src/gromacs/mdlib/sim_util.c

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  *          GROningen MAchine for Chemical Simulations
   9  *
  10  *                        VERSION 3.2.0
  11  * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
  12  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
  13  * Copyright (c) 2001-2004, The GROMACS development team,
  14  * check out http://www.gromacs.org for more information.
  15
  16  * This program is free software; you can redistribute it and/or
  17  * modify it under the terms of the GNU General Public License
  18  * as published by the Free Software Foundation; either version 2
  19  * of the License, or (at your option) any later version.
  20  *
  21  * If you want to redistribute modifications, please consider that
  22  * scientific software is very special. Version control is crucial -
  23  * bugs must be traceable. We will be happy to consider code for
  24  * inclusion in the official distribution, but derived work must not
  25  * be called official GROMACS. Details are found in the README & COPYING
  26  * files - if they are missing, get the official version at www.gromacs.org.
  27  *
  28  * To help us fund GROMACS development, we humbly ask that you cite
  29  * the papers on the package - you can find them in the top README file.
  30  *
  31  * For more info, check our website at http://www.gromacs.org
  32  *
  33  * And Hey:
  34  * GROwing Monsters And Cloning Shrimps
  35  */
  36 #ifdef HAVE_CONFIG_H
  37 #include <config.h>
  38 #endif
  39
  40 #ifdef GMX_CRAY_XT3
  41 #include<catamount/dclock.h>
  42 #endif
  43
  44
  45 #include <stdio.h>
  46 #include <time.h>
  47 #ifdef HAVE_SYS_TIME_H
  48 #include <sys/time.h>
  49 #endif
  50 #include <math.h>
  51 #include "typedefs.h"
  52 #include "string2.h"
  53 #include "gmxfio.h"
  54 #include "smalloc.h"
  55 #include "names.h"
  56 #include "confio.h"
  57 #include "mvdata.h"
  58 #include "txtdump.h"
  59 #include "pbc.h"
  60 #include "chargegroup.h"
  61 #include "vec.h"
  62 #include <time.h>
  63 #include "nrnb.h"
  64 #include "mshift.h"
  65 #include "mdrun.h"
  66 #include "sim_util.h"
  67 #include "update.h"
  68 #include "physics.h"
  69 #include "main.h"
  70 #include "mdatoms.h"
  71 #include "force.h"
  72 #include "bondf.h"
  73 #include "pme.h"
  74 #include "disre.h"
  75 #include "orires.h"
  76 #include "network.h"
  77 #include "calcmu.h"
  78 #include "constr.h"
  79 #include "xvgr.h"
  80 #include "trnio.h"
  81 #include "xtcio.h"
  82 #include "copyrite.h"
  83 #include "pull_rotation.h"
  84 #include "gmx_random.h"
  85 #include "domdec.h"
  86 #include "partdec.h"
  87 #include "gmx_wallcycle.h"
  88 #include "genborn.h"
  89 #include "nbnxn_atomdata.h"
  90 #include "nbnxn_search.h"
  91 #include "nbnxn_kernels/nbnxn_kernel_ref.h"
  92 #include "nbnxn_kernels/nbnxn_kernel_x86_simd128.h"
  93 #include "nbnxn_kernels/nbnxn_kernel_x86_simd256.h"
  94 #include "nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
  95
  96 #ifdef GMX_LIB_MPI
  97 #include <mpi.h>
  98 #endif
  99 #ifdef GMX_THREAD_MPI
 100 #include "tmpi.h"
 101 #endif
 102
 103 #include "adress.h"
 104 #include "qmmm.h"
 105
 106 #include "nbnxn_cuda_data_mgmt.h"
 107 #include "nbnxn_cuda/nbnxn_cuda.h"
 108
 109 #if 0
 110 typedef struct gmx_timeprint {
 111
 112 } t_gmx_timeprint;
 113 #endif
 114
 115 /* Portable version of ctime_r implemented in src/gmxlib/string2.c, but we do not want it declared in public installed headers */
 116 char *
 117 gmx_ctime_r(const time_t *clock,char *buf, int n);
 118
 119
 120 double
 121 gmx_gettime()
 122 {
 123 #ifdef HAVE_GETTIMEOFDAY
 124         struct timeval t;
 125         double seconds;
 126
 127         gettimeofday(&t,NULL);
 128
 129         seconds = (double) t.tv_sec + 1e-6*(double)t.tv_usec;
 130
 131         return seconds;
 132 #else
 133         double  seconds;
 134
 135         seconds = time(NULL);
 136
 137         return seconds;
 138 #endif
 139 }
 140
 141
 142 #define difftime(end,start) ((double)(end)-(double)(start))
 143
 144 void print_time(FILE *out,gmx_runtime_t *runtime,gmx_large_int_t step,
 145                 t_inputrec *ir, t_commrec *cr)
 146 {
 147     time_t finish;
 148     char   timebuf[STRLEN];
 149     double dt;
 150     char buf[48];
 151
 152 #ifndef GMX_THREAD_MPI
 153     if (!PAR(cr))
 154 #endif
 155     {
 156         fprintf(out,"\r");
 157     }
 158     fprintf(out,"step %s",gmx_step_str(step,buf));
 159     if ((step >= ir->nstlist))
 160     {
 161         runtime->last = gmx_gettime();
 162         dt = difftime(runtime->last,runtime->real);
 163         runtime->time_per_step = dt/(step - ir->init_step + 1);
 164
 165         dt = (ir->nsteps + ir->init_step - step)*runtime->time_per_step;
 166
 167         if (ir->nsteps >= 0)
 168         {
 169             if (dt >= 300)
 170             {
 171                 finish = (time_t) (runtime->last + dt);
 172                 gmx_ctime_r(&finish,timebuf,STRLEN);
 173                 sprintf(buf,"%s",timebuf);
 174                 buf[strlen(buf)-1]='\0';
 175                 fprintf(out,", will finish %s",buf);
 176             }
 177             else
 178                 fprintf(out,", remaining runtime: %5d s          ",(int)dt);
 179         }
 180         else
 181         {
 182             fprintf(out," performance: %.1f ns/day    ",
 183                     ir->delta_t/1000*24*60*60/runtime->time_per_step);
 184         }
 185     }
 186 #ifndef GMX_THREAD_MPI
 187     if (PAR(cr))
 188     {
 189         fprintf(out,"\n");
 190     }
 191 #endif
 192
 193     fflush(out);
 194 }
 195
 196 #ifdef NO_CLOCK
 197 #define clock() -1
 198 #endif
 199
 200 static double set_proctime(gmx_runtime_t *runtime)
 201 {
 202     double diff;
 203 #ifdef GMX_CRAY_XT3
 204     double prev;
 205
 206     prev = runtime->proc;
 207     runtime->proc = dclock();
 208
 209     diff = runtime->proc - prev;
 210 #else
 211     clock_t prev;
 212
 213     prev = runtime->proc;
 214     runtime->proc = clock();
 215
 216     diff = (double)(runtime->proc - prev)/(double)CLOCKS_PER_SEC;
 217 #endif
 218     if (diff < 0)
 219     {
 220         /* The counter has probably looped, ignore this data */
 221         diff = 0;
 222     }
 223
 224     return diff;
 225 }
 226
 227 void runtime_start(gmx_runtime_t *runtime)
 228 {
 229     runtime->real = gmx_gettime();
 230     runtime->proc          = 0;
 231     set_proctime(runtime);
 232     runtime->realtime      = 0;
 233     runtime->proctime      = 0;
 234     runtime->last          = 0;
 235     runtime->time_per_step = 0;
 236 }
 237
 238 void runtime_end(gmx_runtime_t *runtime)
 239 {
 240     double now;
 241
 242     now = gmx_gettime();
 243
 244     runtime->proctime += set_proctime(runtime);
 245     runtime->realtime  = now - runtime->real;
 246     runtime->real      = now;
 247 }
 248
 249 void runtime_upd_proc(gmx_runtime_t *runtime)
 250 {
 251     runtime->proctime += set_proctime(runtime);
 252 }
 253
 254 void print_date_and_time(FILE *fplog,int nodeid,const char *title,
 255                          const gmx_runtime_t *runtime)
 256 {
 257     int i;
 258     char timebuf[STRLEN];
 259     char time_string[STRLEN];
 260     time_t tmptime;
 261
 262     if (fplog)
 263     {
 264         if (runtime != NULL)
 265         {
 266             tmptime = (time_t) runtime->real;
 267             gmx_ctime_r(&tmptime,timebuf,STRLEN);
 268         }
 269         else
 270         {
 271             tmptime = (time_t) gmx_gettime();
 272             gmx_ctime_r(&tmptime,timebuf,STRLEN);
 273         }
 274         for(i=0; timebuf[i]>=' '; i++)
 275         {
 276             time_string[i]=timebuf[i];
 277         }
 278         time_string[i]='\0';
 279
 280         fprintf(fplog,"%s on node %d %s\n",title,nodeid,time_string);
 281     }
 282 }
 283
 284 static void sum_forces(int start,int end,rvec f[],rvec flr[])
 285 {
 286   int i;
 287
 288   if (gmx_debug_at) {
 289     pr_rvecs(debug,0,"fsr",f+start,end-start);
 290     pr_rvecs(debug,0,"flr",flr+start,end-start);
 291   }
 292   for(i=start; (i<end); i++)
 293     rvec_inc(f[i],flr[i]);
 294 }
 295
 296 /*
 297  * calc_f_el calculates forces due to an electric field.
 298  *
 299  * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
 300  *
 301  * Et[] contains the parameters for the time dependent
 302  * part of the field (not yet used).
 303  * Ex[] contains the parameters for
 304  * the spatial dependent part of the field. You can have cool periodic
 305  * fields in principle, but only a constant field is supported
 306  * now.
 307  * The function should return the energy due to the electric field
 308  * (if any) but for now returns 0.
 309  *
 310  * WARNING:
 311  * There can be problems with the virial.
 312  * Since the field is not self-consistent this is unavoidable.
 313  * For neutral molecules the virial is correct within this approximation.
 314  * For neutral systems with many charged molecules the error is small.
 315  * But for systems with a net charge or a few charged molecules
 316  * the error can be significant when the field is high.
 317  * Solution: implement a self-consitent electric field into PME.
 318  */
 319 static void calc_f_el(FILE *fp,int  start,int homenr,
 320                       real charge[],rvec x[],rvec f[],
 321                       t_cosines Ex[],t_cosines Et[],double t)
 322 {
 323     rvec Ext;
 324     real t0;
 325     int  i,m;
 326
 327     for(m=0; (m<DIM); m++)
 328     {
 329         if (Et[m].n > 0)
 330         {
 331             if (Et[m].n == 3)
 332             {
 333                 t0 = Et[m].a[1];
 334                 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
 335             }
 336             else
 337             {
 338                 Ext[m] = cos(Et[m].a[0]*t);
 339             }
 340         }
 341         else
 342         {
 343             Ext[m] = 1.0;
 344         }
 345         if (Ex[m].n > 0)
 346         {
 347             /* Convert the field strength from V/nm to MD-units */
 348             Ext[m] *= Ex[m].a[0]*FIELDFAC;
 349             for(i=start; (i<start+homenr); i++)
 350                 f[i][m] += charge[i]*Ext[m];
 351         }
 352         else
 353         {
 354             Ext[m] = 0;
 355         }
 356     }
 357     if (fp != NULL)
 358     {
 359         fprintf(fp,"%10g  %10g  %10g  %10g #FIELD\n",t,
 360                 Ext[XX]/FIELDFAC,Ext[YY]/FIELDFAC,Ext[ZZ]/FIELDFAC);
 361     }
 362 }
 363
 364 static void calc_virial(FILE *fplog,int start,int homenr,rvec x[],rvec f[],
 365                         tensor vir_part,t_graph *graph,matrix box,
 366                         t_nrnb *nrnb,const t_forcerec *fr,int ePBC)
 367 {
 368   int i,j;
 369   tensor virtest;
 370
 371   /* The short-range virial from surrounding boxes */
 372   clear_mat(vir_part);
 373   calc_vir(fplog,SHIFTS,fr->shift_vec,fr->fshift,vir_part,ePBC==epbcSCREW,box);
 374   inc_nrnb(nrnb,eNR_VIRIAL,SHIFTS);
 375
 376   /* Calculate partial virial, for local atoms only, based on short range.
 377    * Total virial is computed in global_stat, called from do_md
 378    */
 379   f_calc_vir(fplog,start,start+homenr,x,f,vir_part,graph,box);
 380   inc_nrnb(nrnb,eNR_VIRIAL,homenr);
 381
 382   /* Add position restraint contribution */
 383   for(i=0; i<DIM; i++) {
 384     vir_part[i][i] += fr->vir_diag_posres[i];
 385   }
 386
 387   /* Add wall contribution */
 388   for(i=0; i<DIM; i++) {
 389     vir_part[i][ZZ] += fr->vir_wall_z[i];
 390   }
 391
 392   if (debug)
 393     pr_rvecs(debug,0,"vir_part",vir_part,DIM);
 394 }
 395
 396 static void posres_wrapper(FILE *fplog,
 397                            int flags,
 398                            gmx_bool bSepDVDL,
 399                            t_inputrec *ir,
 400                            t_nrnb *nrnb,
 401                            gmx_localtop_t *top,
 402                            matrix box,rvec x[],
 403                            rvec f[],
 404                            gmx_enerdata_t *enerd,
 405                            real *lambda,
 406                            t_forcerec *fr)
 407 {
 408     t_pbc pbc;
 409     real  v,dvdl;
 410     int   i;
 411
 412     /* Position restraints always require full pbc */
 413     set_pbc(&pbc,ir->ePBC,box);
 414     dvdl = 0;
 415     v = posres(top->idef.il[F_POSRES].nr,top->idef.il[F_POSRES].iatoms,
 416                top->idef.iparams_posres,
 417                (const rvec*)x,fr->f_novirsum,fr->vir_diag_posres,
 418                ir->ePBC==epbcNONE ? NULL : &pbc,
 419                lambda[efptRESTRAINT],&dvdl,
 420                fr->rc_scaling,fr->ePBC,fr->posres_com,fr->posres_comB);
 421     if (bSepDVDL)
 422     {
 423         fprintf(fplog,sepdvdlformat,
 424                 interaction_function[F_POSRES].longname,v,dvdl);
 425     }
 426     enerd->term[F_POSRES] += v;
 427     /* If just the force constant changes, the FEP term is linear,
 428      * but if k changes, it is not.
 429      */
 430     enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
 431     inc_nrnb(nrnb,eNR_POSRES,top->idef.il[F_POSRES].nr/2);
 432
 433     if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
 434     {
 435         for(i=0; i<enerd->n_lambda; i++)
 436         {
 437             real dvdl_dum,lambda_dum;
 438
 439             lambda_dum = (i==0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
 440             v = posres(top->idef.il[F_POSRES].nr,top->idef.il[F_POSRES].iatoms,
 441                        top->idef.iparams_posres,
 442                        (const rvec*)x,NULL,NULL,
 443                        ir->ePBC==epbcNONE ? NULL : &pbc,lambda_dum,&dvdl,
 444                        fr->rc_scaling,fr->ePBC,fr->posres_com,fr->posres_comB);
 445             enerd->enerpart_lambda[i] += v;
 446         }
 447     }
 448 }
 449
 450 static void pull_potential_wrapper(FILE *fplog,
 451                                    gmx_bool bSepDVDL,
 452                                    t_commrec *cr,
 453                                    t_inputrec *ir,
 454                                    matrix box,rvec x[],
 455                                    rvec f[],
 456                                    tensor vir_force,
 457                                    t_mdatoms *mdatoms,
 458                                    gmx_enerdata_t *enerd,
 459                                    real *lambda,
 460                                    double t)
 461 {
 462     t_pbc  pbc;
 463     real   dvdl;
 464
 465     /* Calculate the center of mass forces, this requires communication,
 466      * which is why pull_potential is called close to other communication.
 467      * The virial contribution is calculated directly,
 468      * which is why we call pull_potential after calc_virial.
 469      */
 470     set_pbc(&pbc,ir->ePBC,box);
 471     dvdl = 0;
 472     enerd->term[F_COM_PULL] +=
 473         pull_potential(ir->ePull,ir->pull,mdatoms,&pbc,
 474                        cr,t,lambda[efptRESTRAINT],x,f,vir_force,&dvdl);
 475     if (bSepDVDL)
 476     {
 477         fprintf(fplog,sepdvdlformat,"Com pull",enerd->term[F_COM_PULL],dvdl);
 478     }
 479     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 480 }
 481
 482 static void pme_receive_force_ener(FILE *fplog,
 483                                    gmx_bool bSepDVDL,
 484                                    t_commrec *cr,
 485                                    gmx_wallcycle_t wcycle,
 486                                    gmx_enerdata_t *enerd,
 487                                    t_forcerec *fr)
 488 {
 489     real   e,v,dvdl;
 490     float  cycles_ppdpme,cycles_seppme;
 491
 492     cycles_ppdpme = wallcycle_stop(wcycle,ewcPPDURINGPME);
 493     dd_cycles_add(cr->dd,cycles_ppdpme,ddCyclPPduringPME);
 494
 495     /* In case of node-splitting, the PP nodes receive the long-range
 496      * forces, virial and energy from the PME nodes here.
 497      */
 498     wallcycle_start(wcycle,ewcPP_PMEWAITRECVF);
 499     dvdl = 0;
 500     gmx_pme_receive_f(cr,fr->f_novirsum,fr->vir_el_recip,&e,&dvdl,
 501                       &cycles_seppme);
 502     if (bSepDVDL)
 503     {
 504         fprintf(fplog,sepdvdlformat,"PME mesh",e,dvdl);
 505     }
 506     enerd->term[F_COUL_RECIP] += e;
 507     enerd->dvdl_lin[efptCOUL] += dvdl;
 508     if (wcycle)
 509     {
 510         dd_cycles_add(cr->dd,cycles_seppme,ddCyclPME);
 511     }
 512     wallcycle_stop(wcycle,ewcPP_PMEWAITRECVF);
 513 }
 514
 515 static void print_large_forces(FILE *fp,t_mdatoms *md,t_commrec *cr,
 516                                gmx_large_int_t step,real pforce,rvec *x,rvec *f)
 517 {
 518   int  i;
 519   real pf2,fn2;
 520   char buf[STEPSTRSIZE];
 521
 522   pf2 = sqr(pforce);
 523   for(i=md->start; i<md->start+md->homenr; i++) {
 524     fn2 = norm2(f[i]);
 525     /* We also catch NAN, if the compiler does not optimize this away. */
 526     if (fn2 >= pf2 || fn2 != fn2) {
 527       fprintf(fp,"step %s  atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 528               gmx_step_str(step,buf),
 529               ddglatnr(cr->dd,i),x[i][XX],x[i][YY],x[i][ZZ],sqrt(fn2));
 530     }
 531   }
 532 }
 533
 534 static void post_process_forces(FILE *fplog,
 535                                 t_commrec *cr,
 536                                 gmx_large_int_t step,
 537                                 t_nrnb *nrnb,gmx_wallcycle_t wcycle,
 538                                 gmx_localtop_t *top,
 539                                 matrix box,rvec x[],
 540                                 rvec f[],
 541                                 tensor vir_force,
 542                                 t_mdatoms *mdatoms,
 543                                 t_graph *graph,
 544                                 t_forcerec *fr,gmx_vsite_t *vsite,
 545                                 int flags)
 546 {
 547     if (fr->bF_NoVirSum)
 548     {
 549         if (vsite)
 550         {
 551             /* Spread the mesh force on virtual sites to the other particles...
 552              * This is parallellized. MPI communication is performed
 553              * if the constructing atoms aren't local.
 554              */
 555             wallcycle_start(wcycle,ewcVSITESPREAD);
 556             spread_vsite_f(fplog,vsite,x,fr->f_novirsum,NULL,
 557                            (flags & GMX_FORCE_VIRIAL),fr->vir_el_recip,
 558                            nrnb,
 559                            &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
 560             wallcycle_stop(wcycle,ewcVSITESPREAD);
 561         }
 562         if (flags & GMX_FORCE_VIRIAL)
 563         {
 564             /* Now add the forces, this is local */
 565             if (fr->bDomDec)
 566             {
 567                 sum_forces(0,fr->f_novirsum_n,f,fr->f_novirsum);
 568             }
 569             else
 570             {
 571                 sum_forces(mdatoms->start,mdatoms->start+mdatoms->homenr,
 572                            f,fr->f_novirsum);
 573             }
 574             if (EEL_FULL(fr->eeltype))
 575             {
 576                 /* Add the mesh contribution to the virial */
 577                 m_add(vir_force,fr->vir_el_recip,vir_force);
 578             }
 579             if (debug)
 580             {
 581                 pr_rvecs(debug,0,"vir_force",vir_force,DIM);
 582             }
 583         }
 584     }
 585
 586     if (fr->print_force >= 0)
 587     {
 588         print_large_forces(stderr,mdatoms,cr,step,fr->print_force,x,f);
 589     }
 590 }
 591
 592 static void do_nb_verlet(t_forcerec *fr,
 593                          interaction_const_t *ic,
 594                          gmx_enerdata_t *enerd,
 595                          int flags, int ilocality,
 596                          int clearF,
 597                          t_nrnb *nrnb,
 598                          gmx_wallcycle_t wcycle)
 599 {
 600     int     nnbl, kernel_type, sh_e;
 601     char    *env;
 602     nonbonded_verlet_group_t  *nbvg;
 603
 604     if (!(flags & GMX_FORCE_NONBONDED))
 605     {
 606         /* skip non-bonded calculation */
 607         return;
 608     }
 609
 610     nbvg = &fr->nbv->grp[ilocality];
 611
 612     /* CUDA kernel launch overhead is already timed separately */
 613     if (fr->cutoff_scheme != ecutsVERLET)
 614     {
 615         gmx_incons("Invalid cut-off scheme passed!");
 616     }
 617
 618     if (nbvg->kernel_type != nbk8x8x8_CUDA)
 619     {
 620         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 621     }
 622     switch (nbvg->kernel_type)
 623     {
 624         case nbk4x4_PlainC:
 625             nbnxn_kernel_ref(&nbvg->nbl_lists,
 626                              nbvg->nbat, ic,
 627                              fr->shift_vec,
 628                              flags,
 629                              clearF,
 630                              fr->fshift[0],
 631                              enerd->grpp.ener[egCOULSR],
 632                              fr->bBHAM ?
 633                              enerd->grpp.ener[egBHAMSR] :
 634                              enerd->grpp.ener[egLJSR]);
 635             break;
 636
 637         case nbk4xN_X86_SIMD128:
 638             nbnxn_kernel_x86_simd128(&nbvg->nbl_lists,
 639                                      nbvg->nbat, ic,
 640                                      fr->shift_vec,
 641                                      flags,
 642                                      clearF,
 643                                      fr->fshift[0],
 644                                      enerd->grpp.ener[egCOULSR],
 645                                      fr->bBHAM ?
 646                                      enerd->grpp.ener[egBHAMSR] :
 647                                      enerd->grpp.ener[egLJSR]);
 648             break;
 649         case nbk4xN_X86_SIMD256:
 650             nbnxn_kernel_x86_simd256(&nbvg->nbl_lists,
 651                                      nbvg->nbat, ic,
 652                                      fr->shift_vec,
 653                                      flags,
 654                                      clearF,
 655                                      fr->fshift[0],
 656                                      enerd->grpp.ener[egCOULSR],
 657                                      fr->bBHAM ?
 658                                      enerd->grpp.ener[egBHAMSR] :
 659                                      enerd->grpp.ener[egLJSR]);
 660             break;
 661
 662         case nbk8x8x8_CUDA:
 663             nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
 664             break;
 665
 666         case nbk8x8x8_PlainC:
 667             nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
 668                                  nbvg->nbat, ic,
 669                                  fr->shift_vec,
 670                                  flags,
 671                                  clearF,
 672                                  nbvg->nbat->out[0].f,
 673                                  fr->fshift[0],
 674                                  enerd->grpp.ener[egCOULSR],
 675                                  fr->bBHAM ?
 676                                  enerd->grpp.ener[egBHAMSR] :
 677                                  enerd->grpp.ener[egLJSR]);
 678             break;
 679
 680         default:
 681             gmx_incons("Invalid nonbonded kernel type passed!");
 682
 683     }
 684     if (nbvg->kernel_type != nbk8x8x8_CUDA)
 685     {
 686         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 687     }
 688
 689     /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 690     sh_e = ((flags & GMX_FORCE_ENERGY) ? 1 : 0);
 691     inc_nrnb(nrnb,
 692              ((EEL_RF(ic->eeltype) || ic->eeltype == eelCUT) ?
 693               eNR_NBNXN_LJ_RF : eNR_NBNXN_LJ_TAB) + sh_e,
 694              nbvg->nbl_lists.natpair_ljq);
 695     inc_nrnb(nrnb,eNR_NBNXN_LJ+sh_e,nbvg->nbl_lists.natpair_lj);
 696     inc_nrnb(nrnb,
 697              ((EEL_RF(ic->eeltype) || ic->eeltype == eelCUT) ?
 698               eNR_NBNXN_RF : eNR_NBNXN_TAB)+sh_e,
 699              nbvg->nbl_lists.natpair_q);
 700 }
 701
 702 void do_force_cutsVERLET(FILE *fplog,t_commrec *cr,
 703               t_inputrec *inputrec,
 704               gmx_large_int_t step,t_nrnb *nrnb,gmx_wallcycle_t wcycle,
 705               gmx_localtop_t *top,
 706               gmx_mtop_t *mtop,
 707               gmx_groups_t *groups,
 708               matrix box,rvec x[],history_t *hist,
 709               rvec f[],
 710               tensor vir_force,
 711               t_mdatoms *mdatoms,
 712               gmx_enerdata_t *enerd,t_fcdata *fcd,
 713               real *lambda,t_graph *graph,
 714               t_forcerec *fr, interaction_const_t *ic,
 715               gmx_vsite_t *vsite,rvec mu_tot,
 716               double t,FILE *field,gmx_edsam_t ed,
 717               gmx_bool bBornRadii,
 718               int flags)
 719 {
 720     int     cg0,cg1,i,j;
 721     int     start,homenr;
 722     int     nb_kernel_type;
 723     double  mu[2*DIM];
 724     gmx_bool   bSepDVDL,bStateChanged,bNS,bFillGrid,bCalcCGCM,bBS;
 725     gmx_bool   bDoLongRange,bDoForces,bSepLRF,bUseGPU,bUseOrEmulGPU;
 726     gmx_bool   bDiffKernels=FALSE;
 727     matrix  boxs;
 728     rvec    vzero,box_diag;
 729     real    e,v,dvdl;
 730     float  cycles_pme,cycles_force;
 731     nonbonded_verlet_t *nbv;
 732
 733     cycles_force = 0;
 734     nbv = fr->nbv;
 735     nb_kernel_type = fr->nbv->grp[0].kernel_type;
 736
 737     start  = mdatoms->start;
 738     homenr = mdatoms->homenr;
 739
 740     bSepDVDL = (fr->bSepDVDL && do_per_step(step,inputrec->nstlog));
 741
 742     clear_mat(vir_force);
 743
 744     cg0 = 0;
 745     if (DOMAINDECOMP(cr))
 746     {
 747         cg1 = cr->dd->ncg_tot;
 748     }
 749     else
 750     {
 751         cg1 = top->cgs.nr;
 752     }
 753     if (fr->n_tpi > 0)
 754     {
 755         cg1--;
 756     }
 757
 758     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
 759     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll==FALSE);
 760     bFillGrid     = (bNS && bStateChanged);
 761     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 762     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
 763     bDoForces     = (flags & GMX_FORCE_FORCES);
 764     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
 765     bUseGPU       = fr->nbv->bUseGPU;
 766     bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbk8x8x8_PlainC);
 767
 768     if (bStateChanged)
 769     {
 770         update_forcerec(fplog,fr,box);
 771
 772         if (NEED_MUTOT(*inputrec))
 773         {
 774             /* Calculate total (local) dipole moment in a temporary common array.
 775              * This makes it possible to sum them over nodes faster.
 776              */
 777             calc_mu(start,homenr,
 778                     x,mdatoms->chargeA,mdatoms->chargeB,mdatoms->nChargePerturbed,
 779                     mu,mu+DIM);
 780         }
 781     }
 782
 783     if (fr->ePBC != epbcNONE) {
 784         /* Compute shift vectors every step,
 785          * because of pressure coupling or box deformation!
 786          */
 787         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 788             calc_shifts(box,fr->shift_vec);
 789
 790         if (bCalcCGCM) {
 791             put_atoms_in_box_omp(fr->ePBC,box,homenr,x);
 792             inc_nrnb(nrnb,eNR_SHIFTX,homenr);
 793         }
 794         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph) {
 795             unshift_self(graph,box,x);
 796         }
 797     }
 798
 799     nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
 800                                   fr->shift_vec,nbv->grp[0].nbat);
 801
 802 #ifdef GMX_MPI
 803     if (!(cr->duty & DUTY_PME)) {
 804         /* Send particle coordinates to the pme nodes.
 805          * Since this is only implemented for domain decomposition
 806          * and domain decomposition does not use the graph,
 807          * we do not need to worry about shifting.
 808          */
 809
 810         wallcycle_start(wcycle,ewcPP_PMESENDX);
 811
 812         bBS = (inputrec->nwall == 2);
 813         if (bBS) {
 814             copy_mat(box,boxs);
 815             svmul(inputrec->wall_ewald_zfac,boxs[ZZ],boxs[ZZ]);
 816         }
 817
 818         gmx_pme_send_x(cr,bBS ? boxs : box,x,
 819                        mdatoms->nChargePerturbed,lambda[efptCOUL],
 820                        (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),step);
 821
 822         wallcycle_stop(wcycle,ewcPP_PMESENDX);
 823     }
 824 #endif /* GMX_MPI */
 825
 826     /* do gridding for pair search */
 827     if (bNS)
 828     {
 829         if (graph && bStateChanged)
 830         {
 831             /* Calculate intramolecular shift vectors to make molecules whole */
 832             mk_mshift(fplog,graph,fr->ePBC,box,x);
 833         }
 834
 835         clear_rvec(vzero);
 836         box_diag[XX] = box[XX][XX];
 837         box_diag[YY] = box[YY][YY];
 838         box_diag[ZZ] = box[ZZ][ZZ];
 839
 840         wallcycle_start(wcycle,ewcNS);
 841         if (!fr->bDomDec)
 842         {
 843             wallcycle_sub_start(wcycle,ewcsNBS_GRID_LOCAL);
 844             nbnxn_put_on_grid(nbv->nbs,fr->ePBC,box,
 845                               0,vzero,box_diag,
 846                               0,mdatoms->homenr,-1,fr->cginfo,x,
 847                               0,NULL,
 848                               nbv->grp[eintLocal].kernel_type,
 849                               nbv->grp[eintLocal].nbat);
 850             wallcycle_sub_stop(wcycle,ewcsNBS_GRID_LOCAL);
 851         }
 852         else
 853         {
 854             wallcycle_sub_start(wcycle,ewcsNBS_GRID_NONLOCAL);
 855             nbnxn_put_on_grid_nonlocal(nbv->nbs,domdec_zones(cr->dd),
 856                                        fr->cginfo,x,
 857                                        nbv->grp[eintNonlocal].kernel_type,
 858                                        nbv->grp[eintNonlocal].nbat);
 859             wallcycle_sub_stop(wcycle,ewcsNBS_GRID_NONLOCAL);
 860         }
 861
 862         if (nbv->ngrp == 1 ||
 863             nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
 864         {
 865             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat,eatAll,
 866                                 nbv->nbs,mdatoms,fr->cginfo);
 867         }
 868         else
 869         {
 870             nbnxn_atomdata_set(nbv->grp[eintLocal].nbat,eatLocal,
 871                                 nbv->nbs,mdatoms,fr->cginfo);
 872             nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat,eatAll,
 873                                 nbv->nbs,mdatoms,fr->cginfo);
 874         }
 875         wallcycle_stop(wcycle, ewcNS);
 876     }
 877
 878     /* initialize the GPU atom data and copy shift vector */
 879     if (bUseGPU)
 880     {
 881         if (bNS)
 882         {
 883             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 884             nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 885             wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 886         }
 887
 888         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
 889         nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
 890         wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
 891     }
 892
 893     /* do local pair search */
 894     if (bNS)
 895     {
 896         wallcycle_start_nocount(wcycle,ewcNS);
 897         wallcycle_sub_start(wcycle,ewcsNBS_SEARCH_LOCAL);
 898         nbnxn_make_pairlist(nbv->nbs,nbv->grp[eintLocal].nbat,
 899                             &top->excls,
 900                             ic->rlist,
 901                             nbv->min_ci_balanced,
 902                             &nbv->grp[eintLocal].nbl_lists,
 903                             eintLocal,
 904                             nbv->grp[eintLocal].kernel_type,
 905                             nrnb);
 906         wallcycle_sub_stop(wcycle,ewcsNBS_SEARCH_LOCAL);
 907
 908         if (bUseGPU)
 909         {
 910             /* initialize local pair-list on the GPU */
 911             nbnxn_cuda_init_pairlist(nbv->cu_nbv,
 912                                      nbv->grp[eintLocal].nbl_lists.nbl[0],
 913                                      eintLocal);
 914         }
 915         wallcycle_stop(wcycle, ewcNS);
 916     }
 917     else
 918     {
 919         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 920         wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
 921         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs,eatLocal,FALSE,x,
 922                                         nbv->grp[eintLocal].nbat);
 923         wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
 924         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
 925     }
 926
 927     if (bUseGPU)
 928     {
 929         wallcycle_start(wcycle,ewcLAUNCH_GPU_NB);
 930         /* launch local nonbonded F on GPU */
 931         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
 932                      nrnb, wcycle);
 933         wallcycle_stop(wcycle,ewcLAUNCH_GPU_NB);
 934     }
 935
 936     /* Communicate coordinates and sum dipole if necessary +
 937        do non-local pair search */
 938     if (DOMAINDECOMP(cr))
 939     {
 940         bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
 941                         nbv->grp[eintLocal].kernel_type);
 942
 943         if (bDiffKernels)
 944         {
 945             /* With GPU+CPU non-bonded calculations we need to copy
 946              * the local coordinates to the non-local nbat struct
 947              * (in CPU format) as the non-local kernel call also
 948              * calculates the local - non-local interactions.
 949              */
 950             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
 951             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
 952             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs,eatLocal,TRUE,x,
 953                                              nbv->grp[eintNonlocal].nbat);
 954             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
 955             wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
 956         }
 957
 958         if (bNS)
 959         {
 960             wallcycle_start_nocount(wcycle,ewcNS);
 961             wallcycle_sub_start(wcycle,ewcsNBS_SEARCH_NONLOCAL);
 962
 963             if (bDiffKernels)
 964             {
 965                 nbnxn_grid_add_simple(nbv->nbs,nbv->grp[eintNonlocal].nbat);
 966             }
 967
 968             nbnxn_make_pairlist(nbv->nbs,nbv->grp[eintNonlocal].nbat,
 969                                 &top->excls,
 970                                 ic->rlist,
 971                                 nbv->min_ci_balanced,
 972                                 &nbv->grp[eintNonlocal].nbl_lists,
 973                                 eintNonlocal,
 974                                 nbv->grp[eintNonlocal].kernel_type,
 975                                 nrnb);
 976
 977             wallcycle_sub_stop(wcycle,ewcsNBS_SEARCH_NONLOCAL);
 978
 979             if (nbv->grp[eintNonlocal].kernel_type == nbk8x8x8_CUDA)
 980             {
 981                 /* initialize non-local pair-list on the GPU */
 982                 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
 983                                          nbv->grp[eintNonlocal].nbl_lists.nbl[0],
 984                                          eintNonlocal);
 985             }
 986             wallcycle_stop(wcycle,ewcNS);
 987         }
 988         else
 989         {
 990             wallcycle_start(wcycle,ewcMOVEX);
 991             dd_move_x(cr->dd,box,x);
 992
 993             /* When we don't need the total dipole we sum it in global_stat */
 994             if (bStateChanged && NEED_MUTOT(*inputrec))
 995             {
 996                 gmx_sumd(2*DIM,mu,cr);
 997             }
 998             wallcycle_stop(wcycle,ewcMOVEX);
 999
1000             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1001             wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1002             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs,eatNonlocal,FALSE,x,
1003                                             nbv->grp[eintNonlocal].nbat);
1004             wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1005             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1006         }
1007
1008         if (bUseGPU && !bDiffKernels)
1009         {
1010             wallcycle_start(wcycle,ewcLAUNCH_GPU_NB);
1011             /* launch non-local nonbonded F on GPU */
1012             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1013                          nrnb, wcycle);
1014             cycles_force += wallcycle_stop(wcycle,ewcLAUNCH_GPU_NB);
1015         }
1016     }
1017
1018     if (bUseGPU)
1019     {
1020         /* launch D2H copy-back F */
1021         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1022         if (DOMAINDECOMP(cr) && !bDiffKernels)
1023         {
1024             nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1025                                       flags, eatNonlocal);
1026         }
1027         nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1028                                   flags, eatLocal);
1029         cycles_force += wallcycle_stop(wcycle,ewcLAUNCH_GPU_NB);
1030     }
1031
1032     if (bStateChanged && NEED_MUTOT(*inputrec))
1033     {
1034         if (PAR(cr))
1035         {
1036             gmx_sumd(2*DIM,mu,cr);
1037         }
1038
1039         for(i=0; i<2; i++)
1040         {
1041             for(j=0;j<DIM;j++)
1042             {
1043                 fr->mu_tot[i][j] = mu[i*DIM + j];
1044             }
1045         }
1046     }
1047     if (fr->efep == efepNO)
1048     {
1049         copy_rvec(fr->mu_tot[0],mu_tot);
1050     }
1051     else
1052     {
1053         for(j=0; j<DIM; j++)
1054         {
1055             mu_tot[j] =
1056                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1057                 lambda[efptCOUL]*fr->mu_tot[1][j];
1058         }
1059     }
1060
1061     /* Reset energies */
1062     reset_enerdata(&(inputrec->opts),fr,bNS,enerd,MASTER(cr));
1063     clear_rvecs(SHIFTS,fr->fshift);
1064
1065     if (DOMAINDECOMP(cr))
1066     {
1067         if (!(cr->duty & DUTY_PME))
1068         {
1069             wallcycle_start(wcycle,ewcPPDURINGPME);
1070             dd_force_flop_start(cr->dd,nrnb);
1071         }
1072     }
1073
1074     /* Start the force cycle counter.
1075      * This counter is stopped in do_forcelow_level.
1076      * No parallel communication should occur while this counter is running,
1077      * since that will interfere with the dynamic load balancing.
1078      */
1079     wallcycle_start(wcycle,ewcFORCE);
1080     if (bDoForces)
1081     {
1082         /* Reset forces for which the virial is calculated separately:
1083          * PME/Ewald forces if necessary */
1084         if (fr->bF_NoVirSum)
1085         {
1086             if (flags & GMX_FORCE_VIRIAL)
1087             {
1088                 fr->f_novirsum = fr->f_novirsum_alloc;
1089                 if (fr->bDomDec)
1090                 {
1091                     clear_rvecs(fr->f_novirsum_n,fr->f_novirsum);
1092                 }
1093                 else
1094                 {
1095                     clear_rvecs(homenr,fr->f_novirsum+start);
1096                 }
1097             }
1098             else
1099             {
1100                 /* We are not calculating the pressure so we do not need
1101                  * a separate array for forces that do not contribute
1102                  * to the pressure.
1103                  */
1104                 fr->f_novirsum = f;
1105             }
1106         }
1107
1108         /* Clear the short- and long-range forces */
1109         clear_rvecs(fr->natoms_force_constr,f);
1110         if(bSepLRF && do_per_step(step,inputrec->nstcalclr))
1111         {
1112             clear_rvecs(fr->natoms_force_constr,fr->f_twin);
1113         }
1114
1115         clear_rvec(fr->vir_diag_posres);
1116     }
1117     if (inputrec->ePull == epullCONSTRAINT)
1118     {
1119         clear_pull_forces(inputrec->pull);
1120     }
1121
1122     /* update QMMMrec, if necessary */
1123     if(fr->bQMMM)
1124     {
1125         update_QMMMrec(cr,fr,x,mdatoms,box,top);
1126     }
1127
1128     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1129     {
1130         posres_wrapper(fplog,flags,bSepDVDL,inputrec,nrnb,top,box,x,
1131                        f,enerd,lambda,fr);
1132     }
1133
1134     /* Compute the bonded and non-bonded energies and optionally forces */
1135     do_force_lowlevel(fplog,step,fr,inputrec,&(top->idef),
1136                       cr,nrnb,wcycle,mdatoms,&(inputrec->opts),
1137                       x,hist,f, bSepLRF ? fr->f_twin : f,enerd,fcd,mtop,top,fr->born,
1138                       &(top->atomtypes),bBornRadii,box,
1139                       inputrec->fepvals,lambda,graph,&(top->excls),fr->mu_tot,
1140                       flags, &cycles_pme);
1141
1142     if(bSepLRF)
1143     {
1144         if (do_per_step(step,inputrec->nstcalclr))
1145         {
1146             /* Add the long range forces to the short range forces */
1147             for(i=0; i<fr->natoms_force_constr; i++)
1148             {
1149                 rvec_add(fr->f_twin[i],f[i],f[i]);
1150             }
1151         }
1152     }
1153
1154     if (!bUseOrEmulGPU)
1155     {
1156         /* Maybe we should move this into do_force_lowlevel */
1157         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1158                      nrnb, wcycle);
1159     }
1160
1161
1162     if (!bUseOrEmulGPU || bDiffKernels)
1163     {
1164         int aloc;
1165
1166         if (DOMAINDECOMP(cr))
1167         {
1168             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1169                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1170                          nrnb, wcycle);
1171         }
1172
1173         if (!bUseOrEmulGPU)
1174         {
1175             aloc = eintLocal;
1176         }
1177         else
1178         {
1179             aloc = eintNonlocal;
1180         }
1181
1182         /* Add all the non-bonded force to the normal force array.
1183          * This can be split into a local a non-local part when overlapping
1184          * communication with calculation with domain decomposition.
1185          */
1186         cycles_force += wallcycle_stop(wcycle,ewcFORCE);
1187         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1188         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1189         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs,eatAll,nbv->grp[aloc].nbat,f);
1190         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1191         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1192         wallcycle_start_nocount(wcycle,ewcFORCE);
1193
1194         /* if there are multiple fshift output buffers reduce them */
1195         if ((flags & GMX_FORCE_VIRIAL) &&
1196             nbv->grp[aloc].nbl_lists.nnbl > 1)
1197         {
1198             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1199                                                       fr->fshift);
1200         }
1201     }
1202
1203     cycles_force += wallcycle_stop(wcycle,ewcFORCE);
1204
1205     if (ed)
1206     {
1207         do_flood(fplog,cr,x,f,ed,box,step,bNS);
1208     }
1209
1210     if (bUseOrEmulGPU && !bDiffKernels)
1211     {
1212         /* wait for non-local forces (or calculate in emulation mode) */
1213         if (DOMAINDECOMP(cr))
1214         {
1215             if (bUseGPU)
1216             {
1217                 wallcycle_start(wcycle,ewcWAIT_GPU_NB_NL);
1218                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1219                                     nbv->grp[eintNonlocal].nbat,
1220                                     flags, eatNonlocal,
1221                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1222                                     fr->fshift);
1223                 cycles_force += wallcycle_stop(wcycle,ewcWAIT_GPU_NB_NL);
1224             }
1225             else
1226             {
1227                 wallcycle_start_nocount(wcycle,ewcFORCE);
1228                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1229                              nrnb, wcycle);
1230                 cycles_force += wallcycle_stop(wcycle,ewcFORCE);
1231             }
1232             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1233             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1234             /* skip the reduction if there was no non-local work to do */
1235             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1236             {
1237                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs,eatNonlocal,
1238                                                nbv->grp[eintNonlocal].nbat,f);
1239             }
1240             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1241             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1242         }
1243     }
1244
1245     if (bDoForces)
1246     {
1247         /* Communicate the forces */
1248         if (PAR(cr))
1249         {
1250             wallcycle_start(wcycle,ewcMOVEF);
1251             if (DOMAINDECOMP(cr))
1252             {
1253                 dd_move_f(cr->dd,f,fr->fshift);
1254                 /* Do we need to communicate the separate force array
1255                  * for terms that do not contribute to the single sum virial?
1256                  * Position restraints and electric fields do not introduce
1257                  * inter-cg forces, only full electrostatics methods do.
1258                  * When we do not calculate the virial, fr->f_novirsum = f,
1259                  * so we have already communicated these forces.
1260                  */
1261                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1262                     (flags & GMX_FORCE_VIRIAL))
1263                 {
1264                     dd_move_f(cr->dd,fr->f_novirsum,NULL);
1265                 }
1266                 if (bSepLRF)
1267                 {
1268                     /* We should not update the shift forces here,
1269                      * since f_twin is already included in f.
1270                      */
1271                     dd_move_f(cr->dd,fr->f_twin,NULL);
1272                 }
1273             }
1274             wallcycle_stop(wcycle,ewcMOVEF);
1275         }
1276     }
1277
1278     if (bUseOrEmulGPU)
1279     {
1280         /* wait for local forces (or calculate in emulation mode) */
1281         if (bUseGPU)
1282         {
1283             wallcycle_start(wcycle,ewcWAIT_GPU_NB_L);
1284             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1285                                 nbv->grp[eintLocal].nbat,
1286                                 flags, eatLocal,
1287                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1288                                 fr->fshift);
1289             wallcycle_stop(wcycle,ewcWAIT_GPU_NB_L);
1290
1291             /* now clear the GPU outputs while we finish the step on the CPU */
1292             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1293         }
1294         else
1295         {
1296             wallcycle_start_nocount(wcycle,ewcFORCE);
1297             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1298                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1299                          nrnb, wcycle);
1300             wallcycle_stop(wcycle,ewcFORCE);
1301         }
1302         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1303         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1304         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1305         {
1306             /* skip the reduction if there was no non-local work to do */
1307             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs,eatLocal,
1308                                            nbv->grp[eintLocal].nbat,f);
1309         }
1310         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1311         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1312     }
1313
1314     if (DOMAINDECOMP(cr))
1315     {
1316         dd_force_flop_stop(cr->dd,nrnb);
1317         if (wcycle)
1318         {
1319             dd_cycles_add(cr->dd,cycles_force-cycles_pme,ddCyclF);
1320         }
1321     }
1322
1323     if (bDoForces)
1324     {
1325         if (IR_ELEC_FIELD(*inputrec))
1326         {
1327             /* Compute forces due to electric field */
1328             calc_f_el(MASTER(cr) ? field : NULL,
1329                       start,homenr,mdatoms->chargeA,x,fr->f_novirsum,
1330                       inputrec->ex,inputrec->et,t);
1331         }
1332
1333         /* If we have NoVirSum forces, but we do not calculate the virial,
1334          * we sum fr->f_novirum=f later.
1335          */
1336         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1337         {
1338             wallcycle_start(wcycle,ewcVSITESPREAD);
1339             spread_vsite_f(fplog,vsite,x,f,fr->fshift,FALSE,NULL,nrnb,
1340                            &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1341             wallcycle_stop(wcycle,ewcVSITESPREAD);
1342
1343             if (bSepLRF)
1344             {
1345                 wallcycle_start(wcycle,ewcVSITESPREAD);
1346                 spread_vsite_f(fplog,vsite,x,fr->f_twin,NULL,FALSE,NULL,
1347                                nrnb,
1348                                &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1349                 wallcycle_stop(wcycle,ewcVSITESPREAD);
1350             }
1351         }
1352
1353         if (flags & GMX_FORCE_VIRIAL)
1354         {
1355             /* Calculation of the virial must be done after vsites! */
1356             calc_virial(fplog,mdatoms->start,mdatoms->homenr,x,f,
1357                         vir_force,graph,box,nrnb,fr,inputrec->ePBC);
1358         }
1359     }
1360
1361     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1362     {
1363         pull_potential_wrapper(fplog,bSepDVDL,cr,inputrec,box,x,
1364                                f,vir_force,mdatoms,enerd,lambda,t);
1365     }
1366
1367     if (PAR(cr) && !(cr->duty & DUTY_PME))
1368     {
1369         /* In case of node-splitting, the PP nodes receive the long-range
1370          * forces, virial and energy from the PME nodes here.
1371          */
1372         pme_receive_force_ener(fplog,bSepDVDL,cr,wcycle,enerd,fr);
1373     }
1374
1375     if (bDoForces)
1376     {
1377         post_process_forces(fplog,cr,step,nrnb,wcycle,
1378                             top,box,x,f,vir_force,mdatoms,graph,fr,vsite,
1379                             flags);
1380     }
1381
1382     /* Sum the potential energy terms from group contributions */
1383     sum_epot(&(inputrec->opts),enerd);
1384 }
1385
1386 void do_force_cutsGROUP(FILE *fplog,t_commrec *cr,
1387               t_inputrec *inputrec,
1388               gmx_large_int_t step,t_nrnb *nrnb,gmx_wallcycle_t wcycle,
1389               gmx_localtop_t *top,
1390               gmx_mtop_t *mtop,
1391               gmx_groups_t *groups,
1392               matrix box,rvec x[],history_t *hist,
1393               rvec f[],
1394               tensor vir_force,
1395               t_mdatoms *mdatoms,
1396               gmx_enerdata_t *enerd,t_fcdata *fcd,
1397               real *lambda,t_graph *graph,
1398               t_forcerec *fr,gmx_vsite_t *vsite,rvec mu_tot,
1399               double t,FILE *field,gmx_edsam_t ed,
1400               gmx_bool bBornRadii,
1401               int flags)
1402 {
1403     int    cg0,cg1,i,j;
1404     int    start,homenr;
1405     double mu[2*DIM];
1406     gmx_bool   bSepDVDL,bStateChanged,bNS,bFillGrid,bCalcCGCM,bBS;
1407     gmx_bool   bDoLongRangeNS,bDoForces,bDoPotential,bSepLRF;
1408     gmx_bool   bDoAdressWF;
1409     matrix boxs;
1410     rvec   vzero,box_diag;
1411     real   e,v,dvdlambda[efptNR];
1412     t_pbc  pbc;
1413     float  cycles_pme,cycles_force;
1414
1415     start  = mdatoms->start;
1416     homenr = mdatoms->homenr;
1417
1418     bSepDVDL = (fr->bSepDVDL && do_per_step(step,inputrec->nstlog));
1419
1420     clear_mat(vir_force);
1421
1422     if (PARTDECOMP(cr))
1423     {
1424         pd_cg_range(cr,&cg0,&cg1);
1425     }
1426     else
1427     {
1428         cg0 = 0;
1429         if (DOMAINDECOMP(cr))
1430         {
1431             cg1 = cr->dd->ncg_tot;
1432         }
1433         else
1434         {
1435             cg1 = top->cgs.nr;
1436         }
1437         if (fr->n_tpi > 0)
1438         {
1439             cg1--;
1440         }
1441     }
1442
1443     bStateChanged  = (flags & GMX_FORCE_STATECHANGED);
1444     bNS            = (flags & GMX_FORCE_NS) && (fr->bAllvsAll==FALSE);
1445     /* Should we update the long-range neighborlists at this step? */
1446     bDoLongRangeNS = fr->bTwinRange && bNS;
1447     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1448     bFillGrid      = (bNS && bStateChanged);
1449     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1450     bDoForces      = (flags & GMX_FORCE_FORCES);
1451     bDoPotential   = (flags & GMX_FORCE_ENERGY);
1452     bSepLRF        = ((inputrec->nstcalclr>1) && bDoForces &&
1453                       (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1454
1455     /* should probably move this to the forcerec since it doesn't change */
1456     bDoAdressWF   = ((fr->adress_type!=eAdressOff));
1457
1458     if (bStateChanged)
1459     {
1460         update_forcerec(fplog,fr,box);
1461
1462         if (NEED_MUTOT(*inputrec))
1463         {
1464             /* Calculate total (local) dipole moment in a temporary common array.
1465              * This makes it possible to sum them over nodes faster.
1466              */
1467             calc_mu(start,homenr,
1468                     x,mdatoms->chargeA,mdatoms->chargeB,mdatoms->nChargePerturbed,
1469                     mu,mu+DIM);
1470         }
1471     }
1472
1473     if (fr->ePBC != epbcNONE) {
1474         /* Compute shift vectors every step,
1475          * because of pressure coupling or box deformation!
1476          */
1477         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1478             calc_shifts(box,fr->shift_vec);
1479
1480         if (bCalcCGCM) {
1481             put_charge_groups_in_box(fplog,cg0,cg1,fr->ePBC,box,
1482                     &(top->cgs),x,fr->cg_cm);
1483             inc_nrnb(nrnb,eNR_CGCM,homenr);
1484             inc_nrnb(nrnb,eNR_RESETX,cg1-cg0);
1485         }
1486         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph) {
1487             unshift_self(graph,box,x);
1488         }
1489     }
1490     else if (bCalcCGCM) {
1491         calc_cgcm(fplog,cg0,cg1,&(top->cgs),x,fr->cg_cm);
1492         inc_nrnb(nrnb,eNR_CGCM,homenr);
1493     }
1494
1495     if (bCalcCGCM) {
1496         if (PAR(cr)) {
1497             move_cgcm(fplog,cr,fr->cg_cm);
1498         }
1499         if (gmx_debug_at)
1500             pr_rvecs(debug,0,"cgcm",fr->cg_cm,top->cgs.nr);
1501     }
1502
1503 #ifdef GMX_MPI
1504     if (!(cr->duty & DUTY_PME)) {
1505         /* Send particle coordinates to the pme nodes.
1506          * Since this is only implemented for domain decomposition
1507          * and domain decomposition does not use the graph,
1508          * we do not need to worry about shifting.
1509          */
1510
1511         wallcycle_start(wcycle,ewcPP_PMESENDX);
1512
1513         bBS = (inputrec->nwall == 2);
1514         if (bBS) {
1515             copy_mat(box,boxs);
1516             svmul(inputrec->wall_ewald_zfac,boxs[ZZ],boxs[ZZ]);
1517         }
1518
1519         gmx_pme_send_x(cr,bBS ? boxs : box,x,
1520                        mdatoms->nChargePerturbed,lambda[efptCOUL],
1521                        (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),step);
1522
1523         wallcycle_stop(wcycle,ewcPP_PMESENDX);
1524     }
1525 #endif /* GMX_MPI */
1526
1527     /* Communicate coordinates and sum dipole if necessary */
1528     if (PAR(cr))
1529     {
1530         wallcycle_start(wcycle,ewcMOVEX);
1531         if (DOMAINDECOMP(cr))
1532         {
1533             dd_move_x(cr->dd,box,x);
1534         }
1535         else
1536         {
1537             move_x(fplog,cr,GMX_LEFT,GMX_RIGHT,x,nrnb);
1538         }
1539         wallcycle_stop(wcycle,ewcMOVEX);
1540     }
1541
1542     /* update adress weight beforehand */
1543     if(bStateChanged && bDoAdressWF)
1544     {
1545         /* need pbc for adress weight calculation with pbc_dx */
1546         set_pbc(&pbc,inputrec->ePBC,box);
1547         if(fr->adress_site == eAdressSITEcog)
1548         {
1549             update_adress_weights_cog(top->idef.iparams,top->idef.il,x,fr,mdatoms,
1550                                       inputrec->ePBC==epbcNONE ? NULL : &pbc);
1551         }
1552         else if (fr->adress_site == eAdressSITEcom)
1553         {
1554             update_adress_weights_com(fplog,cg0,cg1,&(top->cgs),x,fr,mdatoms,
1555                                       inputrec->ePBC==epbcNONE ? NULL : &pbc);
1556         }
1557         else if (fr->adress_site == eAdressSITEatomatom){
1558             update_adress_weights_atom_per_atom(cg0,cg1,&(top->cgs),x,fr,mdatoms,
1559                                                 inputrec->ePBC==epbcNONE ? NULL : &pbc);
1560         }
1561         else
1562         {
1563             update_adress_weights_atom(cg0,cg1,&(top->cgs),x,fr,mdatoms,
1564                                        inputrec->ePBC==epbcNONE ? NULL : &pbc);
1565         }
1566     }
1567
1568     if (NEED_MUTOT(*inputrec))
1569     {
1570
1571         if (bStateChanged)
1572         {
1573             if (PAR(cr))
1574             {
1575                 gmx_sumd(2*DIM,mu,cr);
1576             }
1577             for(i=0; i<2; i++)
1578             {
1579                 for(j=0;j<DIM;j++)
1580                 {
1581                     fr->mu_tot[i][j] = mu[i*DIM + j];
1582                 }
1583             }
1584         }
1585         if (fr->efep == efepNO)
1586         {
1587             copy_rvec(fr->mu_tot[0],mu_tot);
1588         }
1589         else
1590         {
1591             for(j=0; j<DIM; j++)
1592             {
1593                 mu_tot[j] =
1594                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1595             }
1596         }
1597     }
1598
1599     /* Reset energies */
1600     reset_enerdata(&(inputrec->opts),fr,bNS,enerd,MASTER(cr));
1601     clear_rvecs(SHIFTS,fr->fshift);
1602
1603     if (bNS)
1604     {
1605         wallcycle_start(wcycle,ewcNS);
1606
1607         if (graph && bStateChanged)
1608         {
1609             /* Calculate intramolecular shift vectors to make molecules whole */
1610             mk_mshift(fplog,graph,fr->ePBC,box,x);
1611         }
1612
1613         /* Do the actual neighbour searching and if twin range electrostatics
1614          * also do the calculation of long range forces and energies.
1615          */
1616         for (i=0;i<efptNR;i++) {dvdlambda[i] = 0;}
1617         ns(fplog,fr,x,box,
1618            groups,&(inputrec->opts),top,mdatoms,
1619            cr,nrnb,lambda,dvdlambda,&enerd->grpp,bFillGrid,
1620            bDoLongRangeNS);
1621         if (bSepDVDL)
1622         {
1623             fprintf(fplog,sepdvdlformat,"LR non-bonded",0.0,dvdlambda);
1624         }
1625         enerd->dvdl_lin[efptVDW] += dvdlambda[efptVDW];
1626         enerd->dvdl_lin[efptCOUL] += dvdlambda[efptCOUL];
1627
1628         wallcycle_stop(wcycle,ewcNS);
1629     }
1630
1631     if (inputrec->implicit_solvent && bNS)
1632     {
1633         make_gb_nblist(cr,inputrec->gb_algorithm,inputrec->rlist,
1634                        x,box,fr,&top->idef,graph,fr->born);
1635     }
1636
1637     if (DOMAINDECOMP(cr))
1638     {
1639         if (!(cr->duty & DUTY_PME))
1640         {
1641             wallcycle_start(wcycle,ewcPPDURINGPME);
1642             dd_force_flop_start(cr->dd,nrnb);
1643         }
1644     }
1645
1646     if (inputrec->bRot)
1647     {
1648         /* Enforced rotation has its own cycle counter that starts after the collective
1649          * coordinates have been communicated. It is added to ddCyclF to allow
1650          * for proper load-balancing */
1651         wallcycle_start(wcycle,ewcROT);
1652         do_rotation(cr,inputrec,box,x,t,step,wcycle,bNS);
1653         wallcycle_stop(wcycle,ewcROT);
1654     }
1655
1656     /* Start the force cycle counter.
1657      * This counter is stopped in do_forcelow_level.
1658      * No parallel communication should occur while this counter is running,
1659      * since that will interfere with the dynamic load balancing.
1660      */
1661     wallcycle_start(wcycle,ewcFORCE);
1662
1663     if (bDoForces)
1664     {
1665         /* Reset forces for which the virial is calculated separately:
1666          * PME/Ewald forces if necessary */
1667         if (fr->bF_NoVirSum)
1668         {
1669             if (flags & GMX_FORCE_VIRIAL)
1670             {
1671                 fr->f_novirsum = fr->f_novirsum_alloc;
1672                 if (fr->bDomDec)
1673                 {
1674                     clear_rvecs(fr->f_novirsum_n,fr->f_novirsum);
1675                 }
1676                 else
1677                 {
1678                     clear_rvecs(homenr,fr->f_novirsum+start);
1679                 }
1680             }
1681             else
1682             {
1683                 /* We are not calculating the pressure so we do not need
1684                  * a separate array for forces that do not contribute
1685                  * to the pressure.
1686                  */
1687                 fr->f_novirsum = f;
1688             }
1689         }
1690
1691         /* Clear the short- and long-range forces */
1692         clear_rvecs(fr->natoms_force_constr,f);
1693         if(bSepLRF && do_per_step(step,inputrec->nstcalclr))
1694         {
1695             clear_rvecs(fr->natoms_force_constr,fr->f_twin);
1696         }
1697
1698         clear_rvec(fr->vir_diag_posres);
1699     }
1700     if (inputrec->ePull == epullCONSTRAINT)
1701     {
1702         clear_pull_forces(inputrec->pull);
1703     }
1704
1705     /* update QMMMrec, if necessary */
1706     if(fr->bQMMM)
1707     {
1708         update_QMMMrec(cr,fr,x,mdatoms,box,top);
1709     }
1710
1711     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1712     {
1713         posres_wrapper(fplog,flags,bSepDVDL,inputrec,nrnb,top,box,x,
1714                        f,enerd,lambda,fr);
1715     }
1716
1717     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1718     {
1719         /* Flat-bottomed position restraints always require full pbc */
1720         if(!(bStateChanged && bDoAdressWF))
1721         {
1722             set_pbc(&pbc,inputrec->ePBC,box);
1723         }
1724         v = fbposres(top->idef.il[F_FBPOSRES].nr,top->idef.il[F_FBPOSRES].iatoms,
1725                      top->idef.iparams_fbposres,
1726                      (const rvec*)x,fr->f_novirsum,fr->vir_diag_posres,
1727                      inputrec->ePBC==epbcNONE ? NULL : &pbc,
1728                      fr->rc_scaling,fr->ePBC,fr->posres_com);
1729         enerd->term[F_FBPOSRES] += v;
1730         inc_nrnb(nrnb,eNR_FBPOSRES,top->idef.il[F_FBPOSRES].nr/2);
1731     }
1732
1733     /* Compute the bonded and non-bonded energies and optionally forces */
1734     do_force_lowlevel(fplog,step,fr,inputrec,&(top->idef),
1735                       cr,nrnb,wcycle,mdatoms,&(inputrec->opts),
1736                       x,hist,f, bSepLRF ? fr->f_twin : f,enerd,fcd,mtop,top,fr->born,
1737                       &(top->atomtypes),bBornRadii,box,
1738                       inputrec->fepvals,lambda,
1739                       graph,&(top->excls),fr->mu_tot,
1740                       flags,
1741                       &cycles_pme);
1742
1743     if(bSepLRF)
1744     {
1745         if (do_per_step(step,inputrec->nstcalclr))
1746         {
1747             /* Add the long range forces to the short range forces */
1748             for(i=0; i<fr->natoms_force_constr; i++)
1749             {
1750                 rvec_add(fr->f_twin[i],f[i],f[i]);
1751             }
1752         }
1753     }
1754
1755     cycles_force = wallcycle_stop(wcycle,ewcFORCE);
1756
1757     if (ed)
1758     {
1759         do_flood(fplog,cr,x,f,ed,box,step,bNS);
1760     }
1761
1762     if (DOMAINDECOMP(cr))
1763     {
1764         dd_force_flop_stop(cr->dd,nrnb);
1765         if (wcycle)
1766         {
1767             dd_cycles_add(cr->dd,cycles_force-cycles_pme,ddCyclF);
1768         }
1769     }
1770
1771     if (bDoForces)
1772     {
1773         if (IR_ELEC_FIELD(*inputrec))
1774         {
1775             /* Compute forces due to electric field */
1776             calc_f_el(MASTER(cr) ? field : NULL,
1777                       start,homenr,mdatoms->chargeA,x,fr->f_novirsum,
1778                       inputrec->ex,inputrec->et,t);
1779         }
1780
1781         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1782         {
1783             /* Compute thermodynamic force in hybrid AdResS region */
1784             adress_thermo_force(start,homenr,&(top->cgs),x,fr->f_novirsum,fr,mdatoms,
1785                                 inputrec->ePBC==epbcNONE ? NULL : &pbc);
1786         }
1787
1788         /* Communicate the forces */
1789         if (PAR(cr))
1790         {
1791             wallcycle_start(wcycle,ewcMOVEF);
1792             if (DOMAINDECOMP(cr))
1793             {
1794                 dd_move_f(cr->dd,f,fr->fshift);
1795                 /* Do we need to communicate the separate force array
1796                  * for terms that do not contribute to the single sum virial?
1797                  * Position restraints and electric fields do not introduce
1798                  * inter-cg forces, only full electrostatics methods do.
1799                  * When we do not calculate the virial, fr->f_novirsum = f,
1800                  * so we have already communicated these forces.
1801                  */
1802                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1803                     (flags & GMX_FORCE_VIRIAL))
1804                 {
1805                     dd_move_f(cr->dd,fr->f_novirsum,NULL);
1806                 }
1807                 if (bSepLRF)
1808                 {
1809                     /* We should not update the shift forces here,
1810                      * since f_twin is already included in f.
1811                      */
1812                     dd_move_f(cr->dd,fr->f_twin,NULL);
1813                 }
1814             }
1815             else
1816             {
1817                 pd_move_f(cr,f,nrnb);
1818                 if (bSepLRF)
1819                 {
1820                     pd_move_f(cr,fr->f_twin,nrnb);
1821                 }
1822             }
1823             wallcycle_stop(wcycle,ewcMOVEF);
1824         }
1825
1826         /* If we have NoVirSum forces, but we do not calculate the virial,
1827          * we sum fr->f_novirum=f later.
1828          */
1829         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1830         {
1831             wallcycle_start(wcycle,ewcVSITESPREAD);
1832             spread_vsite_f(fplog,vsite,x,f,fr->fshift,FALSE,NULL,nrnb,
1833                            &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1834             wallcycle_stop(wcycle,ewcVSITESPREAD);
1835
1836             if (bSepLRF)
1837             {
1838                 wallcycle_start(wcycle,ewcVSITESPREAD);
1839                 spread_vsite_f(fplog,vsite,x,fr->f_twin,NULL,FALSE,NULL,
1840                                nrnb,
1841                                &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1842                 wallcycle_stop(wcycle,ewcVSITESPREAD);
1843             }
1844         }
1845
1846         if (flags & GMX_FORCE_VIRIAL)
1847         {
1848             /* Calculation of the virial must be done after vsites! */
1849             calc_virial(fplog,mdatoms->start,mdatoms->homenr,x,f,
1850                         vir_force,graph,box,nrnb,fr,inputrec->ePBC);
1851         }
1852     }
1853
1854     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1855     {
1856         pull_potential_wrapper(fplog,bSepDVDL,cr,inputrec,box,x,
1857                                f,vir_force,mdatoms,enerd,lambda,t);
1858     }
1859
1860     /* Add the forces from enforced rotation potentials (if any) */
1861     if (inputrec->bRot)
1862     {
1863         wallcycle_start(wcycle,ewcROTadd);
1864         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr,step,t);
1865         wallcycle_stop(wcycle,ewcROTadd);
1866     }
1867
1868     if (PAR(cr) && !(cr->duty & DUTY_PME))
1869     {
1870         /* In case of node-splitting, the PP nodes receive the long-range
1871          * forces, virial and energy from the PME nodes here.
1872          */
1873         pme_receive_force_ener(fplog,bSepDVDL,cr,wcycle,enerd,fr);
1874     }
1875
1876     if (bDoForces)
1877     {
1878         post_process_forces(fplog,cr,step,nrnb,wcycle,
1879                             top,box,x,f,vir_force,mdatoms,graph,fr,vsite,
1880                             flags);
1881     }
1882
1883     /* Sum the potential energy terms from group contributions */
1884     sum_epot(&(inputrec->opts),enerd);
1885 }
1886
1887 void do_force(FILE *fplog,t_commrec *cr,
1888               t_inputrec *inputrec,
1889               gmx_large_int_t step,t_nrnb *nrnb,gmx_wallcycle_t wcycle,
1890               gmx_localtop_t *top,
1891               gmx_mtop_t *mtop,
1892               gmx_groups_t *groups,
1893               matrix box,rvec x[],history_t *hist,
1894               rvec f[],
1895               tensor vir_force,
1896               t_mdatoms *mdatoms,
1897               gmx_enerdata_t *enerd,t_fcdata *fcd,
1898               real *lambda,t_graph *graph,
1899               t_forcerec *fr,
1900               gmx_vsite_t *vsite,rvec mu_tot,
1901               double t,FILE *field,gmx_edsam_t ed,
1902               gmx_bool bBornRadii,
1903               int flags)
1904 {
1905     /* modify force flag if not doing nonbonded */
1906     if (!fr->bNonbonded)
1907     {
1908         flags &= ~GMX_FORCE_NONBONDED;
1909     }
1910
1911     switch (inputrec->cutoff_scheme)
1912     {
1913         case ecutsVERLET:
1914             do_force_cutsVERLET(fplog, cr, inputrec,
1915                                 step, nrnb, wcycle,
1916                                 top, mtop,
1917                                 groups,
1918                                 box, x, hist,
1919                                 f, vir_force,
1920                                 mdatoms,
1921                                 enerd, fcd,
1922                                 lambda, graph,
1923                                 fr, fr->ic,
1924                                 vsite, mu_tot,
1925                                 t, field, ed,
1926                                 bBornRadii,
1927                                 flags);
1928             break;
1929         case ecutsGROUP:
1930              do_force_cutsGROUP(fplog, cr, inputrec,
1931                                 step, nrnb, wcycle,
1932                                 top, mtop,
1933                                 groups,
1934                                 box, x, hist,
1935                                 f, vir_force,
1936                                 mdatoms,
1937                                 enerd, fcd,
1938                                 lambda, graph,
1939                                 fr, vsite, mu_tot,
1940                                 t, field, ed,
1941                                 bBornRadii,
1942                                 flags);
1943             break;
1944         default:
1945             gmx_incons("Invalid cut-off scheme passed!");
1946     }
1947 }
1948
1949
1950 void do_constrain_first(FILE *fplog,gmx_constr_t constr,
1951                         t_inputrec *ir,t_mdatoms *md,
1952                         t_state *state,rvec *f,
1953                         t_graph *graph,t_commrec *cr,t_nrnb *nrnb,
1954                         t_forcerec *fr, gmx_localtop_t *top, tensor shake_vir)
1955 {
1956     int    i,m,start,end;
1957     gmx_large_int_t step;
1958     real   dt=ir->delta_t;
1959     real   dvdl_dum;
1960     rvec   *savex;
1961
1962     snew(savex,state->natoms);
1963
1964     start = md->start;
1965     end   = md->homenr + start;
1966
1967     if (debug)
1968         fprintf(debug,"vcm: start=%d, homenr=%d, end=%d\n",
1969                 start,md->homenr,end);
1970     /* Do a first constrain to reset particles... */
1971     step = ir->init_step;
1972     if (fplog)
1973     {
1974         char buf[STEPSTRSIZE];
1975         fprintf(fplog,"\nConstraining the starting coordinates (step %s)\n",
1976                 gmx_step_str(step,buf));
1977     }
1978     dvdl_dum = 0;
1979
1980     /* constrain the current position */
1981     constrain(NULL,TRUE,FALSE,constr,&(top->idef),
1982               ir,NULL,cr,step,0,md,
1983               state->x,state->x,NULL,
1984               fr->bMolPBC,state->box,
1985               state->lambda[efptBONDED],&dvdl_dum,
1986               NULL,NULL,nrnb,econqCoord,
1987               ir->epc==epcMTTK,state->veta,state->veta);
1988     if (EI_VV(ir->eI))
1989     {
1990         /* constrain the inital velocity, and save it */
1991         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
1992         /* might not yet treat veta correctly */
1993         constrain(NULL,TRUE,FALSE,constr,&(top->idef),
1994                   ir,NULL,cr,step,0,md,
1995                   state->x,state->v,state->v,
1996                   fr->bMolPBC,state->box,
1997                   state->lambda[efptBONDED],&dvdl_dum,
1998                   NULL,NULL,nrnb,econqVeloc,
1999                   ir->epc==epcMTTK,state->veta,state->veta);
2000     }
2001     /* constrain the inital velocities at t-dt/2 */
2002     if (EI_STATE_VELOCITY(ir->eI) && ir->eI!=eiVV)
2003     {
2004         for(i=start; (i<end); i++)
2005         {
2006             for(m=0; (m<DIM); m++)
2007             {
2008                 /* Reverse the velocity */
2009                 state->v[i][m] = -state->v[i][m];
2010                 /* Store the position at t-dt in buf */
2011                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2012             }
2013         }
2014     /* Shake the positions at t=-dt with the positions at t=0
2015      * as reference coordinates.
2016          */
2017         if (fplog)
2018         {
2019             char buf[STEPSTRSIZE];
2020             fprintf(fplog,"\nConstraining the coordinates at t0-dt (step %s)\n",
2021                     gmx_step_str(step,buf));
2022         }
2023         dvdl_dum = 0;
2024         constrain(NULL,TRUE,FALSE,constr,&(top->idef),
2025                   ir,NULL,cr,step,-1,md,
2026                   state->x,savex,NULL,
2027                   fr->bMolPBC,state->box,
2028                   state->lambda[efptBONDED],&dvdl_dum,
2029                   state->v,NULL,nrnb,econqCoord,
2030                   ir->epc==epcMTTK,state->veta,state->veta);
2031
2032         for(i=start; i<end; i++) {
2033             for(m=0; m<DIM; m++) {
2034                 /* Re-reverse the velocities */
2035                 state->v[i][m] = -state->v[i][m];
2036             }
2037         }
2038     }
2039     sfree(savex);
2040 }
2041
2042 void calc_enervirdiff(FILE *fplog,int eDispCorr,t_forcerec *fr)
2043 {
2044   double eners[2],virs[2],enersum,virsum,y0,f,g,h;
2045   double r0,r1,r,rc3,rc9,ea,eb,ec,pa,pb,pc,pd;
2046   double invscale,invscale2,invscale3;
2047   int    ri0,ri1,ri,i,offstart,offset;
2048   real   scale,*vdwtab,tabfactor,tmp;
2049
2050   fr->enershiftsix = 0;
2051   fr->enershifttwelve = 0;
2052   fr->enerdiffsix = 0;
2053   fr->enerdifftwelve = 0;
2054   fr->virdiffsix = 0;
2055   fr->virdifftwelve = 0;
2056
2057   if (eDispCorr != edispcNO) {
2058     for(i=0; i<2; i++) {
2059       eners[i] = 0;
2060       virs[i]  = 0;
2061     }
2062     if ((fr->vdwtype == evdwSWITCH) || (fr->vdwtype == evdwSHIFT)) {
2063       if (fr->rvdw_switch == 0)
2064         gmx_fatal(FARGS,
2065                   "With dispersion correction rvdw-switch can not be zero "
2066                   "for vdw-type = %s",evdw_names[fr->vdwtype]);
2067
2068       scale  = fr->nblists[0].table_elec_vdw.scale;
2069       vdwtab = fr->nblists[0].table_vdw.data;
2070
2071       /* Round the cut-offs to exact table values for precision */
2072       ri0 = floor(fr->rvdw_switch*scale);
2073       ri1 = ceil(fr->rvdw*scale);
2074       r0  = ri0/scale;
2075       r1  = ri1/scale;
2076       rc3 = r0*r0*r0;
2077       rc9  = rc3*rc3*rc3;
2078
2079       if (fr->vdwtype == evdwSHIFT)
2080       {
2081           /* Determine the constant energy shift below rvdw_switch.
2082            * Table has a scale factor since we have scaled it down to compensate
2083            * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2084            */
2085           fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2086           fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2087       }
2088       /* Add the constant part from 0 to rvdw_switch.
2089        * This integration from 0 to rvdw_switch overcounts the number
2090        * of interactions by 1, as it also counts the self interaction.
2091        * We will correct for this later.
2092        */
2093       eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2094       eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2095
2096       invscale = 1.0/(scale);
2097       invscale2 = invscale*invscale;
2098       invscale3 = invscale*invscale2;
2099
2100       /* following summation derived from cubic spline definition,
2101         Numerical Recipies in C, second edition, p. 113-116.  Exact
2102         for the cubic spline.  We first calculate the negative of
2103         the energy from rvdw to rvdw_switch, assuming that g(r)=1,
2104         and then add the more standard, abrupt cutoff correction to
2105         that result, yielding the long-range correction for a
2106         switched function.  We perform both the pressure and energy
2107         loops at the same time for simplicity, as the computational
2108         cost is low. */
2109
2110       for (i=0;i<2;i++) {
2111         enersum = 0.0; virsum = 0.0;
2112         if (i==0)
2113         {
2114             offstart = 0;
2115             /* Since the dispersion table has been scaled down a factor 6.0 and the repulsion
2116              * a factor 12.0 to compensate for the c6/c12 parameters inside nbfp[] being scaled
2117              * up (to save flops in kernels), we need to correct for this.
2118              */
2119             tabfactor = 6.0;
2120         }
2121         else
2122         {
2123             offstart = 4;
2124             tabfactor = 12.0;
2125         }
2126         for (ri=ri0; ri<ri1; ri++) {
2127           r = ri*invscale;
2128           ea = invscale3;
2129           eb = 2.0*invscale2*r;
2130           ec = invscale*r*r;
2131
2132           pa = invscale3;
2133           pb = 3.0*invscale2*r;
2134           pc = 3.0*invscale*r*r;
2135           pd = r*r*r;
2136
2137           /* this "8" is from the packing in the vdwtab array - perhaps should be #define'ed? */
2138           offset = 8*ri + offstart;
2139           y0 = vdwtab[offset];
2140           f  = vdwtab[offset+1];
2141           g  = vdwtab[offset+2];
2142           h  = vdwtab[offset+3];
2143
2144           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);
2145           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);
2146         }
2147
2148         enersum *= 4.0*M_PI*tabfactor;
2149         virsum  *= 4.0*M_PI*tabfactor;
2150         eners[i] -= enersum;
2151         virs[i]  -= virsum;
2152       }
2153
2154       /* now add the correction for rvdw_switch to infinity */
2155       eners[0] += -4.0*M_PI/(3.0*rc3);
2156       eners[1] +=  4.0*M_PI/(9.0*rc9);
2157       virs[0]  +=  8.0*M_PI/rc3;
2158       virs[1]  += -16.0*M_PI/(3.0*rc9);
2159     }
2160     else if ((fr->vdwtype == evdwCUT) || (fr->vdwtype == evdwUSER)) {
2161       if (fr->vdwtype == evdwUSER && fplog)
2162         fprintf(fplog,
2163                 "WARNING: using dispersion correction with user tables\n");
2164       rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2165       rc9  = rc3*rc3*rc3;
2166       /* Contribution beyond the cut-off */
2167       eners[0] += -4.0*M_PI/(3.0*rc3);
2168       eners[1] +=  4.0*M_PI/(9.0*rc9);
2169       if (fr->vdw_modifier==eintmodPOTSHIFT) {
2170           /* Contribution within the cut-off */
2171           eners[0] += -4.0*M_PI/(3.0*rc3);
2172           eners[1] +=  4.0*M_PI/(3.0*rc9);
2173       }
2174       /* Contribution beyond the cut-off */
2175       virs[0]  +=  8.0*M_PI/rc3;
2176       virs[1]  += -16.0*M_PI/(3.0*rc9);
2177     } else {
2178       gmx_fatal(FARGS,
2179                 "Dispersion correction is not implemented for vdw-type = %s",
2180                 evdw_names[fr->vdwtype]);
2181     }
2182     fr->enerdiffsix    = eners[0];
2183     fr->enerdifftwelve = eners[1];
2184     /* The 0.5 is due to the Gromacs definition of the virial */
2185     fr->virdiffsix     = 0.5*virs[0];
2186     fr->virdifftwelve  = 0.5*virs[1];
2187   }
2188 }
2189
2190 void calc_dispcorr(FILE *fplog,t_inputrec *ir,t_forcerec *fr,
2191                    gmx_large_int_t step,int natoms,
2192                    matrix box,real lambda,tensor pres,tensor virial,
2193                    real *prescorr, real *enercorr, real *dvdlcorr)
2194 {
2195     gmx_bool bCorrAll,bCorrPres;
2196     real dvdlambda,invvol,dens,ninter,avcsix,avctwelve,enerdiff,svir=0,spres=0;
2197     int  m;
2198
2199     *prescorr = 0;
2200     *enercorr = 0;
2201     *dvdlcorr = 0;
2202
2203     clear_mat(virial);
2204     clear_mat(pres);
2205
2206     if (ir->eDispCorr != edispcNO) {
2207         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2208                      ir->eDispCorr == edispcAllEnerPres);
2209         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2210                      ir->eDispCorr == edispcAllEnerPres);
2211
2212         invvol = 1/det(box);
2213         if (fr->n_tpi)
2214         {
2215             /* Only correct for the interactions with the inserted molecule */
2216             dens = (natoms - fr->n_tpi)*invvol;
2217             ninter = fr->n_tpi;
2218         }
2219         else
2220         {
2221             dens = natoms*invvol;
2222             ninter = 0.5*natoms;
2223         }
2224
2225         if (ir->efep == efepNO)
2226         {
2227             avcsix    = fr->avcsix[0];
2228             avctwelve = fr->avctwelve[0];
2229         }
2230         else
2231         {
2232             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2233             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2234         }
2235
2236         enerdiff = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2237         *enercorr += avcsix*enerdiff;
2238         dvdlambda = 0.0;
2239         if (ir->efep != efepNO)
2240         {
2241             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2242         }
2243         if (bCorrAll)
2244         {
2245             enerdiff = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2246             *enercorr += avctwelve*enerdiff;
2247             if (fr->efep != efepNO)
2248             {
2249                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2250             }
2251         }
2252
2253         if (bCorrPres)
2254         {
2255             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2256             if (ir->eDispCorr == edispcAllEnerPres)
2257             {
2258                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2259             }
2260             /* The factor 2 is because of the Gromacs virial definition */
2261             spres = -2.0*invvol*svir*PRESFAC;
2262
2263             for(m=0; m<DIM; m++) {
2264                 virial[m][m] += svir;
2265                 pres[m][m] += spres;
2266             }
2267             *prescorr += spres;
2268         }
2269
2270         /* Can't currently control when it prints, for now, just print when degugging */
2271         if (debug)
2272         {
2273             if (bCorrAll) {
2274                 fprintf(debug,"Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2275                         avcsix,avctwelve);
2276             }
2277             if (bCorrPres)
2278             {
2279                 fprintf(debug,
2280                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2281                         *enercorr,spres,svir);
2282             }
2283             else
2284             {
2285                 fprintf(debug,"Long Range LJ corr.: Epot %10g\n",*enercorr);
2286             }
2287         }
2288
2289         if (fr->bSepDVDL && do_per_step(step,ir->nstlog))
2290         {
2291             fprintf(fplog,sepdvdlformat,"Dispersion correction",
2292                     *enercorr,dvdlambda);
2293         }
2294         if (fr->efep != efepNO)
2295         {
2296             *dvdlcorr += dvdlambda;
2297         }
2298     }
2299 }
2300
2301 void do_pbc_first(FILE *fplog,matrix box,t_forcerec *fr,
2302                   t_graph *graph,rvec x[])
2303 {
2304   if (fplog)
2305     fprintf(fplog,"Removing pbc first time\n");
2306   calc_shifts(box,fr->shift_vec);
2307   if (graph) {
2308     mk_mshift(fplog,graph,fr->ePBC,box,x);
2309     if (gmx_debug_at)
2310       p_graph(debug,"do_pbc_first 1",graph);
2311     shift_self(graph,box,x);
2312     /* By doing an extra mk_mshift the molecules that are broken
2313      * because they were e.g. imported from another software
2314      * will be made whole again. Such are the healing powers
2315      * of GROMACS.
2316      */
2317     mk_mshift(fplog,graph,fr->ePBC,box,x);
2318     if (gmx_debug_at)
2319       p_graph(debug,"do_pbc_first 2",graph);
2320   }
2321   if (fplog)
2322     fprintf(fplog,"Done rmpbc\n");
2323 }
2324
2325 static void low_do_pbc_mtop(FILE *fplog,int ePBC,matrix box,
2326                             gmx_mtop_t *mtop,rvec x[],
2327                             gmx_bool bFirst)
2328 {
2329   t_graph *graph;
2330   int mb,as,mol;
2331   gmx_molblock_t *molb;
2332
2333   if (bFirst && fplog)
2334     fprintf(fplog,"Removing pbc first time\n");
2335
2336   snew(graph,1);
2337   as = 0;
2338   for(mb=0; mb<mtop->nmolblock; mb++) {
2339     molb = &mtop->molblock[mb];
2340     if (molb->natoms_mol == 1 ||
2341         (!bFirst && mtop->moltype[molb->type].cgs.nr == 1)) {
2342       /* Just one atom or charge group in the molecule, no PBC required */
2343       as += molb->nmol*molb->natoms_mol;
2344     } else {
2345       /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2346       mk_graph_ilist(NULL,mtop->moltype[molb->type].ilist,
2347                      0,molb->natoms_mol,FALSE,FALSE,graph);
2348
2349       for(mol=0; mol<molb->nmol; mol++) {
2350         mk_mshift(fplog,graph,ePBC,box,x+as);
2351
2352         shift_self(graph,box,x+as);
2353         /* The molecule is whole now.
2354          * We don't need the second mk_mshift call as in do_pbc_first,
2355          * since we no longer need this graph.
2356          */
2357
2358         as += molb->natoms_mol;
2359       }
2360       done_graph(graph);
2361     }
2362   }
2363   sfree(graph);
2364 }
2365
2366 void do_pbc_first_mtop(FILE *fplog,int ePBC,matrix box,
2367                        gmx_mtop_t *mtop,rvec x[])
2368 {
2369   low_do_pbc_mtop(fplog,ePBC,box,mtop,x,TRUE);
2370 }
2371
2372 void do_pbc_mtop(FILE *fplog,int ePBC,matrix box,
2373                  gmx_mtop_t *mtop,rvec x[])
2374 {
2375   low_do_pbc_mtop(fplog,ePBC,box,mtop,x,FALSE);
2376 }
2377
2378 void finish_run(FILE *fplog,t_commrec *cr,const char *confout,
2379                 t_inputrec *inputrec,
2380                 t_nrnb nrnb[],gmx_wallcycle_t wcycle,
2381                 gmx_runtime_t *runtime,
2382                 wallclock_gpu_t *gputimes,
2383                 int omp_nth_pp,
2384                 gmx_bool bWriteStat)
2385 {
2386     int    i,j;
2387     t_nrnb *nrnb_tot=NULL;
2388     real   delta_t;
2389     double nbfs,mflop;
2390
2391     wallcycle_sum(cr,wcycle);
2392
2393     if (cr->nnodes > 1)
2394     {
2395         snew(nrnb_tot,1);
2396 #ifdef GMX_MPI
2397         MPI_Allreduce(nrnb->n,nrnb_tot->n,eNRNB,MPI_DOUBLE,MPI_SUM,
2398                       cr->mpi_comm_mysim);
2399 #endif
2400     }
2401     else
2402     {
2403         nrnb_tot = nrnb;
2404     }
2405
2406 #if defined(GMX_MPI) && !defined(GMX_THREAD_MPI)
2407     if (cr->nnodes > 1)
2408     {
2409         /* reduce nodetime over all MPI processes in the current simulation */
2410         double sum;
2411         MPI_Allreduce(&runtime->proctime,&sum,1,MPI_DOUBLE,MPI_SUM,
2412                       cr->mpi_comm_mysim);
2413         runtime->proctime = sum;
2414     }
2415 #endif
2416
2417     if (SIMMASTER(cr))
2418     {
2419         print_flop(fplog,nrnb_tot,&nbfs,&mflop);
2420     }
2421     if (cr->nnodes > 1)
2422     {
2423         sfree(nrnb_tot);
2424     }
2425
2426     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2427     {
2428         print_dd_statistics(cr,inputrec,fplog);
2429     }
2430
2431 #ifdef GMX_MPI
2432     if (PARTDECOMP(cr))
2433     {
2434         if (MASTER(cr))
2435         {
2436             t_nrnb     *nrnb_all;
2437             int        s;
2438             MPI_Status stat;
2439
2440             snew(nrnb_all,cr->nnodes);
2441             nrnb_all[0] = *nrnb;
2442             for(s=1; s<cr->nnodes; s++)
2443             {
2444                 MPI_Recv(nrnb_all[s].n,eNRNB,MPI_DOUBLE,s,0,
2445                          cr->mpi_comm_mysim,&stat);
2446             }
2447             pr_load(fplog,cr,nrnb_all);
2448             sfree(nrnb_all);
2449         }
2450         else
2451         {
2452             MPI_Send(nrnb->n,eNRNB,MPI_DOUBLE,MASTERRANK(cr),0,
2453                      cr->mpi_comm_mysim);
2454         }
2455     }
2456 #endif
2457
2458     if (SIMMASTER(cr))
2459     {
2460         wallcycle_print(fplog,cr->nnodes,cr->npmenodes,runtime->realtime,
2461                         wcycle,gputimes);
2462
2463         if (EI_DYNAMICS(inputrec->eI))
2464         {
2465             delta_t = inputrec->delta_t;
2466         }
2467         else
2468         {
2469             delta_t = 0;
2470         }
2471
2472         if (fplog)
2473         {
2474             print_perf(fplog,runtime->proctime,runtime->realtime,
2475                        cr->nnodes-cr->npmenodes,
2476                        runtime->nsteps_done,delta_t,nbfs,mflop,
2477                        omp_nth_pp);
2478         }
2479         if (bWriteStat)
2480         {
2481             print_perf(stderr,runtime->proctime,runtime->realtime,
2482                        cr->nnodes-cr->npmenodes,
2483                        runtime->nsteps_done,delta_t,nbfs,mflop,
2484                        omp_nth_pp);
2485         }
2486     }
2487 }
2488
2489 extern void initialize_lambdas(FILE *fplog,t_inputrec *ir,int *fep_state,real *lambda,double *lam0)
2490 {
2491     /* this function works, but could probably use a logic rewrite to keep all the different
2492        types of efep straight. */
2493
2494     int i;
2495     t_lambda *fep = ir->fepvals;
2496
2497     if ((ir->efep==efepNO) && (ir->bSimTemp == FALSE)) {
2498         for (i=0;i<efptNR;i++)  {
2499             lambda[i] = 0.0;
2500             if (lam0)
2501             {
2502                 lam0[i] = 0.0;
2503             }
2504         }
2505         return;
2506     } else {
2507         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2508                                              if checkpoint is set -- a kludge is in for now
2509                                              to prevent this.*/
2510         for (i=0;i<efptNR;i++)
2511         {
2512             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2513             if (fep->init_lambda>=0) /* if it's -1, it was never initializd */
2514             {
2515                 lambda[i] = fep->init_lambda;
2516                 if (lam0) {
2517                     lam0[i] = lambda[i];
2518                 }
2519             }
2520             else
2521             {
2522                 lambda[i] = fep->all_lambda[i][*fep_state];
2523                 if (lam0) {
2524                     lam0[i] = lambda[i];
2525                 }
2526             }
2527         }
2528         if (ir->bSimTemp) {
2529             /* need to rescale control temperatures to match current state */
2530             for (i=0;i<ir->opts.ngtc;i++) {
2531                 if (ir->opts.ref_t[i] > 0) {
2532                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2533                 }
2534             }
2535         }
2536     }
2537
2538     /* Send to the log the information on the current lambdas */
2539     if (fplog != NULL)
2540     {
2541         fprintf(fplog,"Initial vector of lambda components:[ ");
2542         for (i=0;i<efptNR;i++)
2543         {
2544             fprintf(fplog,"%10.4f ",lambda[i]);
2545         }
2546         fprintf(fplog,"]\n");
2547     }
2548     return;
2549 }
2550
2551
2552 void init_md(FILE *fplog,
2553              t_commrec *cr,t_inputrec *ir,const output_env_t oenv,
2554              double *t,double *t0,
2555              real *lambda, int *fep_state, double *lam0,
2556              t_nrnb *nrnb,gmx_mtop_t *mtop,
2557              gmx_update_t *upd,
2558              int nfile,const t_filenm fnm[],
2559              gmx_mdoutf_t **outf,t_mdebin **mdebin,
2560              tensor force_vir,tensor shake_vir,rvec mu_tot,
2561              gmx_bool *bSimAnn,t_vcm **vcm, t_state *state, unsigned long Flags)
2562 {
2563     int  i,j,n;
2564     real tmpt,mod;
2565
2566     /* Initial values */
2567     *t = *t0       = ir->init_t;
2568
2569     *bSimAnn=FALSE;
2570     for(i=0;i<ir->opts.ngtc;i++)
2571     {
2572         /* set bSimAnn if any group is being annealed */
2573         if(ir->opts.annealing[i]!=eannNO)
2574         {
2575             *bSimAnn = TRUE;
2576         }
2577     }
2578     if (*bSimAnn)
2579     {
2580         update_annealing_target_temp(&(ir->opts),ir->init_t);
2581     }
2582
2583     /* Initialize lambda variables */
2584     initialize_lambdas(fplog,ir,fep_state,lambda,lam0);
2585
2586     if (upd)
2587     {
2588         *upd = init_update(fplog,ir);
2589     }
2590
2591
2592     if (vcm != NULL)
2593     {
2594         *vcm = init_vcm(fplog,&mtop->groups,ir);
2595     }
2596
2597     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2598     {
2599         if (ir->etc == etcBERENDSEN)
2600         {
2601             please_cite(fplog,"Berendsen84a");
2602         }
2603         if (ir->etc == etcVRESCALE)
2604         {
2605             please_cite(fplog,"Bussi2007a");
2606         }
2607     }
2608
2609     init_nrnb(nrnb);
2610
2611     if (nfile != -1)
2612     {
2613         *outf = init_mdoutf(nfile,fnm,Flags,cr,ir,oenv);
2614
2615         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : (*outf)->fp_ene,
2616                               mtop,ir, (*outf)->fp_dhdl);
2617     }
2618
2619     if (ir->bAdress)
2620     {
2621       please_cite(fplog,"Fritsch12");
2622       please_cite(fplog,"Junghans10");
2623     }
2624     /* Initiate variables */
2625     clear_mat(force_vir);
2626     clear_mat(shake_vir);
2627     clear_rvec(mu_tot);
2628
2629     debug_gmx();
2630 }
2631