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         if (bUseGPU)
1417         {
1418             /* We are done with the CPU compute, but the GPU local non-bonded
1419              * kernel can still be running while we communicate the forces.
1420              * We start a counter here, so we can, hopefully, time the rest
1421              * of the GPU kernel execution and data transfer.
1422              */
1423             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L_EST);
1424         }
1425
1426         /* Communicate the forces */
1427         wallcycle_start(wcycle, ewcMOVEF);
1428         dd_move_f(cr->dd, f, fr->fshift);
1429         /* Do we need to communicate the separate force array
1430          * for terms that do not contribute to the single sum virial?
1431          * Position restraints and electric fields do not introduce
1432          * inter-cg forces, only full electrostatics methods do.
1433          * When we do not calculate the virial, fr->f_novirsum = f,
1434          * so we have already communicated these forces.
1435          */
1436         if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1437             (flags & GMX_FORCE_VIRIAL))
1438         {
1439             dd_move_f(cr->dd, fr->f_novirsum, NULL);
1440         }
1441         if (bSepLRF)
1442         {
1443             /* We should not update the shift forces here,
1444              * since f_twin is already included in f.
1445              */
1446             dd_move_f(cr->dd, fr->f_twin, NULL);
1447         }
1448         wallcycle_stop(wcycle, ewcMOVEF);
1449     }
1450
1451     if (bUseOrEmulGPU)
1452     {
1453         /* wait for local forces (or calculate in emulation mode) */
1454         if (bUseGPU)
1455         {
1456             float       cycles_tmp, cycles_wait_est;
1457             const float cuda_api_overhead_margin = 50000.0f; /* cycles */
1458
1459             wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1460             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1461                                 nbv->grp[eintLocal].nbat,
1462                                 flags, eatLocal,
1463                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1464                                 fr->fshift);
1465             cycles_tmp      = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1466
1467             if (bDoForces && DOMAINDECOMP(cr))
1468             {
1469                 cycles_wait_est = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L_EST);
1470
1471                 if (cycles_tmp < cuda_api_overhead_margin)
1472                 {
1473                     /* We measured few cycles, it could be that the kernel
1474                      * and transfer finished earlier and there was no actual
1475                      * wait time, only API call overhead.
1476                      * Then the actual time could be anywhere between 0 and
1477                      * cycles_wait_est. As a compromise, we use half the time.
1478                      */
1479                     cycles_wait_est *= 0.5f;
1480                 }
1481             }
1482             else
1483             {
1484                 /* No force communication so we actually timed the wait */
1485                 cycles_wait_est = cycles_tmp;
1486             }
1487             /* Even though this is after dd_move_f, the actual task we are
1488              * waiting for runs asynchronously with dd_move_f and we usually
1489              * have nothing to balance it with, so we can and should add
1490              * the time to the force time for load balancing.
1491              */
1492             cycles_force    += cycles_wait_est;
1493             cycles_wait_gpu += cycles_wait_est;
1494
1495             /* now clear the GPU outputs while we finish the step on the CPU */
1496
1497             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1498             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1499             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1500         }
1501         else
1502         {
1503             wallcycle_start_nocount(wcycle, ewcFORCE);
1504             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1505                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1506                          nrnb, wcycle);
1507             wallcycle_stop(wcycle, ewcFORCE);
1508         }
1509         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1510         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1511         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1512         {
1513             /* skip the reduction if there was no non-local work to do */
1514             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1515                                            nbv->grp[eintLocal].nbat, f);
1516         }
1517         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1518         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1519     }
1520
1521     if (DOMAINDECOMP(cr))
1522     {
1523         dd_force_flop_stop(cr->dd, nrnb);
1524         if (wcycle)
1525         {
1526             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1527             if (bUseGPU)
1528             {
1529                 dd_cycles_add(cr->dd, cycles_wait_gpu, ddCyclWaitGPU);
1530             }
1531         }
1532     }
1533
1534     if (bDoForces)
1535     {
1536         if (IR_ELEC_FIELD(*inputrec))
1537         {
1538             /* Compute forces due to electric field */
1539             calc_f_el(MASTER(cr) ? field : NULL,
1540                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1541                       inputrec->ex, inputrec->et, t);
1542         }
1543
1544         /* If we have NoVirSum forces, but we do not calculate the virial,
1545          * we sum fr->f_novirum=f later.
1546          */
1547         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1548         {
1549             wallcycle_start(wcycle, ewcVSITESPREAD);
1550             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1551                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1552             wallcycle_stop(wcycle, ewcVSITESPREAD);
1553
1554             if (bSepLRF)
1555             {
1556                 wallcycle_start(wcycle, ewcVSITESPREAD);
1557                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
1558                                nrnb,
1559                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1560                 wallcycle_stop(wcycle, ewcVSITESPREAD);
1561             }
1562         }
1563
1564         if (flags & GMX_FORCE_VIRIAL)
1565         {
1566             /* Calculation of the virial must be done after vsites! */
1567             calc_virial(0, mdatoms->homenr, x, f,
1568                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1569         }
1570     }
1571
1572     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1573     {
1574         /* Since the COM pulling is always done mass-weighted, no forces are
1575          * applied to vsites and this call can be done after vsite spreading.
1576          */
1577         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1578                                f, vir_force, mdatoms, enerd, lambda, t,
1579                                wcycle);
1580     }
1581
1582     /* Add the forces from enforced rotation potentials (if any) */
1583     if (inputrec->bRot)
1584     {
1585         wallcycle_start(wcycle, ewcROTadd);
1586         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1587         wallcycle_stop(wcycle, ewcROTadd);
1588     }
1589
1590     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
1591     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
1592
1593     if (PAR(cr) && !(cr->duty & DUTY_PME))
1594     {
1595         /* In case of node-splitting, the PP nodes receive the long-range
1596          * forces, virial and energy from the PME nodes here.
1597          */
1598         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1599     }
1600
1601     if (bDoForces)
1602     {
1603         post_process_forces(cr, step, nrnb, wcycle,
1604                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1605                             flags);
1606     }
1607
1608     /* Sum the potential energy terms from group contributions */
1609     sum_epot(&(enerd->grpp), enerd->term);
1610 }
1611
1612 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1613                         t_inputrec *inputrec,
1614                         gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1615                         gmx_localtop_t *top,
1616                         gmx_groups_t *groups,
1617                         matrix box, rvec x[], history_t *hist,
1618                         rvec f[],
1619                         tensor vir_force,
1620                         t_mdatoms *mdatoms,
1621                         gmx_enerdata_t *enerd, t_fcdata *fcd,
1622                         real *lambda, t_graph *graph,
1623                         t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1624                         double t, FILE *field, gmx_edsam_t ed,
1625                         gmx_bool bBornRadii,
1626                         int flags)
1627 {
1628     int        cg0, cg1, i, j;
1629     int        start, homenr;
1630     double     mu[2*DIM];
1631     gmx_bool   bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1632     gmx_bool   bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1633     gmx_bool   bDoAdressWF;
1634     matrix     boxs;
1635     rvec       vzero, box_diag;
1636     real       e, v, dvdlambda[efptNR];
1637     t_pbc      pbc;
1638     float      cycles_pme, cycles_force;
1639
1640     start  = 0;
1641     homenr = mdatoms->homenr;
1642
1643     bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
1644
1645     clear_mat(vir_force);
1646
1647     cg0 = 0;
1648     if (DOMAINDECOMP(cr))
1649     {
1650         cg1 = cr->dd->ncg_tot;
1651     }
1652     else
1653     {
1654         cg1 = top->cgs.nr;
1655     }
1656     if (fr->n_tpi > 0)
1657     {
1658         cg1--;
1659     }
1660
1661     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1662     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1663     /* Should we update the long-range neighborlists at this step? */
1664     bDoLongRangeNS = fr->bTwinRange && bNS;
1665     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1666     bFillGrid      = (bNS && bStateChanged);
1667     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1668     bDoForces      = (flags & GMX_FORCE_FORCES);
1669     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1670     bSepLRF        = ((inputrec->nstcalclr > 1) && bDoForces &&
1671                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1672
1673     /* should probably move this to the forcerec since it doesn't change */
1674     bDoAdressWF   = ((fr->adress_type != eAdressOff));
1675
1676     if (bStateChanged)
1677     {
1678         update_forcerec(fr, box);
1679
1680         if (NEED_MUTOT(*inputrec))
1681         {
1682             /* Calculate total (local) dipole moment in a temporary common array.
1683              * This makes it possible to sum them over nodes faster.
1684              */
1685             calc_mu(start, homenr,
1686                     x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1687                     mu, mu+DIM);
1688         }
1689     }
1690
1691     if (fr->ePBC != epbcNONE)
1692     {
1693         /* Compute shift vectors every step,
1694          * because of pressure coupling or box deformation!
1695          */
1696         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1697         {
1698             calc_shifts(box, fr->shift_vec);
1699         }
1700
1701         if (bCalcCGCM)
1702         {
1703             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1704                                      &(top->cgs), x, fr->cg_cm);
1705             inc_nrnb(nrnb, eNR_CGCM, homenr);
1706             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1707         }
1708         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1709         {
1710             unshift_self(graph, box, x);
1711         }
1712     }
1713     else if (bCalcCGCM)
1714     {
1715         calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1716         inc_nrnb(nrnb, eNR_CGCM, homenr);
1717     }
1718
1719     if (bCalcCGCM && gmx_debug_at)
1720     {
1721         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1722     }
1723
1724 #ifdef GMX_MPI
1725     if (!(cr->duty & DUTY_PME))
1726     {
1727         /* Send particle coordinates to the pme nodes.
1728          * Since this is only implemented for domain decomposition
1729          * and domain decomposition does not use the graph,
1730          * we do not need to worry about shifting.
1731          */
1732
1733         int pme_flags = 0;
1734
1735         wallcycle_start(wcycle, ewcPP_PMESENDX);
1736
1737         bBS = (inputrec->nwall == 2);
1738         if (bBS)
1739         {
1740             copy_mat(box, boxs);
1741             svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1742         }
1743
1744         if (EEL_PME(fr->eeltype))
1745         {
1746             pme_flags |= GMX_PME_DO_COULOMB;
1747         }
1748
1749         if (EVDW_PME(fr->vdwtype))
1750         {
1751             pme_flags |= GMX_PME_DO_LJ;
1752         }
1753
1754         gmx_pme_send_coordinates(cr, bBS ? boxs : box, x,
1755                                  mdatoms->nChargePerturbed, mdatoms->nTypePerturbed, lambda[efptCOUL], lambda[efptVDW],
1756                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),
1757                                  pme_flags, step);
1758
1759         wallcycle_stop(wcycle, ewcPP_PMESENDX);
1760     }
1761 #endif /* GMX_MPI */
1762
1763     /* Communicate coordinates and sum dipole if necessary */
1764     if (DOMAINDECOMP(cr))
1765     {
1766         wallcycle_start(wcycle, ewcMOVEX);
1767         dd_move_x(cr->dd, box, x);
1768         wallcycle_stop(wcycle, ewcMOVEX);
1769     }
1770
1771     /* update adress weight beforehand */
1772     if (bStateChanged && bDoAdressWF)
1773     {
1774         /* need pbc for adress weight calculation with pbc_dx */
1775         set_pbc(&pbc, inputrec->ePBC, box);
1776         if (fr->adress_site == eAdressSITEcog)
1777         {
1778             update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1779                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1780         }
1781         else if (fr->adress_site == eAdressSITEcom)
1782         {
1783             update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1784                                       inputrec->ePBC == epbcNONE ? NULL : &pbc);
1785         }
1786         else if (fr->adress_site == eAdressSITEatomatom)
1787         {
1788             update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1789                                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1790         }
1791         else
1792         {
1793             update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1794                                        inputrec->ePBC == epbcNONE ? NULL : &pbc);
1795         }
1796     }
1797
1798     if (NEED_MUTOT(*inputrec))
1799     {
1800
1801         if (bStateChanged)
1802         {
1803             if (PAR(cr))
1804             {
1805                 gmx_sumd(2*DIM, mu, cr);
1806             }
1807             for (i = 0; i < 2; i++)
1808             {
1809                 for (j = 0; j < DIM; j++)
1810                 {
1811                     fr->mu_tot[i][j] = mu[i*DIM + j];
1812                 }
1813             }
1814         }
1815         if (fr->efep == efepNO)
1816         {
1817             copy_rvec(fr->mu_tot[0], mu_tot);
1818         }
1819         else
1820         {
1821             for (j = 0; j < DIM; j++)
1822             {
1823                 mu_tot[j] =
1824                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1825             }
1826         }
1827     }
1828
1829     /* Reset energies */
1830     reset_enerdata(fr, bNS, enerd, MASTER(cr));
1831     clear_rvecs(SHIFTS, fr->fshift);
1832
1833     if (bNS)
1834     {
1835         wallcycle_start(wcycle, ewcNS);
1836
1837         if (graph && bStateChanged)
1838         {
1839             /* Calculate intramolecular shift vectors to make molecules whole */
1840             mk_mshift(fplog, graph, fr->ePBC, box, x);
1841         }
1842
1843         /* Do the actual neighbour searching */
1844         ns(fplog, fr, box,
1845            groups, top, mdatoms,
1846            cr, nrnb, bFillGrid,
1847            bDoLongRangeNS);
1848
1849         wallcycle_stop(wcycle, ewcNS);
1850     }
1851
1852     if (inputrec->implicit_solvent && bNS)
1853     {
1854         make_gb_nblist(cr, inputrec->gb_algorithm,
1855                        x, box, fr, &top->idef, graph, fr->born);
1856     }
1857
1858     if (DOMAINDECOMP(cr) && !(cr->duty & DUTY_PME))
1859     {
1860         wallcycle_start(wcycle, ewcPPDURINGPME);
1861         dd_force_flop_start(cr->dd, nrnb);
1862     }
1863
1864     if (inputrec->bRot)
1865     {
1866         /* Enforced rotation has its own cycle counter that starts after the collective
1867          * coordinates have been communicated. It is added to ddCyclF to allow
1868          * for proper load-balancing */
1869         wallcycle_start(wcycle, ewcROT);
1870         do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1871         wallcycle_stop(wcycle, ewcROT);
1872     }
1873
1874     /* Start the force cycle counter.
1875      * This counter is stopped in do_forcelow_level.
1876      * No parallel communication should occur while this counter is running,
1877      * since that will interfere with the dynamic load balancing.
1878      */
1879     wallcycle_start(wcycle, ewcFORCE);
1880
1881     if (bDoForces)
1882     {
1883         /* Reset forces for which the virial is calculated separately:
1884          * PME/Ewald forces if necessary */
1885         if (fr->bF_NoVirSum)
1886         {
1887             if (flags & GMX_FORCE_VIRIAL)
1888             {
1889                 fr->f_novirsum = fr->f_novirsum_alloc;
1890                 if (fr->bDomDec)
1891                 {
1892                     clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1893                 }
1894                 else
1895                 {
1896                     clear_rvecs(homenr, fr->f_novirsum+start);
1897                 }
1898             }
1899             else
1900             {
1901                 /* We are not calculating the pressure so we do not need
1902                  * a separate array for forces that do not contribute
1903                  * to the pressure.
1904                  */
1905                 fr->f_novirsum = f;
1906             }
1907         }
1908
1909         /* Clear the short- and long-range forces */
1910         clear_rvecs(fr->natoms_force_constr, f);
1911         if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1912         {
1913             clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1914         }
1915
1916         clear_rvec(fr->vir_diag_posres);
1917     }
1918     if (inputrec->ePull == epullCONSTRAINT)
1919     {
1920         clear_pull_forces(inputrec->pull);
1921     }
1922
1923     /* update QMMMrec, if necessary */
1924     if (fr->bQMMM)
1925     {
1926         update_QMMMrec(cr, fr, x, mdatoms, box, top);
1927     }
1928
1929     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1930     {
1931         posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1932                        enerd, lambda, fr);
1933     }
1934
1935     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1936     {
1937         fbposres_wrapper(inputrec, nrnb, top, box, x, enerd, fr);
1938     }
1939
1940     /* Compute the bonded and non-bonded energies and optionally forces */
1941     do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1942                       cr, nrnb, wcycle, mdatoms,
1943                       x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, top, fr->born,
1944                       &(top->atomtypes), bBornRadii, box,
1945                       inputrec->fepvals, lambda,
1946                       graph, &(top->excls), fr->mu_tot,
1947                       flags,
1948                       &cycles_pme);
1949
1950     if (bSepLRF)
1951     {
1952         if (do_per_step(step, inputrec->nstcalclr))
1953         {
1954             /* Add the long range forces to the short range forces */
1955             for (i = 0; i < fr->natoms_force_constr; i++)
1956             {
1957                 rvec_add(fr->f_twin[i], f[i], f[i]);
1958             }
1959         }
1960     }
1961
1962     cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1963
1964     if (ed)
1965     {
1966         do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1967     }
1968
1969     if (DOMAINDECOMP(cr))
1970     {
1971         dd_force_flop_stop(cr->dd, nrnb);
1972         if (wcycle)
1973         {
1974             dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1975         }
1976     }
1977
1978     if (bDoForces)
1979     {
1980         if (IR_ELEC_FIELD(*inputrec))
1981         {
1982             /* Compute forces due to electric field */
1983             calc_f_el(MASTER(cr) ? field : NULL,
1984                       start, homenr, mdatoms->chargeA, fr->f_novirsum,
1985                       inputrec->ex, inputrec->et, t);
1986         }
1987
1988         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1989         {
1990             /* Compute thermodynamic force in hybrid AdResS region */
1991             adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1992                                 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1993         }
1994
1995         /* Communicate the forces */
1996         if (DOMAINDECOMP(cr))
1997         {
1998             wallcycle_start(wcycle, ewcMOVEF);
1999             dd_move_f(cr->dd, f, fr->fshift);
2000             /* Do we need to communicate the separate force array
2001              * for terms that do not contribute to the single sum virial?
2002              * Position restraints and electric fields do not introduce
2003              * inter-cg forces, only full electrostatics methods do.
2004              * When we do not calculate the virial, fr->f_novirsum = f,
2005              * so we have already communicated these forces.
2006              */
2007             if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
2008                 (flags & GMX_FORCE_VIRIAL))
2009             {
2010                 dd_move_f(cr->dd, fr->f_novirsum, NULL);
2011             }
2012             if (bSepLRF)
2013             {
2014                 /* We should not update the shift forces here,
2015                  * since f_twin is already included in f.
2016                  */
2017                 dd_move_f(cr->dd, fr->f_twin, NULL);
2018             }
2019             wallcycle_stop(wcycle, ewcMOVEF);
2020         }
2021
2022         /* If we have NoVirSum forces, but we do not calculate the virial,
2023          * we sum fr->f_novirum=f later.
2024          */
2025         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
2026         {
2027             wallcycle_start(wcycle, ewcVSITESPREAD);
2028             spread_vsite_f(vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
2029                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
2030             wallcycle_stop(wcycle, ewcVSITESPREAD);
2031
2032             if (bSepLRF)
2033             {
2034                 wallcycle_start(wcycle, ewcVSITESPREAD);
2035                 spread_vsite_f(vsite, x, fr->f_twin, NULL, FALSE, NULL,
2036                                nrnb,
2037                                &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
2038                 wallcycle_stop(wcycle, ewcVSITESPREAD);
2039             }
2040         }
2041
2042         if (flags & GMX_FORCE_VIRIAL)
2043         {
2044             /* Calculation of the virial must be done after vsites! */
2045             calc_virial(0, mdatoms->homenr, x, f,
2046                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
2047         }
2048     }
2049
2050     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
2051     {
2052         pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
2053                                f, vir_force, mdatoms, enerd, lambda, t,
2054                                wcycle);
2055     }
2056
2057     /* Add the forces from enforced rotation potentials (if any) */
2058     if (inputrec->bRot)
2059     {
2060         wallcycle_start(wcycle, ewcROTadd);
2061         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
2062         wallcycle_stop(wcycle, ewcROTadd);
2063     }
2064
2065     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
2066     IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
2067
2068     if (PAR(cr) && !(cr->duty & DUTY_PME))
2069     {
2070         /* In case of node-splitting, the PP nodes receive the long-range
2071          * forces, virial and energy from the PME nodes here.
2072          */
2073         pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
2074     }
2075
2076     if (bDoForces)
2077     {
2078         post_process_forces(cr, step, nrnb, wcycle,
2079                             top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
2080                             flags);
2081     }
2082
2083     /* Sum the potential energy terms from group contributions */
2084     sum_epot(&(enerd->grpp), enerd->term);
2085 }
2086
2087 void do_force(FILE *fplog, t_commrec *cr,
2088               t_inputrec *inputrec,
2089               gmx_int64_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
2090               gmx_localtop_t *top,
2091               gmx_groups_t *groups,
2092               matrix box, rvec x[], history_t *hist,
2093               rvec f[],
2094               tensor vir_force,
2095               t_mdatoms *mdatoms,
2096               gmx_enerdata_t *enerd, t_fcdata *fcd,
2097               real *lambda, t_graph *graph,
2098               t_forcerec *fr,
2099               gmx_vsite_t *vsite, rvec mu_tot,
2100               double t, FILE *field, gmx_edsam_t ed,
2101               gmx_bool bBornRadii,
2102               int flags)
2103 {
2104     /* modify force flag if not doing nonbonded */
2105     if (!fr->bNonbonded)
2106     {
2107         flags &= ~GMX_FORCE_NONBONDED;
2108     }
2109
2110     switch (inputrec->cutoff_scheme)
2111     {
2112         case ecutsVERLET:
2113             do_force_cutsVERLET(fplog, cr, inputrec,
2114                                 step, nrnb, wcycle,
2115                                 top,
2116                                 groups,
2117                                 box, x, hist,
2118                                 f, vir_force,
2119                                 mdatoms,
2120                                 enerd, fcd,
2121                                 lambda, graph,
2122                                 fr, fr->ic,
2123                                 vsite, mu_tot,
2124                                 t, field, ed,
2125                                 bBornRadii,
2126                                 flags);
2127             break;
2128         case ecutsGROUP:
2129             do_force_cutsGROUP(fplog, cr, inputrec,
2130                                step, nrnb, wcycle,
2131                                top,
2132                                groups,
2133                                box, x, hist,
2134                                f, vir_force,
2135                                mdatoms,
2136                                enerd, fcd,
2137                                lambda, graph,
2138                                fr, vsite, mu_tot,
2139                                t, field, ed,
2140                                bBornRadii,
2141                                flags);
2142             break;
2143         default:
2144             gmx_incons("Invalid cut-off scheme passed!");
2145     }
2146 }
2147
2148
2149 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2150                         t_inputrec *ir, t_mdatoms *md,
2151                         t_state *state, t_commrec *cr, t_nrnb *nrnb,
2152                         t_forcerec *fr, gmx_localtop_t *top)
2153 {
2154     int             i, m, start, end;
2155     gmx_int64_t     step;
2156     real            dt = ir->delta_t;
2157     real            dvdl_dum;
2158     rvec           *savex;
2159
2160     snew(savex, state->natoms);
2161
2162     start = 0;
2163     end   = md->homenr;
2164
2165     if (debug)
2166     {
2167         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2168                 start, md->homenr, end);
2169     }
2170     /* Do a first constrain to reset particles... */
2171     step = ir->init_step;
2172     if (fplog)
2173     {
2174         char buf[STEPSTRSIZE];
2175         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2176                 gmx_step_str(step, buf));
2177     }
2178     dvdl_dum = 0;
2179
2180     /* constrain the current position */
2181     constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2182               ir, NULL, cr, step, 0, 1.0, md,
2183               state->x, state->x, NULL,
2184               fr->bMolPBC, state->box,
2185               state->lambda[efptBONDED], &dvdl_dum,
2186               NULL, NULL, nrnb, econqCoord,
2187               ir->epc == epcMTTK, state->veta, state->veta);
2188     if (EI_VV(ir->eI))
2189     {
2190         /* constrain the inital velocity, and save it */
2191         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2192         /* might not yet treat veta correctly */
2193         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2194                   ir, NULL, cr, step, 0, 1.0, md,
2195                   state->x, state->v, state->v,
2196                   fr->bMolPBC, state->box,
2197                   state->lambda[efptBONDED], &dvdl_dum,
2198                   NULL, NULL, nrnb, econqVeloc,
2199                   ir->epc == epcMTTK, state->veta, state->veta);
2200     }
2201     /* constrain the inital velocities at t-dt/2 */
2202     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2203     {
2204         for (i = start; (i < end); i++)
2205         {
2206             for (m = 0; (m < DIM); m++)
2207             {
2208                 /* Reverse the velocity */
2209                 state->v[i][m] = -state->v[i][m];
2210                 /* Store the position at t-dt in buf */
2211                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2212             }
2213         }
2214         /* Shake the positions at t=-dt with the positions at t=0
2215          * as reference coordinates.
2216          */
2217         if (fplog)
2218         {
2219             char buf[STEPSTRSIZE];
2220             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2221                     gmx_step_str(step, buf));
2222         }
2223         dvdl_dum = 0;
2224         constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2225                   ir, NULL, cr, step, -1, 1.0, md,
2226                   state->x, savex, NULL,
2227                   fr->bMolPBC, state->box,
2228                   state->lambda[efptBONDED], &dvdl_dum,
2229                   state->v, NULL, nrnb, econqCoord,
2230                   ir->epc == epcMTTK, state->veta, state->veta);
2231
2232         for (i = start; i < end; i++)
2233         {
2234             for (m = 0; m < DIM; m++)
2235             {
2236                 /* Re-reverse the velocities */
2237                 state->v[i][m] = -state->v[i][m];
2238             }
2239         }
2240     }
2241     sfree(savex);
2242 }
2243
2244
2245 static void
2246 integrate_table(real vdwtab[], real scale, int offstart, int rstart, int rend,
2247                 double *enerout, double *virout)
2248 {
2249     double enersum, virsum;
2250     double invscale, invscale2, invscale3;
2251     double r, ea, eb, ec, pa, pb, pc, pd;
2252     double y0, f, g, h;
2253     int    ri, offset, tabfactor;
2254
2255     invscale  = 1.0/scale;
2256     invscale2 = invscale*invscale;
2257     invscale3 = invscale*invscale2;
2258
2259     /* Following summation derived from cubic spline definition,
2260      * Numerical Recipies in C, second edition, p. 113-116.  Exact for
2261      * the cubic spline.  We first calculate the negative of the
2262      * energy from rvdw to rvdw_switch, assuming that g(r)=1, and then
2263      * add the more standard, abrupt cutoff correction to that result,
2264      * yielding the long-range correction for a switched function.  We
2265      * perform both the pressure and energy loops at the same time for
2266      * simplicity, as the computational cost is low. */
2267
2268     if (offstart == 0)
2269     {
2270         /* Since the dispersion table has been scaled down a factor
2271          * 6.0 and the repulsion a factor 12.0 to compensate for the
2272          * c6/c12 parameters inside nbfp[] being scaled up (to save
2273          * flops in kernels), we need to correct for this.
2274          */
2275         tabfactor = 6.0;
2276     }
2277     else
2278     {
2279         tabfactor = 12.0;
2280     }
2281
2282     enersum = 0.0;
2283     virsum  = 0.0;
2284     for (ri = rstart; ri < rend; ++ri)
2285     {
2286         r  = ri*invscale;
2287         ea = invscale3;
2288         eb = 2.0*invscale2*r;
2289         ec = invscale*r*r;
2290
2291         pa = invscale3;
2292         pb = 3.0*invscale2*r;
2293         pc = 3.0*invscale*r*r;
2294         pd = r*r*r;
2295
2296         /* this "8" is from the packing in the vdwtab array - perhaps
2297            should be defined? */
2298
2299         offset = 8*ri + offstart;
2300         y0     = vdwtab[offset];
2301         f      = vdwtab[offset+1];
2302         g      = vdwtab[offset+2];
2303         h      = vdwtab[offset+3];
2304
2305         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);
2306         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);
2307     }
2308     *enerout = 4.0*M_PI*enersum*tabfactor;
2309     *virout  = 4.0*M_PI*virsum*tabfactor;
2310 }
2311
2312 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2313 {
2314     double   eners[2], virs[2], enersum, virsum, y0, f, g, h;
2315     double   r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2316     double   invscale, invscale2, invscale3;
2317     int      ri0, ri1, ri, i, offstart, offset;
2318     real     scale, *vdwtab, tabfactor, tmp;
2319
2320     fr->enershiftsix    = 0;
2321     fr->enershifttwelve = 0;
2322     fr->enerdiffsix     = 0;
2323     fr->enerdifftwelve  = 0;
2324     fr->virdiffsix      = 0;
2325     fr->virdifftwelve   = 0;
2326
2327     if (eDispCorr != edispcNO)
2328     {
2329         for (i = 0; i < 2; i++)
2330         {
2331             eners[i] = 0;
2332             virs[i]  = 0;
2333         }
2334         if ((fr->vdw_modifier == eintmodPOTSHIFT) ||
2335             (fr->vdw_modifier == eintmodPOTSWITCH) ||
2336             (fr->vdw_modifier == eintmodFORCESWITCH) ||
2337             (fr->vdwtype == evdwSHIFT) ||
2338             (fr->vdwtype == evdwSWITCH))
2339         {
2340             if (((fr->vdw_modifier == eintmodPOTSWITCH) ||
2341                  (fr->vdw_modifier == eintmodFORCESWITCH) ||
2342                  (fr->vdwtype == evdwSWITCH)) && fr->rvdw_switch == 0)
2343             {
2344                 gmx_fatal(FARGS,
2345                           "With dispersion correction rvdw-switch can not be zero "
2346                           "for vdw-type = %s", evdw_names[fr->vdwtype]);
2347             }
2348
2349             scale  = fr->nblists[0].table_vdw.scale;
2350             vdwtab = fr->nblists[0].table_vdw.data;
2351
2352             /* Round the cut-offs to exact table values for precision */
2353             ri0  = floor(fr->rvdw_switch*scale);
2354             ri1  = ceil(fr->rvdw*scale);
2355
2356             /* The code below has some support for handling force-switching, i.e.
2357              * when the force (instead of potential) is switched over a limited
2358              * region. This leads to a constant shift in the potential inside the
2359              * switching region, which we can handle by adding a constant energy
2360              * term in the force-switch case just like when we do potential-shift.
2361              *
2362              * For now this is not enabled, but to keep the functionality in the
2363              * code we check separately for switch and shift. When we do force-switch
2364              * the shifting point is rvdw_switch, while it is the cutoff when we
2365              * have a classical potential-shift.
2366              *
2367              * For a pure potential-shift the potential has a constant shift
2368              * all the way out to the cutoff, and that is it. For other forms
2369              * we need to calculate the constant shift up to the point where we
2370              * start modifying the potential.
2371              */
2372             ri0  = (fr->vdw_modifier == eintmodPOTSHIFT) ? ri1 : ri0;
2373
2374             r0   = ri0/scale;
2375             r1   = ri1/scale;
2376             rc3  = r0*r0*r0;
2377             rc9  = rc3*rc3*rc3;
2378
2379             if ((fr->vdw_modifier == eintmodFORCESWITCH) ||
2380                 (fr->vdwtype == evdwSHIFT))
2381             {
2382                 /* Determine the constant energy shift below rvdw_switch.
2383                  * Table has a scale factor since we have scaled it down to compensate
2384                  * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2385                  */
2386                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2387                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2388             }
2389             else if (fr->vdw_modifier == eintmodPOTSHIFT)
2390             {
2391                 fr->enershiftsix    = (real)(-1.0/(rc3*rc3));
2392                 fr->enershifttwelve = (real)( 1.0/(rc9*rc3));
2393             }
2394
2395             /* Add the constant part from 0 to rvdw_switch.
2396              * This integration from 0 to rvdw_switch overcounts the number
2397              * of interactions by 1, as it also counts the self interaction.
2398              * We will correct for this later.
2399              */
2400             eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2401             eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2402
2403             /* Calculate the contribution in the range [r0,r1] where we
2404              * modify the potential. For a pure potential-shift modifier we will
2405              * have ri0==ri1, and there will not be any contribution here.
2406              */
2407             for (i = 0; i < 2; i++)
2408             {
2409                 enersum = 0;
2410                 virsum  = 0;
2411                 integrate_table(vdwtab, scale, (i == 0 ? 0 : 4), ri0, ri1, &enersum, &virsum);
2412                 eners[i] -= enersum;
2413                 virs[i]  -= virsum;
2414             }
2415
2416             /* Alright: Above we compensated by REMOVING the parts outside r0
2417              * corresponding to the ideal VdW 1/r6 and /r12 potentials.
2418              *
2419              * Regardless of whether r0 is the point where we start switching,
2420              * or the cutoff where we calculated the constant shift, we include
2421              * all the parts we are missing out to infinity from r0 by
2422              * calculating the analytical dispersion correction.
2423              */
2424             eners[0] += -4.0*M_PI/(3.0*rc3);
2425             eners[1] +=  4.0*M_PI/(9.0*rc9);
2426             virs[0]  +=  8.0*M_PI/rc3;
2427             virs[1]  += -16.0*M_PI/(3.0*rc9);
2428         }
2429         else if (fr->vdwtype == evdwCUT ||
2430                  EVDW_PME(fr->vdwtype) ||
2431                  fr->vdwtype == evdwUSER)
2432         {
2433             if (fr->vdwtype == evdwUSER && fplog)
2434             {
2435                 fprintf(fplog,
2436                         "WARNING: using dispersion correction with user tables\n");
2437             }
2438
2439             /* Note that with LJ-PME, the dispersion correction is multiplied
2440              * by the difference between the actual C6 and the value of C6
2441              * that would produce the combination rule.
2442              * This means the normal energy and virial difference formulas
2443              * can be used here.
2444              */
2445
2446             rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2447             rc9  = rc3*rc3*rc3;
2448             /* Contribution beyond the cut-off */
2449             eners[0] += -4.0*M_PI/(3.0*rc3);
2450             eners[1] +=  4.0*M_PI/(9.0*rc9);
2451             if (fr->vdw_modifier == eintmodPOTSHIFT)
2452             {
2453                 /* Contribution within the cut-off */
2454                 eners[0] += -4.0*M_PI/(3.0*rc3);
2455                 eners[1] +=  4.0*M_PI/(3.0*rc9);
2456             }
2457             /* Contribution beyond the cut-off */
2458             virs[0]  +=  8.0*M_PI/rc3;
2459             virs[1]  += -16.0*M_PI/(3.0*rc9);
2460         }
2461         else
2462         {
2463             gmx_fatal(FARGS,
2464                       "Dispersion correction is not implemented for vdw-type = %s",
2465                       evdw_names[fr->vdwtype]);
2466         }
2467
2468         /* When we deprecate the group kernels the code below can go too */
2469         if (fr->vdwtype == evdwPME && fr->cutoff_scheme == ecutsGROUP)
2470         {
2471             /* Calculate self-interaction coefficient (assuming that
2472              * the reciprocal-space contribution is constant in the
2473              * region that contributes to the self-interaction).
2474              */
2475             fr->enershiftsix = pow(fr->ewaldcoeff_lj, 6) / 6.0;
2476
2477             eners[0] += -pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3)/3.0;
2478             virs[0]  +=  pow(sqrt(M_PI)*fr->ewaldcoeff_lj, 3);
2479         }
2480
2481         fr->enerdiffsix    = eners[0];
2482         fr->enerdifftwelve = eners[1];
2483         /* The 0.5 is due to the Gromacs definition of the virial */
2484         fr->virdiffsix     = 0.5*virs[0];
2485         fr->virdifftwelve  = 0.5*virs[1];
2486     }
2487 }
2488
2489 void calc_dispcorr(FILE *fplog, t_inputrec *ir, t_forcerec *fr,
2490                    gmx_int64_t step, int natoms,
2491                    matrix box, real lambda, tensor pres, tensor virial,
2492                    real *prescorr, real *enercorr, real *dvdlcorr)
2493 {
2494     gmx_bool bCorrAll, bCorrPres;
2495     real     dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2496     int      m;
2497
2498     *prescorr = 0;
2499     *enercorr = 0;
2500     *dvdlcorr = 0;
2501
2502     clear_mat(virial);
2503     clear_mat(pres);
2504
2505     if (ir->eDispCorr != edispcNO)
2506     {
2507         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2508                      ir->eDispCorr == edispcAllEnerPres);
2509         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2510                      ir->eDispCorr == edispcAllEnerPres);
2511
2512         invvol = 1/det(box);
2513         if (fr->n_tpi)
2514         {
2515             /* Only correct for the interactions with the inserted molecule */
2516             dens   = (natoms - fr->n_tpi)*invvol;
2517             ninter = fr->n_tpi;
2518         }
2519         else
2520         {
2521             dens   = natoms*invvol;
2522             ninter = 0.5*natoms;
2523         }
2524
2525         if (ir->efep == efepNO)
2526         {
2527             avcsix    = fr->avcsix[0];
2528             avctwelve = fr->avctwelve[0];
2529         }
2530         else
2531         {
2532             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2533             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2534         }
2535
2536         enerdiff   = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2537         *enercorr += avcsix*enerdiff;
2538         dvdlambda  = 0.0;
2539         if (ir->efep != efepNO)
2540         {
2541             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2542         }
2543         if (bCorrAll)
2544         {
2545             enerdiff   = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2546             *enercorr += avctwelve*enerdiff;
2547             if (fr->efep != efepNO)
2548             {
2549                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2550             }
2551         }
2552
2553         if (bCorrPres)
2554         {
2555             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2556             if (ir->eDispCorr == edispcAllEnerPres)
2557             {
2558                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2559             }
2560             /* The factor 2 is because of the Gromacs virial definition */
2561             spres = -2.0*invvol*svir*PRESFAC;
2562
2563             for (m = 0; m < DIM; m++)
2564             {
2565                 virial[m][m] += svir;
2566                 pres[m][m]   += spres;
2567             }
2568             *prescorr += spres;
2569         }
2570
2571         /* Can't currently control when it prints, for now, just print when degugging */
2572         if (debug)
2573         {
2574             if (bCorrAll)
2575             {
2576                 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2577                         avcsix, avctwelve);
2578             }
2579             if (bCorrPres)
2580             {
2581                 fprintf(debug,
2582                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2583                         *enercorr, spres, svir);
2584             }
2585             else
2586             {
2587                 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2588             }
2589         }
2590
2591         if (fr->bSepDVDL && do_per_step(step, ir->nstlog))
2592         {
2593             gmx_print_sepdvdl(fplog, "Dispersion correction", *enercorr, dvdlambda);
2594         }
2595         if (fr->efep != efepNO)
2596         {
2597             *dvdlcorr += dvdlambda;
2598         }
2599     }
2600 }
2601
2602 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2603                   t_graph *graph, rvec x[])
2604 {
2605     if (fplog)
2606     {
2607         fprintf(fplog, "Removing pbc first time\n");
2608     }
2609     calc_shifts(box, fr->shift_vec);
2610     if (graph)
2611     {
2612         mk_mshift(fplog, graph, fr->ePBC, box, x);
2613         if (gmx_debug_at)
2614         {
2615             p_graph(debug, "do_pbc_first 1", graph);
2616         }
2617         shift_self(graph, box, x);
2618         /* By doing an extra mk_mshift the molecules that are broken
2619          * because they were e.g. imported from another software
2620          * will be made whole again. Such are the healing powers
2621          * of GROMACS.
2622          */
2623         mk_mshift(fplog, graph, fr->ePBC, box, x);
2624         if (gmx_debug_at)
2625         {
2626             p_graph(debug, "do_pbc_first 2", graph);
2627         }
2628     }
2629     if (fplog)
2630     {
2631         fprintf(fplog, "Done rmpbc\n");
2632     }
2633 }
2634
2635 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2636                             gmx_mtop_t *mtop, rvec x[],
2637                             gmx_bool bFirst)
2638 {
2639     t_graph        *graph;
2640     int             mb, as, mol;
2641     gmx_molblock_t *molb;
2642
2643     if (bFirst && fplog)
2644     {
2645         fprintf(fplog, "Removing pbc first time\n");
2646     }
2647
2648     snew(graph, 1);
2649     as = 0;
2650     for (mb = 0; mb < mtop->nmolblock; mb++)
2651     {
2652         molb = &mtop->molblock[mb];
2653         if (molb->natoms_mol == 1 ||
2654             (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2655         {
2656             /* Just one atom or charge group in the molecule, no PBC required */
2657             as += molb->nmol*molb->natoms_mol;
2658         }
2659         else
2660         {
2661             /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2662             mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2663                            0, molb->natoms_mol, FALSE, FALSE, graph);
2664
2665             for (mol = 0; mol < molb->nmol; mol++)
2666             {
2667                 mk_mshift(fplog, graph, ePBC, box, x+as);
2668
2669                 shift_self(graph, box, x+as);
2670                 /* The molecule is whole now.
2671                  * We don't need the second mk_mshift call as in do_pbc_first,
2672                  * since we no longer need this graph.
2673                  */
2674
2675                 as += molb->natoms_mol;
2676             }
2677             done_graph(graph);
2678         }
2679     }
2680     sfree(graph);
2681 }
2682
2683 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2684                        gmx_mtop_t *mtop, rvec x[])
2685 {
2686     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2687 }
2688
2689 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2690                  gmx_mtop_t *mtop, rvec x[])
2691 {
2692     low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2693 }
2694
2695 void finish_run(FILE *fplog, t_commrec *cr,
2696                 t_inputrec *inputrec,
2697                 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2698                 gmx_walltime_accounting_t walltime_accounting,
2699                 wallclock_gpu_t *gputimes,
2700                 gmx_bool bWriteStat)
2701 {
2702     int     i, j;
2703     t_nrnb *nrnb_tot = NULL;
2704     real    delta_t;
2705     double  nbfs, mflop;
2706     double  elapsed_time,
2707             elapsed_time_over_all_ranks,
2708             elapsed_time_over_all_threads,
2709             elapsed_time_over_all_threads_over_all_ranks;
2710     wallcycle_sum(cr, wcycle);
2711
2712     if (cr->nnodes > 1)
2713     {
2714         snew(nrnb_tot, 1);
2715 #ifdef GMX_MPI
2716         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2717                       cr->mpi_comm_mysim);
2718 #endif
2719     }
2720     else
2721     {
2722         nrnb_tot = nrnb;
2723     }
2724
2725     elapsed_time                                 = walltime_accounting_get_elapsed_time(walltime_accounting);
2726     elapsed_time_over_all_ranks                  = elapsed_time;
2727     elapsed_time_over_all_threads                = walltime_accounting_get_elapsed_time_over_all_threads(walltime_accounting);
2728     elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
2729 #ifdef GMX_MPI
2730     if (cr->nnodes > 1)
2731     {
2732         /* reduce elapsed_time over all MPI ranks in the current simulation */
2733         MPI_Allreduce(&elapsed_time,
2734                       &elapsed_time_over_all_ranks,
2735                       1, MPI_DOUBLE, MPI_SUM,
2736                       cr->mpi_comm_mysim);
2737         elapsed_time_over_all_ranks /= cr->nnodes;
2738         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
2739          * current simulation. */
2740         MPI_Allreduce(&elapsed_time_over_all_threads,
2741                       &elapsed_time_over_all_threads_over_all_ranks,
2742                       1, MPI_DOUBLE, MPI_SUM,
2743                       cr->mpi_comm_mysim);
2744     }
2745 #endif
2746
2747     if (SIMMASTER(cr))
2748     {
2749         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2750     }
2751     if (cr->nnodes > 1)
2752     {
2753         sfree(nrnb_tot);
2754     }
2755
2756     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2757     {
2758         print_dd_statistics(cr, inputrec, fplog);
2759     }
2760
2761     if (SIMMASTER(cr))
2762     {
2763         wallcycle_print(fplog, cr->nnodes, cr->npmenodes,
2764                         elapsed_time_over_all_ranks,
2765                         wcycle, gputimes);
2766
2767         if (EI_DYNAMICS(inputrec->eI))
2768         {
2769             delta_t = inputrec->delta_t;
2770         }
2771         else
2772         {
2773             delta_t = 0;
2774         }
2775
2776         if (fplog)
2777         {
2778             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
2779                        elapsed_time_over_all_ranks,
2780                        walltime_accounting_get_nsteps_done(walltime_accounting),
2781                        delta_t, nbfs, mflop);
2782         }
2783         if (bWriteStat)
2784         {
2785             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
2786                        elapsed_time_over_all_ranks,
2787                        walltime_accounting_get_nsteps_done(walltime_accounting),
2788                        delta_t, nbfs, mflop);
2789         }
2790     }
2791 }
2792
2793 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2794 {
2795     /* this function works, but could probably use a logic rewrite to keep all the different
2796        types of efep straight. */
2797
2798     int       i;
2799     t_lambda *fep = ir->fepvals;
2800
2801     if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2802     {
2803         for (i = 0; i < efptNR; i++)
2804         {
2805             lambda[i] = 0.0;
2806             if (lam0)
2807             {
2808                 lam0[i] = 0.0;
2809             }
2810         }
2811         return;
2812     }
2813     else
2814     {
2815         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2816                                              if checkpoint is set -- a kludge is in for now
2817                                              to prevent this.*/
2818         for (i = 0; i < efptNR; i++)
2819         {
2820             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2821             if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2822             {
2823                 lambda[i] = fep->init_lambda;
2824                 if (lam0)
2825                 {
2826                     lam0[i] = lambda[i];
2827                 }
2828             }
2829             else
2830             {
2831                 lambda[i] = fep->all_lambda[i][*fep_state];
2832                 if (lam0)
2833                 {
2834                     lam0[i] = lambda[i];
2835                 }
2836             }
2837         }
2838         if (ir->bSimTemp)
2839         {
2840             /* need to rescale control temperatures to match current state */
2841             for (i = 0; i < ir->opts.ngtc; i++)
2842             {
2843                 if (ir->opts.ref_t[i] > 0)
2844                 {
2845                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2846                 }
2847             }
2848         }
2849     }
2850
2851     /* Send to the log the information on the current lambdas */
2852     if (fplog != NULL)
2853     {
2854         fprintf(fplog, "Initial vector of lambda components:[ ");
2855         for (i = 0; i < efptNR; i++)
2856         {
2857             fprintf(fplog, "%10.4f ", lambda[i]);
2858         }
2859         fprintf(fplog, "]\n");
2860     }
2861     return;
2862 }
2863
2864
2865 void init_md(FILE *fplog,
2866              t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2867              double *t, double *t0,
2868              real *lambda, int *fep_state, double *lam0,
2869              t_nrnb *nrnb, gmx_mtop_t *mtop,
2870              gmx_update_t *upd,
2871              int nfile, const t_filenm fnm[],
2872              gmx_mdoutf_t *outf, t_mdebin **mdebin,
2873              tensor force_vir, tensor shake_vir, rvec mu_tot,
2874              gmx_bool *bSimAnn, t_vcm **vcm, unsigned long Flags,
2875              gmx_wallcycle_t wcycle)
2876 {
2877     int  i, j, n;
2878     real tmpt, mod;
2879
2880     /* Initial values */
2881     *t = *t0       = ir->init_t;
2882
2883     *bSimAnn = FALSE;
2884     for (i = 0; i < ir->opts.ngtc; i++)
2885     {
2886         /* set bSimAnn if any group is being annealed */
2887         if (ir->opts.annealing[i] != eannNO)
2888         {
2889             *bSimAnn = TRUE;
2890         }
2891     }
2892     if (*bSimAnn)
2893     {
2894         update_annealing_target_temp(&(ir->opts), ir->init_t);
2895     }
2896
2897     /* Initialize lambda variables */
2898     initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2899
2900     if (upd)
2901     {
2902         *upd = init_update(ir);
2903     }
2904
2905
2906     if (vcm != NULL)
2907     {
2908         *vcm = init_vcm(fplog, &mtop->groups, ir);
2909     }
2910
2911     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2912     {
2913         if (ir->etc == etcBERENDSEN)
2914         {
2915             please_cite(fplog, "Berendsen84a");
2916         }
2917         if (ir->etc == etcVRESCALE)
2918         {
2919             please_cite(fplog, "Bussi2007a");
2920         }
2921         if (ir->eI == eiSD1)
2922         {
2923             please_cite(fplog, "Goga2012");
2924         }
2925     }
2926
2927     init_nrnb(nrnb);
2928
2929     if (nfile != -1)
2930     {
2931         *outf = init_mdoutf(fplog, nfile, fnm, Flags, cr, ir, mtop, oenv, wcycle);
2932
2933         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : mdoutf_get_fp_ene(*outf),
2934                               mtop, ir, mdoutf_get_fp_dhdl(*outf));
2935     }
2936
2937     if (ir->bAdress)
2938     {
2939         please_cite(fplog, "Fritsch12");
2940         please_cite(fplog, "Junghans10");
2941     }
2942     /* Initiate variables */
2943     clear_mat(force_vir);
2944     clear_mat(shake_vir);
2945     clear_rvec(mu_tot);
2946
2947     debug_gmx();
2948 }