 518             if (EEL_FULL(fr->eeltype))
 519             {
 520                 /* Add the mesh contribution to the virial */
 521                 m_add(vir_force, fr->vir_el_recip, vir_force);
 522             }
 523             if (EVDW_PME(fr->vdwtype))
 524             {
 525                 /* Add the mesh contribution to the virial */
 526                 m_add(vir_force, fr->vir_lj_recip, vir_force);
 527             }
 528             if (debug)
 529             {
 530                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 531             }
 532         }
 533     }
 534
 535     if (fr->print_force >= 0)
 536     {
 537         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 538     }
 539 }
 540
 541 static void do_nb_verlet(t_forcerec *fr,
 542                          interaction_const_t *ic,
 543                          gmx_enerdata_t *enerd,
 544                          int flags, int ilocality,
 545                          int clearF,
 546                          t_nrnb *nrnb,
 547                          gmx_wallcycle_t wcycle)
 548 {
 549     int                        nnbl, kernel_type, enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
 550     char                      *env;
 551     nonbonded_verlet_group_t  *nbvg;
 552     gmx_bool                   bCUDA;
 553
 554     if (!(flags & GMX_FORCE_NONBONDED))
 555     {
 556         /* skip non-bonded calculation */
 557         return;
 558     }
 559
 560     nbvg = &fr->nbv->grp[ilocality];
 561
 562     /* CUDA kernel launch overhead is already timed separately */
 563     if (fr->cutoff_scheme != ecutsVERLET)
 564     {
 565         gmx_incons("Invalid cut-off scheme passed!");
 566     }
 567
 568     bCUDA = (nbvg->kernel_type == nbnxnk8x8x8_CUDA);
 569
 570     if (!bCUDA)
 571     {
 572         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 573     }
 574     switch (nbvg->kernel_type)
 575     {
 576         case nbnxnk4x4_PlainC:
 577             nbnxn_kernel_ref(&nbvg->nbl_lists,
 578                              nbvg->nbat, ic,
 579                              fr->shift_vec,
 580                              flags,
 581                              clearF,
 582                              fr->fshift[0],
 583                              enerd->grpp.ener[egCOULSR],
 584                              fr->bBHAM ?
 585                              enerd->grpp.ener[egBHAMSR] :
 586                              enerd->grpp.ener[egLJSR]);
 587             break;
 588
 589         case nbnxnk4xN_SIMD_4xN:
 590             nbnxn_kernel_simd_4xn(&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         case nbnxnk4xN_SIMD_2xNN:
 603             nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
 604                                    nbvg->nbat, ic,
 605                                    nbvg->ewald_excl,
 606                                    fr->shift_vec,
 607                                    flags,
 608                                    clearF,
 609                                    fr->fshift[0],
 610                                    enerd->grpp.ener[egCOULSR],
 611                                    fr->bBHAM ?
 612                                    enerd->grpp.ener[egBHAMSR] :
 613                                    enerd->grpp.ener[egLJSR]);
 614             break;
 615
 616         case nbnxnk8x8x8_CUDA:
 617             nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
 618             break;
 619
 620         case nbnxnk8x8x8_PlainC:
 621             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 622                                  nbvg->nbat, ic,
 623                                  fr->shift_vec,
 624                                  flags,
 625                                  clearF,
 626                                  nbvg->nbat->out[0].f,
 627                                  fr->fshift[0],
 628                                  enerd->grpp.ener[egCOULSR],
 629                                  fr->bBHAM ?
 630                                  enerd->grpp.ener[egBHAMSR] :
 631                                  enerd->grpp.ener[egLJSR]);
 632             break;
 633
 634         default:
 635             gmx_incons("Invalid nonbonded kernel type passed!");
 636
 637     }
 638     if (!bCUDA)
 639     {
 640         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 641     }
 642
 643     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 644     {
 645         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 646     }
 647     else if ((!bCUDA && nbvg->ewald_excl == ewaldexclAnalytical) ||
 648              (bCUDA && nbnxn_cuda_is_kernel_ewald_analytical(fr->nbv->cu_nbv)))
 649     {
 650         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 651     }
 652     else
 653     {
 654         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 655     }
 656     enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 657     if (flags & GMX_FORCE_ENERGY)
 658     {
 659         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 660         enr_nbnxn_kernel_ljc += 1;
 661         enr_nbnxn_kernel_lj  += 1;
 662     }
 663
 664     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 665              nbvg->nbl_lists.natpair_ljq);
 666     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 667              nbvg->nbl_lists.natpair_lj);
 668     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 669     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 670              nbvg->nbl_lists.natpair_q);
 671
 672     if (ic->vdw_modifier == eintmodFORCESWITCH)
 673     {
 674         /* We add up the switch cost separately */
 675         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_FSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 676                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 677     }
 678     if (ic->vdw_modifier == eintmodPOTSWITCH)
 679     {
 680         /* We add up the switch cost separately */
 681         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_PSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 682                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 683     }
 684     if (ic->vdwtype == evdwPME)
 685     {
 686         /* We add up the LJ Ewald cost separately */
 687         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_EWALD+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 688                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 689     }
 690 }
 691
 692 static void do_nb_verlet_fep(nbnxn_pairlist_set_t *nbl_lists,
 693                              t_forcerec           *fr,
 694                              rvec                  x[],
 695                              rvec                  f[],
 696                              t_mdatoms            *mdatoms,
 697                              t_lambda             *fepvals,
 698                              real                 *lambda,
 699                              gmx_enerdata_t       *enerd,
 700                              int                   flags,
 701                              t_nrnb               *nrnb,
 702                              gmx_wallcycle_t       wcycle)
 703 {
 704     int              donb_flags;
 705     nb_kernel_data_t kernel_data;
 706     real             lam_i[efptNR];
 707     real             dvdl_nb[efptNR];
 708     int              th;
 709     int              i, j;
 710
 711     donb_flags = 0;
 712     /* Add short-range interactions */
 713     donb_flags |= GMX_NONBONDED_DO_SR;
 714
 715     /* Currently all group scheme kernels always calculate (shift-)forces */
 716     if (flags & GMX_FORCE_FORCES)
 717     {
 718         donb_flags |= GMX_NONBONDED_DO_FORCE;
 719     }
 720     if (flags & GMX_FORCE_VIRIAL)
 721     {
 722         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 723     }
 724     if (flags & GMX_FORCE_ENERGY)
 725     {
 726         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 727     }
 728     if (flags & GMX_FORCE_DO_LR)
 729     {
 730         donb_flags |= GMX_NONBONDED_DO_LR;
 731     }
 732
 733     kernel_data.flags  = donb_flags;
 734     kernel_data.lambda = lambda;
 735     kernel_data.dvdl   = dvdl_nb;
 736
 737     kernel_data.energygrp_elec = enerd->grpp.ener[egCOULSR];
 738     kernel_data.energygrp_vdw  = enerd->grpp.ener[egLJSR];
 739
 740     /* reset free energy components */
 741     for (i = 0; i < efptNR; i++)
 742     {
 743         dvdl_nb[i]  = 0;
 744     }
 745
 746     assert(gmx_omp_nthreads_get(emntNonbonded) == nbl_lists->nnbl);
 747
 748     wallcycle_sub_start(wcycle, ewcsNONBONDED);
 749 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 750     for (th = 0; th < nbl_lists->nnbl; th++)
 751     {
 752         gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 753                                   x, f, fr, mdatoms, &kernel_data, nrnb);
 754     }
 755
 756     if (fepvals->sc_alpha != 0)
 757     {
 758         enerd->dvdl_nonlin[efptVDW]  += dvdl_nb[efptVDW];
 759         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 760     }
 761     else
 762     {
 763         enerd->dvdl_lin[efptVDW]  += dvdl_nb[efptVDW];
 764         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 765     }
 766
 767     /* If we do foreign lambda and we have soft-core interactions
 768      * we have to recalculate the (non-linear) energies contributions.
 769      */
 770     if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
 771     {
 772         kernel_data.flags          = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE)) | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 773         kernel_data.lambda         = lam_i;
 774         kernel_data.energygrp_elec = enerd->foreign_grpp.ener[egCOULSR];
 775         kernel_data.energygrp_vdw  = enerd->foreign_grpp.ener[egLJSR];
 776         /* Note that we add to kernel_data.dvdl, but ignore the result */
 777
 778         for (i = 0; i < enerd->n_lambda; i++)
 779         {
 780             for (j = 0; j < efptNR; j++)
 781             {
 782                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
 783             }
 784             reset_foreign_enerdata(enerd);
 785 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 786             for (th = 0; th < nbl_lists->nnbl; th++)
 787             {
 788                 gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 789                                           x, f, fr, mdatoms, &kernel_data, nrnb);
 790             }
 791
 792             sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
 793             enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
 794         }
 795     }
 796
 797     wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 798 }
 799
 800 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
 801                          t_inputrec *inputrec,
 802                          gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 803                          gmx_localtop_t *top,
 804                          gmx_groups_t gmx_unused *groups,
 805                          matrix box, rvec x[], history_t *hist,
 806                          rvec f[],
 807                          tensor vir_force,
 808                          t_mdatoms *mdatoms,
 809                          gmx_enerdata_t *enerd, t_fcdata *fcd,
 810                          real *lambda, t_graph *graph,
 811                          t_forcerec *fr, interaction_const_t *ic,
 812                          gmx_vsite_t *vsite, rvec mu_tot,
 813                          double t, FILE *field, gmx_edsam_t ed,
 814                          gmx_bool bBornRadii,
 815                          int flags)
 816 {
 817     int                 cg0, cg1, i, j;
 818     int                 start, homenr;
 819     int                 nb_kernel_type;
 820     double              mu[2*DIM];
 821     gmx_bool            bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
 822     gmx_bool            bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
 823     gmx_bool            bDiffKernels = FALSE;
 824     matrix              boxs;
 825     rvec                vzero, box_diag;
 826     real                e, v, dvdl;
 827     float               cycles_pme, cycles_force, cycles_wait_gpu;
 828     nonbonded_verlet_t *nbv;
 829
 830     cycles_force    = 0;
 831     cycles_wait_gpu = 0;
 832     nbv             = fr->nbv;
 833     nb_kernel_type  = fr->nbv->grp[0].kernel_type;
 834
 835     start  = 0;
 836     homenr = mdatoms->homenr;
 837
 838     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
 839
 840     clear_mat(vir_force);
 841
 842     cg0 = 0;
 843     if (DOMAINDECOMP(cr))
 844     {
 845         cg1 = cr->dd->ncg_tot;
 846     }
 847     else
 848     {
 849         cg1 = top->cgs.nr;
 850     }
 851     if (fr->n_tpi > 0)
 852     {
 853         cg1--;
 854     }
 855
 856     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 857     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
 858     bFillGrid     = (bNS && bStateChanged);
 859     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 860     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 861     bDoForces     = (flags & GMX_FORCE_FORCES);
 862     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 863     bUseGPU       = fr->nbv->bUseGPU;
 864     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
 865
 866     if (bStateChanged)
 867     {
 868         update_forcerec(fr, box);
 869
 870         if (NEED_MUTOT(*inputrec))
 871         {
 872             /* Calculate total (local) dipole moment in a temporary common array.
 873              * This makes it possible to sum them over nodes faster.
 874              */
 875             calc_mu(start, homenr,
 876                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 877                     mu, mu+DIM);
 878         }
 879     }
 880
 881     if (fr->ePBC != epbcNONE)
 882     {
 883         /* Compute shift vectors every step,
 884          * because of pressure coupling or box deformation!
 885          */
 886         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 887         {
 888             calc_shifts(box, fr->shift_vec);
 889         }
 890
 891         if (bCalcCGCM)
 892         {
 893             put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
 894             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 895         }
 896         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 897         {
 898             unshift_self(graph, box, x);
 899         }
 900     }
 901
 902     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 903                                  fr->shift_vec, nbv->grp[0].nbat);
 904
 905 #ifdef GMX_MPI
 906     if (!(cr->duty & DUTY_PME))
 907     {
 908         /* Send particle coordinates to the pme nodes.
 909          * Since this is only implemented for domain decomposition
 910          * and domain decomposition does not use the graph,
 911          * we do not need to worry about shifting.
 912          */
 913
 914         int pme_flags = 0;
 915
 916         wallcycle_start(wcycle, ewcPP_PMESENDX);
 917
 918         bBS = (inputrec->nwall == 2);
 919         if (bBS)
 920         {
 921             copy_mat(box, boxs);
 922             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
 923         }
 924
 925         if (EEL_PME(fr->eeltype))
 926         {
 927             pme_flags |= GMX_PME_DO_COULOMB;
 928         }
 929
 930         if (EVDW_PME(fr->vdwtype))
 931         {
 932             pme_flags |= GMX_PME_DO_LJ;
 933         }
 934
 935         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
 936                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
 937                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
 938                                  pme_flags, step);
 939
 940         wallcycle_stop(wcycle, ewcPP_PMESENDX);
 941     }
 942 #endif /* GMX_MPI */
 943
 944     /* do gridding for pair search */
 945     if (bNS)
 946     {
 947         if (graph && bStateChanged)
 948         {
 949             /* Calculate intramolecular shift vectors to make molecules whole */
 950             mk_mshift(fplog, graph, fr->ePBC, box, x);
 951         }
 952
 953         clear_rvec(vzero);
 954         box_diag[XX] = box[XX][XX];
 955         box_diag[YY] = box[YY][YY];
 956         box_diag[ZZ] = box[ZZ][ZZ];
 957
 958         wallcycle_start(wcycle, ewcNS);
 959         if (!fr->bDomDec)
 960         {
 961             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 962             nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
 963                               0, vzero, box_diag,
 964                               0, mdatoms->homenr, -1, fr->cginfo, x,
 965                               0, NULL,
 966                               nbv->grp[eintLocal].kernel_type,
 967                               nbv->grp[eintLocal].nbat);
 968             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 969         }
 970         else
 971         {
 972             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 973             nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
 974                                        fr->cginfo, x,
 975                                        nbv->grp[eintNonlocal].kernel_type,
 976                                        nbv->grp[eintNonlocal].nbat);
 977             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 978         }
 979
 980         if (nbv->ngrp == 1 ||
 981             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 982         {
 983             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
 984                                nbv->nbs, mdatoms, fr->cginfo);
 985         }
 986         else
 987         {
 988             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
 989                                nbv->nbs, mdatoms, fr->cginfo);
 990             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
 991                                nbv->nbs, mdatoms, fr->cginfo);
 992         }
 993         wallcycle_stop(wcycle, ewcNS);
 994     }
 995
 996     /* initialize the GPU atom data and copy shift vector */
 997     if (bUseGPU)
 998     {
 999         if (bNS)
1000         {
1001             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1002             nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
1003             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1004         }
1005
1006         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1007         nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
1008         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1009     }
1010
1011     /* do local pair search */
1012     if (bNS)
1013     {
1014         wallcycle_start_nocount(wcycle, ewcNS);
1015         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1016         nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
1017                             &top->excls,
1018                             ic->rlist,
1019                             nbv->min_ci_balanced,
1020                             &nbv->grp[eintLocal].nbl_lists,
1021                             eintLocal,
1022                             nbv->grp[eintLocal].kernel_type,
1023                             nrnb);
1024         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1025
1026         if (bUseGPU)
1027         {
1028             /* initialize local pair-list on the GPU */
1029             nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1030                                      nbv->grp[eintLocal].nbl_lists.nbl[0],
1031                                      eintLocal);
1032         }
1033         wallcycle_stop(wcycle, ewcNS);
1034     }
1035     else
1036     {
1037         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1038         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1039         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
1040                                         nbv->grp[eintLocal].nbat);
1041         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1042         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1043     }
1044
1045     if (bUseGPU)
1046     {
1047         wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1048         /* launch local nonbonded F on GPU */
1049         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
1050                      nrnb, wcycle);
1051         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1052     }
1053
1054     /* Communicate coordinates and sum dipole if necessary +
1055        do non-local pair search */
1056     if (DOMAINDECOMP(cr))
1057     {
1058         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
1059                         nbv->grp[eintLocal].kernel_type);
1060
1061         if (bDiffKernels)
1062         {
1063             /* With GPU+CPU non-bonded calculations we need to copy
1064              * the local coordinates to the non-local nbat struct
1065              * (in CPU format) as the non-local kernel call also
1066              * calculates the local - non-local interactions.
1067              */
1068             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1069             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1070             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
1071                                             nbv->grp[eintNonlocal].nbat);
1072             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1073             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1074         }
1075
1076         if (bNS)
1077         {
1078             wallcycle_start_nocount(wcycle, ewcNS);
1079             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1080
1081             if (bDiffKernels)
1082             {
1083                 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
1084             }
1085
1086             nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
1087                                 &top->excls,
1088                                 ic->rlist,
1089                                 nbv->min_ci_balanced,
1090                                 &nbv->grp[eintNonlocal].nbl_lists,
1091                                 eintNonlocal,
1092                                 nbv->grp[eintNonlocal].kernel_type,
1093                                 nrnb);
1094
1095             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1096
1097             if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_CUDA)
1098             {
1099                 /* initialize non-local pair-list on the GPU */
1100                 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1101                                          nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1102                                          eintNonlocal);
1103             }
1104             wallcycle_stop(wcycle, ewcNS);
1105         }
1106         else
1107         {
1108             wallcycle_start(wcycle, ewcMOVEX);
1109             dd_move_x(cr->dd, box, x);
1110
1111             /* When we don't need the total dipole we sum it in global_stat */
1112             if (bStateChanged && NEED_MUTOT(*inputrec))
1113             {
1114                 gmx_sumd(2*DIM, mu, cr);
1115             }
1116             wallcycle_stop(wcycle, ewcMOVEX);
1117
1118             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1119             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1120             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1121                                             nbv->grp[eintNonlocal].nbat);
1122             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1123             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1124         }
1125
1126         if (bUseGPU && !bDiffKernels)
1127         {
1128             wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1129             /* launch non-local nonbonded F on GPU */
1130             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1131                          nrnb, wcycle);
1132             cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1133         }
1134     }
1135
1136     if (bUseGPU)
1137     {
1138         /* launch D2H copy-back F */
1139         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1140         if (DOMAINDECOMP(cr) && !bDiffKernels)
1141         {
1142             nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1143                                       flags, eatNonlocal);
1144         }
1145         nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1146                                   flags, eatLocal);
1147         cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1148     }
1149
1150     if (bStateChanged && NEED_MUTOT(*inputrec))
1151     {
1152         if (PAR(cr))
1153         {
1154             gmx_sumd(2*DIM, mu, cr);
1155         }
1156
1157         for (i = 0; i < 2; i++)
1158         {
1159             for (j = 0; j < DIM; j++)
1160             {
1161                 fr->mu_tot[i][j] = mu[i*DIM + j];
1162             }
1163         }
1164     }
1165     if (fr->efep == efepNO)
1166     {
1167         copy_rvec(fr->mu_tot[0], mu_tot);
1168     }
1169     else
1170     {
1171         for (j = 0; j < DIM; j++)
1172         {
1173             mu_tot[j] =
1174                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1175                 lambda[efptCOUL]*fr->mu_tot[1][j];
1176         }
1177     }
1178
1179     /* Reset energies */
1180     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1181     clear_rvecs(SHIFTS, fr->fshift);
1182
1183     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1184     {
1185         wallcycle_start(wcycle, ewcPPDURINGPME);
1186         dd_force_flop_start(cr->dd, nrnb);
1187     }
1188
1189     if (inputrec->bRot)
1190     {
1191         /* Enforced rotation has its own cycle counter that starts after the collective
1192          * coordinates have been communicated. It is added to ddCyclF to allow
1193          * for proper load-balancing */
1194         wallcycle_start(wcycle, ewcROT);
1195         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1196         wallcycle_stop(wcycle, ewcROT);
1197     }
1198
1199     /* Start the force cycle counter.
1200      * This counter is stopped in do_forcelow_level.
1201      * No parallel communication should occur while this counter is running,
1202      * since that will interfere with the dynamic load balancing.
1203      */
1204     wallcycle_start(wcycle, ewcFORCE);
1205     if (bDoForces)
1206     {
1207         /* Reset forces for which the virial is calculated separately:
1208          * PME/Ewald forces if necessary */
1209         if (fr->bF_NoVirSum)
1210         {
1211             if (flags & GMX_FORCE_VIRIAL)
1212             {
1213                 fr->f_novirsum = fr->f_novirsum_alloc;
1214                 if (fr->bDomDec)
1215                 {
1216                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1217                 }
1218                 else
1219                 {
1220                     clear_rvecs(homenr, fr->f_novirsum+start);
1221                 }
1222             }
1223             else
1224             {
1225                 /* We are not calculating the pressure so we do not need
1226                  * a separate array for forces that do not contribute
1227                  * to the pressure.
1228                  */
1229                 fr->f_novirsum = f;
1230             }
1231         }
1232
1233         /* Clear the short- and long-range forces */
1234         clear_rvecs(fr->natoms_force_constr, f);
1235         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1236         {
1237             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1238         }
1239
1240         clear_rvec(fr->vir_diag_posres);
1241     }
1242
1243     if (inputrec->ePull == epullCONSTRAINT)
1244     {
1245         clear_pull_forces(inputrec->pull);
1246     }
1247
1248     /* We calculate the non-bonded forces, when done on the CPU, here.
1249      * We do this before calling do_force_lowlevel, as in there bondeds
1250      * forces are calculated before PME, which does communication.
1251      * With this order, non-bonded and bonded force calculation imbalance
1252      * can be balanced out by the domain decomposition load balancing.
1253      */
1254
1255     if (!bUseOrEmulGPU)
1256     {
1257         /* Maybe we should move this into do_force_lowlevel */
1258         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1259                      nrnb, wcycle);
1260     }
1261
1262     if (fr->efep != efepNO)
1263     {
1264         /* Calculate the local and non-local free energy interactions here.
1265          * Happens here on the CPU both with and without GPU.
1266          */
1267         if (fr->nbv->grp[eintLocal].nbl_lists.nbl_fep[0]->nrj > 0)
1268         {
1269             do_nb_verlet_fep(&fr->nbv->grp[eintLocal].nbl_lists,
1270                              fr, x, f, mdatoms,
1271                              inputrec->fepvals, lambda,
1272                              enerd, flags, nrnb, wcycle);
1273         }
1274
1275         if (DOMAINDECOMP(cr) &&
1276             fr->nbv->grp[eintNonlocal].nbl_lists.nbl_fep[0]->nrj > 0)
1277         {
1278             do_nb_verlet_fep(&fr->nbv->grp[eintNonlocal].nbl_lists,
1279                              fr, x, f, mdatoms,
1280                              inputrec->fepvals, lambda,
1281                              enerd, flags, nrnb, wcycle);
1282         }
1283     }
1284
1285     if (!bUseOrEmulGPU || bDiffKernels)
1286     {
1287         int aloc;
1288
1289         if (DOMAINDECOMP(cr))
1290         {
1291             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1292                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1293                          nrnb, wcycle);
1294         }
1295
1296         if (!bUseOrEmulGPU)
1297         {
1298             aloc = eintLocal;
1299         }
1300         else
1301         {
1302             aloc = eintNonlocal;
1303         }
1304
1305         /* Add all the non-bonded force to the normal force array.
1306          * This can be split into a local a non-local part when overlapping
1307          * communication with calculation with domain decomposition.
1308          */
1309         cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1310         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1311         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1312         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1313         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1314         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1315         wallcycle_start_nocount(wcycle, ewcFORCE);
1316
1317         /* if there are multiple fshift output buffers reduce them */
1318         if ((flags & GMX_FORCE_VIRIAL) &&
1319             nbv->grp[aloc].nbl_lists.nnbl > 1)
1320         {
1321             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1322                                                      fr->fshift);
1323         }
1324     }
1325
1326     /* update QMMMrec, if necessary */
1327     if (fr->bQMMM)
1328     {
1329         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1330     }
1331
1332     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1333     {
1334         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1335                        enerd, lambda, fr);
1336     }
1337
1338     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1339     {
1340         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1341     }
1342
1343     /* Compute the bonded and non-bonded energies and optionally forces */
1344     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1345                       cr, nrnb, wcycle, mdatoms,
1346                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1347                       &(top->atomtypes), bBornRadii, box,
1348                       inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1349                       flags, &cycles_pme);
1350
1351     if (bSepLRF)
1352     {
1353         if (do_per_step(step, inputrec->nstcalclr))
1354         {
1355             /* Add the long range forces to the short range forces */
1356             for (i = 0; i < fr->natoms_force_constr; i++)
1357             {
1358                 rvec_add(fr->f_twin[i], f[i], f[i]);
1359             }
1360         }
1361     }
1362
1363     cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1364
1365     if (ed)
1366     {
1367         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1368     }
1369
1370     if (bUseOrEmulGPU && !bDiffKernels)
1371     {
1372         /* wait for non-local forces (or calculate in emulation mode) */
1373         if (DOMAINDECOMP(cr))
1374         {
1375             if (bUseGPU)
1376             {
1377                 float cycles_tmp;
1378
1379                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1380                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1381                                     nbv->grp[eintNonlocal].nbat,
1382                                     flags, eatNonlocal,
1383                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1384                                     fr->fshift);
1385                 cycles_tmp       = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1386                 cycles_wait_gpu += cycles_tmp;
1387                 cycles_force    += cycles_tmp;
1388             }
1389             else
1390             {
1391                 wallcycle_start_nocount(wcycle, ewcFORCE);
1392                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1393                              nrnb, wcycle);
1394                 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1395             }
1396             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1397             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1398             /* skip the reduction if there was no non-local work to do */
1399             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1400             {
1401                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1402                                                nbv->grp[eintNonlocal].nbat, f);
1403             }
1404             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1405             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1406         }
1407     }
1408
1409     if (bDoForces && DOMAINDECOMP(cr))
1410     {
1411         /* Communicate the forces */
1412         wallcycle_start(wcycle, ewcMOVEF);
1413         dd_move_f(cr->dd, f, fr->fshift);
1414         /* Do we need to communicate the separate force array
1415          * for terms that do not contribute to the single sum virial?
1416          * Position restraints and electric fields do not introduce
1417          * inter-cg forces, only full electrostatics methods do.
1418          * When we do not calculate the virial, fr->f_novirsum = f,
1419          * so we have already communicated these forces.
1420          */
1421         if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1422             (flags & GMX_FORCE_VIRIAL))
1423         {
1424             dd_move_f(cr->dd, fr->f_novirsum, NULL);
1425         }
1426         if (bSepLRF)
1427         {
1428             /* We should not update the shift forces here,
1429              * since f_twin is already included in f.
1430              */
1431             dd_move_f(cr->dd, fr->f_twin, NULL);
1432         }
1433         wallcycle_stop(wcycle, ewcMOVEF);
1434     }
1435
1436     if (bUseOrEmulGPU)
1437     {
1438         /* wait for local forces (or calculate in emulation mode) */
1439         if (bUseGPU)
1440         {
1441             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1442             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1443                                 nbv->grp[eintLocal].nbat,
1444                                 flags, eatLocal,
1445                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1446                                 fr->fshift);
1447             cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1448
1449             /* now clear the GPU outputs while we finish the step on the CPU */
1450
1451             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1452             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1453             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1454         }
1455         else
1456         {
1457             wallcycle_start_nocount(wcycle, ewcFORCE);
1458             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1459                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1460                          nrnb, wcycle);
1461             wallcycle_stop(wcycle, ewcFORCE);
1462         }
1463         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1464         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1465         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1466         {
1467             /* skip the reduction if there was no non-local work to do */
1468             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1469                                            nbv->grp[eintLocal].nbat, f);
1470         }
1471         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1472         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1473     }
1474
1475     if (DOMAINDECOMP(cr))
1476     {
1477         dd_force_flop_stop(cr->dd, nrnb);
1478         if (wcycle)
1479         {
1480             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1481             if (bUseGPU)
1482             {
1483                 dd_cycles_add(cr->dd, cycles_wait_gpu, ddCyclWaitGPU);
1484             }
1485         }
1486     }
1487
1488     if (bDoForces)
1489     {
1490         if (IR_ELEC_FIELD(*inputrec))
1491         {
1492             /* Compute forces due to electric field */
1493             calc_f_el(MASTER(cr) ? field : NULL,
1494                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1495                       inputrec->ex, inputrec->et, t);
1496         }
1497
1498         /* If we have NoVirSum forces, but we do not calculate the virial,
1499          * we sum fr->f_novirum=f later.
1500          */
1501         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1502         {
1503             wallcycle_start(wcycle, ewcVSITESPREAD);
1504             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1505                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1506             wallcycle_stop(wcycle, ewcVSITESPREAD);
1507
1508             if (bSepLRF)
1509             {
1510                 wallcycle_start(wcycle, ewcVSITESPREAD);
1511                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1512                                nrnb,
1513                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1514                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1515             }
1516         }
1517
1518         if (flags & GMX_FORCE_VIRIAL)
1519         {
1520             /* Calculation of the virial must be done after vsites! */
1521             calc_virial(0, mdatoms->homenr, x, f,
1522                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1523         }
1524     }
1525
1526     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1527     {
1528         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1529                                f, vir_force, mdatoms, enerd, lambda, t);
1530     }
1531
1532     /* Add the forces from enforced rotation potentials (if any) */
1533     if (inputrec->bRot)
1534     {
1535         wallcycle_start(wcycle, ewcROTadd);
1536         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1537         wallcycle_stop(wcycle, ewcROTadd);
1538     }
1539
1540     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
1541     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
1542
1543     if (PAR(cr) && !(cr->duty & DUTY_PME))
1544     {
1545         /* In case of node-splitting, the PP nodes receive the long-range
1546          * forces, virial and energy from the PME nodes here.
1547          */
1548         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1549     }
1550
1551     if (bDoForces)
1552     {
1553         post_process_forces(cr, step, nrnb, wcycle,
1554                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1555                             flags);
1556     }
1557
1558     /* Sum the potential energy terms from group contributions */
1559     sum_epot(&(enerd->grpp), enerd->term);
1560 }
1561
1562 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1563                         t_inputrec *inputrec,
1564                         gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1565                         gmx_localtop_t *top,
1566                         gmx_groups_t *groups,
1567                         matrix box, rvec x[], history_t *hist,
1568                         rvec f[],
1569                         tensor vir_force,
1570                         t_mdatoms *mdatoms,
1571                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1572                         real *lambda, t_graph *graph,
1573                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1574                         double t, FILE *field, gmx_edsam_t ed,
1575                         gmx_bool bBornRadii,
1576                         int flags)
1577 {
1578     int        cg0, cg1, i, j;
1579     int        start, homenr;
1580     double     mu[2*DIM];
1581     gmx_bool   bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1582     gmx_bool   bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1583     gmx_bool   bDoAdressWF;
1584     matrix     boxs;
1585     rvec       vzero, box_diag;
1586     real       e, v, dvdlambda[efptNR];
1587     t_pbc      pbc;
1588     float      cycles_pme, cycles_force;
1589
1590     start  = 0;
1591     homenr = mdatoms->homenr;
1592
1593     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
1594
1595     clear_mat(vir_force);
1596
1597     cg0 = 0;
1598     if (DOMAINDECOMP(cr))
1599     {
1600         cg1 = cr->dd->ncg_tot;
1601     }
1602     else
1603     {
1604         cg1 = top->cgs.nr;
1605     }
1606     if (fr->n_tpi > 0)
1607     {
1608         cg1--;
1609     }
1610
1611     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1612     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1613     /* Should we update the long-range neighborlists at this step? */
1614     bDoLongRangeNS = fr->bTwinRange && bNS;
1615     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1616     bFillGrid      = (bNS && bStateChanged);
1617     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1618     bDoForces      = (flags & GMX_FORCE_FORCES);
1619     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1620     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1621                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1622
1623     /* should probably move this to the forcerec since it doesn't change */
1624     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1625
1626     if (bStateChanged)
1627     {
1628         update_forcerec(fr, box);
1629
1630         if (NEED_MUTOT(*inputrec))
1631         {
1632             /* Calculate total (local) dipole moment in a temporary common array.
1633              * This makes it possible to sum them over nodes faster.
1634              */
1635             calc_mu(start, homenr,
1636                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1637                     mu, mu+DIM);
1638         }
1639     }
1640
1641     if (fr->ePBC != epbcNONE)
1642     {
1643         /* Compute shift vectors every step,
1644          * because of pressure coupling or box deformation!
1645          */
1646         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1647         {
1648             calc_shifts(box, fr->shift_vec);
1649         }
1650
1651         if (bCalcCGCM)
1652         {
1653             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1654                                      &(top->cgs), x, fr->cg_cm);
1655             inc_nrnb(nrnb, eNR_CGCM, homenr);
1656             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1657         }
1658         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1659         {
1660             unshift_self(graph, box, x);
1661         }
1662     }
1663     else if (bCalcCGCM)
1664     {
1665         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1666         inc_nrnb(nrnb, eNR_CGCM, homenr);
1667     }
1668
1669     if (bCalcCGCM && gmx_debug_at)
1670     {
1671         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1672     }
1673
1674 #ifdef GMX_MPI
1675     if (!(cr->duty & DUTY_PME))
1676     {
1677         /* Send particle coordinates to the pme nodes.
1678          * Since this is only implemented for domain decomposition
1679          * and domain decomposition does not use the graph,
1680          * we do not need to worry about shifting.
1681          */
1682
1683         int pme_flags = 0;
1684
1685         wallcycle_start(wcycle, ewcPP_PMESENDX);
1686
1687         bBS = (inputrec->nwall == 2);
1688         if (bBS)
1689         {
1690             copy_mat(box, boxs);
1691             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1692         }
1693
1694         if (EEL_PME(fr->eeltype))
1695         {
1696             pme_flags |= GMX_PME_DO_COULOMB;
1697         }
1698
1699         if (EVDW_PME(fr->vdwtype))
1700         {
1701             pme_flags |= GMX_PME_DO_LJ;
1702         }
1703
1704         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
1705                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
1706                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
1707                                  pme_flags, step);
1708
1709         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1710     }
1711 #endif /* GMX_MPI */
1712
1713     /* Communicate coordinates and sum dipole if necessary */
1714     if (DOMAINDECOMP(cr))
1715     {
1716         wallcycle_start(wcycle, ewcMOVEX);
1717         dd_move_x(cr->dd, box, x);
1718         wallcycle_stop(wcycle, ewcMOVEX);
1719     }
1720
1721     /* update adress weight beforehand */
1722     if (bStateChanged && bDoAdressWF)
1723     {
1724         /* need pbc for adress weight calculation with pbc_dx */
1725         set_pbc(&pbc, inputrec->ePBC, box);
1726         if (fr->adress_site == eAdressSITEcog)
1727         {
1728             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1729                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1730         }
1731         else if (fr->adress_site == eAdressSITEcom)
1732         {
1733             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1734                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1735         }
1736         else if (fr->adress_site == eAdressSITEatomatom)
1737         {
1738             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1739                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1740         }
1741         else
1742         {
1743             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1744                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1745         }
1746     }
1747
1748     if (NEED_MUTOT(*inputrec))
1749     {
1750
1751         if (bStateChanged)
1752         {
1753             if (PAR(cr))
1754             {
1755                 gmx_sumd(2*DIM, mu, cr);
1756             }
1757             for (i = 0; i < 2; i++)
1758             {
1759                 for (j = 0; j < DIM; j++)
1760                 {
1761                     fr->mu_tot[i][j] = mu[i*DIM + j];
1762                 }
1763             }
1764         }
1765         if (fr->efep == efepNO)
1766         {
1767             copy_rvec(fr->mu_tot[0], mu_tot);
1768         }
1769         else
1770         {
1771             for (j = 0; j < DIM; j++)
1772             {
1773                 mu_tot[j] =
1774                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1775             }
1776         }
1777     }
1778
1779     /* Reset energies */
1780     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1781     clear_rvecs(SHIFTS, fr->fshift);
1782
1783     if (bNS)
1784     {
1785         wallcycle_start(wcycle, ewcNS);
1786
1787         if (graph && bStateChanged)
1788         {
1789             /* Calculate intramolecular shift vectors to make molecules whole */
1790             mk_mshift(fplog, graph, fr->ePBC, box, x);
1791         }
1792
1793         /* Do the actual neighbour searching */
1794         ns(fplog, fr, box,
1795            groups, top, mdatoms,
1796            cr, nrnb, bFillGrid,
1797            bDoLongRangeNS);
1798
1799         wallcycle_stop(wcycle, ewcNS);
1800     }
1801
1802     if (inputrec->implicit_solvent && bNS)
1803     {
1804         make_gb_nblist(cr, inputrec->gb_algorithm,
1805                        x, box, fr, &top->idef, graph, fr->born);
1806     }
1807
1808     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1809     {
1810         wallcycle_start(wcycle, ewcPPDURINGPME);
1811         dd_force_flop_start(cr->dd, nrnb);
1812     }
1813
1814     if (inputrec->bRot)
1815     {
1816         /* Enforced rotation has its own cycle counter that starts after the collective
1817          * coordinates have been communicated. It is added to ddCyclF to allow
1818          * for proper load-balancing */
1819         wallcycle_start(wcycle, ewcROT);
1820         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1821         wallcycle_stop(wcycle, ewcROT);
1822     }
1823
1824     /* Start the force cycle counter.
1825      * This counter is stopped in do_forcelow_level.
1826      * No parallel communication should occur while this counter is running,
1827      * since that will interfere with the dynamic load balancing.
1828      */
1829     wallcycle_start(wcycle, ewcFORCE);
1830
1831     if (bDoForces)
1832     {
1833         /* Reset forces for which the virial is calculated separately:
1834          * PME/Ewald forces if necessary */
1835         if (fr->bF_NoVirSum)
1836         {
1837             if (flags & GMX_FORCE_VIRIAL)
1838             {
1839                 fr->f_novirsum = fr->f_novirsum_alloc;
1840                 if (fr->bDomDec)
1841                 {
1842                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1843                 }
1844                 else
1845                 {
1846                     clear_rvecs(homenr, fr->f_novirsum+start);
1847                 }
1848             }
1849             else
1850             {
1851                 /* We are not calculating the pressure so we do not need
1852                  * a separate array for forces that do not contribute
1853                  * to the pressure.
1854                  */
1855                 fr->f_novirsum = f;
1856             }
1857         }
1858
1859         /* Clear the short- and long-range forces */
1860         clear_rvecs(fr->natoms_force_constr, f);
1861         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1862         {
1863             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1864         }
1865
1866         clear_rvec(fr->vir_diag_posres);
1867     }
1868     if (inputrec->ePull == epullCONSTRAINT)
1869     {
1870         clear_pull_forces(inputrec->pull);
1871     }
1872
1873     /* update QMMMrec, if necessary */
1874     if (fr->bQMMM)
1875     {
1876         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1877     }
1878
1879     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1880     {
1881         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1882                        enerd, lambda, fr);
1883     }
1884
1885     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1886     {
1887         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1888     }
1889
1890     /* Compute the bonded and non-bonded energies and optionally forces */
1891     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1892                       cr, nrnb, wcycle, mdatoms,
1893                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1894                       &(top->atomtypes), bBornRadii, box,
1895                       inputrec->fepvals, lambda,
1896                       graph, &(top->excls), fr->mu_tot,
1897                       flags,
1898                       &cycles_pme);
1899
1900     if (bSepLRF)
1901     {
1902         if (do_per_step(step, inputrec->nstcalclr))
1903         {
1904             /* Add the long range forces to the short range forces */
1905             for (i = 0; i < fr->natoms_force_constr; i++)
1906             {
1907                 rvec_add(fr->f_twin[i], f[i], f[i]);
1908             }
1909         }
1910     }
1911
1912     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1913
1914     if (ed)
1915     {
1916         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1917     }
1918
1919     if (DOMAINDECOMP(cr))
1920     {
1921         dd_force_flop_stop(cr->dd, nrnb);
1922         if (wcycle)
1923         {
1924             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1925         }
1926     }
1927
1928     if (bDoForces)
1929     {
1930         if (IR_ELEC_FIELD(*inputrec))
1931         {
1932             /* Compute forces due to electric field */
1933             calc_f_el(MASTER(cr) ? field : NULL,
1934                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1935                       inputrec->ex, inputrec->et, t);
1936         }
1937
1938         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1939         {
1940             /* Compute thermodynamic force in hybrid AdResS region */
1941             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1942                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1943         }
1944
1945         /* Communicate the forces */
1946         if (DOMAINDECOMP(cr))
1947         {
1948             wallcycle_start(wcycle, ewcMOVEF);
1949             dd_move_f(cr->dd, f, fr->fshift);
1950             /* Do we need to communicate the separate force array
1951              * for terms that do not contribute to the single sum virial?
1952              * Position restraints and electric fields do not introduce
1953              * inter-cg forces, only full electrostatics methods do.
1954              * When we do not calculate the virial, fr->f_novirsum = f,
1955              * so we have already communicated these forces.
1956              */
1957             if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1958                 (flags & GMX_FORCE_VIRIAL))
1959             {
1960                 dd_move_f(cr->dd, fr->f_novirsum, NULL);
1961             }
1962             if (bSepLRF)
1963             {
1964                 /* We should not update the shift forces here,
1965                  * since f_twin is already included in f.
1966                  */
1967                 dd_move_f(cr->dd, fr->f_twin, NULL);
1968             }
1969             wallcycle_stop(wcycle, ewcMOVEF);
1970         }
1971
1972         /* If we have NoVirSum forces, but we do not calculate the virial,
1973          * we sum fr->f_novirum=f later.
1974          */
1975         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1976         {
1977             wallcycle_start(wcycle, ewcVSITESPREAD);
1978             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1979                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1980             wallcycle_stop(wcycle, ewcVSITESPREAD);
1981
1982             if (bSepLRF)
1983             {
1984                 wallcycle_start(wcycle, ewcVSITESPREAD);
1985                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1986                                nrnb,
1987                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1988                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1989             }
1990         }
1991
1992         if (flags & GMX_FORCE_VIRIAL)
1993         {
1994             /* Calculation of the virial must be done after vsites! */
1995             calc_virial(0, mdatoms->homenr, x, f,
1996                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1997         }
1998     }
1999
2000     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
2001     {
2002         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
2003                                f, vir_force, mdatoms, enerd, lambda, t);
2004     }
2005
2006     /* Add the forces from enforced rotation potentials (if any) */
2007     if (inputrec->bRot)
2008     {
2009         wallcycle_start(wcycle, ewcROTadd);
2010         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
2011         wallcycle_stop(wcycle, ewcROTadd);
2012     }
2013
2014     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
2015     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
2016
2017     if (PAR(cr) && !(cr->duty & DUTY_PME))
2018     {
2019         /* In case of node-splitting, the PP nodes receive the long-range
2020          * forces, virial and energy from the PME nodes here.
2021          */
2022         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
2023     }
2024
2025     if (bDoForces)
2026     {
2027         post_process_forces(cr, step, nrnb, wcycle,
2028                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
2029                             flags);
2030     }
2031
2032     /* Sum the potential energy terms from group contributions */
2033     sum_epot(&(enerd->grpp), enerd->term);
2034 }
2035
2036 void do_force(FILE *fplog, t_commrec *cr,
2037               t_inputrec *inputrec,
2038               gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
2039               gmx_localtop_t *top,
2040               gmx_groups_t *groups,
2041               matrix box, rvec x[], history_t *hist,
2042               rvec f[],
2043               tensor vir_force,
2044               t_mdatoms *mdatoms,
2045               gmx_enerdata_t *enerd, t_fcdata *fcd,
2046               real *lambda, t_graph *graph,
2047               t_forcerec *fr,
2048               gmx_vsite_t *vsite, rvec mu_tot,
2049               double t, FILE *field, gmx_edsam_t ed,
2050               gmx_bool bBornRadii,
2051               int flags)
2052 {
2053     /* modify force flag if not doing nonbonded */
2054     if (!fr->bNonbonded)
2055     {
2056         flags &= ~GMX_FORCE_NONBONDED;
2057     }
2058
2059     switch (inputrec->cutoff_scheme)
2060     {
2061         case ecutsVERLET:
2062             do_force_cutsVERLET(fplog, cr, inputrec,
2063                                 step, nrnb, wcycle,
2064                                 top,
2065                                 groups,
2066                                 box, x, hist,
2067                                 f, vir_force,
2068                                 mdatoms,
2069                                 enerd, fcd,
2070                                 lambda, graph,
2071                                 fr, fr->ic,
2072                                 vsite, mu_tot,
2073                                 t, field, ed,
2074                                 bBornRadii,
2075                                 flags);
2076             break;
2077         case ecutsGROUP:
2078             do_force_cutsGROUP(fplog, cr, inputrec,
2079                                step, nrnb, wcycle,
2080                                top,
2081                                groups,
2082                                box, x, hist,
2083                                f, vir_force,
2084                                mdatoms,
2085                                enerd, fcd,
2086                                lambda, graph,
2087                                fr, vsite, mu_tot,
2088                                t, field, ed,
2089                                bBornRadii,
2090                                flags);
2091             break;
2092         default:
2093             gmx_incons("Invalid cut-off scheme passed!");
2094     }
2095 }
2096
2097
2098 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2099                         t_inputrec *ir, t_mdatoms *md,
2100                         t_state *state, t_commrec *cr, t_nrnb *nrnb,
2101                         t_forcerec *fr, gmx_localtop_t *top)
2102 {
2103     int             i, m, start, end;
2104     gmx_int64_t     step;
2105     real            dt = ir->delta_t;
2106     real            dvdl_dum;
2107     rvec           *savex;
2108
2109     snew(savex, state->natoms);
2110
2111     start = 0;
2112     end   = md->homenr;
2113
2114     if (debug)
2115     {
2116         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2117                 start, md->homenr, end);
2118     }
2119     /* Do a first constrain to reset particles... */
2120     step = ir->init_step;
2121     if (fplog)
2122     {
2123         char buf[STEPSTRSIZE];
2124         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2125                 gmx_step_str(step, buf));
2126     }
2127     dvdl_dum = 0;
2128
2129     /* constrain the current position */
2130     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2131               ir, NULL, cr, step, 0, 1.0, md,
2132               state->x, state->x, NULL,
2133               fr->bMolPBC, state->box,
2134               state->lambda[efptBONDED], &dvdl_dum,
2135               NULL, NULL, nrnb, econqCoord,
2136               ir->epc == epcMTTK, state->veta, state->veta);
2137     if (EI_VV(ir->eI))
2138     {
2139         /* constrain the inital velocity, and save it */
2140         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2141         /* might not yet treat veta correctly */
2142         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2143                   ir, NULL, cr, step, 0, 1.0, md,
2144                   state->x, state->v, state->v,
2145                   fr->bMolPBC, state->box,
2146                   state->lambda[efptBONDED], &dvdl_dum,
2147                   NULL, NULL, nrnb, econqVeloc,
2148                   ir->epc == epcMTTK, state->veta, state->veta);
2149     }
2150     /* constrain the inital velocities at t-dt/2 */
2151     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2152     {
2153         for (i = start; (i < end); i++)
2154         {
2155             for (m = 0; (m < DIM); m++)
2156             {
2157                 /* Reverse the velocity */
2158                 state->v[i][m] = -state->v[i][m];
2159                 /* Store the position at t-dt in buf */
2160                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2161             }
2162         }
2163         /* Shake the positions at t=-dt with the positions at t=0
2164          * as reference coordinates.
2165          */
2166         if (fplog)
2167         {
2168             char buf[STEPSTRSIZE];
2169             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2170                     gmx_step_str(step, buf));
2171         }
2172         dvdl_dum = 0;
2173         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2174                   ir, NULL, cr, step, -1, 1.0, md,
2175                   state->x, savex, NULL,
2176                   fr->bMolPBC, state->box,
2177                   state->lambda[efptBONDED], &dvdl_dum,
2178                   state->v, NULL, nrnb, econqCoord,
2179                   ir->epc == epcMTTK, state->veta, state->veta);
2180
2181         for (i = start; i < end; i++)
2182         {
2183             for (m = 0; m < DIM; m++)
2184             {
2185                 /* Re-reverse the velocities */
2186                 state->v[i][m] = -state->v[i][m];
2187             }
2188         }
2189     }
2190     sfree(savex);
2191 }
2192
2193
2194 static void
2195 integrate_table(real vdwtab[], real scale, int offstart, int rstart, int rend,
2196                 double *enerout, double *virout)
2197 {
2198     double enersum, virsum;
2199     double invscale, invscale2, invscale3;
2200     double r, ea, eb, ec, pa, pb, pc, pd;
2201     double y0, f, g, h;
2202     int    ri, offset, tabfactor;
2203
2204     invscale  = 1.0/scale;
2205     invscale2 = invscale*invscale;
2206     invscale3 = invscale*invscale2;
2207
2208     /* Following summation derived from cubic spline definition,
2209      * Numerical Recipies in C, second edition, p. 113-116.  Exact for
2210      * the cubic spline.  We first calculate the negative of the
2211      * energy from rvdw to rvdw_switch, assuming that g(r)=1, and then
2212      * add the more standard, abrupt cutoff correction to that result,
2213      * yielding the long-range correction for a switched function.  We
2214      * perform both the pressure and energy loops at the same time for
2215      * simplicity, as the computational cost is low. */
2216
2217     if (offstart == 0)
2218     {
2219         /* Since the dispersion table has been scaled down a factor
2220          * 6.0 and the repulsion a factor 12.0 to compensate for the
2221          * c6/c12 parameters inside nbfp[] being scaled up (to save
2222          * flops in kernels), we need to correct for this.
2223          */
2224         tabfactor = 6.0;
2225     }
2226     else
2227     {
2228         tabfactor = 12.0;
2229     }
2230
2231     enersum = 0.0;
2232     virsum  = 0.0;
2233     for (ri = rstart; ri < rend; ++ri)
2234     {
2235         r  = ri*invscale;
2236         ea = invscale3;
2237         eb = 2.0*invscale2*r;
2238         ec = invscale*r*r;
2239
2240         pa = invscale3;
2241         pb = 3.0*invscale2*r;
2242         pc = 3.0*invscale*r*r;
2243         pd = r*r*r;
2244
2245         /* this "8" is from the packing in the vdwtab array - perhaps
2246            should be defined? */
2247
2248         offset = 8*ri + offstart;
2249         y0     = vdwtab[offset];
2250         f      = vdwtab[offset+1];
2251         g      = vdwtab[offset+2];
2252         h      = vdwtab[offset+3];
2253
2254         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);
2255         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);
2256     }
2257     *enerout = 4.0*M_PI*enersum*tabfactor;
2258     *virout  = 4.0*M_PI*virsum*tabfactor;
2259 }
2260
2261 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2262 {
2263     double   eners[2], virs[2], enersum, virsum, y0, f, g, h;
2264     double   r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2265     double   invscale, invscale2, invscale3;
2266     int      ri0, ri1, ri, i, offstart, offset;
2267     real     scale, *vdwtab, tabfactor, tmp;
2268
2269     fr->enershiftsix    = 0;
2270     fr->enershifttwelve = 0;
2271     fr->enerdiffsix     = 0;
2272     fr->enerdifftwelve  = 0;
2273     fr->virdiffsix      = 0;
2274     fr->virdifftwelve   = 0;
2275
2276     if (eDispCorr != edispcNO)
2277     {
2278         for (i = 0; i < 2; i++)
2279         {
2280             eners[i] = 0;
2281             virs[i]  = 0;
2282         }
2283         if ((fr->vdw_modifier == eintmodPOTSHIFT) ||
2284             (fr->vdw_modifier == eintmodPOTSWITCH) ||
2285             (fr->vdw_modifier == eintmodFORCESWITCH) ||
2286             (fr->vdwtype == evdwSHIFT) ||
2287             (fr->vdwtype == evdwSWITCH))
2288         {
2289             if (((fr->vdw_modifier == eintmodPOTSWITCH) ||
2290                  (fr->vdw_modifier == eintmodFORCESWITCH) ||
2291                  (fr->vdwtype == evdwSWITCH)) && fr->rvdw_switch == 0)
2292             {
2293                 gmx_fatal(FARGS,
2294                           "With dispersion correction rvdw-switch can not be zero "
2295                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2296             }
2297
2298             scale  = fr->nblists[0].table_vdw.scale;
2299             vdwtab = fr->nblists[0].table_vdw.data;
2300
2301             /* Round the cut-offs to exact table values for precision */
2302             ri0  = floor(fr->rvdw_switch*scale);
2303             ri1  = ceil(fr->rvdw*scale);
2304
2305             /* The code below has some support for handling force-switching, i.e.
2306              * when the force (instead of potential) is switched over a limited
2307              * region. This leads to a constant shift in the potential inside the
2308              * switching region, which we can handle by adding a constant energy
2309              * term in the force-switch case just like when we do potential-shift.
2310              *
2311              * For now this is not enabled, but to keep the functionality in the
2312              * code we check separately for switch and shift. When we do force-switch
2313              * the shifting point is rvdw_switch, while it is the cutoff when we
2314              * have a classical potential-shift.
2315              *
2316              * For a pure potential-shift the potential has a constant shift
2317              * all the way out to the cutoff, and that is it. For other forms
2318              * we need to calculate the constant shift up to the point where we
2319              * start modifying the potential.
2320              */
2321             ri0  = (fr->vdw_modifier == eintmodPOTSHIFT) ? ri1 : ri0;
2322
2323             r0   = ri0/scale;
2324             r1   = ri1/scale;
2325             rc3  = r0*r0*r0;
2326             rc9  = rc3*rc3*rc3;
2327
2328             if ((fr->vdw_modifier == eintmodFORCESWITCH) ||
2329                 (fr->vdwtype == evdwSHIFT))
2330             {
2331                 /* Determine the constant energy shift below rvdw_switch.
2332                  * Table has a scale factor since we have scaled it down to compensate
2333                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2334                  */
2335                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2336                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2337             }
2338             else if (fr->vdw_modifier == eintmodPOTSHIFT)
2339             {
2340                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3));
2341                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3));
2342             }
2343
2344             /* Add the constant part from 0 to rvdw_switch.
2345              * This integration from 0 to rvdw_switch overcounts the number
2346              * of interactions by 1, as it also counts the self interaction.
2347              * We will correct for this later.
2348              */
2349             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2350             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2351
2352             /* Calculate the contribution in the range [r0,r1] where we
2353              * modify the potential. For a pure potential-shift modifier we will
2354              * have ri0==ri1, and there will not be any contribution here.
2355              */
2356             for (i = 0; i < 2; i++)
2357             {
2358                 enersum = 0;
2359                 virsum  = 0;
2360                 integrate_table(vdwtab, scale, (i == 0 ? 0 : 4), ri0, ri1, &enersum, &virsum);
2361                 eners[i] -= enersum;
2362                 virs[i]  -= virsum;
2363             }
2364
2365             /* Alright: Above we compensated by REMOVING the parts outside r0
2366              * corresponding to the ideal VdW 1/r6 and /r12 potentials.
2367              *
2368              * Regardless of whether r0 is the point where we start switching,
2369              * or the cutoff where we calculated the constant shift, we include
2370              * all the parts we are missing out to infinity from r0 by
2371              * calculating the analytical dispersion correction.
2372              */
2373             eners[0] += -4.0*M_PI/(3.0*rc3);
2374             eners[1] +=  4.0*M_PI/(9.0*rc9);
2375             virs[0]  +=  8.0*M_PI/rc3;
2376             virs[1]  += -16.0*M_PI/(3.0*rc9);
2377         }
2378         else if (fr->vdwtype == evdwCUT ||
2379                  EVDW_PME(fr->vdwtype) ||
2380                  fr->vdwtype == evdwUSER)
2381         {
2382             if (fr->vdwtype == evdwUSER && fplog)
2383             {
2384                 fprintf(fplog,
2385                         "WARNING: using dispersion correction with user tables\n");
2386             }
2387
2388             /* Note that with LJ-PME, the dispersion correction is multiplied
2389              * by the difference between the actual C6 and the value of C6
2390              * that would produce the combination rule.
2391              * This means the normal energy and virial difference formulas
2392              * can be used here.
2393              */
2394
2395             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2396             rc9  = rc3*rc3*rc3;
2397             /* Contribution beyond the cut-off */
2398             eners[0] += -4.0*M_PI/(3.0*rc3);
2399             eners[1] +=  4.0*M_PI/(9.0*rc9);
2400             if (fr->vdw_modifier == eintmodPOTSHIFT)
2401             {
2402                 /* Contribution within the cut-off */
2403                 eners[0] += -4.0*M_PI/(3.0*rc3);
2404                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2405             }
2406             /* Contribution beyond the cut-off */
2407             virs[0]  +=  8.0*M_PI/rc3;
2408             virs[1]  += -16.0*M_PI/(3.0*rc9);
2409         }
2410         else
2411         {
2412             gmx_fatal(FARGS,
2413                       "Dispersion correction is not implemented for vdw-type = %s",
2414                       evdw_names[fr->vdwtype]);
2415         }
2416
2417         /* When we deprecate the group kernels the code below can go too */
2418         if (fr->vdwtype == evdwPME && fr->cutoff_scheme == ecutsGROUP)
2419         {
2420             /* Calculate self-interaction coefficient (assuming that
2421              * the reciprocal-space contribution is constant in the
2422              * region that contributes to the self-interaction).
2423              */
2424             fr->enershiftsix = pow(fr->ewaldcoeff_lj, 6) / 6.0;
2425
2426             eners[0] += -pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3)/3.0;
2427             virs[0]  +=  pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3);
2428         }
2429
2430         fr->enerdiffsix    = eners[0];
2431         fr->enerdifftwelve = eners[1];
2432         /* The 0.5 is due to the Gromacs definition of the virial */
2433         fr->virdiffsix     = 0.5*virs[0];
2434         fr->virdifftwelve  = 0.5*virs[1];
2435     }
2436 }
2437
2438 void calc_dispcorr(FILE *fplog, t_inputrec *ir, t_forcerec *fr,
2439                    gmx_int64_t step, int natoms,
2440                    matrix box, real lambda, tensor pres, tensor virial,
2441                    real *prescorr, real *enercorr, real *dvdlcorr)
2442 {
2443     gmx_bool bCorrAll, bCorrPres;
2444     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2445     int      m;
2446
2447     *prescorr = 0;
2448     *enercorr = 0;
2449     *dvdlcorr = 0;
2450
2451     clear_mat(virial);
2452     clear_mat(pres);
2453
2454     if (ir->eDispCorr != edispcNO)
2455     {
2456         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2457                      ir->eDispCorr == edispcAllEnerPres);
2458         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2459                      ir->eDispCorr == edispcAllEnerPres);
2460
2461         invvol = 1/det(box);
2462         if (fr->n_tpi)
2463         {
2464             /* Only correct for the interactions with the inserted molecule */
2465             dens   = (natoms - fr->n_tpi)*invvol;
2466             ninter = fr->n_tpi;
2467         }
2468         else
2469         {
2470             dens   = natoms*invvol;
2471             ninter = 0.5*natoms;
2472         }
2473
2474         if (ir->efep == efepNO)
2475         {
2476             avcsix    = fr->avcsix[0];
2477             avctwelve = fr->avctwelve[0];
2478         }
2479         else
2480         {
2481             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2482             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2483         }
2484
2485         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2486         *enercorr += avcsix*enerdiff;
2487         dvdlambda  = 0.0;
2488         if (ir->efep != efepNO)
2489         {
2490             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2491         }
2492         if (bCorrAll)
2493         {
2494             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2495             *enercorr += avctwelve*enerdiff;
2496             if (fr->efep != efepNO)
2497             {
2498                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2499             }
2500         }
2501
2502         if (bCorrPres)
2503         {
2504             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2505             if (ir->eDispCorr == edispcAllEnerPres)
2506             {
2507                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2508             }
2509             /* The factor 2 is because of the Gromacs virial definition */
2510             spres = -2.0*invvol*svir*PRESFAC;
2511
2512             for (m = 0; m < DIM; m++)
2513             {
2514                 virial[m][m] += svir;
2515                 pres[m][m]   += spres;
2516             }
2517             *prescorr += spres;
2518         }
2519
2520         /* Can't currently control when it prints, for now, just print when degugging */
2521         if (debug)
2522         {
2523             if (bCorrAll)
2524             {
2525                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2526                         avcsix, avctwelve);
2527             }
2528             if (bCorrPres)
2529             {
2530                 fprintf(debug,
2531                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2532                         *enercorr, spres, svir);
2533             }
2534             else
2535             {
2536                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2537             }
2538         }
2539
2540         if (fr->bSepDVDL && do_per_step(step, ir->nstlog))
2541         {
2542             gmx_print_sepdvdl(fplog, "Dispersion correction", *enercorr, dvdlambda);
2543         }
2544         if (fr->efep != efepNO)
2545         {
2546             *dvdlcorr += dvdlambda;
2547         }
2548     }
2549 }
2550
2551 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2552                   t_graph *graph, rvec x[])
2553 {
2554     if (fplog)
2555     {
2556         fprintf(fplog, "Removing pbc first time\n");
2557     }
2558     calc_shifts(box, fr->shift_vec);
2559     if (graph)
2560     {
2561         mk_mshift(fplog, graph, fr->ePBC, box, x);
2562         if (gmx_debug_at)
2563         {
2564             p_graph(debug, "do_pbc_first 1", graph);
2565         }
2566         shift_self(graph, box, x);
2567         /* By doing an extra mk_mshift the molecules that are broken
2568          * because they were e.g. imported from another software
2569          * will be made whole again. Such are the healing powers
2570          * of GROMACS.
2571          */
2572         mk_mshift(fplog, graph, fr->ePBC, box, x);
2573         if (gmx_debug_at)
2574         {
2575             p_graph(debug, "do_pbc_first 2", graph);
2576         }
2577     }
2578     if (fplog)
2579     {
2580         fprintf(fplog, "Done rmpbc\n");
2581     }
2582 }
2583
2584 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2585                             gmx_mtop_t *mtop, rvec x[],
2586                             gmx_bool bFirst)
2587 {
2588     t_graph        *graph;
2589     int             mb, as, mol;
2590     gmx_molblock_t *molb;
2591
2592     if (bFirst && fplog)
2593     {
2594         fprintf(fplog, "Removing pbc first time\n");
2595     }
2596
2597     snew(graph, 1);
2598     as = 0;
2599     for (mb = 0; mb < mtop->nmolblock; mb++)
2600     {
2601         molb = &mtop->molblock[mb];
2602         if (molb->natoms_mol == 1 ||
2603             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2604         {
2605             /* Just one atom or charge group in the molecule, no PBC required */
2606             as += molb->nmol*molb->natoms_mol;
2607         }
2608         else
2609         {
2610             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2611             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2612                            0, molb->natoms_mol, FALSE, FALSE, graph);
2613
2614             for (mol = 0; mol < molb->nmol; mol++)
2615             {
2616                 mk_mshift(fplog, graph, ePBC, box, x+as);
2617
2618                 shift_self(graph, box, x+as);
2619                 /* The molecule is whole now.
2620                  * We don't need the second mk_mshift call as in do_pbc_first,
2621                  * since we no longer need this graph.
2622                  */
2623
2624                 as += molb->natoms_mol;
2625             }
2626             done_graph(graph);
2627         }
2628     }
2629     sfree(graph);
2630 }
2631
2632 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2633                        gmx_mtop_t *mtop, rvec x[])
2634 {
2635     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2636 }
2637
2638 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2639                  gmx_mtop_t *mtop, rvec x[])
2640 {
2641     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2642 }
2643
2644 void finish_run(FILE *fplog, t_commrec *cr,
2645                 t_inputrec *inputrec,
2646                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2647                 gmx_walltime_accounting_t walltime_accounting,
2648                 wallclock_gpu_t *gputimes,
2649                 gmx_bool bWriteStat)
2650 {
2651     int     i, j;
2652     t_nrnb *nrnb_tot = NULL;
2653     real    delta_t;
2654     double  nbfs, mflop;
2655     double  elapsed_time,
2656             elapsed_time_over_all_ranks,
2657             elapsed_time_over_all_threads,
2658             elapsed_time_over_all_threads_over_all_ranks;
2659     wallcycle_sum(cr, wcycle);
2660
2661     if (cr->nnodes > 1)
2662     {
2663         snew(nrnb_tot, 1);
2664 #ifdef GMX_MPI
2665         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2666                       cr->mpi_comm_mysim);
2667 #endif
2668     }
2669     else
2670     {
2671         nrnb_tot = nrnb;
2672     }
2673
2674     elapsed_time                                 = walltime_accounting_get_elapsed_time(walltime_accounting);
2675     elapsed_time_over_all_ranks                  = elapsed_time;
2676     elapsed_time_over_all_threads                = walltime_accounting_get_elapsed_time_over_all_threads(walltime_accounting);
2677     elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
2678 #ifdef GMX_MPI
2679     if (cr->nnodes > 1)
2680     {
2681         /* reduce elapsed_time over all MPI ranks in the current simulation */
2682         MPI_Allreduce(&elapsed_time,
2683                       &elapsed_time_over_all_ranks,
2684                       1, MPI_DOUBLE, MPI_SUM,
2685                       cr->mpi_comm_mysim);
2686         elapsed_time_over_all_ranks /= cr->nnodes;
2687         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
2688          * current simulation. */
2689         MPI_Allreduce(&elapsed_time_over_all_threads,
2690                       &elapsed_time_over_all_threads_over_all_ranks,
2691                       1, MPI_DOUBLE, MPI_SUM,
2692                       cr->mpi_comm_mysim);
2693     }
2694 #endif
2695
2696     if (SIMMASTER(cr))
2697     {
2698         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2699     }
2700     if (cr->nnodes > 1)
2701     {
2702         sfree(nrnb_tot);
2703     }
2704
2705     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2706     {
2707         print_dd_statistics(cr, inputrec, fplog);
2708     }
2709
2710     if (SIMMASTER(cr))
2711     {
2712         wallcycle_print(fplog, cr->nnodes, cr->npmenodes,
2713                         elapsed_time_over_all_ranks,
2714                         wcycle, gputimes);
2715
2716         if (EI_DYNAMICS(inputrec->eI))
2717         {
2718             delta_t = inputrec->delta_t;
2719         }
2720         else
2721         {
2722             delta_t = 0;
2723         }
2724
2725         if (fplog)
2726         {
2727             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
2728                        elapsed_time_over_all_ranks,
2729                        walltime_accounting_get_nsteps_done(walltime_accounting),
2730                        delta_t, nbfs, mflop);
2731         }
2732         if (bWriteStat)
2733         {
2734             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
2735                        elapsed_time_over_all_ranks,
2736                        walltime_accounting_get_nsteps_done(walltime_accounting),
2737                        delta_t, nbfs, mflop);
2738         }
2739     }
2740 }
2741
2742 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2743 {
2744     /* this function works, but could probably use a logic rewrite to keep all the different
2745        types of efep straight. */
2746
2747     int       i;
2748     t_lambda *fep = ir->fepvals;
2749
2750     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2751     {
2752         for (i = 0; i < efptNR; i++)
2753         {
2754             lambda[i] = 0.0;
2755             if (lam0)
2756             {
2757                 lam0[i] = 0.0;
2758             }
2759         }
2760         return;
2761     }
2762     else
2763     {
2764         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2765                                              if checkpoint is set -- a kludge is in for now
2766                                              to prevent this.*/
2767         for (i = 0; i < efptNR; i++)
2768         {
2769             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2770             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2771             {
2772                 lambda[i] = fep->init_lambda;
2773                 if (lam0)
2774                 {
2775                     lam0[i] = lambda[i];
2776                 }
2777             }
2778             else
2779             {
2780                 lambda[i] = fep->all_lambda[i][*fep_state];
2781                 if (lam0)
2782                 {
2783                     lam0[i] = lambda[i];
2784                 }
2785             }
2786         }
2787         if (ir->bSimTemp)
2788         {
2789             /* need to rescale control temperatures to match current state */
2790             for (i = 0; i < ir->opts.ngtc; i++)
2791             {
2792                 if (ir->opts.ref_t[i] > 0)
2793                 {
2794                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2795                 }
2796             }
2797         }
2798     }
2799
2800     /* Send to the log the information on the current lambdas */
2801     if (fplog != NULL)
2802     {
2803         fprintf(fplog, "Initial vector of lambda components:[ ");
2804         for (i = 0; i < efptNR; i++)
2805         {
2806             fprintf(fplog, "%10.4f ", lambda[i]);
2807         }
2808         fprintf(fplog, "]\n");
2809     }
2810     return;
2811 }
2812
2813
2814 void init_md(FILE *fplog,
2815              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2816              double *t, double *t0,
2817              real *lambda, int *fep_state, double *lam0,
2818              t_nrnb *nrnb, gmx_mtop_t *mtop,
2819              gmx_update_t *upd,
2820              int nfile, const t_filenm fnm[],
2821              gmx_mdoutf_t *outf, t_mdebin **mdebin,
2822              tensor force_vir, tensor shake_vir, rvec mu_tot,
2823              gmx_bool *bSimAnn, t_vcm **vcm, unsigned long Flags)
2824 {
2825     int  i, j, n;
2826     real tmpt, mod;
2827
2828     /* Initial values */
2829     *t = *t0       = ir->init_t;
2830
2831     *bSimAnn = FALSE;
2832     for (i = 0; i < ir->opts.ngtc; i++)
2833     {
2834         /* set bSimAnn if any group is being annealed */
2835         if (ir->opts.annealing[i] != eannNO)
2836         {
2837             *bSimAnn = TRUE;
2838         }
2839     }
2840     if (*bSimAnn)
2841     {
2842         update_annealing_target_temp(&(ir->opts), ir->init_t);
2843     }
2844
2845     /* Initialize lambda variables */
2846     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2847
2848     if (upd)
2849     {
2850         *upd = init_update(ir);
2851     }
2852
2853
2854     if (vcm != NULL)
2855     {
2856         *vcm = init_vcm(fplog, &mtop->groups, ir);
2857     }
2858
2859     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2860     {
2861         if (ir->etc == etcBERENDSEN)
2862         {
2863             please_cite(fplog, "Berendsen84a");
2864         }
2865         if (ir->etc == etcVRESCALE)
2866         {
2867             please_cite(fplog, "Bussi2007a");
2868         }
2869         if (ir->eI == eiSD1)
2870         {
2871             please_cite(fplog, "Goga2012");
2872         }
2873     }
2874
2875     init_nrnb(nrnb);
2876
2877     if (nfile != -1)
2878     {
2879         *outf = init_mdoutf(fplog, nfile, fnm, Flags, cr, ir, mtop, oenv);
2880
2881         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : mdoutf_get_fp_ene(*outf),
2882                               mtop, ir, mdoutf_get_fp_dhdl(*outf));
2883     }
2884
2885     if (ir->bAdress)
2886     {
2887         please_cite(fplog, "Fritsch12");
2888         please_cite(fplog, "Junghans10");
2889     }
2890     /* Initiate variables */
2891     clear_mat(force_vir);
2892     clear_mat(shake_vir);
2893     clear_rvec(mu_tot);
2894
2895     debug_gmx();
2896 }