src/gromacs/mdlib/sim_util.cpp

   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) 2013,2014,2015, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "gromacs/legacyheaders/sim_util.h"
  40
  41 #include "config.h"
  42
  43 #include <assert.h>
  44 #include <math.h>
  45 #include <stdio.h>
  46 #include <string.h>
  47
  48 #include "gromacs/domdec/domdec.h"
  49 #include "gromacs/essentialdynamics/edsam.h"
  50 #include "gromacs/ewald/pme.h"
  51 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  52 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  53 #include "gromacs/imd/imd.h"
  54 #include "gromacs/legacyheaders/calcmu.h"
  55 #include "gromacs/legacyheaders/chargegroup.h"
  56 #include "gromacs/legacyheaders/constr.h"
  57 #include "gromacs/legacyheaders/copyrite.h"
  58 #include "gromacs/legacyheaders/disre.h"
  59 #include "gromacs/legacyheaders/force.h"
  60 #include "gromacs/legacyheaders/genborn.h"
  61 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
  62 #include "gromacs/legacyheaders/mdatoms.h"
  63 #include "gromacs/legacyheaders/mdrun.h"
  64 #include "gromacs/legacyheaders/names.h"
  65 #include "gromacs/legacyheaders/network.h"
  66 #include "gromacs/legacyheaders/nonbonded.h"
  67 #include "gromacs/legacyheaders/nrnb.h"
  68 #include "gromacs/legacyheaders/orires.h"
  69 #include "gromacs/legacyheaders/qmmm.h"
  70 #include "gromacs/legacyheaders/txtdump.h"
  71 #include "gromacs/legacyheaders/typedefs.h"
  72 #include "gromacs/legacyheaders/update.h"
  73 #include "gromacs/legacyheaders/types/commrec.h"
  74 #include "gromacs/listed-forces/bonded.h"
  75 #include "gromacs/math/units.h"
  76 #include "gromacs/math/vec.h"
  77 #include "gromacs/mdlib/nb_verlet.h"
  78 #include "gromacs/mdlib/nbnxn_atomdata.h"
  79 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
  80 #include "gromacs/mdlib/nbnxn_search.h"
  81 #include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  82 #include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_ref.h"
  83 #include "gromacs/mdlib/nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn.h"
  84 #include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.h"
  85 #include "gromacs/pbcutil/ishift.h"
  86 #include "gromacs/pbcutil/mshift.h"
  87 #include "gromacs/pbcutil/pbc.h"
  88 #include "gromacs/pulling/pull.h"
  89 #include "gromacs/pulling/pull_rotation.h"
  90 #include "gromacs/timing/gpu_timing.h"
  91 #include "gromacs/timing/wallcycle.h"
  92 #include "gromacs/timing/walltime_accounting.h"
  93 #include "gromacs/utility/cstringutil.h"
  94 #include "gromacs/utility/gmxmpi.h"
  95 #include "gromacs/utility/smalloc.h"
  96 #include "gromacs/utility/sysinfo.h"
  97
  98 #include "adress.h"
  99 #include "nbnxn_gpu.h"
 100
 101 void print_time(FILE                     *out,
 102                 gmx_walltime_accounting_t walltime_accounting,
 103                 gmx_int64_t               step,
 104                 t_inputrec               *ir,
 105                 t_commrec gmx_unused     *cr)
 106 {
 107     time_t finish;
 108     char   timebuf[STRLEN];
 109     double dt, elapsed_seconds, time_per_step;
 110     char   buf[48];
 111
 112 #ifndef GMX_THREAD_MPI
 113     if (!PAR(cr))
 114 #endif
 115     {
 116         fprintf(out, "\r");
 117     }
 118     fprintf(out, "step %s", gmx_step_str(step, buf));
 119     if ((step >= ir->nstlist))
 120     {
 121         double seconds_since_epoch = gmx_gettime();
 122         elapsed_seconds = seconds_since_epoch - walltime_accounting_get_start_time_stamp(walltime_accounting);
 123         time_per_step   = elapsed_seconds/(step - ir->init_step + 1);
 124         dt              = (ir->nsteps + ir->init_step - step) * time_per_step;
 125
 126         if (ir->nsteps >= 0)
 127         {
 128             if (dt >= 300)
 129             {
 130                 finish = (time_t) (seconds_since_epoch + dt);
 131                 gmx_ctime_r(&finish, timebuf, STRLEN);
 132                 sprintf(buf, "%s", timebuf);
 133                 buf[strlen(buf)-1] = '\0';
 134                 fprintf(out, ", will finish %s", buf);
 135             }
 136             else
 137             {
 138                 fprintf(out, ", remaining wall clock time: %5d s          ", (int)dt);
 139             }
 140         }
 141         else
 142         {
 143             fprintf(out, " performance: %.1f ns/day    ",
 144                     ir->delta_t/1000*24*60*60/time_per_step);
 145         }
 146     }
 147 #ifndef GMX_THREAD_MPI
 148     if (PAR(cr))
 149     {
 150         fprintf(out, "\n");
 151     }
 152 #endif
 153
 154     fflush(out);
 155 }
 156
 157 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
 158                          double the_time)
 159 {
 160     char   time_string[STRLEN];
 161
 162     if (!fplog)
 163     {
 164         return;
 165     }
 166
 167     {
 168         int    i;
 169         char   timebuf[STRLEN];
 170         time_t temp_time = (time_t) the_time;
 171
 172         gmx_ctime_r(&temp_time, timebuf, STRLEN);
 173         for (i = 0; timebuf[i] >= ' '; i++)
 174         {
 175             time_string[i] = timebuf[i];
 176         }
 177         time_string[i] = '\0';
 178     }
 179
 180     fprintf(fplog, "%s on rank %d %s\n", title, nodeid, time_string);
 181 }
 182
 183 void print_start(FILE *fplog, t_commrec *cr,
 184                  gmx_walltime_accounting_t walltime_accounting,
 185                  const char *name)
 186 {
 187     char buf[STRLEN];
 188
 189     sprintf(buf, "Started %s", name);
 190     print_date_and_time(fplog, cr->nodeid, buf,
 191                         walltime_accounting_get_start_time_stamp(walltime_accounting));
 192 }
 193
 194 static void sum_forces(int start, int end, rvec f[], rvec flr[])
 195 {
 196     int i;
 197
 198     if (gmx_debug_at)
 199     {
 200         pr_rvecs(debug, 0, "fsr", f+start, end-start);
 201         pr_rvecs(debug, 0, "flr", flr+start, end-start);
 202     }
 203     for (i = start; (i < end); i++)
 204     {
 205         rvec_inc(f[i], flr[i]);
 206     }
 207 }
 208
 209 /*
 210  * calc_f_el calculates forces due to an electric field.
 211  *
 212  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 213  *
 214  * Et[] contains the parameters for the time dependent
 215  * part of the field.
 216  * Ex[] contains the parameters for
 217  * the spatial dependent part of the field.
 218  * The function should return the energy due to the electric field
 219  * (if any) but for now returns 0.
 220  *
 221  * WARNING:
 222  * There can be problems with the virial.
 223  * Since the field is not self-consistent this is unavoidable.
 224  * For neutral molecules the virial is correct within this approximation.
 225  * For neutral systems with many charged molecules the error is small.
 226  * But for systems with a net charge or a few charged molecules
 227  * the error can be significant when the field is high.
 228  * Solution: implement a self-consistent electric field into PME.
 229  */
 230 static void calc_f_el(FILE *fp, int  start, int homenr,
 231                       real charge[], rvec f[],
 232                       t_cosines Ex[], t_cosines Et[], double t)
 233 {
 234     rvec Ext;
 235     real t0;
 236     int  i, m;
 237
 238     for (m = 0; (m < DIM); m++)
 239     {
 240         if (Et[m].n > 0)
 241         {
 242             if (Et[m].n == 3)
 243             {
 244                 t0     = Et[m].a[1];
 245                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 246             }
 247             else
 248             {
 249                 Ext[m] = cos(Et[m].a[0]*t);
 250             }
 251         }
 252         else
 253         {
 254             Ext[m] = 1.0;
 255         }
 256         if (Ex[m].n > 0)
 257         {
 258             /* Convert the field strength from V/nm to MD-units */
 259             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 260             for (i = start; (i < start+homenr); i++)
 261             {
 262                 f[i][m] += charge[i]*Ext[m];
 263             }
 264         }
 265         else
 266         {
 267             Ext[m] = 0;
 268         }
 269     }
 270     if (fp != NULL)
 271     {
 272         fprintf(fp, "%10g  %10g  %10g  %10g #FIELD\n", t,
 273                 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
 274     }
 275 }
 276
 277 static void calc_virial(int start, int homenr, rvec x[], rvec f[],
 278                         tensor vir_part, t_graph *graph, matrix box,
 279                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 280 {
 281     int    i;
 282
 283     /* The short-range virial from surrounding boxes */
 284     clear_mat(vir_part);
 285     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 286     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 287
 288     /* Calculate partial virial, for local atoms only, based on short range.
 289      * Total virial is computed in global_stat, called from do_md
 290      */
 291     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 292     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 293
 294     /* Add position restraint contribution */
 295     for (i = 0; i < DIM; i++)
 296     {
 297         vir_part[i][i] += fr->vir_diag_posres[i];
 298     }
 299
 300     /* Add wall contribution */
 301     for (i = 0; i < DIM; i++)
 302     {
 303         vir_part[i][ZZ] += fr->vir_wall_z[i];
 304     }
 305
 306     if (debug)
 307     {
 308         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 309     }
 310 }
 311
 312 static void pull_potential_wrapper(t_commrec *cr,
 313                                    t_inputrec *ir,
 314                                    matrix box, rvec x[],
 315                                    rvec f[],
 316                                    tensor vir_force,
 317                                    t_mdatoms *mdatoms,
 318                                    gmx_enerdata_t *enerd,
 319                                    real *lambda,
 320                                    double t,
 321                                    gmx_wallcycle_t wcycle)
 322 {
 323     t_pbc  pbc;
 324     real   dvdl;
 325
 326     /* Calculate the center of mass forces, this requires communication,
 327      * which is why pull_potential is called close to other communication.
 328      * The virial contribution is calculated directly,
 329      * which is why we call pull_potential after calc_virial.
 330      */
 331     wallcycle_start(wcycle, ewcPULLPOT);
 332     set_pbc(&pbc, ir->ePBC, box);
 333     dvdl                     = 0;
 334     enerd->term[F_COM_PULL] +=
 335         pull_potential(ir->pull_work, mdatoms, &pbc,
 336                        cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
 337     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 338     wallcycle_stop(wcycle, ewcPULLPOT);
 339 }
 340
 341 static void pme_receive_force_ener(t_commrec      *cr,
 342                                    gmx_wallcycle_t wcycle,
 343                                    gmx_enerdata_t *enerd,
 344                                    t_forcerec     *fr)
 345 {
 346     real   e_q, e_lj, dvdl_q, dvdl_lj;
 347     float  cycles_ppdpme, cycles_seppme;
 348
 349     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 350     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 351
 352     /* In case of node-splitting, the PP nodes receive the long-range
 353      * forces, virial and energy from the PME nodes here.
 354      */
 355     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 356     dvdl_q  = 0;
 357     dvdl_lj = 0;
 358     gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e_q,
 359                       fr->vir_lj_recip, &e_lj, &dvdl_q, &dvdl_lj,
 360                       &cycles_seppme);
 361     enerd->term[F_COUL_RECIP] += e_q;
 362     enerd->term[F_LJ_RECIP]   += e_lj;
 363     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 364     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 365
 366     if (wcycle)
 367     {
 368         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 369     }
 370     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 371 }
 372
 373 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
 374                                gmx_int64_t step, real pforce, rvec *x, rvec *f)
 375 {
 376     int  i;
 377     real pf2, fn2;
 378     char buf[STEPSTRSIZE];
 379
 380     pf2 = sqr(pforce);
 381     for (i = 0; i < md->homenr; i++)
 382     {
 383         fn2 = norm2(f[i]);
 384         /* We also catch NAN, if the compiler does not optimize this away. */
 385         if (fn2 >= pf2 || fn2 != fn2)
 386         {
 387             fprintf(fp, "step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 388                     gmx_step_str(step, buf),
 389                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
 390         }
 391     }
 392 }
 393
 394 static void post_process_forces(t_commrec *cr,
 395                                 gmx_int64_t step,
 396                                 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 397                                 gmx_localtop_t *top,
 398                                 matrix box, rvec x[],
 399                                 rvec f[],
 400                                 tensor vir_force,
 401                                 t_mdatoms *mdatoms,
 402                                 t_graph *graph,
 403                                 t_forcerec *fr, gmx_vsite_t *vsite,
 404                                 int flags)
 405 {
 406     if (fr->bF_NoVirSum)
 407     {
 408         if (vsite)
 409         {
 410             /* Spread the mesh force on virtual sites to the other particles...
 411              * This is parallellized. MPI communication is performed
 412              * if the constructing atoms aren't local.
 413              */
 414             wallcycle_start(wcycle, ewcVSITESPREAD);
 415             spread_vsite_f(vsite, x, fr->f_novirsum, NULL,
 416                            (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
 417                            nrnb,
 418                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
 419             wallcycle_stop(wcycle, ewcVSITESPREAD);
 420         }
 421         if (flags & GMX_FORCE_VIRIAL)
 422         {
 423             /* Now add the forces, this is local */
 424             if (fr->bDomDec)
 425             {
 426                 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
 427             }
 428             else
 429             {
 430                 sum_forces(0, mdatoms->homenr,
 431                            f, fr->f_novirsum);
 432             }
 433             if (EEL_FULL(fr->eeltype))
 434             {
 435                 /* Add the mesh contribution to the virial */
 436                 m_add(vir_force, fr->vir_el_recip, vir_force);
 437             }
 438             if (EVDW_PME(fr->vdwtype))
 439             {
 440                 /* Add the mesh contribution to the virial */
 441                 m_add(vir_force, fr->vir_lj_recip, vir_force);
 442             }
 443             if (debug)
 444             {
 445                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 446             }
 447         }
 448     }
 449
 450     if (fr->print_force >= 0)
 451     {
 452         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 453     }
 454 }
 455
 456 static void do_nb_verlet(t_forcerec *fr,
 457                          interaction_const_t *ic,
 458                          gmx_enerdata_t *enerd,
 459                          int flags, int ilocality,
 460                          int clearF,
 461                          t_nrnb *nrnb,
 462                          gmx_wallcycle_t wcycle)
 463 {
 464     int                        enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
 465     nonbonded_verlet_group_t  *nbvg;
 466     gmx_bool                   bUsingGpuKernels;
 467
 468     if (!(flags & GMX_FORCE_NONBONDED))
 469     {
 470         /* skip non-bonded calculation */
 471         return;
 472     }
 473
 474     nbvg = &fr->nbv->grp[ilocality];
 475
 476     /* GPU kernel launch overhead is already timed separately */
 477     if (fr->cutoff_scheme != ecutsVERLET)
 478     {
 479         gmx_incons("Invalid cut-off scheme passed!");
 480     }
 481
 482     bUsingGpuKernels = (nbvg->kernel_type == nbnxnk8x8x8_GPU);
 483
 484     if (!bUsingGpuKernels)
 485     {
 486         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 487     }
 488     switch (nbvg->kernel_type)
 489     {
 490         case nbnxnk4x4_PlainC:
 491             nbnxn_kernel_ref(&nbvg->nbl_lists,
 492                              nbvg->nbat, ic,
 493                              fr->shift_vec,
 494                              flags,
 495                              clearF,
 496                              fr->fshift[0],
 497                              enerd->grpp.ener[egCOULSR],
 498                              fr->bBHAM ?
 499                              enerd->grpp.ener[egBHAMSR] :
 500                              enerd->grpp.ener[egLJSR]);
 501             break;
 502
 503         case nbnxnk4xN_SIMD_4xN:
 504             nbnxn_kernel_simd_4xn(&nbvg->nbl_lists,
 505                                   nbvg->nbat, ic,
 506                                   nbvg->ewald_excl,
 507                                   fr->shift_vec,
 508                                   flags,
 509                                   clearF,
 510                                   fr->fshift[0],
 511                                   enerd->grpp.ener[egCOULSR],
 512                                   fr->bBHAM ?
 513                                   enerd->grpp.ener[egBHAMSR] :
 514                                   enerd->grpp.ener[egLJSR]);
 515             break;
 516         case nbnxnk4xN_SIMD_2xNN:
 517             nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
 518                                    nbvg->nbat, ic,
 519                                    nbvg->ewald_excl,
 520                                    fr->shift_vec,
 521                                    flags,
 522                                    clearF,
 523                                    fr->fshift[0],
 524                                    enerd->grpp.ener[egCOULSR],
 525                                    fr->bBHAM ?
 526                                    enerd->grpp.ener[egBHAMSR] :
 527                                    enerd->grpp.ener[egLJSR]);
 528             break;
 529
 530         case nbnxnk8x8x8_GPU:
 531             nbnxn_gpu_launch_kernel(fr->nbv->gpu_nbv, nbvg->nbat, flags, ilocality);
 532             break;
 533
 534         case nbnxnk8x8x8_PlainC:
 535             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 536                                  nbvg->nbat, ic,
 537                                  fr->shift_vec,
 538                                  flags,
 539                                  clearF,
 540                                  nbvg->nbat->out[0].f,
 541                                  fr->fshift[0],
 542                                  enerd->grpp.ener[egCOULSR],
 543                                  fr->bBHAM ?
 544                                  enerd->grpp.ener[egBHAMSR] :
 545                                  enerd->grpp.ener[egLJSR]);
 546             break;
 547
 548         default:
 549             gmx_incons("Invalid nonbonded kernel type passed!");
 550
 551     }
 552     if (!bUsingGpuKernels)
 553     {
 554         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 555     }
 556
 557     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 558     {
 559         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 560     }
 561     else if ((!bUsingGpuKernels && nbvg->ewald_excl == ewaldexclAnalytical) ||
 562              (bUsingGpuKernels && nbnxn_gpu_is_kernel_ewald_analytical(fr->nbv->gpu_nbv)))
 563     {
 564         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 565     }
 566     else
 567     {
 568         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 569     }
 570     enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 571     if (flags & GMX_FORCE_ENERGY)
 572     {
 573         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 574         enr_nbnxn_kernel_ljc += 1;
 575         enr_nbnxn_kernel_lj  += 1;
 576     }
 577
 578     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 579              nbvg->nbl_lists.natpair_ljq);
 580     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 581              nbvg->nbl_lists.natpair_lj);
 582     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 583     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 584              nbvg->nbl_lists.natpair_q);
 585
 586     if (ic->vdw_modifier == eintmodFORCESWITCH)
 587     {
 588         /* We add up the switch cost separately */
 589         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_FSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 590                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 591     }
 592     if (ic->vdw_modifier == eintmodPOTSWITCH)
 593     {
 594         /* We add up the switch cost separately */
 595         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_PSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 596                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 597     }
 598     if (ic->vdwtype == evdwPME)
 599     {
 600         /* We add up the LJ Ewald cost separately */
 601         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_EWALD+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 602                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 603     }
 604 }
 605
 606 static void do_nb_verlet_fep(nbnxn_pairlist_set_t *nbl_lists,
 607                              t_forcerec           *fr,
 608                              rvec                  x[],
 609                              rvec                  f[],
 610                              t_mdatoms            *mdatoms,
 611                              t_lambda             *fepvals,
 612                              real                 *lambda,
 613                              gmx_enerdata_t       *enerd,
 614                              int                   flags,
 615                              t_nrnb               *nrnb,
 616                              gmx_wallcycle_t       wcycle)
 617 {
 618     int              donb_flags;
 619     nb_kernel_data_t kernel_data;
 620     real             lam_i[efptNR];
 621     real             dvdl_nb[efptNR];
 622     int              th;
 623     int              i, j;
 624
 625     donb_flags = 0;
 626     /* Add short-range interactions */
 627     donb_flags |= GMX_NONBONDED_DO_SR;
 628
 629     /* Currently all group scheme kernels always calculate (shift-)forces */
 630     if (flags & GMX_FORCE_FORCES)
 631     {
 632         donb_flags |= GMX_NONBONDED_DO_FORCE;
 633     }
 634     if (flags & GMX_FORCE_VIRIAL)
 635     {
 636         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 637     }
 638     if (flags & GMX_FORCE_ENERGY)
 639     {
 640         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 641     }
 642     if (flags & GMX_FORCE_DO_LR)
 643     {
 644         donb_flags |= GMX_NONBONDED_DO_LR;
 645     }
 646
 647     kernel_data.flags  = donb_flags;
 648     kernel_data.lambda = lambda;
 649     kernel_data.dvdl   = dvdl_nb;
 650
 651     kernel_data.energygrp_elec = enerd->grpp.ener[egCOULSR];
 652     kernel_data.energygrp_vdw  = enerd->grpp.ener[egLJSR];
 653
 654     /* reset free energy components */
 655     for (i = 0; i < efptNR; i++)
 656     {
 657         dvdl_nb[i]  = 0;
 658     }
 659
 660     assert(gmx_omp_nthreads_get(emntNonbonded) == nbl_lists->nnbl);
 661
 662     wallcycle_sub_start(wcycle, ewcsNONBONDED);
 663 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 664     for (th = 0; th < nbl_lists->nnbl; th++)
 665     {
 666         gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 667                                   x, f, fr, mdatoms, &kernel_data, nrnb);
 668     }
 669
 670     if (fepvals->sc_alpha != 0)
 671     {
 672         enerd->dvdl_nonlin[efptVDW]  += dvdl_nb[efptVDW];
 673         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 674     }
 675     else
 676     {
 677         enerd->dvdl_lin[efptVDW]  += dvdl_nb[efptVDW];
 678         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 679     }
 680
 681     /* If we do foreign lambda and we have soft-core interactions
 682      * we have to recalculate the (non-linear) energies contributions.
 683      */
 684     if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
 685     {
 686         kernel_data.flags          = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE)) | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 687         kernel_data.lambda         = lam_i;
 688         kernel_data.energygrp_elec = enerd->foreign_grpp.ener[egCOULSR];
 689         kernel_data.energygrp_vdw  = enerd->foreign_grpp.ener[egLJSR];
 690         /* Note that we add to kernel_data.dvdl, but ignore the result */
 691
 692         for (i = 0; i < enerd->n_lambda; i++)
 693         {
 694             for (j = 0; j < efptNR; j++)
 695             {
 696                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
 697             }
 698             reset_foreign_enerdata(enerd);
 699 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 700             for (th = 0; th < nbl_lists->nnbl; th++)
 701             {
 702                 gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 703                                           x, f, fr, mdatoms, &kernel_data, nrnb);
 704             }
 705
 706             sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
 707             enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
 708         }
 709     }
 710
 711     wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 712 }
 713
 714 gmx_bool use_GPU(const nonbonded_verlet_t *nbv)
 715 {
 716     return nbv != NULL && nbv->bUseGPU;
 717 }
 718
 719 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
 720                          t_inputrec *inputrec,
 721                          gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 722                          gmx_localtop_t *top,
 723                          gmx_groups_t gmx_unused *groups,
 724                          matrix box, rvec x[], history_t *hist,
 725                          rvec f[],
 726                          tensor vir_force,
 727                          t_mdatoms *mdatoms,
 728                          gmx_enerdata_t *enerd, t_fcdata *fcd,
 729                          real *lambda, t_graph *graph,
 730                          t_forcerec *fr, interaction_const_t *ic,
 731                          gmx_vsite_t *vsite, rvec mu_tot,
 732                          double t, FILE *field, gmx_edsam_t ed,
 733                          gmx_bool bBornRadii,
 734                          int flags)
 735 {
 736     int                 cg1, i, j;
 737     int                 start, homenr;
 738     double              mu[2*DIM];
 739     gmx_bool            bStateChanged, bNS, bFillGrid, bCalcCGCM;
 740     gmx_bool            bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
 741     gmx_bool            bDiffKernels = FALSE;
 742     rvec                vzero, box_diag;
 743     float               cycles_pme, cycles_force, cycles_wait_gpu;
 744     nonbonded_verlet_t *nbv;
 745
 746     cycles_force    = 0;
 747     cycles_wait_gpu = 0;
 748     nbv             = fr->nbv;
 749
 750     start  = 0;
 751     homenr = mdatoms->homenr;
 752
 753     clear_mat(vir_force);
 754
 755     if (DOMAINDECOMP(cr))
 756     {
 757         cg1 = cr->dd->ncg_tot;
 758     }
 759     else
 760     {
 761         cg1 = top->cgs.nr;
 762     }
 763     if (fr->n_tpi > 0)
 764     {
 765         cg1--;
 766     }
 767
 768     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 769     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
 770     bFillGrid     = (bNS && bStateChanged);
 771     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 772     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 773     bDoForces     = (flags & GMX_FORCE_FORCES);
 774     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 775     bUseGPU       = fr->nbv->bUseGPU;
 776     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
 777
 778     if (bStateChanged)
 779     {
 780         update_forcerec(fr, box);
 781
 782         if (NEED_MUTOT(*inputrec))
 783         {
 784             /* Calculate total (local) dipole moment in a temporary common array.
 785              * This makes it possible to sum them over nodes faster.
 786              */
 787             calc_mu(start, homenr,
 788                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 789                     mu, mu+DIM);
 790         }
 791     }
 792
 793     if (fr->ePBC != epbcNONE)
 794     {
 795         /* Compute shift vectors every step,
 796          * because of pressure coupling or box deformation!
 797          */
 798         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 799         {
 800             calc_shifts(box, fr->shift_vec);
 801         }
 802
 803         if (bCalcCGCM)
 804         {
 805             put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
 806             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 807         }
 808         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 809         {
 810             unshift_self(graph, box, x);
 811         }
 812     }
 813
 814     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 815                                  fr->shift_vec, nbv->grp[0].nbat);
 816
 817 #ifdef GMX_MPI
 818     if (!(cr->duty & DUTY_PME))
 819     {
 820         gmx_bool bBS;
 821         matrix   boxs;
 822
 823         /* Send particle coordinates to the pme nodes.
 824          * Since this is only implemented for domain decomposition
 825          * and domain decomposition does not use the graph,
 826          * we do not need to worry about shifting.
 827          */
 828
 829         int pme_flags = 0;
 830
 831         wallcycle_start(wcycle, ewcPP_PMESENDX);
 832
 833         bBS = (inputrec->nwall == 2);
 834         if (bBS)
 835         {
 836             copy_mat(box, boxs);
 837             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
 838         }
 839
 840         if (EEL_PME(fr->eeltype))
 841         {
 842             pme_flags |= GMX_PME_DO_COULOMB;
 843         }
 844
 845         if (EVDW_PME(fr->vdwtype))
 846         {
 847             pme_flags |= GMX_PME_DO_LJ;
 848         }
 849
 850         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
 851                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
 852                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
 853                                  pme_flags, step);
 854
 855         wallcycle_stop(wcycle, ewcPP_PMESENDX);
 856     }
 857 #endif /* GMX_MPI */
 858
 859     /* do gridding for pair search */
 860     if (bNS)
 861     {
 862         if (graph && bStateChanged)
 863         {
 864             /* Calculate intramolecular shift vectors to make molecules whole */
 865             mk_mshift(fplog, graph, fr->ePBC, box, x);
 866         }
 867
 868         clear_rvec(vzero);
 869         box_diag[XX] = box[XX][XX];
 870         box_diag[YY] = box[YY][YY];
 871         box_diag[ZZ] = box[ZZ][ZZ];
 872
 873         wallcycle_start(wcycle, ewcNS);
 874         if (!fr->bDomDec)
 875         {
 876             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 877             nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
 878                               0, vzero, box_diag,
 879                               0, mdatoms->homenr, -1, fr->cginfo, x,
 880                               0, NULL,
 881                               nbv->grp[eintLocal].kernel_type,
 882                               nbv->grp[eintLocal].nbat);
 883             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 884         }
 885         else
 886         {
 887             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 888             nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
 889                                        fr->cginfo, x,
 890                                        nbv->grp[eintNonlocal].kernel_type,
 891                                        nbv->grp[eintNonlocal].nbat);
 892             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 893         }
 894
 895         if (nbv->ngrp == 1 ||
 896             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 897         {
 898             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
 899                                nbv->nbs, mdatoms, fr->cginfo);
 900         }
 901         else
 902         {
 903             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
 904                                nbv->nbs, mdatoms, fr->cginfo);
 905             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
 906                                nbv->nbs, mdatoms, fr->cginfo);
 907         }
 908         wallcycle_stop(wcycle, ewcNS);
 909     }
 910
 911     /* initialize the GPU atom data and copy shift vector */
 912     if (bUseGPU)
 913     {
 914         if (bNS)
 915         {
 916             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 917             nbnxn_gpu_init_atomdata(nbv->gpu_nbv, nbv->grp[eintLocal].nbat);
 918             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 919         }
 920
 921         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 922         nbnxn_gpu_upload_shiftvec(nbv->gpu_nbv, nbv->grp[eintLocal].nbat);
 923         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 924     }
 925
 926     /* do local pair search */
 927     if (bNS)
 928     {
 929         wallcycle_start_nocount(wcycle, ewcNS);
 930         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
 931         nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
 932                             &top->excls,
 933                             ic->rlist,
 934                             nbv->min_ci_balanced,
 935                             &nbv->grp[eintLocal].nbl_lists,
 936                             eintLocal,
 937                             nbv->grp[eintLocal].kernel_type,
 938                             nrnb);
 939         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
 940
 941         if (bUseGPU)
 942         {
 943             /* initialize local pair-list on the GPU */
 944             nbnxn_gpu_init_pairlist(nbv->gpu_nbv,
 945                                     nbv->grp[eintLocal].nbl_lists.nbl[0],
 946                                     eintLocal);
 947         }
 948         wallcycle_stop(wcycle, ewcNS);
 949     }
 950     else
 951     {
 952         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 953         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
 954         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
 955                                         nbv->grp[eintLocal].nbat);
 956         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
 957         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
 958     }
 959
 960     if (bUseGPU)
 961     {
 962         wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
 963         /* launch local nonbonded F on GPU */
 964         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
 965                      nrnb, wcycle);
 966         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 967     }
 968
 969     /* Communicate coordinates and sum dipole if necessary +
 970        do non-local pair search */
 971     if (DOMAINDECOMP(cr))
 972     {
 973         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
 974                         nbv->grp[eintLocal].kernel_type);
 975
 976         if (bDiffKernels)
 977         {
 978             /* With GPU+CPU non-bonded calculations we need to copy
 979              * the local coordinates to the non-local nbat struct
 980              * (in CPU format) as the non-local kernel call also
 981              * calculates the local - non-local interactions.
 982              */
 983             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 984             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
 985             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
 986                                             nbv->grp[eintNonlocal].nbat);
 987             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
 988             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
 989         }
 990
 991         if (bNS)
 992         {
 993             wallcycle_start_nocount(wcycle, ewcNS);
 994             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
 995
 996             if (bDiffKernels)
 997             {
 998                 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
 999             }
1000
1001             nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
1002                                 &top->excls,
1003                                 ic->rlist,
1004                                 nbv->min_ci_balanced,
1005                                 &nbv->grp[eintNonlocal].nbl_lists,
1006                                 eintNonlocal,
1007                                 nbv->grp[eintNonlocal].kernel_type,
1008                                 nrnb);
1009
1010             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1011
1012             if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_GPU)
1013             {
1014                 /* initialize non-local pair-list on the GPU */
1015                 nbnxn_gpu_init_pairlist(nbv->gpu_nbv,
1016                                         nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1017                                         eintNonlocal);
1018             }
1019             wallcycle_stop(wcycle, ewcNS);
1020         }
1021         else
1022         {
1023             wallcycle_start(wcycle, ewcMOVEX);
1024             dd_move_x(cr->dd, box, x);
1025
1026             /* When we don't need the total dipole we sum it in global_stat */
1027             if (bStateChanged && NEED_MUTOT(*inputrec))
1028             {
1029                 gmx_sumd(2*DIM, mu, cr);
1030             }
1031             wallcycle_stop(wcycle, ewcMOVEX);
1032
1033             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1034             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1035             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1036                                             nbv->grp[eintNonlocal].nbat);
1037             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1038             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1039         }
1040
1041         if (bUseGPU && !bDiffKernels)
1042         {
1043             wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1044             /* launch non-local nonbonded F on GPU */
1045             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1046                          nrnb, wcycle);
1047             cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1048         }
1049     }
1050
1051     if (bUseGPU)
1052     {
1053         /* launch D2H copy-back F */
1054         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1055         if (DOMAINDECOMP(cr) && !bDiffKernels)
1056         {
1057             nbnxn_gpu_launch_cpyback(nbv->gpu_nbv, nbv->grp[eintNonlocal].nbat,
1058                                      flags, eatNonlocal);
1059         }
1060         nbnxn_gpu_launch_cpyback(nbv->gpu_nbv, nbv->grp[eintLocal].nbat,
1061                                  flags, eatLocal);
1062         cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1063     }
1064
1065     if (bStateChanged && NEED_MUTOT(*inputrec))
1066     {
1067         if (PAR(cr))
1068         {
1069             gmx_sumd(2*DIM, mu, cr);
1070         }
1071
1072         for (i = 0; i < 2; i++)
1073         {
1074             for (j = 0; j < DIM; j++)
1075             {
1076                 fr->mu_tot[i][j] = mu[i*DIM + j];
1077             }
1078         }
1079     }
1080     if (fr->efep == efepNO)
1081     {
1082         copy_rvec(fr->mu_tot[0], mu_tot);
1083     }
1084     else
1085     {
1086         for (j = 0; j < DIM; j++)
1087         {
1088             mu_tot[j] =
1089                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1090                 lambda[efptCOUL]*fr->mu_tot[1][j];
1091         }
1092     }
1093
1094     /* Reset energies */
1095     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1096     clear_rvecs(SHIFTS, fr->fshift);
1097
1098     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1099     {
1100         wallcycle_start(wcycle, ewcPPDURINGPME);
1101         dd_force_flop_start(cr->dd, nrnb);
1102     }
1103
1104     if (inputrec->bRot)
1105     {
1106         /* Enforced rotation has its own cycle counter that starts after the collective
1107          * coordinates have been communicated. It is added to ddCyclF to allow
1108          * for proper load-balancing */
1109         wallcycle_start(wcycle, ewcROT);
1110         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1111         wallcycle_stop(wcycle, ewcROT);
1112     }
1113
1114     /* Start the force cycle counter.
1115      * This counter is stopped after do_force_lowlevel.
1116      * No parallel communication should occur while this counter is running,
1117      * since that will interfere with the dynamic load balancing.
1118      */
1119     wallcycle_start(wcycle, ewcFORCE);
1120     if (bDoForces)
1121     {
1122         /* Reset forces for which the virial is calculated separately:
1123          * PME/Ewald forces if necessary */
1124         if (fr->bF_NoVirSum)
1125         {
1126             if (flags & GMX_FORCE_VIRIAL)
1127             {
1128                 fr->f_novirsum = fr->f_novirsum_alloc;
1129                 if (fr->bDomDec)
1130                 {
1131                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1132                 }
1133                 else
1134                 {
1135                     clear_rvecs(homenr, fr->f_novirsum+start);
1136                 }
1137             }
1138             else
1139             {
1140                 /* We are not calculating the pressure so we do not need
1141                  * a separate array for forces that do not contribute
1142                  * to the pressure.
1143                  */
1144                 fr->f_novirsum = f;
1145             }
1146         }
1147
1148         /* Clear the short- and long-range forces */
1149         clear_rvecs(fr->natoms_force_constr, f);
1150         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1151         {
1152             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1153         }
1154
1155         clear_rvec(fr->vir_diag_posres);
1156     }
1157
1158     if (inputrec->bPull && pull_have_constraint(inputrec->pull_work))
1159     {
1160         clear_pull_forces(inputrec->pull_work);
1161     }
1162
1163     /* We calculate the non-bonded forces, when done on the CPU, here.
1164      * We do this before calling do_force_lowlevel, because in that
1165      * function, the listed forces are calculated before PME, which
1166      * does communication.  With this order, non-bonded and listed
1167      * force calculation imbalance can be balanced out by the domain
1168      * decomposition load balancing.
1169      */
1170
1171     if (!bUseOrEmulGPU)
1172     {
1173         /* Maybe we should move this into do_force_lowlevel */
1174         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1175                      nrnb, wcycle);
1176     }
1177
1178     if (fr->efep != efepNO)
1179     {
1180         /* Calculate the local and non-local free energy interactions here.
1181          * Happens here on the CPU both with and without GPU.
1182          */
1183         if (fr->nbv->grp[eintLocal].nbl_lists.nbl_fep[0]->nrj > 0)
1184         {
1185             do_nb_verlet_fep(&fr->nbv->grp[eintLocal].nbl_lists,
1186                              fr, x, f, mdatoms,
1187                              inputrec->fepvals, lambda,
1188                              enerd, flags, nrnb, wcycle);
1189         }
1190
1191         if (DOMAINDECOMP(cr) &&
1192             fr->nbv->grp[eintNonlocal].nbl_lists.nbl_fep[0]->nrj > 0)
1193         {
1194             do_nb_verlet_fep(&fr->nbv->grp[eintNonlocal].nbl_lists,
1195                              fr, x, f, mdatoms,
1196                              inputrec->fepvals, lambda,
1197                              enerd, flags, nrnb, wcycle);
1198         }
1199     }
1200
1201     if (!bUseOrEmulGPU || bDiffKernels)
1202     {
1203         int aloc;
1204
1205         if (DOMAINDECOMP(cr))
1206         {
1207             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1208                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1209                          nrnb, wcycle);
1210         }
1211
1212         if (!bUseOrEmulGPU)
1213         {
1214             aloc = eintLocal;
1215         }
1216         else
1217         {
1218             aloc = eintNonlocal;
1219         }
1220
1221         /* Add all the non-bonded force to the normal force array.
1222          * This can be split into a local and a non-local part when overlapping
1223          * communication with calculation with domain decomposition.
1224          */
1225         cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1226         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1227         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1228         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1229         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1230         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1231         wallcycle_start_nocount(wcycle, ewcFORCE);
1232
1233         /* if there are multiple fshift output buffers reduce them */
1234         if ((flags & GMX_FORCE_VIRIAL) &&
1235             nbv->grp[aloc].nbl_lists.nnbl > 1)
1236         {
1237             /* This is not in a subcounter because it takes a
1238                negligible and constant-sized amount of time */
1239             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1240                                                      fr->fshift);
1241         }
1242     }
1243
1244     /* update QMMMrec, if necessary */
1245     if (fr->bQMMM)
1246     {
1247         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1248     }
1249
1250     /* Compute the bonded and non-bonded energies and optionally forces */
1251     do_force_lowlevel(fr, inputrec, &(top->idef),
1252                       cr, nrnb, wcycle, mdatoms,
1253                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1254                       bBornRadii, box,
1255                       inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1256                       flags, &cycles_pme);
1257
1258     if (bSepLRF)
1259     {
1260         if (do_per_step(step, inputrec->nstcalclr))
1261         {
1262             /* Add the long range forces to the short range forces */
1263             for (i = 0; i < fr->natoms_force_constr; i++)
1264             {
1265                 rvec_add(fr->f_twin[i], f[i], f[i]);
1266             }
1267         }
1268     }
1269
1270     cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1271
1272     if (ed)
1273     {
1274         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1275     }
1276
1277     if (bUseOrEmulGPU && !bDiffKernels)
1278     {
1279         /* wait for non-local forces (or calculate in emulation mode) */
1280         if (DOMAINDECOMP(cr))
1281         {
1282             if (bUseGPU)
1283             {
1284                 float cycles_tmp;
1285
1286                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1287                 nbnxn_gpu_wait_for_gpu(nbv->gpu_nbv,
1288                                        nbv->grp[eintNonlocal].nbat,
1289                                        flags, eatNonlocal,
1290                                        enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1291                                        fr->fshift);
1292                 cycles_tmp       = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1293                 cycles_wait_gpu += cycles_tmp;
1294                 cycles_force    += cycles_tmp;
1295             }
1296             else
1297             {
1298                 wallcycle_start_nocount(wcycle, ewcFORCE);
1299                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1300                              nrnb, wcycle);
1301                 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1302             }
1303             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1304             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1305             /* skip the reduction if there was no non-local work to do */
1306             if (nbv->grp[eintNonlocal].nbl_lists.nbl[0]->nsci > 0)
1307             {
1308                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1309                                                nbv->grp[eintNonlocal].nbat, f);
1310             }
1311             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1312             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1313         }
1314     }
1315
1316     if (bDoForces && DOMAINDECOMP(cr))
1317     {
1318         if (bUseGPU)
1319         {
1320             /* We are done with the CPU compute, but the GPU local non-bonded
1321              * kernel can still be running while we communicate the forces.
1322              * We start a counter here, so we can, hopefully, time the rest
1323              * of the GPU kernel execution and data transfer.
1324              */
1325             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L_EST);
1326         }
1327
1328         /* Communicate the forces */
1329         wallcycle_start(wcycle, ewcMOVEF);
1330         dd_move_f(cr->dd, f, fr->fshift);
1331         if (bSepLRF)
1332         {
1333             /* We should not update the shift forces here,
1334              * since f_twin is already included in f.
1335              */
1336             dd_move_f(cr->dd, fr->f_twin, NULL);
1337         }
1338         wallcycle_stop(wcycle, ewcMOVEF);
1339     }
1340
1341     if (bUseOrEmulGPU)
1342     {
1343         /* wait for local forces (or calculate in emulation mode) */
1344         if (bUseGPU)
1345         {
1346             float       cycles_tmp, cycles_wait_est;
1347             const float cuda_api_overhead_margin = 50000.0f; /* cycles */
1348
1349             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1350             nbnxn_gpu_wait_for_gpu(nbv->gpu_nbv,
1351                                    nbv->grp[eintLocal].nbat,
1352                                    flags, eatLocal,
1353                                    enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1354                                    fr->fshift);
1355             cycles_tmp      = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1356
1357             if (bDoForces && DOMAINDECOMP(cr))
1358             {
1359                 cycles_wait_est = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L_EST);
1360
1361                 if (cycles_tmp < cuda_api_overhead_margin)
1362                 {
1363                     /* We measured few cycles, it could be that the kernel
1364                      * and transfer finished earlier and there was no actual
1365                      * wait time, only API call overhead.
1366                      * Then the actual time could be anywhere between 0 and
1367                      * cycles_wait_est. As a compromise, we use half the time.
1368                      */
1369                     cycles_wait_est *= 0.5f;
1370                 }
1371             }
1372             else
1373             {
1374                 /* No force communication so we actually timed the wait */
1375                 cycles_wait_est = cycles_tmp;
1376             }
1377             /* Even though this is after dd_move_f, the actual task we are
1378              * waiting for runs asynchronously with dd_move_f and we usually
1379              * have nothing to balance it with, so we can and should add
1380              * the time to the force time for load balancing.
1381              */
1382             cycles_force    += cycles_wait_est;
1383             cycles_wait_gpu += cycles_wait_est;
1384
1385             /* now clear the GPU outputs while we finish the step on the CPU */
1386
1387             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1388             nbnxn_gpu_clear_outputs(nbv->gpu_nbv, flags);
1389             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1390         }
1391         else
1392         {
1393             wallcycle_start_nocount(wcycle, ewcFORCE);
1394             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1395                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1396                          nrnb, wcycle);
1397             wallcycle_stop(wcycle, ewcFORCE);
1398         }
1399         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1400         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1401         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1402                                        nbv->grp[eintLocal].nbat, f);
1403         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1404         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1405     }
1406
1407     if (DOMAINDECOMP(cr))
1408     {
1409         dd_force_flop_stop(cr->dd, nrnb);
1410         if (wcycle)
1411         {
1412             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1413             if (bUseGPU)
1414             {
1415                 dd_cycles_add(cr->dd, cycles_wait_gpu, ddCyclWaitGPU);
1416             }
1417         }
1418     }
1419
1420     if (bDoForces)
1421     {
1422         if (IR_ELEC_FIELD(*inputrec))
1423         {
1424             /* Compute forces due to electric field */
1425             calc_f_el(MASTER(cr) ? field : NULL,
1426                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1427                       inputrec->ex, inputrec->et, t);
1428         }
1429
1430         /* If we have NoVirSum forces, but we do not calculate the virial,
1431          * we sum fr->f_novirsum=f later.
1432          */
1433         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1434         {
1435             wallcycle_start(wcycle, ewcVSITESPREAD);
1436             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1437                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1438             wallcycle_stop(wcycle, ewcVSITESPREAD);
1439
1440             if (bSepLRF)
1441             {
1442                 wallcycle_start(wcycle, ewcVSITESPREAD);
1443                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1444                                nrnb,
1445                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1446                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1447             }
1448         }
1449
1450         if (flags & GMX_FORCE_VIRIAL)
1451         {
1452             /* Calculation of the virial must be done after vsites! */
1453             calc_virial(0, mdatoms->homenr, x, f,
1454                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1455         }
1456     }
1457
1458     if (inputrec->bPull && pull_have_potential(inputrec->pull_work))
1459     {
1460         /* Since the COM pulling is always done mass-weighted, no forces are
1461          * applied to vsites and this call can be done after vsite spreading.
1462          */
1463         pull_potential_wrapper(cr, inputrec, box, x,
1464                                f, vir_force, mdatoms, enerd, lambda, t,
1465                                wcycle);
1466     }
1467
1468     /* Add the forces from enforced rotation potentials (if any) */
1469     if (inputrec->bRot)
1470     {
1471         wallcycle_start(wcycle, ewcROTadd);
1472         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1473         wallcycle_stop(wcycle, ewcROTadd);
1474     }
1475
1476     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
1477     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
1478
1479     if (PAR(cr) && !(cr->duty & DUTY_PME))
1480     {
1481         /* In case of node-splitting, the PP nodes receive the long-range
1482          * forces, virial and energy from the PME nodes here.
1483          */
1484         pme_receive_force_ener(cr, wcycle, enerd, fr);
1485     }
1486
1487     if (bDoForces)
1488     {
1489         post_process_forces(cr, step, nrnb, wcycle,
1490                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1491                             flags);
1492     }
1493
1494     /* Sum the potential energy terms from group contributions */
1495     sum_epot(&(enerd->grpp), enerd->term);
1496 }
1497
1498 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1499                         t_inputrec *inputrec,
1500                         gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1501                         gmx_localtop_t *top,
1502                         gmx_groups_t *groups,
1503                         matrix box, rvec x[], history_t *hist,
1504                         rvec f[],
1505                         tensor vir_force,
1506                         t_mdatoms *mdatoms,
1507                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1508                         real *lambda, t_graph *graph,
1509                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1510                         double t, FILE *field, gmx_edsam_t ed,
1511                         gmx_bool bBornRadii,
1512                         int flags)
1513 {
1514     int        cg0, cg1, i, j;
1515     int        start, homenr;
1516     double     mu[2*DIM];
1517     gmx_bool   bStateChanged, bNS, bFillGrid, bCalcCGCM;
1518     gmx_bool   bDoLongRangeNS, bDoForces, bSepLRF;
1519     gmx_bool   bDoAdressWF;
1520     t_pbc      pbc;
1521     float      cycles_pme, cycles_force;
1522
1523     start  = 0;
1524     homenr = mdatoms->homenr;
1525
1526     clear_mat(vir_force);
1527
1528     cg0 = 0;
1529     if (DOMAINDECOMP(cr))
1530     {
1531         cg1 = cr->dd->ncg_tot;
1532     }
1533     else
1534     {
1535         cg1 = top->cgs.nr;
1536     }
1537     if (fr->n_tpi > 0)
1538     {
1539         cg1--;
1540     }
1541
1542     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1543     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1544     /* Should we update the long-range neighborlists at this step? */
1545     bDoLongRangeNS = fr->bTwinRange && bNS;
1546     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1547     bFillGrid      = (bNS && bStateChanged);
1548     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1549     bDoForces      = (flags & GMX_FORCE_FORCES);
1550     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1551                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1552
1553     /* should probably move this to the forcerec since it doesn't change */
1554     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1555
1556     if (bStateChanged)
1557     {
1558         update_forcerec(fr, box);
1559
1560         if (NEED_MUTOT(*inputrec))
1561         {
1562             /* Calculate total (local) dipole moment in a temporary common array.
1563              * This makes it possible to sum them over nodes faster.
1564              */
1565             calc_mu(start, homenr,
1566                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1567                     mu, mu+DIM);
1568         }
1569     }
1570
1571     if (fr->ePBC != epbcNONE)
1572     {
1573         /* Compute shift vectors every step,
1574          * because of pressure coupling or box deformation!
1575          */
1576         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1577         {
1578             calc_shifts(box, fr->shift_vec);
1579         }
1580
1581         if (bCalcCGCM)
1582         {
1583             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1584                                      &(top->cgs), x, fr->cg_cm);
1585             inc_nrnb(nrnb, eNR_CGCM, homenr);
1586             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1587         }
1588         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1589         {
1590             unshift_self(graph, box, x);
1591         }
1592     }
1593     else if (bCalcCGCM)
1594     {
1595         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1596         inc_nrnb(nrnb, eNR_CGCM, homenr);
1597     }
1598
1599     if (bCalcCGCM && gmx_debug_at)
1600     {
1601         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1602     }
1603
1604 #ifdef GMX_MPI
1605     if (!(cr->duty & DUTY_PME))
1606     {
1607         gmx_bool bBS;
1608         matrix   boxs;
1609
1610         /* Send particle coordinates to the pme nodes.
1611          * Since this is only implemented for domain decomposition
1612          * and domain decomposition does not use the graph,
1613          * we do not need to worry about shifting.
1614          */
1615
1616         int pme_flags = 0;
1617
1618         wallcycle_start(wcycle, ewcPP_PMESENDX);
1619
1620         bBS = (inputrec->nwall == 2);
1621         if (bBS)
1622         {
1623             copy_mat(box, boxs);
1624             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1625         }
1626
1627         if (EEL_PME(fr->eeltype))
1628         {
1629             pme_flags |= GMX_PME_DO_COULOMB;
1630         }
1631
1632         if (EVDW_PME(fr->vdwtype))
1633         {
1634             pme_flags |= GMX_PME_DO_LJ;
1635         }
1636
1637         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
1638                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
1639                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
1640                                  pme_flags, step);
1641
1642         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1643     }
1644 #endif /* GMX_MPI */
1645
1646     /* Communicate coordinates and sum dipole if necessary */
1647     if (DOMAINDECOMP(cr))
1648     {
1649         wallcycle_start(wcycle, ewcMOVEX);
1650         dd_move_x(cr->dd, box, x);
1651         wallcycle_stop(wcycle, ewcMOVEX);
1652     }
1653
1654     /* update adress weight beforehand */
1655     if (bStateChanged && bDoAdressWF)
1656     {
1657         /* need pbc for adress weight calculation with pbc_dx */
1658         set_pbc(&pbc, inputrec->ePBC, box);
1659         if (fr->adress_site == eAdressSITEcog)
1660         {
1661             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1662                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1663         }
1664         else if (fr->adress_site == eAdressSITEcom)
1665         {
1666             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1667                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1668         }
1669         else if (fr->adress_site == eAdressSITEatomatom)
1670         {
1671             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1672                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1673         }
1674         else
1675         {
1676             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1677                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1678         }
1679     }
1680
1681     if (NEED_MUTOT(*inputrec))
1682     {
1683
1684         if (bStateChanged)
1685         {
1686             if (PAR(cr))
1687             {
1688                 gmx_sumd(2*DIM, mu, cr);
1689             }
1690             for (i = 0; i < 2; i++)
1691             {
1692                 for (j = 0; j < DIM; j++)
1693                 {
1694                     fr->mu_tot[i][j] = mu[i*DIM + j];
1695                 }
1696             }
1697         }
1698         if (fr->efep == efepNO)
1699         {
1700             copy_rvec(fr->mu_tot[0], mu_tot);
1701         }
1702         else
1703         {
1704             for (j = 0; j < DIM; j++)
1705             {
1706                 mu_tot[j] =
1707                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1708             }
1709         }
1710     }
1711
1712     /* Reset energies */
1713     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1714     clear_rvecs(SHIFTS, fr->fshift);
1715
1716     if (bNS)
1717     {
1718         wallcycle_start(wcycle, ewcNS);
1719
1720         if (graph && bStateChanged)
1721         {
1722             /* Calculate intramolecular shift vectors to make molecules whole */
1723             mk_mshift(fplog, graph, fr->ePBC, box, x);
1724         }
1725
1726         /* Do the actual neighbour searching */
1727         ns(fplog, fr, box,
1728            groups, top, mdatoms,
1729            cr, nrnb, bFillGrid,
1730            bDoLongRangeNS);
1731
1732         wallcycle_stop(wcycle, ewcNS);
1733     }
1734
1735     if (inputrec->implicit_solvent && bNS)
1736     {
1737         make_gb_nblist(cr, inputrec->gb_algorithm,
1738                        x, box, fr, &top->idef, graph, fr->born);
1739     }
1740
1741     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1742     {
1743         wallcycle_start(wcycle, ewcPPDURINGPME);
1744         dd_force_flop_start(cr->dd, nrnb);
1745     }
1746
1747     if (inputrec->bRot)
1748     {
1749         /* Enforced rotation has its own cycle counter that starts after the collective
1750          * coordinates have been communicated. It is added to ddCyclF to allow
1751          * for proper load-balancing */
1752         wallcycle_start(wcycle, ewcROT);
1753         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1754         wallcycle_stop(wcycle, ewcROT);
1755     }
1756
1757     /* Start the force cycle counter.
1758      * This counter is stopped after do_force_lowlevel.
1759      * No parallel communication should occur while this counter is running,
1760      * since that will interfere with the dynamic load balancing.
1761      */
1762     wallcycle_start(wcycle, ewcFORCE);
1763
1764     if (bDoForces)
1765     {
1766         /* Reset forces for which the virial is calculated separately:
1767          * PME/Ewald forces if necessary */
1768         if (fr->bF_NoVirSum)
1769         {
1770             if (flags & GMX_FORCE_VIRIAL)
1771             {
1772                 fr->f_novirsum = fr->f_novirsum_alloc;
1773                 if (fr->bDomDec)
1774                 {
1775                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1776                 }
1777                 else
1778                 {
1779                     clear_rvecs(homenr, fr->f_novirsum+start);
1780                 }
1781             }
1782             else
1783             {
1784                 /* We are not calculating the pressure so we do not need
1785                  * a separate array for forces that do not contribute
1786                  * to the pressure.
1787                  */
1788                 fr->f_novirsum = f;
1789             }
1790         }
1791
1792         /* Clear the short- and long-range forces */
1793         clear_rvecs(fr->natoms_force_constr, f);
1794         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1795         {
1796             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1797         }
1798
1799         clear_rvec(fr->vir_diag_posres);
1800     }
1801     if (inputrec->bPull && pull_have_constraint(inputrec->pull_work))
1802     {
1803         clear_pull_forces(inputrec->pull_work);
1804     }
1805
1806     /* update QMMMrec, if necessary */
1807     if (fr->bQMMM)
1808     {
1809         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1810     }
1811
1812     /* Compute the bonded and non-bonded energies and optionally forces */
1813     do_force_lowlevel(fr, inputrec, &(top->idef),
1814                       cr, nrnb, wcycle, mdatoms,
1815                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1816                       bBornRadii, box,
1817                       inputrec->fepvals, lambda,
1818                       graph, &(top->excls), fr->mu_tot,
1819                       flags,
1820                       &cycles_pme);
1821
1822     if (bSepLRF)
1823     {
1824         if (do_per_step(step, inputrec->nstcalclr))
1825         {
1826             /* Add the long range forces to the short range forces */
1827             for (i = 0; i < fr->natoms_force_constr; i++)
1828             {
1829                 rvec_add(fr->f_twin[i], f[i], f[i]);
1830             }
1831         }
1832     }
1833
1834     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1835
1836     if (ed)
1837     {
1838         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1839     }
1840
1841     if (DOMAINDECOMP(cr))
1842     {
1843         dd_force_flop_stop(cr->dd, nrnb);
1844         if (wcycle)
1845         {
1846             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1847         }
1848     }
1849
1850     if (bDoForces)
1851     {
1852         if (IR_ELEC_FIELD(*inputrec))
1853         {
1854             /* Compute forces due to electric field */
1855             calc_f_el(MASTER(cr) ? field : NULL,
1856                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1857                       inputrec->ex, inputrec->et, t);
1858         }
1859
1860         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1861         {
1862             /* Compute thermodynamic force in hybrid AdResS region */
1863             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1864                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1865         }
1866
1867         /* Communicate the forces */
1868         if (DOMAINDECOMP(cr))
1869         {
1870             wallcycle_start(wcycle, ewcMOVEF);
1871             dd_move_f(cr->dd, f, fr->fshift);
1872             /* Do we need to communicate the separate force array
1873              * for terms that do not contribute to the single sum virial?
1874              * Position restraints and electric fields do not introduce
1875              * inter-cg forces, only full electrostatics methods do.
1876              * When we do not calculate the virial, fr->f_novirsum = f,
1877              * so we have already communicated these forces.
1878              */
1879             if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1880                 (flags & GMX_FORCE_VIRIAL))
1881             {
1882                 dd_move_f(cr->dd, fr->f_novirsum, NULL);
1883             }
1884             if (bSepLRF)
1885             {
1886                 /* We should not update the shift forces here,
1887                  * since f_twin is already included in f.
1888                  */
1889                 dd_move_f(cr->dd, fr->f_twin, NULL);
1890             }
1891             wallcycle_stop(wcycle, ewcMOVEF);
1892         }
1893
1894         /* If we have NoVirSum forces, but we do not calculate the virial,
1895          * we sum fr->f_novirsum=f later.
1896          */
1897         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1898         {
1899             wallcycle_start(wcycle, ewcVSITESPREAD);
1900             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1901                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1902             wallcycle_stop(wcycle, ewcVSITESPREAD);
1903
1904             if (bSepLRF)
1905             {
1906                 wallcycle_start(wcycle, ewcVSITESPREAD);
1907                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1908                                nrnb,
1909                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1910                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1911             }
1912         }
1913
1914         if (flags & GMX_FORCE_VIRIAL)
1915         {
1916             /* Calculation of the virial must be done after vsites! */
1917             calc_virial(0, mdatoms->homenr, x, f,
1918                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1919         }
1920     }
1921
1922     if (inputrec->bPull && pull_have_potential(inputrec->pull_work))
1923     {
1924         pull_potential_wrapper(cr, inputrec, box, x,
1925                                f, vir_force, mdatoms, enerd, lambda, t,
1926                                wcycle);
1927     }
1928
1929     /* Add the forces from enforced rotation potentials (if any) */
1930     if (inputrec->bRot)
1931     {
1932         wallcycle_start(wcycle, ewcROTadd);
1933         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1934         wallcycle_stop(wcycle, ewcROTadd);
1935     }
1936
1937     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
1938     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
1939
1940     if (PAR(cr) && !(cr->duty & DUTY_PME))
1941     {
1942         /* In case of node-splitting, the PP nodes receive the long-range
1943          * forces, virial and energy from the PME nodes here.
1944          */
1945         pme_receive_force_ener(cr, wcycle, enerd, fr);
1946     }
1947
1948     if (bDoForces)
1949     {
1950         post_process_forces(cr, step, nrnb, wcycle,
1951                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1952                             flags);
1953     }
1954
1955     /* Sum the potential energy terms from group contributions */
1956     sum_epot(&(enerd->grpp), enerd->term);
1957 }
1958
1959 void do_force(FILE *fplog, t_commrec *cr,
1960               t_inputrec *inputrec,
1961               gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1962               gmx_localtop_t *top,
1963               gmx_groups_t *groups,
1964               matrix box, rvec x[], history_t *hist,
1965               rvec f[],
1966               tensor vir_force,
1967               t_mdatoms *mdatoms,
1968               gmx_enerdata_t *enerd, t_fcdata *fcd,
1969               real *lambda, t_graph *graph,
1970               t_forcerec *fr,
1971               gmx_vsite_t *vsite, rvec mu_tot,
1972               double t, FILE *field, gmx_edsam_t ed,
1973               gmx_bool bBornRadii,
1974               int flags)
1975 {
1976     /* modify force flag if not doing nonbonded */
1977     if (!fr->bNonbonded)
1978     {
1979         flags &= ~GMX_FORCE_NONBONDED;
1980     }
1981
1982     switch (inputrec->cutoff_scheme)
1983     {
1984         case ecutsVERLET:
1985             do_force_cutsVERLET(fplog, cr, inputrec,
1986                                 step, nrnb, wcycle,
1987                                 top,
1988                                 groups,
1989                                 box, x, hist,
1990                                 f, vir_force,
1991                                 mdatoms,
1992                                 enerd, fcd,
1993                                 lambda, graph,
1994                                 fr, fr->ic,
1995                                 vsite, mu_tot,
1996                                 t, field, ed,
1997                                 bBornRadii,
1998                                 flags);
1999             break;
2000         case ecutsGROUP:
2001             do_force_cutsGROUP(fplog, cr, inputrec,
2002                                step, nrnb, wcycle,
2003                                top,
2004                                groups,
2005                                box, x, hist,
2006                                f, vir_force,
2007                                mdatoms,
2008                                enerd, fcd,
2009                                lambda, graph,
2010                                fr, vsite, mu_tot,
2011                                t, field, ed,
2012                                bBornRadii,
2013                                flags);
2014             break;
2015         default:
2016             gmx_incons("Invalid cut-off scheme passed!");
2017     }
2018 }
2019
2020
2021 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2022                         t_inputrec *ir, t_mdatoms *md,
2023                         t_state *state, t_commrec *cr, t_nrnb *nrnb,
2024                         t_forcerec *fr, gmx_localtop_t *top)
2025 {
2026     int             i, m, start, end;
2027     gmx_int64_t     step;
2028     real            dt = ir->delta_t;
2029     real            dvdl_dum;
2030     rvec           *savex;
2031
2032     snew(savex, state->natoms);
2033
2034     start = 0;
2035     end   = md->homenr;
2036
2037     if (debug)
2038     {
2039         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2040                 start, md->homenr, end);
2041     }
2042     /* Do a first constrain to reset particles... */
2043     step = ir->init_step;
2044     if (fplog)
2045     {
2046         char buf[STEPSTRSIZE];
2047         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2048                 gmx_step_str(step, buf));
2049     }
2050     dvdl_dum = 0;
2051
2052     /* constrain the current position */
2053     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2054               ir, cr, step, 0, 1.0, md,
2055               state->x, state->x, NULL,
2056               fr->bMolPBC, state->box,
2057               state->lambda[efptBONDED], &dvdl_dum,
2058               NULL, NULL, nrnb, econqCoord);
2059     if (EI_VV(ir->eI))
2060     {
2061         /* constrain the inital velocity, and save it */
2062         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2063         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2064                   ir, cr, step, 0, 1.0, md,
2065                   state->x, state->v, state->v,
2066                   fr->bMolPBC, state->box,
2067                   state->lambda[efptBONDED], &dvdl_dum,
2068                   NULL, NULL, nrnb, econqVeloc);
2069     }
2070     /* constrain the inital velocities at t-dt/2 */
2071     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2072     {
2073         for (i = start; (i < end); i++)
2074         {
2075             for (m = 0; (m < DIM); m++)
2076             {
2077                 /* Reverse the velocity */
2078                 state->v[i][m] = -state->v[i][m];
2079                 /* Store the position at t-dt in buf */
2080                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2081             }
2082         }
2083         /* Shake the positions at t=-dt with the positions at t=0
2084          * as reference coordinates.
2085          */
2086         if (fplog)
2087         {
2088             char buf[STEPSTRSIZE];
2089             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2090                     gmx_step_str(step, buf));
2091         }
2092         dvdl_dum = 0;
2093         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2094                   ir, cr, step, -1, 1.0, md,
2095                   state->x, savex, NULL,
2096                   fr->bMolPBC, state->box,
2097                   state->lambda[efptBONDED], &dvdl_dum,
2098                   state->v, NULL, nrnb, econqCoord);
2099
2100         for (i = start; i < end; i++)
2101         {
2102             for (m = 0; m < DIM; m++)
2103             {
2104                 /* Re-reverse the velocities */
2105                 state->v[i][m] = -state->v[i][m];
2106             }
2107         }
2108     }
2109     sfree(savex);
2110 }
2111
2112
2113 static void
2114 integrate_table(real vdwtab[], real scale, int offstart, int rstart, int rend,
2115                 double *enerout, double *virout)
2116 {
2117     double enersum, virsum;
2118     double invscale, invscale2, invscale3;
2119     double r, ea, eb, ec, pa, pb, pc, pd;
2120     double y0, f, g, h;
2121     int    ri, offset;
2122     double tabfactor;
2123
2124     invscale  = 1.0/scale;
2125     invscale2 = invscale*invscale;
2126     invscale3 = invscale*invscale2;
2127
2128     /* Following summation derived from cubic spline definition,
2129      * Numerical Recipies in C, second edition, p. 113-116.  Exact for
2130      * the cubic spline.  We first calculate the negative of the
2131      * energy from rvdw to rvdw_switch, assuming that g(r)=1, and then
2132      * add the more standard, abrupt cutoff correction to that result,
2133      * yielding the long-range correction for a switched function.  We
2134      * perform both the pressure and energy loops at the same time for
2135      * simplicity, as the computational cost is low. */
2136
2137     if (offstart == 0)
2138     {
2139         /* Since the dispersion table has been scaled down a factor
2140          * 6.0 and the repulsion a factor 12.0 to compensate for the
2141          * c6/c12 parameters inside nbfp[] being scaled up (to save
2142          * flops in kernels), we need to correct for this.
2143          */
2144         tabfactor = 6.0;
2145     }
2146     else
2147     {
2148         tabfactor = 12.0;
2149     }
2150
2151     enersum = 0.0;
2152     virsum  = 0.0;
2153     for (ri = rstart; ri < rend; ++ri)
2154     {
2155         r  = ri*invscale;
2156         ea = invscale3;
2157         eb = 2.0*invscale2*r;
2158         ec = invscale*r*r;
2159
2160         pa = invscale3;
2161         pb = 3.0*invscale2*r;
2162         pc = 3.0*invscale*r*r;
2163         pd = r*r*r;
2164
2165         /* this "8" is from the packing in the vdwtab array - perhaps
2166            should be defined? */
2167
2168         offset = 8*ri + offstart;
2169         y0     = vdwtab[offset];
2170         f      = vdwtab[offset+1];
2171         g      = vdwtab[offset+2];
2172         h      = vdwtab[offset+3];
2173
2174         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);
2175         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);
2176     }
2177     *enerout = 4.0*M_PI*enersum*tabfactor;
2178     *virout  = 4.0*M_PI*virsum*tabfactor;
2179 }
2180
2181 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2182 {
2183     double   eners[2], virs[2], enersum, virsum;
2184     double   r0, rc3, rc9;
2185     int      ri0, ri1, i;
2186     real     scale, *vdwtab;
2187
2188     fr->enershiftsix    = 0;
2189     fr->enershifttwelve = 0;
2190     fr->enerdiffsix     = 0;
2191     fr->enerdifftwelve  = 0;
2192     fr->virdiffsix      = 0;
2193     fr->virdifftwelve   = 0;
2194
2195     if (eDispCorr != edispcNO)
2196     {
2197         for (i = 0; i < 2; i++)
2198         {
2199             eners[i] = 0;
2200             virs[i]  = 0;
2201         }
2202         if ((fr->vdw_modifier == eintmodPOTSHIFT) ||
2203             (fr->vdw_modifier == eintmodPOTSWITCH) ||
2204             (fr->vdw_modifier == eintmodFORCESWITCH) ||
2205             (fr->vdwtype == evdwSHIFT) ||
2206             (fr->vdwtype == evdwSWITCH))
2207         {
2208             if (((fr->vdw_modifier == eintmodPOTSWITCH) ||
2209                  (fr->vdw_modifier == eintmodFORCESWITCH) ||
2210                  (fr->vdwtype == evdwSWITCH)) && fr->rvdw_switch == 0)
2211             {
2212                 gmx_fatal(FARGS,
2213                           "With dispersion correction rvdw-switch can not be zero "
2214                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2215             }
2216
2217             scale  = fr->nblists[0].table_vdw.scale;
2218             vdwtab = fr->nblists[0].table_vdw.data;
2219
2220             /* Round the cut-offs to exact table values for precision */
2221             ri0  = static_cast<int>(floor(fr->rvdw_switch*scale));
2222             ri1  = static_cast<int>(ceil(fr->rvdw*scale));
2223
2224             /* The code below has some support for handling force-switching, i.e.
2225              * when the force (instead of potential) is switched over a limited
2226              * region. This leads to a constant shift in the potential inside the
2227              * switching region, which we can handle by adding a constant energy
2228              * term in the force-switch case just like when we do potential-shift.
2229              *
2230              * For now this is not enabled, but to keep the functionality in the
2231              * code we check separately for switch and shift. When we do force-switch
2232              * the shifting point is rvdw_switch, while it is the cutoff when we
2233              * have a classical potential-shift.
2234              *
2235              * For a pure potential-shift the potential has a constant shift
2236              * all the way out to the cutoff, and that is it. For other forms
2237              * we need to calculate the constant shift up to the point where we
2238              * start modifying the potential.
2239              */
2240             ri0  = (fr->vdw_modifier == eintmodPOTSHIFT) ? ri1 : ri0;
2241
2242             r0   = ri0/scale;
2243             rc3  = r0*r0*r0;
2244             rc9  = rc3*rc3*rc3;
2245
2246             if ((fr->vdw_modifier == eintmodFORCESWITCH) ||
2247                 (fr->vdwtype == evdwSHIFT))
2248             {
2249                 /* Determine the constant energy shift below rvdw_switch.
2250                  * Table has a scale factor since we have scaled it down to compensate
2251                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2252                  */
2253                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2254                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2255             }
2256             else if (fr->vdw_modifier == eintmodPOTSHIFT)
2257             {
2258                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3));
2259                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3));
2260             }
2261
2262             /* Add the constant part from 0 to rvdw_switch.
2263              * This integration from 0 to rvdw_switch overcounts the number
2264              * of interactions by 1, as it also counts the self interaction.
2265              * We will correct for this later.
2266              */
2267             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2268             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2269
2270             /* Calculate the contribution in the range [r0,r1] where we
2271              * modify the potential. For a pure potential-shift modifier we will
2272              * have ri0==ri1, and there will not be any contribution here.
2273              */
2274             for (i = 0; i < 2; i++)
2275             {
2276                 enersum = 0;
2277                 virsum  = 0;
2278                 integrate_table(vdwtab, scale, (i == 0 ? 0 : 4), ri0, ri1, &enersum, &virsum);
2279                 eners[i] -= enersum;
2280                 virs[i]  -= virsum;
2281             }
2282
2283             /* Alright: Above we compensated by REMOVING the parts outside r0
2284              * corresponding to the ideal VdW 1/r6 and /r12 potentials.
2285              *
2286              * Regardless of whether r0 is the point where we start switching,
2287              * or the cutoff where we calculated the constant shift, we include
2288              * all the parts we are missing out to infinity from r0 by
2289              * calculating the analytical dispersion correction.
2290              */
2291             eners[0] += -4.0*M_PI/(3.0*rc3);
2292             eners[1] +=  4.0*M_PI/(9.0*rc9);
2293             virs[0]  +=  8.0*M_PI/rc3;
2294             virs[1]  += -16.0*M_PI/(3.0*rc9);
2295         }
2296         else if (fr->vdwtype == evdwCUT ||
2297                  EVDW_PME(fr->vdwtype) ||
2298                  fr->vdwtype == evdwUSER)
2299         {
2300             if (fr->vdwtype == evdwUSER && fplog)
2301             {
2302                 fprintf(fplog,
2303                         "WARNING: using dispersion correction with user tables\n");
2304             }
2305
2306             /* Note that with LJ-PME, the dispersion correction is multiplied
2307              * by the difference between the actual C6 and the value of C6
2308              * that would produce the combination rule.
2309              * This means the normal energy and virial difference formulas
2310              * can be used here.
2311              */
2312
2313             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2314             rc9  = rc3*rc3*rc3;
2315             /* Contribution beyond the cut-off */
2316             eners[0] += -4.0*M_PI/(3.0*rc3);
2317             eners[1] +=  4.0*M_PI/(9.0*rc9);
2318             if (fr->vdw_modifier == eintmodPOTSHIFT)
2319             {
2320                 /* Contribution within the cut-off */
2321                 eners[0] += -4.0*M_PI/(3.0*rc3);
2322                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2323             }
2324             /* Contribution beyond the cut-off */
2325             virs[0]  +=  8.0*M_PI/rc3;
2326             virs[1]  += -16.0*M_PI/(3.0*rc9);
2327         }
2328         else
2329         {
2330             gmx_fatal(FARGS,
2331                       "Dispersion correction is not implemented for vdw-type = %s",
2332                       evdw_names[fr->vdwtype]);
2333         }
2334
2335         /* When we deprecate the group kernels the code below can go too */
2336         if (fr->vdwtype == evdwPME && fr->cutoff_scheme == ecutsGROUP)
2337         {
2338             /* Calculate self-interaction coefficient (assuming that
2339              * the reciprocal-space contribution is constant in the
2340              * region that contributes to the self-interaction).
2341              */
2342             fr->enershiftsix = pow(fr->ewaldcoeff_lj, 6) / 6.0;
2343
2344             eners[0] += -pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3)/3.0;
2345             virs[0]  +=  pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3);
2346         }
2347
2348         fr->enerdiffsix    = eners[0];
2349         fr->enerdifftwelve = eners[1];
2350         /* The 0.5 is due to the Gromacs definition of the virial */
2351         fr->virdiffsix     = 0.5*virs[0];
2352         fr->virdifftwelve  = 0.5*virs[1];
2353     }
2354 }
2355
2356 void calc_dispcorr(t_inputrec *ir, t_forcerec *fr,
2357                    int natoms,
2358                    matrix box, real lambda, tensor pres, tensor virial,
2359                    real *prescorr, real *enercorr, real *dvdlcorr)
2360 {
2361     gmx_bool bCorrAll, bCorrPres;
2362     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2363     int      m;
2364
2365     *prescorr = 0;
2366     *enercorr = 0;
2367     *dvdlcorr = 0;
2368
2369     clear_mat(virial);
2370     clear_mat(pres);
2371
2372     if (ir->eDispCorr != edispcNO)
2373     {
2374         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2375                      ir->eDispCorr == edispcAllEnerPres);
2376         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2377                      ir->eDispCorr == edispcAllEnerPres);
2378
2379         invvol = 1/det(box);
2380         if (fr->n_tpi)
2381         {
2382             /* Only correct for the interactions with the inserted molecule */
2383             dens   = (natoms - fr->n_tpi)*invvol;
2384             ninter = fr->n_tpi;
2385         }
2386         else
2387         {
2388             dens   = natoms*invvol;
2389             ninter = 0.5*natoms;
2390         }
2391
2392         if (ir->efep == efepNO)
2393         {
2394             avcsix    = fr->avcsix[0];
2395             avctwelve = fr->avctwelve[0];
2396         }
2397         else
2398         {
2399             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2400             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2401         }
2402
2403         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2404         *enercorr += avcsix*enerdiff;
2405         dvdlambda  = 0.0;
2406         if (ir->efep != efepNO)
2407         {
2408             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2409         }
2410         if (bCorrAll)
2411         {
2412             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2413             *enercorr += avctwelve*enerdiff;
2414             if (fr->efep != efepNO)
2415             {
2416                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2417             }
2418         }
2419
2420         if (bCorrPres)
2421         {
2422             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2423             if (ir->eDispCorr == edispcAllEnerPres)
2424             {
2425                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2426             }
2427             /* The factor 2 is because of the Gromacs virial definition */
2428             spres = -2.0*invvol*svir*PRESFAC;
2429
2430             for (m = 0; m < DIM; m++)
2431             {
2432                 virial[m][m] += svir;
2433                 pres[m][m]   += spres;
2434             }
2435             *prescorr += spres;
2436         }
2437
2438         /* Can't currently control when it prints, for now, just print when degugging */
2439         if (debug)
2440         {
2441             if (bCorrAll)
2442             {
2443                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2444                         avcsix, avctwelve);
2445             }
2446             if (bCorrPres)
2447             {
2448                 fprintf(debug,
2449                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2450                         *enercorr, spres, svir);
2451             }
2452             else
2453             {
2454                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2455             }
2456         }
2457
2458         if (fr->efep != efepNO)
2459         {
2460             *dvdlcorr += dvdlambda;
2461         }
2462     }
2463 }
2464
2465 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2466                   t_graph *graph, rvec x[])
2467 {
2468     if (fplog)
2469     {
2470         fprintf(fplog, "Removing pbc first time\n");
2471     }
2472     calc_shifts(box, fr->shift_vec);
2473     if (graph)
2474     {
2475         mk_mshift(fplog, graph, fr->ePBC, box, x);
2476         if (gmx_debug_at)
2477         {
2478             p_graph(debug, "do_pbc_first 1", graph);
2479         }
2480         shift_self(graph, box, x);
2481         /* By doing an extra mk_mshift the molecules that are broken
2482          * because they were e.g. imported from another software
2483          * will be made whole again. Such are the healing powers
2484          * of GROMACS.
2485          */
2486         mk_mshift(fplog, graph, fr->ePBC, box, x);
2487         if (gmx_debug_at)
2488         {
2489             p_graph(debug, "do_pbc_first 2", graph);
2490         }
2491     }
2492     if (fplog)
2493     {
2494         fprintf(fplog, "Done rmpbc\n");
2495     }
2496 }
2497
2498 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2499                             gmx_mtop_t *mtop, rvec x[],
2500                             gmx_bool bFirst)
2501 {
2502     t_graph        *graph;
2503     int             mb, as, mol;
2504     gmx_molblock_t *molb;
2505
2506     if (bFirst && fplog)
2507     {
2508         fprintf(fplog, "Removing pbc first time\n");
2509     }
2510
2511     snew(graph, 1);
2512     as = 0;
2513     for (mb = 0; mb < mtop->nmolblock; mb++)
2514     {
2515         molb = &mtop->molblock[mb];
2516         if (molb->natoms_mol == 1 ||
2517             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2518         {
2519             /* Just one atom or charge group in the molecule, no PBC required */
2520             as += molb->nmol*molb->natoms_mol;
2521         }
2522         else
2523         {
2524             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2525             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2526                            0, molb->natoms_mol, FALSE, FALSE, graph);
2527
2528             for (mol = 0; mol < molb->nmol; mol++)
2529             {
2530                 mk_mshift(fplog, graph, ePBC, box, x+as);
2531
2532                 shift_self(graph, box, x+as);
2533                 /* The molecule is whole now.
2534                  * We don't need the second mk_mshift call as in do_pbc_first,
2535                  * since we no longer need this graph.
2536                  */
2537
2538                 as += molb->natoms_mol;
2539             }
2540             done_graph(graph);
2541         }
2542     }
2543     sfree(graph);
2544 }
2545
2546 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2547                        gmx_mtop_t *mtop, rvec x[])
2548 {
2549     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2550 }
2551
2552 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2553                  gmx_mtop_t *mtop, rvec x[])
2554 {
2555     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2556 }
2557
2558 void finish_run(FILE *fplog, t_commrec *cr,
2559                 t_inputrec *inputrec,
2560                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2561                 gmx_walltime_accounting_t walltime_accounting,
2562                 nonbonded_verlet_t *nbv,
2563                 gmx_bool bWriteStat)
2564 {
2565     t_nrnb *nrnb_tot = NULL;
2566     double  delta_t  = 0;
2567     double  nbfs     = 0, mflop = 0;
2568     double  elapsed_time,
2569             elapsed_time_over_all_ranks,
2570             elapsed_time_over_all_threads,
2571             elapsed_time_over_all_threads_over_all_ranks;
2572     wallcycle_sum(cr, wcycle);
2573
2574     if (cr->nnodes > 1)
2575     {
2576         snew(nrnb_tot, 1);
2577 #ifdef GMX_MPI
2578         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2579                       cr->mpi_comm_mysim);
2580 #endif
2581     }
2582     else
2583     {
2584         nrnb_tot = nrnb;
2585     }
2586
2587     elapsed_time                                 = walltime_accounting_get_elapsed_time(walltime_accounting);
2588     elapsed_time_over_all_ranks                  = elapsed_time;
2589     elapsed_time_over_all_threads                = walltime_accounting_get_elapsed_time_over_all_threads(walltime_accounting);
2590     elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
2591 #ifdef GMX_MPI
2592     if (cr->nnodes > 1)
2593     {
2594         /* reduce elapsed_time over all MPI ranks in the current simulation */
2595         MPI_Allreduce(&elapsed_time,
2596                       &elapsed_time_over_all_ranks,
2597                       1, MPI_DOUBLE, MPI_SUM,
2598                       cr->mpi_comm_mysim);
2599         elapsed_time_over_all_ranks /= cr->nnodes;
2600         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
2601          * current simulation. */
2602         MPI_Allreduce(&elapsed_time_over_all_threads,
2603                       &elapsed_time_over_all_threads_over_all_ranks,
2604                       1, MPI_DOUBLE, MPI_SUM,
2605                       cr->mpi_comm_mysim);
2606     }
2607 #endif
2608
2609     if (SIMMASTER(cr))
2610     {
2611         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2612     }
2613     if (cr->nnodes > 1)
2614     {
2615         sfree(nrnb_tot);
2616     }
2617
2618     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2619     {
2620         print_dd_statistics(cr, inputrec, fplog);
2621     }
2622
2623     if (SIMMASTER(cr))
2624     {
2625         struct gmx_wallclock_gpu_t* gputimes = use_GPU(nbv) ? nbnxn_gpu_get_timings(nbv->gpu_nbv) : NULL;
2626
2627         wallcycle_print(fplog, cr->nnodes, cr->npmenodes,
2628                         elapsed_time_over_all_ranks,
2629                         wcycle, gputimes);
2630
2631         if (EI_DYNAMICS(inputrec->eI))
2632         {
2633             delta_t = inputrec->delta_t;
2634         }
2635
2636         if (fplog)
2637         {
2638             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
2639                        elapsed_time_over_all_ranks,
2640                        walltime_accounting_get_nsteps_done(walltime_accounting),
2641                        delta_t, nbfs, mflop);
2642         }
2643         if (bWriteStat)
2644         {
2645             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
2646                        elapsed_time_over_all_ranks,
2647                        walltime_accounting_get_nsteps_done(walltime_accounting),
2648                        delta_t, nbfs, mflop);
2649         }
2650     }
2651 }
2652
2653 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2654 {
2655     /* this function works, but could probably use a logic rewrite to keep all the different
2656        types of efep straight. */
2657
2658     int       i;
2659     t_lambda *fep = ir->fepvals;
2660
2661     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2662     {
2663         for (i = 0; i < efptNR; i++)
2664         {
2665             lambda[i] = 0.0;
2666             if (lam0)
2667             {
2668                 lam0[i] = 0.0;
2669             }
2670         }
2671         return;
2672     }
2673     else
2674     {
2675         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2676                                              if checkpoint is set -- a kludge is in for now
2677                                              to prevent this.*/
2678         for (i = 0; i < efptNR; i++)
2679         {
2680             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2681             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2682             {
2683                 lambda[i] = fep->init_lambda;
2684                 if (lam0)
2685                 {
2686                     lam0[i] = lambda[i];
2687                 }
2688             }
2689             else
2690             {
2691                 lambda[i] = fep->all_lambda[i][*fep_state];
2692                 if (lam0)
2693                 {
2694                     lam0[i] = lambda[i];
2695                 }
2696             }
2697         }
2698         if (ir->bSimTemp)
2699         {
2700             /* need to rescale control temperatures to match current state */
2701             for (i = 0; i < ir->opts.ngtc; i++)
2702             {
2703                 if (ir->opts.ref_t[i] > 0)
2704                 {
2705                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2706                 }
2707             }
2708         }
2709     }
2710
2711     /* Send to the log the information on the current lambdas */
2712     if (fplog != NULL)
2713     {
2714         fprintf(fplog, "Initial vector of lambda components:[ ");
2715         for (i = 0; i < efptNR; i++)
2716         {
2717             fprintf(fplog, "%10.4f ", lambda[i]);
2718         }
2719         fprintf(fplog, "]\n");
2720     }
2721     return;
2722 }
2723
2724
2725 void init_md(FILE *fplog,
2726              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2727              double *t, double *t0,
2728              real *lambda, int *fep_state, double *lam0,
2729              t_nrnb *nrnb, gmx_mtop_t *mtop,
2730              gmx_update_t *upd,
2731              int nfile, const t_filenm fnm[],
2732              gmx_mdoutf_t *outf, t_mdebin **mdebin,
2733              tensor force_vir, tensor shake_vir, rvec mu_tot,
2734              gmx_bool *bSimAnn, t_vcm **vcm, unsigned long Flags,
2735              gmx_wallcycle_t wcycle)
2736 {
2737     int  i;
2738
2739     /* Initial values */
2740     *t = *t0       = ir->init_t;
2741
2742     *bSimAnn = FALSE;
2743     for (i = 0; i < ir->opts.ngtc; i++)
2744     {
2745         /* set bSimAnn if any group is being annealed */
2746         if (ir->opts.annealing[i] != eannNO)
2747         {
2748             *bSimAnn = TRUE;
2749         }
2750     }
2751     if (*bSimAnn)
2752     {
2753         update_annealing_target_temp(&(ir->opts), ir->init_t);
2754     }
2755
2756     /* Initialize lambda variables */
2757     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2758
2759     if (upd)
2760     {
2761         *upd = init_update(ir);
2762     }
2763
2764
2765     if (vcm != NULL)
2766     {
2767         *vcm = init_vcm(fplog, &mtop->groups, ir);
2768     }
2769
2770     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2771     {
2772         if (ir->etc == etcBERENDSEN)
2773         {
2774             please_cite(fplog, "Berendsen84a");
2775         }
2776         if (ir->etc == etcVRESCALE)
2777         {
2778             please_cite(fplog, "Bussi2007a");
2779         }
2780         if (ir->eI == eiSD1)
2781         {
2782             please_cite(fplog, "Goga2012");
2783         }
2784     }
2785     if ((ir->et[XX].n > 0) || (ir->et[YY].n > 0) || (ir->et[ZZ].n > 0))
2786     {
2787         please_cite(fplog, "Caleman2008a");
2788     }
2789     init_nrnb(nrnb);
2790
2791     if (nfile != -1)
2792     {
2793         *outf = init_mdoutf(fplog, nfile, fnm, Flags, cr, ir, mtop, oenv, wcycle);
2794
2795         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : mdoutf_get_fp_ene(*outf),
2796                               mtop, ir, mdoutf_get_fp_dhdl(*outf));
2797     }
2798
2799     if (ir->bAdress)
2800     {
2801         please_cite(fplog, "Fritsch12");
2802         please_cite(fplog, "Junghans10");
2803     }
2804     /* Initiate variables */
2805     clear_mat(force_vir);
2806     clear_mat(shake_vir);
2807     clear_rvec(mu_tot);
2808
2809     debug_gmx();
2810 }