src/gromacs/mdlib/sim_util.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #ifdef HAVE_CONFIG_H
  38 #include <config.h>
  39 #endif
  40
  41 #include <assert.h>
  42 #include <math.h>
  43 #include <stdio.h>
  44 #include <string.h>
  45 #ifdef HAVE_SYS_TIME_H
  46 #include <sys/time.h>
  47 #endif
  48
  49 #include "typedefs.h"
  50 #include "gromacs/utility/cstringutil.h"
  51 #include "gromacs/utility/smalloc.h"
  52 #include "names.h"
  53 #include "txtdump.h"
  54 #include "pbc.h"
  55 #include "chargegroup.h"
  56 #include "vec.h"
  57 #include "nrnb.h"
  58 #include "mshift.h"
  59 #include "mdrun.h"
  60 #include "sim_util.h"
  61 #include "update.h"
  62 #include "physics.h"
  63 #include "main.h"
  64 #include "mdatoms.h"
  65 #include "force.h"
  66 #include "bondf.h"
  67 #include "pme.h"
  68 #include "disre.h"
  69 #include "orires.h"
  70 #include "network.h"
  71 #include "calcmu.h"
  72 #include "constr.h"
  73 #include "xvgr.h"
  74 #include "copyrite.h"
  75 #include "domdec.h"
  76 #include "genborn.h"
  77 #include "nbnxn_atomdata.h"
  78 #include "nbnxn_search.h"
  79 #include "nbnxn_kernels/nbnxn_kernel_ref.h"
  80 #include "nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.h"
  81 #include "nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn.h"
  82 #include "nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  83 #include "nonbonded.h"
  84 #include "../gmxlib/nonbonded/nb_kernel.h"
  85 #include "../gmxlib/nonbonded/nb_free_energy.h"
  86
  87 #include "gromacs/timing/wallcycle.h"
  88 #include "gromacs/timing/walltime_accounting.h"
  89 #include "gromacs/utility/gmxmpi.h"
  90 #include "gromacs/essentialdynamics/edsam.h"
  91 #include "gromacs/pulling/pull.h"
  92 #include "gromacs/pulling/pull_rotation.h"
  93 #include "gromacs/imd/imd.h"
  94 #include "adress.h"
  95 #include "qmmm.h"
  96
  97 #include "gmx_omp_nthreads.h"
  98
  99 #include "nbnxn_cuda_data_mgmt.h"
 100 #include "nbnxn_cuda/nbnxn_cuda.h"
 101
 102 void print_time(FILE                     *out,
 103                 gmx_walltime_accounting_t walltime_accounting,
 104                 gmx_int64_t               step,
 105                 t_inputrec               *ir,
 106                 t_commrec gmx_unused     *cr)
 107 {
 108     time_t finish;
 109     char   timebuf[STRLEN];
 110     double dt, elapsed_seconds, time_per_step;
 111     char   buf[48];
 112
 113 #ifndef GMX_THREAD_MPI
 114     if (!PAR(cr))
 115 #endif
 116     {
 117         fprintf(out, "\r");
 118     }
 119     fprintf(out, "step %s", gmx_step_str(step, buf));
 120     if ((step >= ir->nstlist))
 121     {
 122         double seconds_since_epoch = gmx_gettime();
 123         elapsed_seconds = seconds_since_epoch - walltime_accounting_get_start_time_stamp(walltime_accounting);
 124         time_per_step   = elapsed_seconds/(step - ir->init_step + 1);
 125         dt              = (ir->nsteps + ir->init_step - step) * time_per_step;
 126
 127         if (ir->nsteps >= 0)
 128         {
 129             if (dt >= 300)
 130             {
 131                 finish = (time_t) (seconds_since_epoch + dt);
 132                 gmx_ctime_r(&finish, timebuf, STRLEN);
 133                 sprintf(buf, "%s", timebuf);
 134                 buf[strlen(buf)-1] = '\0';
 135                 fprintf(out, ", will finish %s", buf);
 136             }
 137             else
 138             {
 139                 fprintf(out, ", remaining wall clock time: %5d s          ", (int)dt);
 140             }
 141         }
 142         else
 143         {
 144             fprintf(out, " performance: %.1f ns/day    ",
 145                     ir->delta_t/1000*24*60*60/time_per_step);
 146         }
 147     }
 148 #ifndef GMX_THREAD_MPI
 149     if (PAR(cr))
 150     {
 151         fprintf(out, "\n");
 152     }
 153 #endif
 154
 155     fflush(out);
 156 }
 157
 158 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
 159                          double the_time)
 160 {
 161     char   time_string[STRLEN];
 162
 163     if (!fplog)
 164     {
 165         return;
 166     }
 167
 168     {
 169         int    i;
 170         char   timebuf[STRLEN];
 171         time_t temp_time = (time_t) the_time;
 172
 173         gmx_ctime_r(&temp_time, timebuf, STRLEN);
 174         for (i = 0; timebuf[i] >= ' '; i++)
 175         {
 176             time_string[i] = timebuf[i];
 177         }
 178         time_string[i] = '\0';
 179     }
 180
 181     fprintf(fplog, "%s on rank %d %s\n", title, nodeid, time_string);
 182 }
 183
 184 void print_start(FILE *fplog, t_commrec *cr,
 185                  gmx_walltime_accounting_t walltime_accounting,
 186                  const char *name)
 187 {
 188     char buf[STRLEN];
 189
 190     sprintf(buf, "Started %s", name);
 191     print_date_and_time(fplog, cr->nodeid, buf,
 192                         walltime_accounting_get_start_time_stamp(walltime_accounting));
 193 }
 194
 195 static void sum_forces(int start, int end, rvec f[], rvec flr[])
 196 {
 197     int i;
 198
 199     if (gmx_debug_at)
 200     {
 201         pr_rvecs(debug, 0, "fsr", f+start, end-start);
 202         pr_rvecs(debug, 0, "flr", flr+start, end-start);
 203     }
 204     for (i = start; (i < end); i++)
 205     {
 206         rvec_inc(f[i], flr[i]);
 207     }
 208 }
 209
 210 /*
 211  * calc_f_el calculates forces due to an electric field.
 212  *
 213  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 214  *
 215  * Et[] contains the parameters for the time dependent
 216  * part of the field (not yet used).
 217  * Ex[] contains the parameters for
 218  * the spatial dependent part of the field. You can have cool periodic
 219  * fields in principle, but only a constant field is supported
 220  * now.
 221  * The function should return the energy due to the electric field
 222  * (if any) but for now returns 0.
 223  *
 224  * WARNING:
 225  * There can be problems with the virial.
 226  * Since the field is not self-consistent this is unavoidable.
 227  * For neutral molecules the virial is correct within this approximation.
 228  * For neutral systems with many charged molecules the error is small.
 229  * But for systems with a net charge or a few charged molecules
 230  * the error can be significant when the field is high.
 231  * Solution: implement a self-consitent electric field into PME.
 232  */
 233 static void calc_f_el(FILE *fp, int  start, int homenr,
 234                       real charge[], rvec f[],
 235                       t_cosines Ex[], t_cosines Et[], double t)
 236 {
 237     rvec Ext;
 238     real t0;
 239     int  i, m;
 240
 241     for (m = 0; (m < DIM); m++)
 242     {
 243         if (Et[m].n > 0)
 244         {
 245             if (Et[m].n == 3)
 246             {
 247                 t0     = Et[m].a[1];
 248                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 249             }
 250             else
 251             {
 252                 Ext[m] = cos(Et[m].a[0]*t);
 253             }
 254         }
 255         else
 256         {
 257             Ext[m] = 1.0;
 258         }
 259         if (Ex[m].n > 0)
 260         {
 261             /* Convert the field strength from V/nm to MD-units */
 262             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 263             for (i = start; (i < start+homenr); i++)
 264             {
 265                 f[i][m] += charge[i]*Ext[m];
 266             }
 267         }
 268         else
 269         {
 270             Ext[m] = 0;
 271         }
 272     }
 273     if (fp != NULL)
 274     {
 275         fprintf(fp, "%10g  %10g  %10g  %10g #FIELD\n", t,
 276                 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
 277     }
 278 }
 279
 280 static void calc_virial(int start, int homenr, rvec x[], rvec f[],
 281                         tensor vir_part, t_graph *graph, matrix box,
 282                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 283 {
 284     int    i, j;
 285     tensor virtest;
 286
 287     /* The short-range virial from surrounding boxes */
 288     clear_mat(vir_part);
 289     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 290     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 291
 292     /* Calculate partial virial, for local atoms only, based on short range.
 293      * Total virial is computed in global_stat, called from do_md
 294      */
 295     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 296     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 297
 298     /* Add position restraint contribution */
 299     for (i = 0; i < DIM; i++)
 300     {
 301         vir_part[i][i] += fr->vir_diag_posres[i];
 302     }
 303
 304     /* Add wall contribution */
 305     for (i = 0; i < DIM; i++)
 306     {
 307         vir_part[i][ZZ] += fr->vir_wall_z[i];
 308     }
 309
 310     if (debug)
 311     {
 312         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 313     }
 314 }
 315
 316 static void posres_wrapper(FILE *fplog,
 317                            int flags,
 318                            gmx_bool bSepDVDL,
 319                            t_inputrec *ir,
 320                            t_nrnb *nrnb,
 321                            gmx_localtop_t *top,
 322                            matrix box, rvec x[],
 323                            gmx_enerdata_t *enerd,
 324                            real *lambda,
 325                            t_forcerec *fr)
 326 {
 327     t_pbc pbc;
 328     real  v, dvdl;
 329     int   i;
 330
 331     /* Position restraints always require full pbc */
 332     set_pbc(&pbc, ir->ePBC, box);
 333     dvdl = 0;
 334     v    = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 335                   top->idef.iparams_posres,
 336                   (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 337                   ir->ePBC == epbcNONE ? NULL : &pbc,
 338                   lambda[efptRESTRAINT], &dvdl,
 339                   fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 340     if (bSepDVDL)
 341     {
 342         gmx_print_sepdvdl(fplog, interaction_function[F_POSRES].longname, v, dvdl);
 343     }
 344     enerd->term[F_POSRES] += v;
 345     /* If just the force constant changes, the FEP term is linear,
 346      * but if k changes, it is not.
 347      */
 348     enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
 349     inc_nrnb(nrnb, eNR_POSRES, top->idef.il[F_POSRES].nr/2);
 350
 351     if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
 352     {
 353         for (i = 0; i < enerd->n_lambda; i++)
 354         {
 355             real dvdl_dum, lambda_dum;
 356
 357             lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
 358             v          = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
 359                                 top->idef.iparams_posres,
 360                                 (const rvec*)x, NULL, NULL,
 361                                 ir->ePBC == epbcNONE ? NULL : &pbc, lambda_dum, &dvdl,
 362                                 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
 363             enerd->enerpart_lambda[i] += v;
 364         }
 365     }
 366 }
 367
 368 static void fbposres_wrapper(t_inputrec *ir,
 369                              t_nrnb *nrnb,
 370                              gmx_localtop_t *top,
 371                              matrix box, rvec x[],
 372                              gmx_enerdata_t *enerd,
 373                              t_forcerec *fr)
 374 {
 375     t_pbc pbc;
 376     real  v;
 377
 378     /* Flat-bottomed position restraints always require full pbc */
 379     set_pbc(&pbc, ir->ePBC, box);
 380     v = fbposres(top->idef.il[F_FBPOSRES].nr, top->idef.il[F_FBPOSRES].iatoms,
 381                  top->idef.iparams_fbposres,
 382                  (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
 383                  ir->ePBC == epbcNONE ? NULL : &pbc,
 384                  fr->rc_scaling, fr->ePBC, fr->posres_com);
 385     enerd->term[F_FBPOSRES] += v;
 386     inc_nrnb(nrnb, eNR_FBPOSRES, top->idef.il[F_FBPOSRES].nr/2);
 387 }
 388
 389 static void pull_potential_wrapper(FILE *fplog,
 390                                    gmx_bool bSepDVDL,
 391                                    t_commrec *cr,
 392                                    t_inputrec *ir,
 393                                    matrix box, rvec x[],
 394                                    rvec f[],
 395                                    tensor vir_force,
 396                                    t_mdatoms *mdatoms,
 397                                    gmx_enerdata_t *enerd,
 398                                    real *lambda,
 399                                    double t,
 400                                    gmx_wallcycle_t wcycle)
 401 {
 402     t_pbc  pbc;
 403     real   dvdl;
 404
 405     /* Calculate the center of mass forces, this requires communication,
 406      * which is why pull_potential is called close to other communication.
 407      * The virial contribution is calculated directly,
 408      * which is why we call pull_potential after calc_virial.
 409      */
 410     wallcycle_start(wcycle, ewcPULLPOT);
 411     set_pbc(&pbc, ir->ePBC, box);
 412     dvdl                     = 0;
 413     enerd->term[F_COM_PULL] +=
 414         pull_potential(ir->ePull, ir->pull, mdatoms, &pbc,
 415                        cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
 416     if (bSepDVDL)
 417     {
 418         gmx_print_sepdvdl(fplog, "Com pull", enerd->term[F_COM_PULL], dvdl);
 419     }
 420     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 421     wallcycle_stop(wcycle, ewcPULLPOT);
 422 }
 423
 424 static void pme_receive_force_ener(FILE           *fplog,
 425                                    gmx_bool        bSepDVDL,
 426                                    t_commrec      *cr,
 427                                    gmx_wallcycle_t wcycle,
 428                                    gmx_enerdata_t *enerd,
 429                                    t_forcerec     *fr)
 430 {
 431     real   e_q, e_lj, v, dvdl_q, dvdl_lj;
 432     float  cycles_ppdpme, cycles_seppme;
 433
 434     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 435     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 436
 437     /* In case of node-splitting, the PP nodes receive the long-range
 438      * forces, virial and energy from the PME nodes here.
 439      */
 440     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 441     dvdl_q  = 0;
 442     dvdl_lj = 0;
 443     gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e_q,
 444                       fr->vir_lj_recip, &e_lj, &dvdl_q, &dvdl_lj,
 445                       &cycles_seppme);
 446     if (bSepDVDL)
 447     {
 448         gmx_print_sepdvdl(fplog, "Electrostatic PME mesh", e_q, dvdl_q);
 449         gmx_print_sepdvdl(fplog, "Lennard-Jones PME mesh", e_lj, dvdl_lj);
 450     }
 451     enerd->term[F_COUL_RECIP] += e_q;
 452     enerd->term[F_LJ_RECIP]   += e_lj;
 453     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 454     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 455
 456     if (wcycle)
 457     {
 458         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 459     }
 460     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 461 }
 462
 463 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
 464                                gmx_int64_t step, real pforce, rvec *x, rvec *f)
 465 {
 466     int  i;
 467     real pf2, fn2;
 468     char buf[STEPSTRSIZE];
 469
 470     pf2 = sqr(pforce);
 471     for (i = 0; i < md->homenr; i++)
 472     {
 473         fn2 = norm2(f[i]);
 474         /* We also catch NAN, if the compiler does not optimize this away. */
 475         if (fn2 >= pf2 || fn2 != fn2)
 476         {
 477             fprintf(fp, "step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 478                     gmx_step_str(step, buf),
 479                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
 480         }
 481     }
 482 }
 483
 484 static void post_process_forces(t_commrec *cr,
 485                                 gmx_int64_t step,
 486                                 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 487                                 gmx_localtop_t *top,
 488                                 matrix box, rvec x[],
 489                                 rvec f[],
 490                                 tensor vir_force,
 491                                 t_mdatoms *mdatoms,
 492                                 t_graph *graph,
 493                                 t_forcerec *fr, gmx_vsite_t *vsite,
 494                                 int flags)
 495 {
 496     if (fr->bF_NoVirSum)
 497     {
 498         if (vsite)
 499         {
 500             /* Spread the mesh force on virtual sites to the other particles...
 501              * This is parallellized. MPI communication is performed
 502              * if the constructing atoms aren't local.
 503              */
 504             wallcycle_start(wcycle, ewcVSITESPREAD);
 505             spread_vsite_f(vsite, x, fr->f_novirsum, NULL,
 506                            (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
 507                            nrnb,
 508                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
 509             wallcycle_stop(wcycle, ewcVSITESPREAD);
 510         }
 511         if (flags & GMX_FORCE_VIRIAL)
 512         {
 513             /* Now add the forces, this is local */
 514             if (fr->bDomDec)
 515             {
 516                 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
 517             }
 518             else
 519             {
 520                 sum_forces(0, mdatoms->homenr,
 521                            f, fr->f_novirsum);
 522             }
 523             if (EEL_FULL(fr->eeltype))
 524             {
 525                 /* Add the mesh contribution to the virial */
 526                 m_add(vir_force, fr->vir_el_recip, vir_force);
 527             }
 528             if (EVDW_PME(fr->vdwtype))
 529             {
 530                 /* Add the mesh contribution to the virial */
 531                 m_add(vir_force, fr->vir_lj_recip, vir_force);
 532             }
 533             if (debug)
 534             {
 535                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 536             }
 537         }
 538     }
 539
 540     if (fr->print_force >= 0)
 541     {
 542         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 543     }
 544 }
 545
 546 static void do_nb_verlet(t_forcerec *fr,
 547                          interaction_const_t *ic,
 548                          gmx_enerdata_t *enerd,
 549                          int flags, int ilocality,
 550                          int clearF,
 551                          t_nrnb *nrnb,
 552                          gmx_wallcycle_t wcycle)
 553 {
 554     int                        nnbl, kernel_type, enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
 555     char                      *env;
 556     nonbonded_verlet_group_t  *nbvg;
 557     gmx_bool                   bCUDA;
 558
 559     if (!(flags & GMX_FORCE_NONBONDED))
 560     {
 561         /* skip non-bonded calculation */
 562         return;
 563     }
 564
 565     nbvg = &fr->nbv->grp[ilocality];
 566
 567     /* CUDA kernel launch overhead is already timed separately */
 568     if (fr->cutoff_scheme != ecutsVERLET)
 569     {
 570         gmx_incons("Invalid cut-off scheme passed!");
 571     }
 572
 573     bCUDA = (nbvg->kernel_type == nbnxnk8x8x8_CUDA);
 574
 575     if (!bCUDA)
 576     {
 577         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 578     }
 579     switch (nbvg->kernel_type)
 580     {
 581         case nbnxnk4x4_PlainC:
 582             nbnxn_kernel_ref(&nbvg->nbl_lists,
 583                              nbvg->nbat, ic,
 584                              fr->shift_vec,
 585                              flags,
 586                              clearF,
 587                              fr->fshift[0],
 588                              enerd->grpp.ener[egCOULSR],
 589                              fr->bBHAM ?
 590                              enerd->grpp.ener[egBHAMSR] :
 591                              enerd->grpp.ener[egLJSR]);
 592             break;
 593
 594         case nbnxnk4xN_SIMD_4xN:
 595             nbnxn_kernel_simd_4xn(&nbvg->nbl_lists,
 596                                   nbvg->nbat, ic,
 597                                   nbvg->ewald_excl,
 598                                   fr->shift_vec,
 599                                   flags,
 600                                   clearF,
 601                                   fr->fshift[0],
 602                                   enerd->grpp.ener[egCOULSR],
 603                                   fr->bBHAM ?
 604                                   enerd->grpp.ener[egBHAMSR] :
 605                                   enerd->grpp.ener[egLJSR]);
 606             break;
 607         case nbnxnk4xN_SIMD_2xNN:
 608             nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
 609                                    nbvg->nbat, ic,
 610                                    nbvg->ewald_excl,
 611                                    fr->shift_vec,
 612                                    flags,
 613                                    clearF,
 614                                    fr->fshift[0],
 615                                    enerd->grpp.ener[egCOULSR],
 616                                    fr->bBHAM ?
 617                                    enerd->grpp.ener[egBHAMSR] :
 618                                    enerd->grpp.ener[egLJSR]);
 619             break;
 620
 621         case nbnxnk8x8x8_CUDA:
 622             nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
 623             break;
 624
 625         case nbnxnk8x8x8_PlainC:
 626             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 627                                  nbvg->nbat, ic,
 628                                  fr->shift_vec,
 629                                  flags,
 630                                  clearF,
 631                                  nbvg->nbat->out[0].f,
 632                                  fr->fshift[0],
 633                                  enerd->grpp.ener[egCOULSR],
 634                                  fr->bBHAM ?
 635                                  enerd->grpp.ener[egBHAMSR] :
 636                                  enerd->grpp.ener[egLJSR]);
 637             break;
 638
 639         default:
 640             gmx_incons("Invalid nonbonded kernel type passed!");
 641
 642     }
 643     if (!bCUDA)
 644     {
 645         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 646     }
 647
 648     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 649     {
 650         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 651     }
 652     else if ((!bCUDA && nbvg->ewald_excl == ewaldexclAnalytical) ||
 653              (bCUDA && nbnxn_cuda_is_kernel_ewald_analytical(fr->nbv->cu_nbv)))
 654     {
 655         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 656     }
 657     else
 658     {
 659         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 660     }
 661     enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 662     if (flags & GMX_FORCE_ENERGY)
 663     {
 664         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 665         enr_nbnxn_kernel_ljc += 1;
 666         enr_nbnxn_kernel_lj  += 1;
 667     }
 668
 669     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 670              nbvg->nbl_lists.natpair_ljq);
 671     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 672              nbvg->nbl_lists.natpair_lj);
 673     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 674     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 675              nbvg->nbl_lists.natpair_q);
 676
 677     if (ic->vdw_modifier == eintmodFORCESWITCH)
 678     {
 679         /* We add up the switch cost separately */
 680         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_FSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 681                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 682     }
 683     if (ic->vdw_modifier == eintmodPOTSWITCH)
 684     {
 685         /* We add up the switch cost separately */
 686         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_PSW+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 687                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 688     }
 689     if (ic->vdwtype == evdwPME)
 690     {
 691         /* We add up the LJ Ewald cost separately */
 692         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_EWALD+((flags & GMX_FORCE_ENERGY) ? 1 : 0),
 693                  nbvg->nbl_lists.natpair_ljq + nbvg->nbl_lists.natpair_lj);
 694     }
 695 }
 696
 697 static void do_nb_verlet_fep(nbnxn_pairlist_set_t *nbl_lists,
 698                              t_forcerec           *fr,
 699                              rvec                  x[],
 700                              rvec                  f[],
 701                              t_mdatoms            *mdatoms,
 702                              t_lambda             *fepvals,
 703                              real                 *lambda,
 704                              gmx_enerdata_t       *enerd,
 705                              int                   flags,
 706                              t_nrnb               *nrnb,
 707                              gmx_wallcycle_t       wcycle)
 708 {
 709     int              donb_flags;
 710     nb_kernel_data_t kernel_data;
 711     real             lam_i[efptNR];
 712     real             dvdl_nb[efptNR];
 713     int              th;
 714     int              i, j;
 715
 716     donb_flags = 0;
 717     /* Add short-range interactions */
 718     donb_flags |= GMX_NONBONDED_DO_SR;
 719
 720     /* Currently all group scheme kernels always calculate (shift-)forces */
 721     if (flags & GMX_FORCE_FORCES)
 722     {
 723         donb_flags |= GMX_NONBONDED_DO_FORCE;
 724     }
 725     if (flags & GMX_FORCE_VIRIAL)
 726     {
 727         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 728     }
 729     if (flags & GMX_FORCE_ENERGY)
 730     {
 731         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 732     }
 733     if (flags & GMX_FORCE_DO_LR)
 734     {
 735         donb_flags |= GMX_NONBONDED_DO_LR;
 736     }
 737
 738     kernel_data.flags  = donb_flags;
 739     kernel_data.lambda = lambda;
 740     kernel_data.dvdl   = dvdl_nb;
 741
 742     kernel_data.energygrp_elec = enerd->grpp.ener[egCOULSR];
 743     kernel_data.energygrp_vdw  = enerd->grpp.ener[egLJSR];
 744
 745     /* reset free energy components */
 746     for (i = 0; i < efptNR; i++)
 747     {
 748         dvdl_nb[i]  = 0;
 749     }
 750
 751     assert(gmx_omp_nthreads_get(emntNonbonded) == nbl_lists->nnbl);
 752
 753     wallcycle_sub_start(wcycle, ewcsNONBONDED);
 754 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 755     for (th = 0; th < nbl_lists->nnbl; th++)
 756     {
 757         gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 758                                   x, f, fr, mdatoms, &kernel_data, nrnb);
 759     }
 760
 761     if (fepvals->sc_alpha != 0)
 762     {
 763         enerd->dvdl_nonlin[efptVDW]  += dvdl_nb[efptVDW];
 764         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 765     }
 766     else
 767     {
 768         enerd->dvdl_lin[efptVDW]  += dvdl_nb[efptVDW];
 769         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 770     }
 771
 772     /* If we do foreign lambda and we have soft-core interactions
 773      * we have to recalculate the (non-linear) energies contributions.
 774      */
 775     if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
 776     {
 777         kernel_data.flags          = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE)) | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 778         kernel_data.lambda         = lam_i;
 779         kernel_data.energygrp_elec = enerd->foreign_grpp.ener[egCOULSR];
 780         kernel_data.energygrp_vdw  = enerd->foreign_grpp.ener[egLJSR];
 781         /* Note that we add to kernel_data.dvdl, but ignore the result */
 782
 783         for (i = 0; i < enerd->n_lambda; i++)
 784         {
 785             for (j = 0; j < efptNR; j++)
 786             {
 787                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
 788             }
 789             reset_foreign_enerdata(enerd);
 790 #pragma omp parallel for schedule(static) num_threads(nbl_lists->nnbl)
 791             for (th = 0; th < nbl_lists->nnbl; th++)
 792             {
 793                 gmx_nb_free_energy_kernel(nbl_lists->nbl_fep[th],
 794                                           x, f, fr, mdatoms, &kernel_data, nrnb);
 795             }
 796
 797             sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
 798             enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
 799         }
 800     }
 801
 802     wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 803 }
 804
 805 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
 806                          t_inputrec *inputrec,
 807                          gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
 808                          gmx_localtop_t *top,
 809                          gmx_groups_t gmx_unused *groups,
 810                          matrix box, rvec x[], history_t *hist,
 811                          rvec f[],
 812                          tensor vir_force,
 813                          t_mdatoms *mdatoms,
 814                          gmx_enerdata_t *enerd, t_fcdata *fcd,
 815                          real *lambda, t_graph *graph,
 816                          t_forcerec *fr, interaction_const_t *ic,
 817                          gmx_vsite_t *vsite, rvec mu_tot,
 818                          double t, FILE *field, gmx_edsam_t ed,
 819                          gmx_bool bBornRadii,
 820                          int flags)
 821 {
 822     int                 cg0, cg1, i, j;
 823     int                 start, homenr;
 824     int                 nb_kernel_type;
 825     double              mu[2*DIM];
 826     gmx_bool            bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
 827     gmx_bool            bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
 828     gmx_bool            bDiffKernels = FALSE;
 829     matrix              boxs;
 830     rvec                vzero, box_diag;
 831     real                e, v, dvdl;
 832     float               cycles_pme, cycles_force, cycles_wait_gpu;
 833     nonbonded_verlet_t *nbv;
 834
 835     cycles_force    = 0;
 836     cycles_wait_gpu = 0;
 837     nbv             = fr->nbv;
 838     nb_kernel_type  = fr->nbv->grp[0].kernel_type;
 839
 840     start  = 0;
 841     homenr = mdatoms->homenr;
 842
 843     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
 844
 845     clear_mat(vir_force);
 846
 847     cg0 = 0;
 848     if (DOMAINDECOMP(cr))
 849     {
 850         cg1 = cr->dd->ncg_tot;
 851     }
 852     else
 853     {
 854         cg1 = top->cgs.nr;
 855     }
 856     if (fr->n_tpi > 0)
 857     {
 858         cg1--;
 859     }
 860
 861     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 862     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
 863     bFillGrid     = (bNS && bStateChanged);
 864     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 865     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 866     bDoForces     = (flags & GMX_FORCE_FORCES);
 867     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 868     bUseGPU       = fr->nbv->bUseGPU;
 869     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
 870
 871     if (bStateChanged)
 872     {
 873         update_forcerec(fr, box);
 874
 875         if (NEED_MUTOT(*inputrec))
 876         {
 877             /* Calculate total (local) dipole moment in a temporary common array.
 878              * This makes it possible to sum them over nodes faster.
 879              */
 880             calc_mu(start, homenr,
 881                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 882                     mu, mu+DIM);
 883         }
 884     }
 885
 886     if (fr->ePBC != epbcNONE)
 887     {
 888         /* Compute shift vectors every step,
 889          * because of pressure coupling or box deformation!
 890          */
 891         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 892         {
 893             calc_shifts(box, fr->shift_vec);
 894         }
 895
 896         if (bCalcCGCM)
 897         {
 898             put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
 899             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 900         }
 901         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 902         {
 903             unshift_self(graph, box, x);
 904         }
 905     }
 906
 907     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 908                                  fr->shift_vec, nbv->grp[0].nbat);
 909
 910 #ifdef GMX_MPI
 911     if (!(cr->duty & DUTY_PME))
 912     {
 913         /* Send particle coordinates to the pme nodes.
 914          * Since this is only implemented for domain decomposition
 915          * and domain decomposition does not use the graph,
 916          * we do not need to worry about shifting.
 917          */
 918
 919         int pme_flags = 0;
 920
 921         wallcycle_start(wcycle, ewcPP_PMESENDX);
 922
 923         bBS = (inputrec->nwall == 2);
 924         if (bBS)
 925         {
 926             copy_mat(box, boxs);
 927             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
 928         }
 929
 930         if (EEL_PME(fr->eeltype))
 931         {
 932             pme_flags |= GMX_PME_DO_COULOMB;
 933         }
 934
 935         if (EVDW_PME(fr->vdwtype))
 936         {
 937             pme_flags |= GMX_PME_DO_LJ;
 938         }
 939
 940         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
 941                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
 942                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
 943                                  pme_flags, step);
 944
 945         wallcycle_stop(wcycle, ewcPP_PMESENDX);
 946     }
 947 #endif /* GMX_MPI */
 948
 949     /* do gridding for pair search */
 950     if (bNS)
 951     {
 952         if (graph && bStateChanged)
 953         {
 954             /* Calculate intramolecular shift vectors to make molecules whole */
 955             mk_mshift(fplog, graph, fr->ePBC, box, x);
 956         }
 957
 958         clear_rvec(vzero);
 959         box_diag[XX] = box[XX][XX];
 960         box_diag[YY] = box[YY][YY];
 961         box_diag[ZZ] = box[ZZ][ZZ];
 962
 963         wallcycle_start(wcycle, ewcNS);
 964         if (!fr->bDomDec)
 965         {
 966             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 967             nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
 968                               0, vzero, box_diag,
 969                               0, mdatoms->homenr, -1, fr->cginfo, x,
 970                               0, NULL,
 971                               nbv->grp[eintLocal].kernel_type,
 972                               nbv->grp[eintLocal].nbat);
 973             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 974         }
 975         else
 976         {
 977             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 978             nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
 979                                        fr->cginfo, x,
 980                                        nbv->grp[eintNonlocal].kernel_type,
 981                                        nbv->grp[eintNonlocal].nbat);
 982             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 983         }
 984
 985         if (nbv->ngrp == 1 ||
 986             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 987         {
 988             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
 989                                nbv->nbs, mdatoms, fr->cginfo);
 990         }
 991         else
 992         {
 993             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
 994                                nbv->nbs, mdatoms, fr->cginfo);
 995             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
 996                                nbv->nbs, mdatoms, fr->cginfo);
 997         }
 998         wallcycle_stop(wcycle, ewcNS);
 999     }
1000
1001     /* initialize the GPU atom data and copy shift vector */
1002     if (bUseGPU)
1003     {
1004         if (bNS)
1005         {
1006             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1007             nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
1008             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1009         }
1010
1011         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1012         nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
1013         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1014     }
1015
1016     /* do local pair search */
1017     if (bNS)
1018     {
1019         wallcycle_start_nocount(wcycle, ewcNS);
1020         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1021         nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
1022                             &top->excls,
1023                             ic->rlist,
1024                             nbv->min_ci_balanced,
1025                             &nbv->grp[eintLocal].nbl_lists,
1026                             eintLocal,
1027                             nbv->grp[eintLocal].kernel_type,
1028                             nrnb);
1029         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1030
1031         if (bUseGPU)
1032         {
1033             /* initialize local pair-list on the GPU */
1034             nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1035                                      nbv->grp[eintLocal].nbl_lists.nbl[0],
1036                                      eintLocal);
1037         }
1038         wallcycle_stop(wcycle, ewcNS);
1039     }
1040     else
1041     {
1042         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1043         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1044         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
1045                                         nbv->grp[eintLocal].nbat);
1046         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1047         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1048     }
1049
1050     if (bUseGPU)
1051     {
1052         wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1053         /* launch local nonbonded F on GPU */
1054         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
1055                      nrnb, wcycle);
1056         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1057     }
1058
1059     /* Communicate coordinates and sum dipole if necessary +
1060        do non-local pair search */
1061     if (DOMAINDECOMP(cr))
1062     {
1063         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
1064                         nbv->grp[eintLocal].kernel_type);
1065
1066         if (bDiffKernels)
1067         {
1068             /* With GPU+CPU non-bonded calculations we need to copy
1069              * the local coordinates to the non-local nbat struct
1070              * (in CPU format) as the non-local kernel call also
1071              * calculates the local - non-local interactions.
1072              */
1073             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1074             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1075             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
1076                                             nbv->grp[eintNonlocal].nbat);
1077             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1078             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1079         }
1080
1081         if (bNS)
1082         {
1083             wallcycle_start_nocount(wcycle, ewcNS);
1084             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1085
1086             if (bDiffKernels)
1087             {
1088                 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
1089             }
1090
1091             nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
1092                                 &top->excls,
1093                                 ic->rlist,
1094                                 nbv->min_ci_balanced,
1095                                 &nbv->grp[eintNonlocal].nbl_lists,
1096                                 eintNonlocal,
1097                                 nbv->grp[eintNonlocal].kernel_type,
1098                                 nrnb);
1099
1100             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1101
1102             if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_CUDA)
1103             {
1104                 /* initialize non-local pair-list on the GPU */
1105                 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1106                                          nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1107                                          eintNonlocal);
1108             }
1109             wallcycle_stop(wcycle, ewcNS);
1110         }
1111         else
1112         {
1113             wallcycle_start(wcycle, ewcMOVEX);
1114             dd_move_x(cr->dd, box, x);
1115
1116             /* When we don't need the total dipole we sum it in global_stat */
1117             if (bStateChanged && NEED_MUTOT(*inputrec))
1118             {
1119                 gmx_sumd(2*DIM, mu, cr);
1120             }
1121             wallcycle_stop(wcycle, ewcMOVEX);
1122
1123             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1124             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1125             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1126                                             nbv->grp[eintNonlocal].nbat);
1127             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1128             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1129         }
1130
1131         if (bUseGPU && !bDiffKernels)
1132         {
1133             wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1134             /* launch non-local nonbonded F on GPU */
1135             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1136                          nrnb, wcycle);
1137             cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1138         }
1139     }
1140
1141     if (bUseGPU)
1142     {
1143         /* launch D2H copy-back F */
1144         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1145         if (DOMAINDECOMP(cr) && !bDiffKernels)
1146         {
1147             nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1148                                       flags, eatNonlocal);
1149         }
1150         nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1151                                   flags, eatLocal);
1152         cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1153     }
1154
1155     if (bStateChanged && NEED_MUTOT(*inputrec))
1156     {
1157         if (PAR(cr))
1158         {
1159             gmx_sumd(2*DIM, mu, cr);
1160         }
1161
1162         for (i = 0; i < 2; i++)
1163         {
1164             for (j = 0; j < DIM; j++)
1165             {
1166                 fr->mu_tot[i][j] = mu[i*DIM + j];
1167             }
1168         }
1169     }
1170     if (fr->efep == efepNO)
1171     {
1172         copy_rvec(fr->mu_tot[0], mu_tot);
1173     }
1174     else
1175     {
1176         for (j = 0; j < DIM; j++)
1177         {
1178             mu_tot[j] =
1179                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1180                 lambda[efptCOUL]*fr->mu_tot[1][j];
1181         }
1182     }
1183
1184     /* Reset energies */
1185     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1186     clear_rvecs(SHIFTS, fr->fshift);
1187
1188     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1189     {
1190         wallcycle_start(wcycle, ewcPPDURINGPME);
1191         dd_force_flop_start(cr->dd, nrnb);
1192     }
1193
1194     if (inputrec->bRot)
1195     {
1196         /* Enforced rotation has its own cycle counter that starts after the collective
1197          * coordinates have been communicated. It is added to ddCyclF to allow
1198          * for proper load-balancing */
1199         wallcycle_start(wcycle, ewcROT);
1200         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1201         wallcycle_stop(wcycle, ewcROT);
1202     }
1203
1204     /* Start the force cycle counter.
1205      * This counter is stopped in do_forcelow_level.
1206      * No parallel communication should occur while this counter is running,
1207      * since that will interfere with the dynamic load balancing.
1208      */
1209     wallcycle_start(wcycle, ewcFORCE);
1210     if (bDoForces)
1211     {
1212         /* Reset forces for which the virial is calculated separately:
1213          * PME/Ewald forces if necessary */
1214         if (fr->bF_NoVirSum)
1215         {
1216             if (flags & GMX_FORCE_VIRIAL)
1217             {
1218                 fr->f_novirsum = fr->f_novirsum_alloc;
1219                 if (fr->bDomDec)
1220                 {
1221                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1222                 }
1223                 else
1224                 {
1225                     clear_rvecs(homenr, fr->f_novirsum+start);
1226                 }
1227             }
1228             else
1229             {
1230                 /* We are not calculating the pressure so we do not need
1231                  * a separate array for forces that do not contribute
1232                  * to the pressure.
1233                  */
1234                 fr->f_novirsum = f;
1235             }
1236         }
1237
1238         /* Clear the short- and long-range forces */
1239         clear_rvecs(fr->natoms_force_constr, f);
1240         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1241         {
1242             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1243         }
1244
1245         clear_rvec(fr->vir_diag_posres);
1246     }
1247
1248     if (inputrec->ePull == epullCONSTRAINT)
1249     {
1250         clear_pull_forces(inputrec->pull);
1251     }
1252
1253     /* We calculate the non-bonded forces, when done on the CPU, here.
1254      * We do this before calling do_force_lowlevel, as in there bondeds
1255      * forces are calculated before PME, which does communication.
1256      * With this order, non-bonded and bonded force calculation imbalance
1257      * can be balanced out by the domain decomposition load balancing.
1258      */
1259
1260     if (!bUseOrEmulGPU)
1261     {
1262         /* Maybe we should move this into do_force_lowlevel */
1263         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1264                      nrnb, wcycle);
1265     }
1266
1267     if (fr->efep != efepNO)
1268     {
1269         /* Calculate the local and non-local free energy interactions here.
1270          * Happens here on the CPU both with and without GPU.
1271          */
1272         if (fr->nbv->grp[eintLocal].nbl_lists.nbl_fep[0]->nrj > 0)
1273         {
1274             do_nb_verlet_fep(&fr->nbv->grp[eintLocal].nbl_lists,
1275                              fr, x, f, mdatoms,
1276                              inputrec->fepvals, lambda,
1277                              enerd, flags, nrnb, wcycle);
1278         }
1279
1280         if (DOMAINDECOMP(cr) &&
1281             fr->nbv->grp[eintNonlocal].nbl_lists.nbl_fep[0]->nrj > 0)
1282         {
1283             do_nb_verlet_fep(&fr->nbv->grp[eintNonlocal].nbl_lists,
1284                              fr, x, f, mdatoms,
1285                              inputrec->fepvals, lambda,
1286                              enerd, flags, nrnb, wcycle);
1287         }
1288     }
1289
1290     if (!bUseOrEmulGPU || bDiffKernels)
1291     {
1292         int aloc;
1293
1294         if (DOMAINDECOMP(cr))
1295         {
1296             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1297                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1298                          nrnb, wcycle);
1299         }
1300
1301         if (!bUseOrEmulGPU)
1302         {
1303             aloc = eintLocal;
1304         }
1305         else
1306         {
1307             aloc = eintNonlocal;
1308         }
1309
1310         /* Add all the non-bonded force to the normal force array.
1311          * This can be split into a local a non-local part when overlapping
1312          * communication with calculation with domain decomposition.
1313          */
1314         cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1315         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1316         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1317         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1318         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1319         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1320         wallcycle_start_nocount(wcycle, ewcFORCE);
1321
1322         /* if there are multiple fshift output buffers reduce them */
1323         if ((flags & GMX_FORCE_VIRIAL) &&
1324             nbv->grp[aloc].nbl_lists.nnbl > 1)
1325         {
1326             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1327                                                      fr->fshift);
1328         }
1329     }
1330
1331     /* update QMMMrec, if necessary */
1332     if (fr->bQMMM)
1333     {
1334         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1335     }
1336
1337     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1338     {
1339         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1340                        enerd, lambda, fr);
1341     }
1342
1343     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1344     {
1345         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1346     }
1347
1348     /* Compute the bonded and non-bonded energies and optionally forces */
1349     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1350                       cr, nrnb, wcycle, mdatoms,
1351                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1352                       &(top->atomtypes), bBornRadii, box,
1353                       inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1354                       flags, &cycles_pme);
1355
1356     if (bSepLRF)
1357     {
1358         if (do_per_step(step, inputrec->nstcalclr))
1359         {
1360             /* Add the long range forces to the short range forces */
1361             for (i = 0; i < fr->natoms_force_constr; i++)
1362             {
1363                 rvec_add(fr->f_twin[i], f[i], f[i]);
1364             }
1365         }
1366     }
1367
1368     cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1369
1370     if (ed)
1371     {
1372         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1373     }
1374
1375     if (bUseOrEmulGPU && !bDiffKernels)
1376     {
1377         /* wait for non-local forces (or calculate in emulation mode) */
1378         if (DOMAINDECOMP(cr))
1379         {
1380             if (bUseGPU)
1381             {
1382                 float cycles_tmp;
1383
1384                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1385                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1386                                     nbv->grp[eintNonlocal].nbat,
1387                                     flags, eatNonlocal,
1388                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1389                                     fr->fshift);
1390                 cycles_tmp       = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1391                 cycles_wait_gpu += cycles_tmp;
1392                 cycles_force    += cycles_tmp;
1393             }
1394             else
1395             {
1396                 wallcycle_start_nocount(wcycle, ewcFORCE);
1397                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1398                              nrnb, wcycle);
1399                 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1400             }
1401             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1402             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1403             /* skip the reduction if there was no non-local work to do */
1404             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1405             {
1406                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1407                                                nbv->grp[eintNonlocal].nbat, f);
1408             }
1409             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1410             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1411         }
1412     }
1413
1414     if (bDoForces && DOMAINDECOMP(cr))
1415     {
1416         /* Communicate the forces */
1417         wallcycle_start(wcycle, ewcMOVEF);
1418         dd_move_f(cr->dd, f, fr->fshift);
1419         /* Do we need to communicate the separate force array
1420          * for terms that do not contribute to the single sum virial?
1421          * Position restraints and electric fields do not introduce
1422          * inter-cg forces, only full electrostatics methods do.
1423          * When we do not calculate the virial, fr->f_novirsum = f,
1424          * so we have already communicated these forces.
1425          */
1426         if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1427             (flags & GMX_FORCE_VIRIAL))
1428         {
1429             dd_move_f(cr->dd, fr->f_novirsum, NULL);
1430         }
1431         if (bSepLRF)
1432         {
1433             /* We should not update the shift forces here,
1434              * since f_twin is already included in f.
1435              */
1436             dd_move_f(cr->dd, fr->f_twin, NULL);
1437         }
1438         wallcycle_stop(wcycle, ewcMOVEF);
1439     }
1440
1441     if (bUseOrEmulGPU)
1442     {
1443         /* wait for local forces (or calculate in emulation mode) */
1444         if (bUseGPU)
1445         {
1446             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1447             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1448                                 nbv->grp[eintLocal].nbat,
1449                                 flags, eatLocal,
1450                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1451                                 fr->fshift);
1452             cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1453
1454             /* now clear the GPU outputs while we finish the step on the CPU */
1455
1456             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1457             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1458             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1459         }
1460         else
1461         {
1462             wallcycle_start_nocount(wcycle, ewcFORCE);
1463             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1464                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1465                          nrnb, wcycle);
1466             wallcycle_stop(wcycle, ewcFORCE);
1467         }
1468         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1469         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1470         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1471         {
1472             /* skip the reduction if there was no non-local work to do */
1473             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1474                                            nbv->grp[eintLocal].nbat, f);
1475         }
1476         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1477         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1478     }
1479
1480     if (DOMAINDECOMP(cr))
1481     {
1482         dd_force_flop_stop(cr->dd, nrnb);
1483         if (wcycle)
1484         {
1485             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1486             if (bUseGPU)
1487             {
1488                 dd_cycles_add(cr->dd, cycles_wait_gpu, ddCyclWaitGPU);
1489             }
1490         }
1491     }
1492
1493     if (bDoForces)
1494     {
1495         if (IR_ELEC_FIELD(*inputrec))
1496         {
1497             /* Compute forces due to electric field */
1498             calc_f_el(MASTER(cr) ? field : NULL,
1499                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1500                       inputrec->ex, inputrec->et, t);
1501         }
1502
1503         /* If we have NoVirSum forces, but we do not calculate the virial,
1504          * we sum fr->f_novirum=f later.
1505          */
1506         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1507         {
1508             wallcycle_start(wcycle, ewcVSITESPREAD);
1509             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1510                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1511             wallcycle_stop(wcycle, ewcVSITESPREAD);
1512
1513             if (bSepLRF)
1514             {
1515                 wallcycle_start(wcycle, ewcVSITESPREAD);
1516                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1517                                nrnb,
1518                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1519                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1520             }
1521         }
1522
1523         if (flags & GMX_FORCE_VIRIAL)
1524         {
1525             /* Calculation of the virial must be done after vsites! */
1526             calc_virial(0, mdatoms->homenr, x, f,
1527                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1528         }
1529     }
1530
1531     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1532     {
1533         /* Since the COM pulling is always done mass-weighted, no forces are
1534          * applied to vsites and this call can be done after vsite spreading.
1535          */
1536         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1537                                f, vir_force, mdatoms, enerd, lambda, t,
1538                                wcycle);
1539     }
1540
1541     /* Add the forces from enforced rotation potentials (if any) */
1542     if (inputrec->bRot)
1543     {
1544         wallcycle_start(wcycle, ewcROTadd);
1545         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1546         wallcycle_stop(wcycle, ewcROTadd);
1547     }
1548
1549     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
1550     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
1551
1552     if (PAR(cr) && !(cr->duty & DUTY_PME))
1553     {
1554         /* In case of node-splitting, the PP nodes receive the long-range
1555          * forces, virial and energy from the PME nodes here.
1556          */
1557         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1558     }
1559
1560     if (bDoForces)
1561     {
1562         post_process_forces(cr, step, nrnb, wcycle,
1563                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1564                             flags);
1565     }
1566
1567     /* Sum the potential energy terms from group contributions */
1568     sum_epot(&(enerd->grpp), enerd->term);
1569 }
1570
1571 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1572                         t_inputrec *inputrec,
1573                         gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1574                         gmx_localtop_t *top,
1575                         gmx_groups_t *groups,
1576                         matrix box, rvec x[], history_t *hist,
1577                         rvec f[],
1578                         tensor vir_force,
1579                         t_mdatoms *mdatoms,
1580                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1581                         real *lambda, t_graph *graph,
1582                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1583                         double t, FILE *field, gmx_edsam_t ed,
1584                         gmx_bool bBornRadii,
1585                         int flags)
1586 {
1587     int        cg0, cg1, i, j;
1588     int        start, homenr;
1589     double     mu[2*DIM];
1590     gmx_bool   bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1591     gmx_bool   bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1592     gmx_bool   bDoAdressWF;
1593     matrix     boxs;
1594     rvec       vzero, box_diag;
1595     real       e, v, dvdlambda[efptNR];
1596     t_pbc      pbc;
1597     float      cycles_pme, cycles_force;
1598
1599     start  = 0;
1600     homenr = mdatoms->homenr;
1601
1602     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
1603
1604     clear_mat(vir_force);
1605
1606     cg0 = 0;
1607     if (DOMAINDECOMP(cr))
1608     {
1609         cg1 = cr->dd->ncg_tot;
1610     }
1611     else
1612     {
1613         cg1 = top->cgs.nr;
1614     }
1615     if (fr->n_tpi > 0)
1616     {
1617         cg1--;
1618     }
1619
1620     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1621     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1622     /* Should we update the long-range neighborlists at this step? */
1623     bDoLongRangeNS = fr->bTwinRange && bNS;
1624     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1625     bFillGrid      = (bNS && bStateChanged);
1626     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1627     bDoForces      = (flags & GMX_FORCE_FORCES);
1628     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1629     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1630                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1631
1632     /* should probably move this to the forcerec since it doesn't change */
1633     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1634
1635     if (bStateChanged)
1636     {
1637         update_forcerec(fr, box);
1638
1639         if (NEED_MUTOT(*inputrec))
1640         {
1641             /* Calculate total (local) dipole moment in a temporary common array.
1642              * This makes it possible to sum them over nodes faster.
1643              */
1644             calc_mu(start, homenr,
1645                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1646                     mu, mu+DIM);
1647         }
1648     }
1649
1650     if (fr->ePBC != epbcNONE)
1651     {
1652         /* Compute shift vectors every step,
1653          * because of pressure coupling or box deformation!
1654          */
1655         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1656         {
1657             calc_shifts(box, fr->shift_vec);
1658         }
1659
1660         if (bCalcCGCM)
1661         {
1662             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1663                                      &(top->cgs), x, fr->cg_cm);
1664             inc_nrnb(nrnb, eNR_CGCM, homenr);
1665             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1666         }
1667         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1668         {
1669             unshift_self(graph, box, x);
1670         }
1671     }
1672     else if (bCalcCGCM)
1673     {
1674         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1675         inc_nrnb(nrnb, eNR_CGCM, homenr);
1676     }
1677
1678     if (bCalcCGCM && gmx_debug_at)
1679     {
1680         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1681     }
1682
1683 #ifdef GMX_MPI
1684     if (!(cr->duty & DUTY_PME))
1685     {
1686         /* Send particle coordinates to the pme nodes.
1687          * Since this is only implemented for domain decomposition
1688          * and domain decomposition does not use the graph,
1689          * we do not need to worry about shifting.
1690          */
1691
1692         int pme_flags = 0;
1693
1694         wallcycle_start(wcycle, ewcPP_PMESENDX);
1695
1696         bBS = (inputrec->nwall == 2);
1697         if (bBS)
1698         {
1699             copy_mat(box, boxs);
1700             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1701         }
1702
1703         if (EEL_PME(fr->eeltype))
1704         {
1705             pme_flags |= GMX_PME_DO_COULOMB;
1706         }
1707
1708         if (EVDW_PME(fr->vdwtype))
1709         {
1710             pme_flags |= GMX_PME_DO_LJ;
1711         }
1712
1713         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
1714                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
1715                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
1716                                  pme_flags, step);
1717
1718         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1719     }
1720 #endif /* GMX_MPI */
1721
1722     /* Communicate coordinates and sum dipole if necessary */
1723     if (DOMAINDECOMP(cr))
1724     {
1725         wallcycle_start(wcycle, ewcMOVEX);
1726         dd_move_x(cr->dd, box, x);
1727         wallcycle_stop(wcycle, ewcMOVEX);
1728     }
1729
1730     /* update adress weight beforehand */
1731     if (bStateChanged && bDoAdressWF)
1732     {
1733         /* need pbc for adress weight calculation with pbc_dx */
1734         set_pbc(&pbc, inputrec->ePBC, box);
1735         if (fr->adress_site == eAdressSITEcog)
1736         {
1737             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1738                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1739         }
1740         else if (fr->adress_site == eAdressSITEcom)
1741         {
1742             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1743                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1744         }
1745         else if (fr->adress_site == eAdressSITEatomatom)
1746         {
1747             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1748                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1749         }
1750         else
1751         {
1752             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1753                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1754         }
1755     }
1756
1757     if (NEED_MUTOT(*inputrec))
1758     {
1759
1760         if (bStateChanged)
1761         {
1762             if (PAR(cr))
1763             {
1764                 gmx_sumd(2*DIM, mu, cr);
1765             }
1766             for (i = 0; i < 2; i++)
1767             {
1768                 for (j = 0; j < DIM; j++)
1769                 {
1770                     fr->mu_tot[i][j] = mu[i*DIM + j];
1771                 }
1772             }
1773         }
1774         if (fr->efep == efepNO)
1775         {
1776             copy_rvec(fr->mu_tot[0], mu_tot);
1777         }
1778         else
1779         {
1780             for (j = 0; j < DIM; j++)
1781             {
1782                 mu_tot[j] =
1783                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1784             }
1785         }
1786     }
1787
1788     /* Reset energies */
1789     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1790     clear_rvecs(SHIFTS, fr->fshift);
1791
1792     if (bNS)
1793     {
1794         wallcycle_start(wcycle, ewcNS);
1795
1796         if (graph && bStateChanged)
1797         {
1798             /* Calculate intramolecular shift vectors to make molecules whole */
1799             mk_mshift(fplog, graph, fr->ePBC, box, x);
1800         }
1801
1802         /* Do the actual neighbour searching */
1803         ns(fplog, fr, box,
1804            groups, top, mdatoms,
1805            cr, nrnb, bFillGrid,
1806            bDoLongRangeNS);
1807
1808         wallcycle_stop(wcycle, ewcNS);
1809     }
1810
1811     if (inputrec->implicit_solvent && bNS)
1812     {
1813         make_gb_nblist(cr, inputrec->gb_algorithm,
1814                        x, box, fr, &top->idef, graph, fr->born);
1815     }
1816
1817     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1818     {
1819         wallcycle_start(wcycle, ewcPPDURINGPME);
1820         dd_force_flop_start(cr->dd, nrnb);
1821     }
1822
1823     if (inputrec->bRot)
1824     {
1825         /* Enforced rotation has its own cycle counter that starts after the collective
1826          * coordinates have been communicated. It is added to ddCyclF to allow
1827          * for proper load-balancing */
1828         wallcycle_start(wcycle, ewcROT);
1829         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1830         wallcycle_stop(wcycle, ewcROT);
1831     }
1832
1833     /* Start the force cycle counter.
1834      * This counter is stopped in do_forcelow_level.
1835      * No parallel communication should occur while this counter is running,
1836      * since that will interfere with the dynamic load balancing.
1837      */
1838     wallcycle_start(wcycle, ewcFORCE);
1839
1840     if (bDoForces)
1841     {
1842         /* Reset forces for which the virial is calculated separately:
1843          * PME/Ewald forces if necessary */
1844         if (fr->bF_NoVirSum)
1845         {
1846             if (flags & GMX_FORCE_VIRIAL)
1847             {
1848                 fr->f_novirsum = fr->f_novirsum_alloc;
1849                 if (fr->bDomDec)
1850                 {
1851                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1852                 }
1853                 else
1854                 {
1855                     clear_rvecs(homenr, fr->f_novirsum+start);
1856                 }
1857             }
1858             else
1859             {
1860                 /* We are not calculating the pressure so we do not need
1861                  * a separate array for forces that do not contribute
1862                  * to the pressure.
1863                  */
1864                 fr->f_novirsum = f;
1865             }
1866         }
1867
1868         /* Clear the short- and long-range forces */
1869         clear_rvecs(fr->natoms_force_constr, f);
1870         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1871         {
1872             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1873         }
1874
1875         clear_rvec(fr->vir_diag_posres);
1876     }
1877     if (inputrec->ePull == epullCONSTRAINT)
1878     {
1879         clear_pull_forces(inputrec->pull);
1880     }
1881
1882     /* update QMMMrec, if necessary */
1883     if (fr->bQMMM)
1884     {
1885         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1886     }
1887
1888     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1889     {
1890         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1891                        enerd, lambda, fr);
1892     }
1893
1894     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1895     {
1896         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1897     }
1898
1899     /* Compute the bonded and non-bonded energies and optionally forces */
1900     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1901                       cr, nrnb, wcycle, mdatoms,
1902                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1903                       &(top->atomtypes), bBornRadii, box,
1904                       inputrec->fepvals, lambda,
1905                       graph, &(top->excls), fr->mu_tot,
1906                       flags,
1907                       &cycles_pme);
1908
1909     if (bSepLRF)
1910     {
1911         if (do_per_step(step, inputrec->nstcalclr))
1912         {
1913             /* Add the long range forces to the short range forces */
1914             for (i = 0; i < fr->natoms_force_constr; i++)
1915             {
1916                 rvec_add(fr->f_twin[i], f[i], f[i]);
1917             }
1918         }
1919     }
1920
1921     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1922
1923     if (ed)
1924     {
1925         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1926     }
1927
1928     if (DOMAINDECOMP(cr))
1929     {
1930         dd_force_flop_stop(cr->dd, nrnb);
1931         if (wcycle)
1932         {
1933             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1934         }
1935     }
1936
1937     if (bDoForces)
1938     {
1939         if (IR_ELEC_FIELD(*inputrec))
1940         {
1941             /* Compute forces due to electric field */
1942             calc_f_el(MASTER(cr) ? field : NULL,
1943                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1944                       inputrec->ex, inputrec->et, t);
1945         }
1946
1947         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1948         {
1949             /* Compute thermodynamic force in hybrid AdResS region */
1950             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1951                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1952         }
1953
1954         /* Communicate the forces */
1955         if (DOMAINDECOMP(cr))
1956         {
1957             wallcycle_start(wcycle, ewcMOVEF);
1958             dd_move_f(cr->dd, f, fr->fshift);
1959             /* Do we need to communicate the separate force array
1960              * for terms that do not contribute to the single sum virial?
1961              * Position restraints and electric fields do not introduce
1962              * inter-cg forces, only full electrostatics methods do.
1963              * When we do not calculate the virial, fr->f_novirsum = f,
1964              * so we have already communicated these forces.
1965              */
1966             if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1967                 (flags & GMX_FORCE_VIRIAL))
1968             {
1969                 dd_move_f(cr->dd, fr->f_novirsum, NULL);
1970             }
1971             if (bSepLRF)
1972             {
1973                 /* We should not update the shift forces here,
1974                  * since f_twin is already included in f.
1975                  */
1976                 dd_move_f(cr->dd, fr->f_twin, NULL);
1977             }
1978             wallcycle_stop(wcycle, ewcMOVEF);
1979         }
1980
1981         /* If we have NoVirSum forces, but we do not calculate the virial,
1982          * we sum fr->f_novirum=f later.
1983          */
1984         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1985         {
1986             wallcycle_start(wcycle, ewcVSITESPREAD);
1987             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1988                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1989             wallcycle_stop(wcycle, ewcVSITESPREAD);
1990
1991             if (bSepLRF)
1992             {
1993                 wallcycle_start(wcycle, ewcVSITESPREAD);
1994                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1995                                nrnb,
1996                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1997                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1998             }
1999         }
2000
2001         if (flags & GMX_FORCE_VIRIAL)
2002         {
2003             /* Calculation of the virial must be done after vsites! */
2004             calc_virial(0, mdatoms->homenr, x, f,
2005                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
2006         }
2007     }
2008
2009     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
2010     {
2011         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
2012                                f, vir_force, mdatoms, enerd, lambda, t,
2013                                wcycle);
2014     }
2015
2016     /* Add the forces from enforced rotation potentials (if any) */
2017     if (inputrec->bRot)
2018     {
2019         wallcycle_start(wcycle, ewcROTadd);
2020         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
2021         wallcycle_stop(wcycle, ewcROTadd);
2022     }
2023
2024     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
2025     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
2026
2027     if (PAR(cr) && !(cr->duty & DUTY_PME))
2028     {
2029         /* In case of node-splitting, the PP nodes receive the long-range
2030          * forces, virial and energy from the PME nodes here.
2031          */
2032         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
2033     }
2034
2035     if (bDoForces)
2036     {
2037         post_process_forces(cr, step, nrnb, wcycle,
2038                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
2039                             flags);
2040     }
2041
2042     /* Sum the potential energy terms from group contributions */
2043     sum_epot(&(enerd->grpp), enerd->term);
2044 }
2045
2046 void do_force(FILE *fplog, t_commrec *cr,
2047               t_inputrec *inputrec,
2048               gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
2049               gmx_localtop_t *top,
2050               gmx_groups_t *groups,
2051               matrix box, rvec x[], history_t *hist,
2052               rvec f[],
2053               tensor vir_force,
2054               t_mdatoms *mdatoms,
2055               gmx_enerdata_t *enerd, t_fcdata *fcd,
2056               real *lambda, t_graph *graph,
2057               t_forcerec *fr,
2058               gmx_vsite_t *vsite, rvec mu_tot,
2059               double t, FILE *field, gmx_edsam_t ed,
2060               gmx_bool bBornRadii,
2061               int flags)
2062 {
2063     /* modify force flag if not doing nonbonded */
2064     if (!fr->bNonbonded)
2065     {
2066         flags &= ~GMX_FORCE_NONBONDED;
2067     }
2068
2069     switch (inputrec->cutoff_scheme)
2070     {
2071         case ecutsVERLET:
2072             do_force_cutsVERLET(fplog, cr, inputrec,
2073                                 step, nrnb, wcycle,
2074                                 top,
2075                                 groups,
2076                                 box, x, hist,
2077                                 f, vir_force,
2078                                 mdatoms,
2079                                 enerd, fcd,
2080                                 lambda, graph,
2081                                 fr, fr->ic,
2082                                 vsite, mu_tot,
2083                                 t, field, ed,
2084                                 bBornRadii,
2085                                 flags);
2086             break;
2087         case ecutsGROUP:
2088             do_force_cutsGROUP(fplog, cr, inputrec,
2089                                step, nrnb, wcycle,
2090                                top,
2091                                groups,
2092                                box, x, hist,
2093                                f, vir_force,
2094                                mdatoms,
2095                                enerd, fcd,
2096                                lambda, graph,
2097                                fr, vsite, mu_tot,
2098                                t, field, ed,
2099                                bBornRadii,
2100                                flags);
2101             break;
2102         default:
2103             gmx_incons("Invalid cut-off scheme passed!");
2104     }
2105 }
2106
2107
2108 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2109                         t_inputrec *ir, t_mdatoms *md,
2110                         t_state *state, t_commrec *cr, t_nrnb *nrnb,
2111                         t_forcerec *fr, gmx_localtop_t *top)
2112 {
2113     int             i, m, start, end;
2114     gmx_int64_t     step;
2115     real            dt = ir->delta_t;
2116     real            dvdl_dum;
2117     rvec           *savex;
2118
2119     snew(savex, state->natoms);
2120
2121     start = 0;
2122     end   = md->homenr;
2123
2124     if (debug)
2125     {
2126         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2127                 start, md->homenr, end);
2128     }
2129     /* Do a first constrain to reset particles... */
2130     step = ir->init_step;
2131     if (fplog)
2132     {
2133         char buf[STEPSTRSIZE];
2134         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2135                 gmx_step_str(step, buf));
2136     }
2137     dvdl_dum = 0;
2138
2139     /* constrain the current position */
2140     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2141               ir, NULL, cr, step, 0, 1.0, md,
2142               state->x, state->x, NULL,
2143               fr->bMolPBC, state->box,
2144               state->lambda[efptBONDED], &dvdl_dum,
2145               NULL, NULL, nrnb, econqCoord,
2146               ir->epc == epcMTTK, state->veta, state->veta);
2147     if (EI_VV(ir->eI))
2148     {
2149         /* constrain the inital velocity, and save it */
2150         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2151         /* might not yet treat veta correctly */
2152         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2153                   ir, NULL, cr, step, 0, 1.0, md,
2154                   state->x, state->v, state->v,
2155                   fr->bMolPBC, state->box,
2156                   state->lambda[efptBONDED], &dvdl_dum,
2157                   NULL, NULL, nrnb, econqVeloc,
2158                   ir->epc == epcMTTK, state->veta, state->veta);
2159     }
2160     /* constrain the inital velocities at t-dt/2 */
2161     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2162     {
2163         for (i = start; (i < end); i++)
2164         {
2165             for (m = 0; (m < DIM); m++)
2166             {
2167                 /* Reverse the velocity */
2168                 state->v[i][m] = -state->v[i][m];
2169                 /* Store the position at t-dt in buf */
2170                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2171             }
2172         }
2173         /* Shake the positions at t=-dt with the positions at t=0
2174          * as reference coordinates.
2175          */
2176         if (fplog)
2177         {
2178             char buf[STEPSTRSIZE];
2179             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2180                     gmx_step_str(step, buf));
2181         }
2182         dvdl_dum = 0;
2183         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2184                   ir, NULL, cr, step, -1, 1.0, md,
2185                   state->x, savex, NULL,
2186                   fr->bMolPBC, state->box,
2187                   state->lambda[efptBONDED], &dvdl_dum,
2188                   state->v, NULL, nrnb, econqCoord,
2189                   ir->epc == epcMTTK, state->veta, state->veta);
2190
2191         for (i = start; i < end; i++)
2192         {
2193             for (m = 0; m < DIM; m++)
2194             {
2195                 /* Re-reverse the velocities */
2196                 state->v[i][m] = -state->v[i][m];
2197             }
2198         }
2199     }
2200     sfree(savex);
2201 }
2202
2203
2204 static void
2205 integrate_table(real vdwtab[], real scale, int offstart, int rstart, int rend,
2206                 double *enerout, double *virout)
2207 {
2208     double enersum, virsum;
2209     double invscale, invscale2, invscale3;
2210     double r, ea, eb, ec, pa, pb, pc, pd;
2211     double y0, f, g, h;
2212     int    ri, offset, tabfactor;
2213
2214     invscale  = 1.0/scale;
2215     invscale2 = invscale*invscale;
2216     invscale3 = invscale*invscale2;
2217
2218     /* Following summation derived from cubic spline definition,
2219      * Numerical Recipies in C, second edition, p. 113-116.  Exact for
2220      * the cubic spline.  We first calculate the negative of the
2221      * energy from rvdw to rvdw_switch, assuming that g(r)=1, and then
2222      * add the more standard, abrupt cutoff correction to that result,
2223      * yielding the long-range correction for a switched function.  We
2224      * perform both the pressure and energy loops at the same time for
2225      * simplicity, as the computational cost is low. */
2226
2227     if (offstart == 0)
2228     {
2229         /* Since the dispersion table has been scaled down a factor
2230          * 6.0 and the repulsion a factor 12.0 to compensate for the
2231          * c6/c12 parameters inside nbfp[] being scaled up (to save
2232          * flops in kernels), we need to correct for this.
2233          */
2234         tabfactor = 6.0;
2235     }
2236     else
2237     {
2238         tabfactor = 12.0;
2239     }
2240
2241     enersum = 0.0;
2242     virsum  = 0.0;
2243     for (ri = rstart; ri < rend; ++ri)
2244     {
2245         r  = ri*invscale;
2246         ea = invscale3;
2247         eb = 2.0*invscale2*r;
2248         ec = invscale*r*r;
2249
2250         pa = invscale3;
2251         pb = 3.0*invscale2*r;
2252         pc = 3.0*invscale*r*r;
2253         pd = r*r*r;
2254
2255         /* this "8" is from the packing in the vdwtab array - perhaps
2256            should be defined? */
2257
2258         offset = 8*ri + offstart;
2259         y0     = vdwtab[offset];
2260         f      = vdwtab[offset+1];
2261         g      = vdwtab[offset+2];
2262         h      = vdwtab[offset+3];
2263
2264         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);
2265         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);
2266     }
2267     *enerout = 4.0*M_PI*enersum*tabfactor;
2268     *virout  = 4.0*M_PI*virsum*tabfactor;
2269 }
2270
2271 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2272 {
2273     double   eners[2], virs[2], enersum, virsum, y0, f, g, h;
2274     double   r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2275     double   invscale, invscale2, invscale3;
2276     int      ri0, ri1, ri, i, offstart, offset;
2277     real     scale, *vdwtab, tabfactor, tmp;
2278
2279     fr->enershiftsix    = 0;
2280     fr->enershifttwelve = 0;
2281     fr->enerdiffsix     = 0;
2282     fr->enerdifftwelve  = 0;
2283     fr->virdiffsix      = 0;
2284     fr->virdifftwelve   = 0;
2285
2286     if (eDispCorr != edispcNO)
2287     {
2288         for (i = 0; i < 2; i++)
2289         {
2290             eners[i] = 0;
2291             virs[i]  = 0;
2292         }
2293         if ((fr->vdw_modifier == eintmodPOTSHIFT) ||
2294             (fr->vdw_modifier == eintmodPOTSWITCH) ||
2295             (fr->vdw_modifier == eintmodFORCESWITCH) ||
2296             (fr->vdwtype == evdwSHIFT) ||
2297             (fr->vdwtype == evdwSWITCH))
2298         {
2299             if (((fr->vdw_modifier == eintmodPOTSWITCH) ||
2300                  (fr->vdw_modifier == eintmodFORCESWITCH) ||
2301                  (fr->vdwtype == evdwSWITCH)) && fr->rvdw_switch == 0)
2302             {
2303                 gmx_fatal(FARGS,
2304                           "With dispersion correction rvdw-switch can not be zero "
2305                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2306             }
2307
2308             scale  = fr->nblists[0].table_vdw.scale;
2309             vdwtab = fr->nblists[0].table_vdw.data;
2310
2311             /* Round the cut-offs to exact table values for precision */
2312             ri0  = floor(fr->rvdw_switch*scale);
2313             ri1  = ceil(fr->rvdw*scale);
2314
2315             /* The code below has some support for handling force-switching, i.e.
2316              * when the force (instead of potential) is switched over a limited
2317              * region. This leads to a constant shift in the potential inside the
2318              * switching region, which we can handle by adding a constant energy
2319              * term in the force-switch case just like when we do potential-shift.
2320              *
2321              * For now this is not enabled, but to keep the functionality in the
2322              * code we check separately for switch and shift. When we do force-switch
2323              * the shifting point is rvdw_switch, while it is the cutoff when we
2324              * have a classical potential-shift.
2325              *
2326              * For a pure potential-shift the potential has a constant shift
2327              * all the way out to the cutoff, and that is it. For other forms
2328              * we need to calculate the constant shift up to the point where we
2329              * start modifying the potential.
2330              */
2331             ri0  = (fr->vdw_modifier == eintmodPOTSHIFT) ? ri1 : ri0;
2332
2333             r0   = ri0/scale;
2334             r1   = ri1/scale;
2335             rc3  = r0*r0*r0;
2336             rc9  = rc3*rc3*rc3;
2337
2338             if ((fr->vdw_modifier == eintmodFORCESWITCH) ||
2339                 (fr->vdwtype == evdwSHIFT))
2340             {
2341                 /* Determine the constant energy shift below rvdw_switch.
2342                  * Table has a scale factor since we have scaled it down to compensate
2343                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2344                  */
2345                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2346                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2347             }
2348             else if (fr->vdw_modifier == eintmodPOTSHIFT)
2349             {
2350                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3));
2351                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3));
2352             }
2353
2354             /* Add the constant part from 0 to rvdw_switch.
2355              * This integration from 0 to rvdw_switch overcounts the number
2356              * of interactions by 1, as it also counts the self interaction.
2357              * We will correct for this later.
2358              */
2359             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2360             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2361
2362             /* Calculate the contribution in the range [r0,r1] where we
2363              * modify the potential. For a pure potential-shift modifier we will
2364              * have ri0==ri1, and there will not be any contribution here.
2365              */
2366             for (i = 0; i < 2; i++)
2367             {
2368                 enersum = 0;
2369                 virsum  = 0;
2370                 integrate_table(vdwtab, scale, (i == 0 ? 0 : 4), ri0, ri1, &enersum, &virsum);
2371                 eners[i] -= enersum;
2372                 virs[i]  -= virsum;
2373             }
2374
2375             /* Alright: Above we compensated by REMOVING the parts outside r0
2376              * corresponding to the ideal VdW 1/r6 and /r12 potentials.
2377              *
2378              * Regardless of whether r0 is the point where we start switching,
2379              * or the cutoff where we calculated the constant shift, we include
2380              * all the parts we are missing out to infinity from r0 by
2381              * calculating the analytical dispersion correction.
2382              */
2383             eners[0] += -4.0*M_PI/(3.0*rc3);
2384             eners[1] +=  4.0*M_PI/(9.0*rc9);
2385             virs[0]  +=  8.0*M_PI/rc3;
2386             virs[1]  += -16.0*M_PI/(3.0*rc9);
2387         }
2388         else if (fr->vdwtype == evdwCUT ||
2389                  EVDW_PME(fr->vdwtype) ||
2390                  fr->vdwtype == evdwUSER)
2391         {
2392             if (fr->vdwtype == evdwUSER && fplog)
2393             {
2394                 fprintf(fplog,
2395                         "WARNING: using dispersion correction with user tables\n");
2396             }
2397
2398             /* Note that with LJ-PME, the dispersion correction is multiplied
2399              * by the difference between the actual C6 and the value of C6
2400              * that would produce the combination rule.
2401              * This means the normal energy and virial difference formulas
2402              * can be used here.
2403              */
2404
2405             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2406             rc9  = rc3*rc3*rc3;
2407             /* Contribution beyond the cut-off */
2408             eners[0] += -4.0*M_PI/(3.0*rc3);
2409             eners[1] +=  4.0*M_PI/(9.0*rc9);
2410             if (fr->vdw_modifier == eintmodPOTSHIFT)
2411             {
2412                 /* Contribution within the cut-off */
2413                 eners[0] += -4.0*M_PI/(3.0*rc3);
2414                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2415             }
2416             /* Contribution beyond the cut-off */
2417             virs[0]  +=  8.0*M_PI/rc3;
2418             virs[1]  += -16.0*M_PI/(3.0*rc9);
2419         }
2420         else
2421         {
2422             gmx_fatal(FARGS,
2423                       "Dispersion correction is not implemented for vdw-type = %s",
2424                       evdw_names[fr->vdwtype]);
2425         }
2426
2427         /* When we deprecate the group kernels the code below can go too */
2428         if (fr->vdwtype == evdwPME && fr->cutoff_scheme == ecutsGROUP)
2429         {
2430             /* Calculate self-interaction coefficient (assuming that
2431              * the reciprocal-space contribution is constant in the
2432              * region that contributes to the self-interaction).
2433              */
2434             fr->enershiftsix = pow(fr->ewaldcoeff_lj, 6) / 6.0;
2435
2436             eners[0] += -pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3)/3.0;
2437             virs[0]  +=  pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3);
2438         }
2439
2440         fr->enerdiffsix    = eners[0];
2441         fr->enerdifftwelve = eners[1];
2442         /* The 0.5 is due to the Gromacs definition of the virial */
2443         fr->virdiffsix     = 0.5*virs[0];
2444         fr->virdifftwelve  = 0.5*virs[1];
2445     }
2446 }
2447
2448 void calc_dispcorr(FILE *fplog, t_inputrec *ir, t_forcerec *fr,
2449                    gmx_int64_t step, int natoms,
2450                    matrix box, real lambda, tensor pres, tensor virial,
2451                    real *prescorr, real *enercorr, real *dvdlcorr)
2452 {
2453     gmx_bool bCorrAll, bCorrPres;
2454     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2455     int      m;
2456
2457     *prescorr = 0;
2458     *enercorr = 0;
2459     *dvdlcorr = 0;
2460
2461     clear_mat(virial);
2462     clear_mat(pres);
2463
2464     if (ir->eDispCorr != edispcNO)
2465     {
2466         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2467                      ir->eDispCorr == edispcAllEnerPres);
2468         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2469                      ir->eDispCorr == edispcAllEnerPres);
2470
2471         invvol = 1/det(box);
2472         if (fr->n_tpi)
2473         {
2474             /* Only correct for the interactions with the inserted molecule */
2475             dens   = (natoms - fr->n_tpi)*invvol;
2476             ninter = fr->n_tpi;
2477         }
2478         else
2479         {
2480             dens   = natoms*invvol;
2481             ninter = 0.5*natoms;
2482         }
2483
2484         if (ir->efep == efepNO)
2485         {
2486             avcsix    = fr->avcsix[0];
2487             avctwelve = fr->avctwelve[0];
2488         }
2489         else
2490         {
2491             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2492             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2493         }
2494
2495         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2496         *enercorr += avcsix*enerdiff;
2497         dvdlambda  = 0.0;
2498         if (ir->efep != efepNO)
2499         {
2500             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2501         }
2502         if (bCorrAll)
2503         {
2504             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2505             *enercorr += avctwelve*enerdiff;
2506             if (fr->efep != efepNO)
2507             {
2508                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2509             }
2510         }
2511
2512         if (bCorrPres)
2513         {
2514             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2515             if (ir->eDispCorr == edispcAllEnerPres)
2516             {
2517                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2518             }
2519             /* The factor 2 is because of the Gromacs virial definition */
2520             spres = -2.0*invvol*svir*PRESFAC;
2521
2522             for (m = 0; m < DIM; m++)
2523             {
2524                 virial[m][m] += svir;
2525                 pres[m][m]   += spres;
2526             }
2527             *prescorr += spres;
2528         }
2529
2530         /* Can't currently control when it prints, for now, just print when degugging */
2531         if (debug)
2532         {
2533             if (bCorrAll)
2534             {
2535                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2536                         avcsix, avctwelve);
2537             }
2538             if (bCorrPres)
2539             {
2540                 fprintf(debug,
2541                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2542                         *enercorr, spres, svir);
2543             }
2544             else
2545             {
2546                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2547             }
2548         }
2549
2550         if (fr->bSepDVDL && do_per_step(step, ir->nstlog))
2551         {
2552             gmx_print_sepdvdl(fplog, "Dispersion correction", *enercorr, dvdlambda);
2553         }
2554         if (fr->efep != efepNO)
2555         {
2556             *dvdlcorr += dvdlambda;
2557         }
2558     }
2559 }
2560
2561 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2562                   t_graph *graph, rvec x[])
2563 {
2564     if (fplog)
2565     {
2566         fprintf(fplog, "Removing pbc first time\n");
2567     }
2568     calc_shifts(box, fr->shift_vec);
2569     if (graph)
2570     {
2571         mk_mshift(fplog, graph, fr->ePBC, box, x);
2572         if (gmx_debug_at)
2573         {
2574             p_graph(debug, "do_pbc_first 1", graph);
2575         }
2576         shift_self(graph, box, x);
2577         /* By doing an extra mk_mshift the molecules that are broken
2578          * because they were e.g. imported from another software
2579          * will be made whole again. Such are the healing powers
2580          * of GROMACS.
2581          */
2582         mk_mshift(fplog, graph, fr->ePBC, box, x);
2583         if (gmx_debug_at)
2584         {
2585             p_graph(debug, "do_pbc_first 2", graph);
2586         }
2587     }
2588     if (fplog)
2589     {
2590         fprintf(fplog, "Done rmpbc\n");
2591     }
2592 }
2593
2594 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2595                             gmx_mtop_t *mtop, rvec x[],
2596                             gmx_bool bFirst)
2597 {
2598     t_graph        *graph;
2599     int             mb, as, mol;
2600     gmx_molblock_t *molb;
2601
2602     if (bFirst && fplog)
2603     {
2604         fprintf(fplog, "Removing pbc first time\n");
2605     }
2606
2607     snew(graph, 1);
2608     as = 0;
2609     for (mb = 0; mb < mtop->nmolblock; mb++)
2610     {
2611         molb = &mtop->molblock[mb];
2612         if (molb->natoms_mol == 1 ||
2613             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2614         {
2615             /* Just one atom or charge group in the molecule, no PBC required */
2616             as += molb->nmol*molb->natoms_mol;
2617         }
2618         else
2619         {
2620             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2621             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2622                            0, molb->natoms_mol, FALSE, FALSE, graph);
2623
2624             for (mol = 0; mol < molb->nmol; mol++)
2625             {
2626                 mk_mshift(fplog, graph, ePBC, box, x+as);
2627
2628                 shift_self(graph, box, x+as);
2629                 /* The molecule is whole now.
2630                  * We don't need the second mk_mshift call as in do_pbc_first,
2631                  * since we no longer need this graph.
2632                  */
2633
2634                 as += molb->natoms_mol;
2635             }
2636             done_graph(graph);
2637         }
2638     }
2639     sfree(graph);
2640 }
2641
2642 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2643                        gmx_mtop_t *mtop, rvec x[])
2644 {
2645     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2646 }
2647
2648 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2649                  gmx_mtop_t *mtop, rvec x[])
2650 {
2651     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2652 }
2653
2654 void finish_run(FILE *fplog, t_commrec *cr,
2655                 t_inputrec *inputrec,
2656                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2657                 gmx_walltime_accounting_t walltime_accounting,
2658                 wallclock_gpu_t *gputimes,
2659                 gmx_bool bWriteStat)
2660 {
2661     int     i, j;
2662     t_nrnb *nrnb_tot = NULL;
2663     real    delta_t;
2664     double  nbfs, mflop;
2665     double  elapsed_time,
2666             elapsed_time_over_all_ranks,
2667             elapsed_time_over_all_threads,
2668             elapsed_time_over_all_threads_over_all_ranks;
2669     wallcycle_sum(cr, wcycle);
2670
2671     if (cr->nnodes > 1)
2672     {
2673         snew(nrnb_tot, 1);
2674 #ifdef GMX_MPI
2675         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2676                       cr->mpi_comm_mysim);
2677 #endif
2678     }
2679     else
2680     {
2681         nrnb_tot = nrnb;
2682     }
2683
2684     elapsed_time                                 = walltime_accounting_get_elapsed_time(walltime_accounting);
2685     elapsed_time_over_all_ranks                  = elapsed_time;
2686     elapsed_time_over_all_threads                = walltime_accounting_get_elapsed_time_over_all_threads(walltime_accounting);
2687     elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
2688 #ifdef GMX_MPI
2689     if (cr->nnodes > 1)
2690     {
2691         /* reduce elapsed_time over all MPI ranks in the current simulation */
2692         MPI_Allreduce(&elapsed_time,
2693                       &elapsed_time_over_all_ranks,
2694                       1, MPI_DOUBLE, MPI_SUM,
2695                       cr->mpi_comm_mysim);
2696         elapsed_time_over_all_ranks /= cr->nnodes;
2697         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
2698          * current simulation. */
2699         MPI_Allreduce(&elapsed_time_over_all_threads,
2700                       &elapsed_time_over_all_threads_over_all_ranks,
2701                       1, MPI_DOUBLE, MPI_SUM,
2702                       cr->mpi_comm_mysim);
2703     }
2704 #endif
2705
2706     if (SIMMASTER(cr))
2707     {
2708         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2709     }
2710     if (cr->nnodes > 1)
2711     {
2712         sfree(nrnb_tot);
2713     }
2714
2715     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2716     {
2717         print_dd_statistics(cr, inputrec, fplog);
2718     }
2719
2720     if (SIMMASTER(cr))
2721     {
2722         wallcycle_print(fplog, cr->nnodes, cr->npmenodes,
2723                         elapsed_time_over_all_ranks,
2724                         wcycle, gputimes);
2725
2726         if (EI_DYNAMICS(inputrec->eI))
2727         {
2728             delta_t = inputrec->delta_t;
2729         }
2730         else
2731         {
2732             delta_t = 0;
2733         }
2734
2735         if (fplog)
2736         {
2737             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
2738                        elapsed_time_over_all_ranks,
2739                        walltime_accounting_get_nsteps_done(walltime_accounting),
2740                        delta_t, nbfs, mflop);
2741         }
2742         if (bWriteStat)
2743         {
2744             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
2745                        elapsed_time_over_all_ranks,
2746                        walltime_accounting_get_nsteps_done(walltime_accounting),
2747                        delta_t, nbfs, mflop);
2748         }
2749     }
2750 }
2751
2752 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2753 {
2754     /* this function works, but could probably use a logic rewrite to keep all the different
2755        types of efep straight. */
2756
2757     int       i;
2758     t_lambda *fep = ir->fepvals;
2759
2760     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2761     {
2762         for (i = 0; i < efptNR; i++)
2763         {
2764             lambda[i] = 0.0;
2765             if (lam0)
2766             {
2767                 lam0[i] = 0.0;
2768             }
2769         }
2770         return;
2771     }
2772     else
2773     {
2774         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2775                                              if checkpoint is set -- a kludge is in for now
2776                                              to prevent this.*/
2777         for (i = 0; i < efptNR; i++)
2778         {
2779             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2780             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2781             {
2782                 lambda[i] = fep->init_lambda;
2783                 if (lam0)
2784                 {
2785                     lam0[i] = lambda[i];
2786                 }
2787             }
2788             else
2789             {
2790                 lambda[i] = fep->all_lambda[i][*fep_state];
2791                 if (lam0)
2792                 {
2793                     lam0[i] = lambda[i];
2794                 }
2795             }
2796         }
2797         if (ir->bSimTemp)
2798         {
2799             /* need to rescale control temperatures to match current state */
2800             for (i = 0; i < ir->opts.ngtc; i++)
2801             {
2802                 if (ir->opts.ref_t[i] > 0)
2803                 {
2804                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2805                 }
2806             }
2807         }
2808     }
2809
2810     /* Send to the log the information on the current lambdas */
2811     if (fplog != NULL)
2812     {
2813         fprintf(fplog, "Initial vector of lambda components:[ ");
2814         for (i = 0; i < efptNR; i++)
2815         {
2816             fprintf(fplog, "%10.4f ", lambda[i]);
2817         }
2818         fprintf(fplog, "]\n");
2819     }
2820     return;
2821 }
2822
2823
2824 void init_md(FILE *fplog,
2825              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2826              double *t, double *t0,
2827              real *lambda, int *fep_state, double *lam0,
2828              t_nrnb *nrnb, gmx_mtop_t *mtop,
2829              gmx_update_t *upd,
2830              int nfile, const t_filenm fnm[],
2831              gmx_mdoutf_t *outf, t_mdebin **mdebin,
2832              tensor force_vir, tensor shake_vir, rvec mu_tot,
2833              gmx_bool *bSimAnn, t_vcm **vcm, unsigned long Flags,
2834              gmx_wallcycle_t wcycle)
2835 {
2836     int  i, j, n;
2837     real tmpt, mod;
2838
2839     /* Initial values */
2840     *t = *t0       = ir->init_t;
2841
2842     *bSimAnn = FALSE;
2843     for (i = 0; i < ir->opts.ngtc; i++)
2844     {
2845         /* set bSimAnn if any group is being annealed */
2846         if (ir->opts.annealing[i] != eannNO)
2847         {
2848             *bSimAnn = TRUE;
2849         }
2850     }
2851     if (*bSimAnn)
2852     {
2853         update_annealing_target_temp(&(ir->opts), ir->init_t);
2854     }
2855
2856     /* Initialize lambda variables */
2857     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2858
2859     if (upd)
2860     {
2861         *upd = init_update(ir);
2862     }
2863
2864
2865     if (vcm != NULL)
2866     {
2867         *vcm = init_vcm(fplog, &mtop->groups, ir);
2868     }
2869
2870     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2871     {
2872         if (ir->etc == etcBERENDSEN)
2873         {
2874             please_cite(fplog, "Berendsen84a");
2875         }
2876         if (ir->etc == etcVRESCALE)
2877         {
2878             please_cite(fplog, "Bussi2007a");
2879         }
2880         if (ir->eI == eiSD1)
2881         {
2882             please_cite(fplog, "Goga2012");
2883         }
2884     }
2885
2886     init_nrnb(nrnb);
2887
2888     if (nfile != -1)
2889     {
2890         *outf = init_mdoutf(fplog, nfile, fnm, Flags, cr, ir, mtop, oenv, wcycle);
2891
2892         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : mdoutf_get_fp_ene(*outf),
2893                               mtop, ir, mdoutf_get_fp_dhdl(*outf));
2894     }
2895
2896     if (ir->bAdress)
2897     {
2898         please_cite(fplog, "Fritsch12");
2899         please_cite(fplog, "Junghans10");
2900     }
2901     /* Initiate variables */
2902     clear_mat(force_vir);
2903     clear_mat(shake_vir);
2904     clear_rvec(mu_tot);
2905
2906     debug_gmx();
2907 }