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_DOLR));
 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         if (bSepLRF)
1109         {
1110             /* Add the long range forces to the short range forces */
1111             for(i=0; i<fr->natoms_force_constr; i++)
1112             {
1113                 copy_rvec(fr->f_twin[i],f[i]);
1114             }
1115         }
1116         else if (!(fr->bTwinRange && bNS))
1117         {
1118             /* Clear the short-range forces */
1119             clear_rvecs(fr->natoms_force_constr,f);
1120         }
1121
1122         clear_rvec(fr->vir_diag_posres);
1123     }
1124     if (inputrec->ePull == epullCONSTRAINT)
1125     {
1126         clear_pull_forces(inputrec->pull);
1127     }
1128
1129     /* update QMMMrec, if necessary */
1130     if(fr->bQMMM)
1131     {
1132         update_QMMMrec(cr,fr,x,mdatoms,box,top);
1133     }
1134
1135     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1136     {
1137         posres_wrapper(fplog,flags,bSepDVDL,inputrec,nrnb,top,box,x,
1138                        f,enerd,lambda,fr);
1139     }
1140
1141     /* Compute the bonded and non-bonded energies and optionally forces */
1142     /* if we use the GPU turn off the nonbonded */
1143     do_force_lowlevel(fplog,step,fr,inputrec,&(top->idef),
1144                       cr,nrnb,wcycle,mdatoms,&(inputrec->opts),
1145                       x,hist,f,enerd,fcd,mtop,top,fr->born,
1146                       &(top->atomtypes),bBornRadii,box,
1147                       inputrec->fepvals,lambda,graph,&(top->excls),fr->mu_tot,
1148                       ((nb_kernel_type == nbk8x8x8_CUDA || nb_kernel_type == nbk8x8x8_PlainC)
1149                         ? flags&~GMX_FORCE_NONBONDED : flags),
1150                       &cycles_pme);
1151
1152     if (!bUseOrEmulGPU)
1153     {
1154         /* Maybe we should move this into do_force_lowlevel */
1155         do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1156                      nrnb, wcycle);
1157     }
1158
1159
1160     if (!bUseOrEmulGPU || bDiffKernels)
1161     {
1162         int aloc;
1163
1164         if (DOMAINDECOMP(cr))
1165         {
1166             do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1167                          bDiffKernels ? enbvClearFYes : enbvClearFNo,
1168                          nrnb, wcycle);
1169         }
1170
1171         if (!bUseOrEmulGPU)
1172         {
1173             aloc = eintLocal;
1174         }
1175         else
1176         {
1177             aloc = eintNonlocal;
1178         }
1179
1180         /* Add all the non-bonded force to the normal force array.
1181          * This can be split into a local a non-local part when overlapping
1182          * communication with calculation with domain decomposition.
1183          */
1184         cycles_force += wallcycle_stop(wcycle,ewcFORCE);
1185         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1186         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1187         nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs,eatAll,nbv->grp[aloc].nbat,f);
1188         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1189         cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1190         wallcycle_start_nocount(wcycle,ewcFORCE);
1191
1192         /* if there are multiple fshift output buffers reduce them */
1193         if ((flags & GMX_FORCE_VIRIAL) &&
1194             nbv->grp[aloc].nbl_lists.nnbl > 1)
1195         {
1196             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1197                                                       fr->fshift);
1198         }
1199     }
1200
1201     cycles_force += wallcycle_stop(wcycle,ewcFORCE);
1202
1203     if (ed)
1204     {
1205         do_flood(fplog,cr,x,f,ed,box,step,bNS);
1206     }
1207
1208     if (bUseOrEmulGPU && !bDiffKernels)
1209     {
1210         /* wait for non-local forces (or calculate in emulation mode) */
1211         if (DOMAINDECOMP(cr))
1212         {
1213             if (bUseGPU)
1214             {
1215                 wallcycle_start(wcycle,ewcWAIT_GPU_NB_NL);
1216                 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1217                                     nbv->grp[eintNonlocal].nbat,
1218                                     flags, eatNonlocal,
1219                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1220                                     fr->fshift);
1221                 cycles_force += wallcycle_stop(wcycle,ewcWAIT_GPU_NB_NL);
1222             }
1223             else
1224             {
1225                 wallcycle_start_nocount(wcycle,ewcFORCE);
1226                 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1227                              nrnb, wcycle);
1228                 cycles_force += wallcycle_stop(wcycle,ewcFORCE);
1229             }
1230             wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1231             wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1232             /* skip the reduction if there was no non-local work to do */
1233             if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1234             {
1235                 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs,eatNonlocal,
1236                                                nbv->grp[eintNonlocal].nbat,f);
1237             }
1238             wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1239             cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1240         }
1241     }
1242
1243     if (bDoForces)
1244     {
1245         /* Communicate the forces */
1246         if (PAR(cr))
1247         {
1248             wallcycle_start(wcycle,ewcMOVEF);
1249             if (DOMAINDECOMP(cr))
1250             {
1251                 dd_move_f(cr->dd,f,fr->fshift);
1252                 /* Do we need to communicate the separate force array
1253                  * for terms that do not contribute to the single sum virial?
1254                  * Position restraints and electric fields do not introduce
1255                  * inter-cg forces, only full electrostatics methods do.
1256                  * When we do not calculate the virial, fr->f_novirsum = f,
1257                  * so we have already communicated these forces.
1258                  */
1259                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1260                     (flags & GMX_FORCE_VIRIAL))
1261                 {
1262                     dd_move_f(cr->dd,fr->f_novirsum,NULL);
1263                 }
1264                 if (bSepLRF)
1265                 {
1266                     /* We should not update the shift forces here,
1267                      * since f_twin is already included in f.
1268                      */
1269                     dd_move_f(cr->dd,fr->f_twin,NULL);
1270                 }
1271             }
1272             wallcycle_stop(wcycle,ewcMOVEF);
1273         }
1274     }
1275
1276     if (bUseOrEmulGPU)
1277     {
1278         /* wait for local forces (or calculate in emulation mode) */
1279         if (bUseGPU)
1280         {
1281             wallcycle_start(wcycle,ewcWAIT_GPU_NB_L);
1282             nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1283                                 nbv->grp[eintLocal].nbat,
1284                                 flags, eatLocal,
1285                                 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1286                                 fr->fshift);
1287             wallcycle_stop(wcycle,ewcWAIT_GPU_NB_L);
1288
1289             /* now clear the GPU outputs while we finish the step on the CPU */
1290             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1291         }
1292         else
1293         {
1294             wallcycle_start_nocount(wcycle,ewcFORCE);
1295             do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1296                          DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1297                          nrnb, wcycle);
1298             wallcycle_stop(wcycle,ewcFORCE);
1299         }
1300         wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1301         wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1302         if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1303         {
1304             /* skip the reduction if there was no non-local work to do */
1305             nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs,eatLocal,
1306                                            nbv->grp[eintLocal].nbat,f);
1307         }
1308         wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1309         wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1310     }
1311
1312     if (DOMAINDECOMP(cr))
1313     {
1314         dd_force_flop_stop(cr->dd,nrnb);
1315         if (wcycle)
1316         {
1317             dd_cycles_add(cr->dd,cycles_force-cycles_pme,ddCyclF);
1318         }
1319     }
1320
1321     if (bDoForces)
1322     {
1323         if (IR_ELEC_FIELD(*inputrec))
1324         {
1325             /* Compute forces due to electric field */
1326             calc_f_el(MASTER(cr) ? field : NULL,
1327                       start,homenr,mdatoms->chargeA,x,fr->f_novirsum,
1328                       inputrec->ex,inputrec->et,t);
1329         }
1330
1331         /* If we have NoVirSum forces, but we do not calculate the virial,
1332          * we sum fr->f_novirum=f later.
1333          */
1334         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1335         {
1336             wallcycle_start(wcycle,ewcVSITESPREAD);
1337             spread_vsite_f(fplog,vsite,x,f,fr->fshift,FALSE,NULL,nrnb,
1338                            &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1339             wallcycle_stop(wcycle,ewcVSITESPREAD);
1340
1341             if (bSepLRF)
1342             {
1343                 wallcycle_start(wcycle,ewcVSITESPREAD);
1344                 spread_vsite_f(fplog,vsite,x,fr->f_twin,NULL,FALSE,NULL,
1345                                nrnb,
1346                                &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1347                 wallcycle_stop(wcycle,ewcVSITESPREAD);
1348             }
1349         }
1350
1351         if (flags & GMX_FORCE_VIRIAL)
1352         {
1353             /* Calculation of the virial must be done after vsites! */
1354             calc_virial(fplog,mdatoms->start,mdatoms->homenr,x,f,
1355                         vir_force,graph,box,nrnb,fr,inputrec->ePBC);
1356         }
1357     }
1358
1359     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1360     {
1361         pull_potential_wrapper(fplog,bSepDVDL,cr,inputrec,box,x,
1362                                f,vir_force,mdatoms,enerd,lambda,t);
1363     }
1364
1365     if (PAR(cr) && !(cr->duty & DUTY_PME))
1366     {
1367         /* In case of node-splitting, the PP nodes receive the long-range
1368          * forces, virial and energy from the PME nodes here.
1369          */
1370         pme_receive_force_ener(fplog,bSepDVDL,cr,wcycle,enerd,fr);
1371     }
1372
1373     if (bDoForces)
1374     {
1375         post_process_forces(fplog,cr,step,nrnb,wcycle,
1376                             top,box,x,f,vir_force,mdatoms,graph,fr,vsite,
1377                             flags);
1378     }
1379
1380     /* Sum the potential energy terms from group contributions */
1381     sum_epot(&(inputrec->opts),enerd);
1382 }
1383
1384 void do_force_cutsGROUP(FILE *fplog,t_commrec *cr,
1385               t_inputrec *inputrec,
1386               gmx_large_int_t step,t_nrnb *nrnb,gmx_wallcycle_t wcycle,
1387               gmx_localtop_t *top,
1388               gmx_mtop_t *mtop,
1389               gmx_groups_t *groups,
1390               matrix box,rvec x[],history_t *hist,
1391               rvec f[],
1392               tensor vir_force,
1393               t_mdatoms *mdatoms,
1394               gmx_enerdata_t *enerd,t_fcdata *fcd,
1395               real *lambda,t_graph *graph,
1396               t_forcerec *fr,gmx_vsite_t *vsite,rvec mu_tot,
1397               double t,FILE *field,gmx_edsam_t ed,
1398               gmx_bool bBornRadii,
1399               int flags)
1400 {
1401     int    cg0,cg1,i,j;
1402     int    start,homenr;
1403     double mu[2*DIM];
1404     gmx_bool   bSepDVDL,bStateChanged,bNS,bFillGrid,bCalcCGCM,bBS;
1405     gmx_bool   bDoLongRange,bDoForces,bSepLRF;
1406     gmx_bool   bDoAdressWF;
1407     matrix boxs;
1408     rvec   vzero,box_diag;
1409     real   e,v,dvdlambda[efptNR];
1410     t_pbc  pbc;
1411     float  cycles_pme,cycles_force;
1412
1413     start  = mdatoms->start;
1414     homenr = mdatoms->homenr;
1415
1416     bSepDVDL = (fr->bSepDVDL && do_per_step(step,inputrec->nstlog));
1417
1418     clear_mat(vir_force);
1419
1420     if (PARTDECOMP(cr))
1421     {
1422         pd_cg_range(cr,&cg0,&cg1);
1423     }
1424     else
1425     {
1426         cg0 = 0;
1427         if (DOMAINDECOMP(cr))
1428         {
1429             cg1 = cr->dd->ncg_tot;
1430         }
1431         else
1432         {
1433             cg1 = top->cgs.nr;
1434         }
1435         if (fr->n_tpi > 0)
1436         {
1437             cg1--;
1438         }
1439     }
1440
1441     bStateChanged = (flags & GMX_FORCE_STATECHANGED);
1442     bNS           = (flags & GMX_FORCE_NS) && (fr->bAllvsAll==FALSE);
1443     bFillGrid     = (bNS && bStateChanged);
1444     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
1445     bDoLongRange  = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DOLR));
1446     bDoForces     = (flags & GMX_FORCE_FORCES);
1447     bSepLRF       = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
1448     /* should probably move this to the forcerec since it doesn't change */
1449     bDoAdressWF   = ((fr->adress_type!=eAdressOff));
1450
1451     if (bStateChanged)
1452     {
1453         update_forcerec(fplog,fr,box);
1454
1455         if (NEED_MUTOT(*inputrec))
1456         {
1457             /* Calculate total (local) dipole moment in a temporary common array.
1458              * This makes it possible to sum them over nodes faster.
1459              */
1460             calc_mu(start,homenr,
1461                     x,mdatoms->chargeA,mdatoms->chargeB,mdatoms->nChargePerturbed,
1462                     mu,mu+DIM);
1463         }
1464     }
1465
1466     if (fr->ePBC != epbcNONE) {
1467         /* Compute shift vectors every step,
1468          * because of pressure coupling or box deformation!
1469          */
1470         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1471             calc_shifts(box,fr->shift_vec);
1472
1473         if (bCalcCGCM) {
1474             put_charge_groups_in_box(fplog,cg0,cg1,fr->ePBC,box,
1475                     &(top->cgs),x,fr->cg_cm);
1476             inc_nrnb(nrnb,eNR_CGCM,homenr);
1477             inc_nrnb(nrnb,eNR_RESETX,cg1-cg0);
1478         }
1479         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph) {
1480             unshift_self(graph,box,x);
1481         }
1482     }
1483     else if (bCalcCGCM) {
1484         calc_cgcm(fplog,cg0,cg1,&(top->cgs),x,fr->cg_cm);
1485         inc_nrnb(nrnb,eNR_CGCM,homenr);
1486     }
1487
1488     if (bCalcCGCM) {
1489         if (PAR(cr)) {
1490             move_cgcm(fplog,cr,fr->cg_cm);
1491         }
1492         if (gmx_debug_at)
1493             pr_rvecs(debug,0,"cgcm",fr->cg_cm,top->cgs.nr);
1494     }
1495
1496 #ifdef GMX_MPI
1497     if (!(cr->duty & DUTY_PME)) {
1498         /* Send particle coordinates to the pme nodes.
1499          * Since this is only implemented for domain decomposition
1500          * and domain decomposition does not use the graph,
1501          * we do not need to worry about shifting.
1502          */
1503
1504         wallcycle_start(wcycle,ewcPP_PMESENDX);
1505
1506         bBS = (inputrec->nwall == 2);
1507         if (bBS) {
1508             copy_mat(box,boxs);
1509             svmul(inputrec->wall_ewald_zfac,boxs[ZZ],boxs[ZZ]);
1510         }
1511
1512         gmx_pme_send_x(cr,bBS ? boxs : box,x,
1513                        mdatoms->nChargePerturbed,lambda[efptCOUL],
1514                        (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)),step);
1515
1516         wallcycle_stop(wcycle,ewcPP_PMESENDX);
1517     }
1518 #endif /* GMX_MPI */
1519
1520     /* Communicate coordinates and sum dipole if necessary */
1521     if (PAR(cr))
1522     {
1523         wallcycle_start(wcycle,ewcMOVEX);
1524         if (DOMAINDECOMP(cr))
1525         {
1526             dd_move_x(cr->dd,box,x);
1527         }
1528         else
1529         {
1530             move_x(fplog,cr,GMX_LEFT,GMX_RIGHT,x,nrnb);
1531         }
1532         wallcycle_stop(wcycle,ewcMOVEX);
1533     }
1534
1535     /* update adress weight beforehand */
1536     if(bStateChanged && bDoAdressWF)
1537     {
1538         /* need pbc for adress weight calculation with pbc_dx */
1539         set_pbc(&pbc,inputrec->ePBC,box);
1540         if(fr->adress_site == eAdressSITEcog)
1541         {
1542             update_adress_weights_cog(top->idef.iparams,top->idef.il,x,fr,mdatoms,
1543                                       inputrec->ePBC==epbcNONE ? NULL : &pbc);
1544         }
1545         else if (fr->adress_site == eAdressSITEcom)
1546         {
1547             update_adress_weights_com(fplog,cg0,cg1,&(top->cgs),x,fr,mdatoms,
1548                                       inputrec->ePBC==epbcNONE ? NULL : &pbc);
1549         }
1550         else if (fr->adress_site == eAdressSITEatomatom){
1551             update_adress_weights_atom_per_atom(cg0,cg1,&(top->cgs),x,fr,mdatoms,
1552                                                 inputrec->ePBC==epbcNONE ? NULL : &pbc);
1553         }
1554         else
1555         {
1556             update_adress_weights_atom(cg0,cg1,&(top->cgs),x,fr,mdatoms,
1557                                        inputrec->ePBC==epbcNONE ? NULL : &pbc);
1558         }
1559     }
1560
1561     if (NEED_MUTOT(*inputrec))
1562     {
1563
1564         if (bStateChanged)
1565         {
1566             if (PAR(cr))
1567             {
1568                 gmx_sumd(2*DIM,mu,cr);
1569             }
1570             for(i=0; i<2; i++)
1571             {
1572                 for(j=0;j<DIM;j++)
1573                 {
1574                     fr->mu_tot[i][j] = mu[i*DIM + j];
1575                 }
1576             }
1577         }
1578         if (fr->efep == efepNO)
1579         {
1580             copy_rvec(fr->mu_tot[0],mu_tot);
1581         }
1582         else
1583         {
1584             for(j=0; j<DIM; j++)
1585             {
1586                 mu_tot[j] =
1587                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1588             }
1589         }
1590     }
1591
1592     /* Reset energies */
1593     reset_enerdata(&(inputrec->opts),fr,bNS,enerd,MASTER(cr));
1594     clear_rvecs(SHIFTS,fr->fshift);
1595
1596     if (bNS)
1597     {
1598         wallcycle_start(wcycle,ewcNS);
1599
1600         if (graph && bStateChanged)
1601         {
1602             /* Calculate intramolecular shift vectors to make molecules whole */
1603             mk_mshift(fplog,graph,fr->ePBC,box,x);
1604         }
1605
1606         /* Reset long range forces if necessary */
1607         if (fr->bTwinRange)
1608         {
1609             /* Reset the (long-range) forces if necessary */
1610             clear_rvecs(fr->natoms_force_constr,bSepLRF ? fr->f_twin : f);
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            bDoLongRange,bDoForces,bSepLRF ? fr->f_twin : f);
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         if (bSepLRF)
1692         {
1693             /* Add the long range forces to the short range forces */
1694             for(i=0; i<fr->natoms_force_constr; i++)
1695             {
1696                 copy_rvec(fr->f_twin[i],f[i]);
1697             }
1698         }
1699         else if (!(fr->bTwinRange && bNS))
1700         {
1701             /* Clear the short-range forces */
1702             clear_rvecs(fr->natoms_force_constr,f);
1703         }
1704
1705         clear_rvec(fr->vir_diag_posres);
1706     }
1707     if (inputrec->ePull == epullCONSTRAINT)
1708     {
1709         clear_pull_forces(inputrec->pull);
1710     }
1711
1712     /* update QMMMrec, if necessary */
1713     if(fr->bQMMM)
1714     {
1715         update_QMMMrec(cr,fr,x,mdatoms,box,top);
1716     }
1717
1718     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1719     {
1720         posres_wrapper(fplog,flags,bSepDVDL,inputrec,nrnb,top,box,x,
1721                        f,enerd,lambda,fr);
1722     }
1723
1724     if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_FBPOSRES].nr > 0)
1725     {
1726         /* Flat-bottomed position restraints always require full pbc */
1727         if(!(bStateChanged && bDoAdressWF))
1728         {
1729             set_pbc(&pbc,inputrec->ePBC,box);
1730         }
1731         v = fbposres(top->idef.il[F_FBPOSRES].nr,top->idef.il[F_FBPOSRES].iatoms,
1732                      top->idef.iparams_fbposres,
1733                      (const rvec*)x,fr->f_novirsum,fr->vir_diag_posres,
1734                      inputrec->ePBC==epbcNONE ? NULL : &pbc,
1735                      fr->rc_scaling,fr->ePBC,fr->posres_com);
1736         enerd->term[F_FBPOSRES] += v;
1737         inc_nrnb(nrnb,eNR_FBPOSRES,top->idef.il[F_FBPOSRES].nr/2);
1738     }
1739
1740     /* Compute the bonded and non-bonded energies and optionally forces */
1741     do_force_lowlevel(fplog,step,fr,inputrec,&(top->idef),
1742                       cr,nrnb,wcycle,mdatoms,&(inputrec->opts),
1743                       x,hist,f,enerd,fcd,mtop,top,fr->born,
1744                       &(top->atomtypes),bBornRadii,box,
1745                       inputrec->fepvals,lambda,
1746                       graph,&(top->excls),fr->mu_tot,
1747                       flags,
1748                       &cycles_pme);
1749
1750     cycles_force = wallcycle_stop(wcycle,ewcFORCE);
1751
1752     if (ed)
1753     {
1754         do_flood(fplog,cr,x,f,ed,box,step,bNS);
1755     }
1756
1757     if (DOMAINDECOMP(cr))
1758     {
1759         dd_force_flop_stop(cr->dd,nrnb);
1760         if (wcycle)
1761         {
1762             dd_cycles_add(cr->dd,cycles_force-cycles_pme,ddCyclF);
1763         }
1764     }
1765
1766     if (bDoForces)
1767     {
1768         if (IR_ELEC_FIELD(*inputrec))
1769         {
1770             /* Compute forces due to electric field */
1771             calc_f_el(MASTER(cr) ? field : NULL,
1772                       start,homenr,mdatoms->chargeA,x,fr->f_novirsum,
1773                       inputrec->ex,inputrec->et,t);
1774         }
1775
1776         if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1777         {
1778             /* Compute thermodynamic force in hybrid AdResS region */
1779             adress_thermo_force(start,homenr,&(top->cgs),x,fr->f_novirsum,fr,mdatoms,
1780                                 inputrec->ePBC==epbcNONE ? NULL : &pbc);
1781         }
1782
1783         /* Communicate the forces */
1784         if (PAR(cr))
1785         {
1786             wallcycle_start(wcycle,ewcMOVEF);
1787             if (DOMAINDECOMP(cr))
1788             {
1789                 dd_move_f(cr->dd,f,fr->fshift);
1790                 /* Do we need to communicate the separate force array
1791                  * for terms that do not contribute to the single sum virial?
1792                  * Position restraints and electric fields do not introduce
1793                  * inter-cg forces, only full electrostatics methods do.
1794                  * When we do not calculate the virial, fr->f_novirsum = f,
1795                  * so we have already communicated these forces.
1796                  */
1797                 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1798                     (flags & GMX_FORCE_VIRIAL))
1799                 {
1800                     dd_move_f(cr->dd,fr->f_novirsum,NULL);
1801                 }
1802                 if (bSepLRF)
1803                 {
1804                     /* We should not update the shift forces here,
1805                      * since f_twin is already included in f.
1806                      */
1807                     dd_move_f(cr->dd,fr->f_twin,NULL);
1808                 }
1809             }
1810             else
1811             {
1812                 pd_move_f(cr,f,nrnb);
1813                 if (bSepLRF)
1814                 {
1815                     pd_move_f(cr,fr->f_twin,nrnb);
1816                 }
1817             }
1818             wallcycle_stop(wcycle,ewcMOVEF);
1819         }
1820
1821         /* If we have NoVirSum forces, but we do not calculate the virial,
1822          * we sum fr->f_novirum=f later.
1823          */
1824         if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1825         {
1826             wallcycle_start(wcycle,ewcVSITESPREAD);
1827             spread_vsite_f(fplog,vsite,x,f,fr->fshift,FALSE,NULL,nrnb,
1828                            &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1829             wallcycle_stop(wcycle,ewcVSITESPREAD);
1830
1831             if (bSepLRF)
1832             {
1833                 wallcycle_start(wcycle,ewcVSITESPREAD);
1834                 spread_vsite_f(fplog,vsite,x,fr->f_twin,NULL,FALSE,NULL,
1835                                nrnb,
1836                                &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
1837                 wallcycle_stop(wcycle,ewcVSITESPREAD);
1838             }
1839         }
1840
1841         if (flags & GMX_FORCE_VIRIAL)
1842         {
1843             /* Calculation of the virial must be done after vsites! */
1844             calc_virial(fplog,mdatoms->start,mdatoms->homenr,x,f,
1845                         vir_force,graph,box,nrnb,fr,inputrec->ePBC);
1846         }
1847     }
1848
1849     if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1850     {
1851         pull_potential_wrapper(fplog,bSepDVDL,cr,inputrec,box,x,
1852                                f,vir_force,mdatoms,enerd,lambda,t);
1853     }
1854
1855     /* Add the forces from enforced rotation potentials (if any) */
1856     if (inputrec->bRot)
1857     {
1858         wallcycle_start(wcycle,ewcROTadd);
1859         enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr,step,t);
1860         wallcycle_stop(wcycle,ewcROTadd);
1861     }
1862
1863     if (PAR(cr) && !(cr->duty & DUTY_PME))
1864     {
1865         /* In case of node-splitting, the PP nodes receive the long-range
1866          * forces, virial and energy from the PME nodes here.
1867          */
1868         pme_receive_force_ener(fplog,bSepDVDL,cr,wcycle,enerd,fr);
1869     }
1870
1871     if (bDoForces)
1872     {
1873         post_process_forces(fplog,cr,step,nrnb,wcycle,
1874                             top,box,x,f,vir_force,mdatoms,graph,fr,vsite,
1875                             flags);
1876     }
1877
1878     /* Sum the potential energy terms from group contributions */
1879     sum_epot(&(inputrec->opts),enerd);
1880 }
1881
1882 void do_force(FILE *fplog,t_commrec *cr,
1883               t_inputrec *inputrec,
1884               gmx_large_int_t step,t_nrnb *nrnb,gmx_wallcycle_t wcycle,
1885               gmx_localtop_t *top,
1886               gmx_mtop_t *mtop,
1887               gmx_groups_t *groups,
1888               matrix box,rvec x[],history_t *hist,
1889               rvec f[],
1890               tensor vir_force,
1891               t_mdatoms *mdatoms,
1892               gmx_enerdata_t *enerd,t_fcdata *fcd,
1893               real *lambda,t_graph *graph,
1894               t_forcerec *fr,
1895               gmx_vsite_t *vsite,rvec mu_tot,
1896               double t,FILE *field,gmx_edsam_t ed,
1897               gmx_bool bBornRadii,
1898               int flags)
1899 {
1900     /* modify force flag if not doing nonbonded */
1901     if (!fr->bNonbonded)
1902     {
1903         flags &= ~GMX_FORCE_NONBONDED;
1904     }
1905
1906     switch (inputrec->cutoff_scheme)
1907     {
1908         case ecutsVERLET:
1909             do_force_cutsVERLET(fplog, cr, inputrec,
1910                                 step, nrnb, wcycle,
1911                                 top, mtop,
1912                                 groups,
1913                                 box, x, hist,
1914                                 f, vir_force,
1915                                 mdatoms,
1916                                 enerd, fcd,
1917                                 lambda, graph,
1918                                 fr, fr->ic,
1919                                 vsite, mu_tot,
1920                                 t, field, ed,
1921                                 bBornRadii,
1922                                 flags);
1923             break;
1924         case ecutsGROUP:
1925              do_force_cutsGROUP(fplog, cr, inputrec,
1926                                 step, nrnb, wcycle,
1927                                 top, mtop,
1928                                 groups,
1929                                 box, x, hist,
1930                                 f, vir_force,
1931                                 mdatoms,
1932                                 enerd, fcd,
1933                                 lambda, graph,
1934                                 fr, vsite, mu_tot,
1935                                 t, field, ed,
1936                                 bBornRadii,
1937                                 flags);
1938             break;
1939         default:
1940             gmx_incons("Invalid cut-off scheme passed!");
1941     }
1942 }
1943
1944
1945 void do_constrain_first(FILE *fplog,gmx_constr_t constr,
1946                         t_inputrec *ir,t_mdatoms *md,
1947                         t_state *state,rvec *f,
1948                         t_graph *graph,t_commrec *cr,t_nrnb *nrnb,
1949                         t_forcerec *fr, gmx_localtop_t *top, tensor shake_vir)
1950 {
1951     int    i,m,start,end;
1952     gmx_large_int_t step;
1953     real   dt=ir->delta_t;
1954     real   dvdl_dum;
1955     rvec   *savex;
1956
1957     snew(savex,state->natoms);
1958
1959     start = md->start;
1960     end   = md->homenr + start;
1961
1962     if (debug)
1963         fprintf(debug,"vcm: start=%d, homenr=%d, end=%d\n",
1964                 start,md->homenr,end);
1965     /* Do a first constrain to reset particles... */
1966     step = ir->init_step;
1967     if (fplog)
1968     {
1969         char buf[STEPSTRSIZE];
1970         fprintf(fplog,"\nConstraining the starting coordinates (step %s)\n",
1971                 gmx_step_str(step,buf));
1972     }
1973     dvdl_dum = 0;
1974
1975     /* constrain the current position */
1976     constrain(NULL,TRUE,FALSE,constr,&(top->idef),
1977               ir,NULL,cr,step,0,md,
1978               state->x,state->x,NULL,
1979               fr->bMolPBC,state->box,
1980               state->lambda[efptBONDED],&dvdl_dum,
1981               NULL,NULL,nrnb,econqCoord,
1982               ir->epc==epcMTTK,state->veta,state->veta);
1983     if (EI_VV(ir->eI))
1984     {
1985         /* constrain the inital velocity, and save it */
1986         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
1987         /* might not yet treat veta correctly */
1988         constrain(NULL,TRUE,FALSE,constr,&(top->idef),
1989                   ir,NULL,cr,step,0,md,
1990                   state->x,state->v,state->v,
1991                   fr->bMolPBC,state->box,
1992                   state->lambda[efptBONDED],&dvdl_dum,
1993                   NULL,NULL,nrnb,econqVeloc,
1994                   ir->epc==epcMTTK,state->veta,state->veta);
1995     }
1996     /* constrain the inital velocities at t-dt/2 */
1997     if (EI_STATE_VELOCITY(ir->eI) && ir->eI!=eiVV)
1998     {
1999         for(i=start; (i<end); i++)
2000         {
2001             for(m=0; (m<DIM); m++)
2002             {
2003                 /* Reverse the velocity */
2004                 state->v[i][m] = -state->v[i][m];
2005                 /* Store the position at t-dt in buf */
2006                 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2007             }
2008         }
2009     /* Shake the positions at t=-dt with the positions at t=0
2010      * as reference coordinates.
2011          */
2012         if (fplog)
2013         {
2014             char buf[STEPSTRSIZE];
2015             fprintf(fplog,"\nConstraining the coordinates at t0-dt (step %s)\n",
2016                     gmx_step_str(step,buf));
2017         }
2018         dvdl_dum = 0;
2019         constrain(NULL,TRUE,FALSE,constr,&(top->idef),
2020                   ir,NULL,cr,step,-1,md,
2021                   state->x,savex,NULL,
2022                   fr->bMolPBC,state->box,
2023                   state->lambda[efptBONDED],&dvdl_dum,
2024                   state->v,NULL,nrnb,econqCoord,
2025                   ir->epc==epcMTTK,state->veta,state->veta);
2026
2027         for(i=start; i<end; i++) {
2028             for(m=0; m<DIM; m++) {
2029                 /* Re-reverse the velocities */
2030                 state->v[i][m] = -state->v[i][m];
2031             }
2032         }
2033     }
2034     sfree(savex);
2035 }
2036
2037 void calc_enervirdiff(FILE *fplog,int eDispCorr,t_forcerec *fr)
2038 {
2039   double eners[2],virs[2],enersum,virsum,y0,f,g,h;
2040   double r0,r1,r,rc3,rc9,ea,eb,ec,pa,pb,pc,pd;
2041   double invscale,invscale2,invscale3;
2042   int    ri0,ri1,ri,i,offstart,offset;
2043   real   scale,*vdwtab;
2044
2045   fr->enershiftsix = 0;
2046   fr->enershifttwelve = 0;
2047   fr->enerdiffsix = 0;
2048   fr->enerdifftwelve = 0;
2049   fr->virdiffsix = 0;
2050   fr->virdifftwelve = 0;
2051
2052   if (eDispCorr != edispcNO) {
2053     for(i=0; i<2; i++) {
2054       eners[i] = 0;
2055       virs[i]  = 0;
2056     }
2057     if ((fr->vdwtype == evdwSWITCH) || (fr->vdwtype == evdwSHIFT)) {
2058       if (fr->rvdw_switch == 0)
2059         gmx_fatal(FARGS,
2060                   "With dispersion correction rvdw-switch can not be zero "
2061                   "for vdw-type = %s",evdw_names[fr->vdwtype]);
2062
2063       scale  = fr->nblists[0].tab.scale;
2064       vdwtab = fr->nblists[0].vdwtab;
2065
2066       /* Round the cut-offs to exact table values for precision */
2067       ri0 = floor(fr->rvdw_switch*scale);
2068       ri1 = ceil(fr->rvdw*scale);
2069       r0  = ri0/scale;
2070       r1  = ri1/scale;
2071       rc3 = r0*r0*r0;
2072       rc9  = rc3*rc3*rc3;
2073
2074       if (fr->vdwtype == evdwSHIFT) {
2075         /* Determine the constant energy shift below rvdw_switch */
2076         fr->enershiftsix    = (real)(-1.0/(rc3*rc3)) - vdwtab[8*ri0];
2077         fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - vdwtab[8*ri0 + 4];
2078       }
2079       /* Add the constant part from 0 to rvdw_switch.
2080        * This integration from 0 to rvdw_switch overcounts the number
2081        * of interactions by 1, as it also counts the self interaction.
2082        * We will correct for this later.
2083        */
2084       eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2085       eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2086
2087       invscale = 1.0/(scale);
2088       invscale2 = invscale*invscale;
2089       invscale3 = invscale*invscale2;
2090
2091       /* following summation derived from cubic spline definition,
2092         Numerical Recipies in C, second edition, p. 113-116.  Exact
2093         for the cubic spline.  We first calculate the negative of
2094         the energy from rvdw to rvdw_switch, assuming that g(r)=1,
2095         and then add the more standard, abrupt cutoff correction to
2096         that result, yielding the long-range correction for a
2097         switched function.  We perform both the pressure and energy
2098         loops at the same time for simplicity, as the computational
2099         cost is low. */
2100
2101       for (i=0;i<2;i++) {
2102         enersum = 0.0; virsum = 0.0;
2103         if (i==0)
2104           offstart = 0;
2105         else
2106           offstart = 4;
2107         for (ri=ri0; ri<ri1; ri++) {
2108           r = ri*invscale;
2109           ea = invscale3;
2110           eb = 2.0*invscale2*r;
2111           ec = invscale*r*r;
2112
2113           pa = invscale3;
2114           pb = 3.0*invscale2*r;
2115           pc = 3.0*invscale*r*r;
2116           pd = r*r*r;
2117
2118           /* this "8" is from the packing in the vdwtab array - perhaps
2119             should be #define'ed? */
2120           offset = 8*ri + offstart;
2121           y0 = vdwtab[offset];
2122           f = vdwtab[offset+1];
2123           g = vdwtab[offset+2];
2124           h = vdwtab[offset+3];
2125
2126           enersum += y0*(ea/3 + eb/2 + ec) + f*(ea/4 + eb/3 + ec/2)+
2127             g*(ea/5 + eb/4 + ec/3) + h*(ea/6 + eb/5 + ec/4);
2128           virsum  +=  f*(pa/4 + pb/3 + pc/2 + pd) +
2129             2*g*(pa/5 + pb/4 + pc/3 + pd/2) + 3*h*(pa/6 + pb/5 + pc/4 + pd/3);
2130
2131         }
2132         enersum *= 4.0*M_PI;
2133         virsum  *= 4.0*M_PI;
2134         eners[i] -= enersum;
2135         virs[i]  -= virsum;
2136       }
2137
2138       /* now add the correction for rvdw_switch to infinity */
2139       eners[0] += -4.0*M_PI/(3.0*rc3);
2140       eners[1] +=  4.0*M_PI/(9.0*rc9);
2141       virs[0]  +=  8.0*M_PI/rc3;
2142       virs[1]  += -16.0*M_PI/(3.0*rc9);
2143     }
2144     else if ((fr->vdwtype == evdwCUT) || (fr->vdwtype == evdwUSER)) {
2145       if (fr->vdwtype == evdwUSER && fplog)
2146         fprintf(fplog,
2147                 "WARNING: using dispersion correction with user tables\n");
2148       rc3  = fr->rvdw*fr->rvdw*fr->rvdw;
2149       rc9  = rc3*rc3*rc3;
2150       /* Contribution beyond the cut-off */
2151       eners[0] += -4.0*M_PI/(3.0*rc3);
2152       eners[1] +=  4.0*M_PI/(9.0*rc9);
2153       if (fr->vdw_pot_shift) {
2154           /* Contribution within the cut-off */
2155           eners[0] += -4.0*M_PI/(3.0*rc3);
2156           eners[1] +=  4.0*M_PI/(3.0*rc9);
2157       }
2158       /* Contribution beyond the cut-off */
2159       virs[0]  +=  8.0*M_PI/rc3;
2160       virs[1]  += -16.0*M_PI/(3.0*rc9);
2161     } else {
2162       gmx_fatal(FARGS,
2163                 "Dispersion correction is not implemented for vdw-type = %s",
2164                 evdw_names[fr->vdwtype]);
2165     }
2166     fr->enerdiffsix    = eners[0];
2167     fr->enerdifftwelve = eners[1];
2168     /* The 0.5 is due to the Gromacs definition of the virial */
2169     fr->virdiffsix     = 0.5*virs[0];
2170     fr->virdifftwelve  = 0.5*virs[1];
2171   }
2172 }
2173
2174 void calc_dispcorr(FILE *fplog,t_inputrec *ir,t_forcerec *fr,
2175                    gmx_large_int_t step,int natoms,
2176                    matrix box,real lambda,tensor pres,tensor virial,
2177                    real *prescorr, real *enercorr, real *dvdlcorr)
2178 {
2179     gmx_bool bCorrAll,bCorrPres;
2180     real dvdlambda,invvol,dens,ninter,avcsix,avctwelve,enerdiff,svir=0,spres=0;
2181     int  m;
2182
2183     *prescorr = 0;
2184     *enercorr = 0;
2185     *dvdlcorr = 0;
2186
2187     clear_mat(virial);
2188     clear_mat(pres);
2189
2190     if (ir->eDispCorr != edispcNO) {
2191         bCorrAll  = (ir->eDispCorr == edispcAllEner ||
2192                      ir->eDispCorr == edispcAllEnerPres);
2193         bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2194                      ir->eDispCorr == edispcAllEnerPres);
2195
2196         invvol = 1/det(box);
2197         if (fr->n_tpi)
2198         {
2199             /* Only correct for the interactions with the inserted molecule */
2200             dens = (natoms - fr->n_tpi)*invvol;
2201             ninter = fr->n_tpi;
2202         }
2203         else
2204         {
2205             dens = natoms*invvol;
2206             ninter = 0.5*natoms;
2207         }
2208
2209         if (ir->efep == efepNO)
2210         {
2211             avcsix    = fr->avcsix[0];
2212             avctwelve = fr->avctwelve[0];
2213         }
2214         else
2215         {
2216             avcsix    = (1 - lambda)*fr->avcsix[0]    + lambda*fr->avcsix[1];
2217             avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2218         }
2219
2220         enerdiff = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2221         *enercorr += avcsix*enerdiff;
2222         dvdlambda = 0.0;
2223         if (ir->efep != efepNO)
2224         {
2225             dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2226         }
2227         if (bCorrAll)
2228         {
2229             enerdiff = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2230             *enercorr += avctwelve*enerdiff;
2231             if (fr->efep != efepNO)
2232             {
2233                 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2234             }
2235         }
2236
2237         if (bCorrPres)
2238         {
2239             svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2240             if (ir->eDispCorr == edispcAllEnerPres)
2241             {
2242                 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2243             }
2244             /* The factor 2 is because of the Gromacs virial definition */
2245             spres = -2.0*invvol*svir*PRESFAC;
2246
2247             for(m=0; m<DIM; m++) {
2248                 virial[m][m] += svir;
2249                 pres[m][m] += spres;
2250             }
2251             *prescorr += spres;
2252         }
2253
2254         /* Can't currently control when it prints, for now, just print when degugging */
2255         if (debug)
2256         {
2257             if (bCorrAll) {
2258                 fprintf(debug,"Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2259                         avcsix,avctwelve);
2260             }
2261             if (bCorrPres)
2262             {
2263                 fprintf(debug,
2264                         "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2265                         *enercorr,spres,svir);
2266             }
2267             else
2268             {
2269                 fprintf(debug,"Long Range LJ corr.: Epot %10g\n",*enercorr);
2270             }
2271         }
2272
2273         if (fr->bSepDVDL && do_per_step(step,ir->nstlog))
2274         {
2275             fprintf(fplog,sepdvdlformat,"Dispersion correction",
2276                     *enercorr,dvdlambda);
2277         }
2278         if (fr->efep != efepNO)
2279         {
2280             *dvdlcorr += dvdlambda;
2281         }
2282     }
2283 }
2284
2285 void do_pbc_first(FILE *fplog,matrix box,t_forcerec *fr,
2286                   t_graph *graph,rvec x[])
2287 {
2288   if (fplog)
2289     fprintf(fplog,"Removing pbc first time\n");
2290   calc_shifts(box,fr->shift_vec);
2291   if (graph) {
2292     mk_mshift(fplog,graph,fr->ePBC,box,x);
2293     if (gmx_debug_at)
2294       p_graph(debug,"do_pbc_first 1",graph);
2295     shift_self(graph,box,x);
2296     /* By doing an extra mk_mshift the molecules that are broken
2297      * because they were e.g. imported from another software
2298      * will be made whole again. Such are the healing powers
2299      * of GROMACS.
2300      */
2301     mk_mshift(fplog,graph,fr->ePBC,box,x);
2302     if (gmx_debug_at)
2303       p_graph(debug,"do_pbc_first 2",graph);
2304   }
2305   if (fplog)
2306     fprintf(fplog,"Done rmpbc\n");
2307 }
2308
2309 static void low_do_pbc_mtop(FILE *fplog,int ePBC,matrix box,
2310                             gmx_mtop_t *mtop,rvec x[],
2311                             gmx_bool bFirst)
2312 {
2313   t_graph *graph;
2314   int mb,as,mol;
2315   gmx_molblock_t *molb;
2316
2317   if (bFirst && fplog)
2318     fprintf(fplog,"Removing pbc first time\n");
2319
2320   snew(graph,1);
2321   as = 0;
2322   for(mb=0; mb<mtop->nmolblock; mb++) {
2323     molb = &mtop->molblock[mb];
2324     if (molb->natoms_mol == 1 ||
2325         (!bFirst && mtop->moltype[molb->type].cgs.nr == 1)) {
2326       /* Just one atom or charge group in the molecule, no PBC required */
2327       as += molb->nmol*molb->natoms_mol;
2328     } else {
2329       /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2330       mk_graph_ilist(NULL,mtop->moltype[molb->type].ilist,
2331                      0,molb->natoms_mol,FALSE,FALSE,graph);
2332
2333       for(mol=0; mol<molb->nmol; mol++) {
2334         mk_mshift(fplog,graph,ePBC,box,x+as);
2335
2336         shift_self(graph,box,x+as);
2337         /* The molecule is whole now.
2338          * We don't need the second mk_mshift call as in do_pbc_first,
2339          * since we no longer need this graph.
2340          */
2341
2342         as += molb->natoms_mol;
2343       }
2344       done_graph(graph);
2345     }
2346   }
2347   sfree(graph);
2348 }
2349
2350 void do_pbc_first_mtop(FILE *fplog,int ePBC,matrix box,
2351                        gmx_mtop_t *mtop,rvec x[])
2352 {
2353   low_do_pbc_mtop(fplog,ePBC,box,mtop,x,TRUE);
2354 }
2355
2356 void do_pbc_mtop(FILE *fplog,int ePBC,matrix box,
2357                  gmx_mtop_t *mtop,rvec x[])
2358 {
2359   low_do_pbc_mtop(fplog,ePBC,box,mtop,x,FALSE);
2360 }
2361
2362 void finish_run(FILE *fplog,t_commrec *cr,const char *confout,
2363                 t_inputrec *inputrec,
2364                 t_nrnb nrnb[],gmx_wallcycle_t wcycle,
2365                 gmx_runtime_t *runtime,
2366                 wallclock_gpu_t *gputimes,
2367                 int omp_nth_pp,
2368                 gmx_bool bWriteStat)
2369 {
2370     int    i,j;
2371     t_nrnb *nrnb_tot=NULL;
2372     real   delta_t;
2373     double nbfs,mflop;
2374
2375     wallcycle_sum(cr,wcycle);
2376
2377     if (cr->nnodes > 1)
2378     {
2379         snew(nrnb_tot,1);
2380 #ifdef GMX_MPI
2381         MPI_Allreduce(nrnb->n,nrnb_tot->n,eNRNB,MPI_DOUBLE,MPI_SUM,
2382                       cr->mpi_comm_mysim);
2383 #endif
2384     }
2385     else
2386     {
2387         nrnb_tot = nrnb;
2388     }
2389
2390 #if defined(GMX_MPI) && !defined(GMX_THREAD_MPI)
2391     if (cr->nnodes > 1)
2392     {
2393         /* reduce nodetime over all MPI processes in the current simulation */
2394         double sum;
2395         MPI_Allreduce(&runtime->proctime,&sum,1,MPI_DOUBLE,MPI_SUM,
2396                       cr->mpi_comm_mysim);
2397         runtime->proctime = sum;
2398     }
2399 #endif
2400
2401     if (SIMMASTER(cr))
2402     {
2403         print_flop(fplog,nrnb_tot,&nbfs,&mflop);
2404     }
2405     if (cr->nnodes > 1)
2406     {
2407         sfree(nrnb_tot);
2408     }
2409
2410     if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2411     {
2412         print_dd_statistics(cr,inputrec,fplog);
2413     }
2414
2415 #ifdef GMX_MPI
2416     if (PARTDECOMP(cr))
2417     {
2418         if (MASTER(cr))
2419         {
2420             t_nrnb     *nrnb_all;
2421             int        s;
2422             MPI_Status stat;
2423
2424             snew(nrnb_all,cr->nnodes);
2425             nrnb_all[0] = *nrnb;
2426             for(s=1; s<cr->nnodes; s++)
2427             {
2428                 MPI_Recv(nrnb_all[s].n,eNRNB,MPI_DOUBLE,s,0,
2429                          cr->mpi_comm_mysim,&stat);
2430             }
2431             pr_load(fplog,cr,nrnb_all);
2432             sfree(nrnb_all);
2433         }
2434         else
2435         {
2436             MPI_Send(nrnb->n,eNRNB,MPI_DOUBLE,MASTERRANK(cr),0,
2437                      cr->mpi_comm_mysim);
2438         }
2439     }
2440 #endif
2441
2442     if (SIMMASTER(cr))
2443     {
2444         wallcycle_print(fplog,cr->nnodes,cr->npmenodes,runtime->realtime,
2445                         wcycle,gputimes);
2446
2447         if (EI_DYNAMICS(inputrec->eI))
2448         {
2449             delta_t = inputrec->delta_t;
2450         }
2451         else
2452         {
2453             delta_t = 0;
2454         }
2455
2456         if (fplog)
2457         {
2458             print_perf(fplog,runtime->proctime,runtime->realtime,
2459                        cr->nnodes-cr->npmenodes,
2460                        runtime->nsteps_done,delta_t,nbfs,mflop,
2461                        omp_nth_pp);
2462         }
2463         if (bWriteStat)
2464         {
2465             print_perf(stderr,runtime->proctime,runtime->realtime,
2466                        cr->nnodes-cr->npmenodes,
2467                        runtime->nsteps_done,delta_t,nbfs,mflop,
2468                        omp_nth_pp);
2469         }
2470     }
2471 }
2472
2473 extern void initialize_lambdas(FILE *fplog,t_inputrec *ir,int *fep_state,real *lambda,double *lam0)
2474 {
2475     /* this function works, but could probably use a logic rewrite to keep all the different
2476        types of efep straight. */
2477
2478     int i;
2479     t_lambda *fep = ir->fepvals;
2480
2481     if ((ir->efep==efepNO) && (ir->bSimTemp == FALSE)) {
2482         for (i=0;i<efptNR;i++)  {
2483             lambda[i] = 0.0;
2484             if (lam0)
2485             {
2486                 lam0[i] = 0.0;
2487             }
2488         }
2489         return;
2490     } else {
2491         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2492                                              if checkpoint is set -- a kludge is in for now
2493                                              to prevent this.*/
2494         for (i=0;i<efptNR;i++)
2495         {
2496             /* overwrite lambda state with init_lambda for now for backwards compatibility */
2497             if (fep->init_lambda>=0) /* if it's -1, it was never initializd */
2498             {
2499                 lambda[i] = fep->init_lambda;
2500                 if (lam0) {
2501                     lam0[i] = lambda[i];
2502                 }
2503             }
2504             else
2505             {
2506                 lambda[i] = fep->all_lambda[i][*fep_state];
2507                 if (lam0) {
2508                     lam0[i] = lambda[i];
2509                 }
2510             }
2511         }
2512         if (ir->bSimTemp) {
2513             /* need to rescale control temperatures to match current state */
2514             for (i=0;i<ir->opts.ngtc;i++) {
2515                 if (ir->opts.ref_t[i] > 0) {
2516                     ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2517                 }
2518             }
2519         }
2520     }
2521
2522     /* Send to the log the information on the current lambdas */
2523     if (fplog != NULL)
2524     {
2525         fprintf(fplog,"Initial vector of lambda components:[ ");
2526         for (i=0;i<efptNR;i++)
2527         {
2528             fprintf(fplog,"%10.4f ",lambda[i]);
2529         }
2530         fprintf(fplog,"]\n");
2531     }
2532     return;
2533 }
2534
2535
2536 void init_md(FILE *fplog,
2537              t_commrec *cr,t_inputrec *ir,const output_env_t oenv,
2538              double *t,double *t0,
2539              real *lambda, int *fep_state, double *lam0,
2540              t_nrnb *nrnb,gmx_mtop_t *mtop,
2541              gmx_update_t *upd,
2542              int nfile,const t_filenm fnm[],
2543              gmx_mdoutf_t **outf,t_mdebin **mdebin,
2544              tensor force_vir,tensor shake_vir,rvec mu_tot,
2545              gmx_bool *bSimAnn,t_vcm **vcm, t_state *state, unsigned long Flags)
2546 {
2547     int  i,j,n;
2548     real tmpt,mod;
2549
2550     /* Initial values */
2551     *t = *t0       = ir->init_t;
2552
2553     *bSimAnn=FALSE;
2554     for(i=0;i<ir->opts.ngtc;i++)
2555     {
2556         /* set bSimAnn if any group is being annealed */
2557         if(ir->opts.annealing[i]!=eannNO)
2558         {
2559             *bSimAnn = TRUE;
2560         }
2561     }
2562     if (*bSimAnn)
2563     {
2564         update_annealing_target_temp(&(ir->opts),ir->init_t);
2565     }
2566
2567     /* Initialize lambda variables */
2568     initialize_lambdas(fplog,ir,fep_state,lambda,lam0);
2569
2570     if (upd)
2571     {
2572         *upd = init_update(fplog,ir);
2573     }
2574
2575
2576     if (vcm != NULL)
2577     {
2578         *vcm = init_vcm(fplog,&mtop->groups,ir);
2579     }
2580
2581     if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2582     {
2583         if (ir->etc == etcBERENDSEN)
2584         {
2585             please_cite(fplog,"Berendsen84a");
2586         }
2587         if (ir->etc == etcVRESCALE)
2588         {
2589             please_cite(fplog,"Bussi2007a");
2590         }
2591     }
2592
2593     init_nrnb(nrnb);
2594
2595     if (nfile != -1)
2596     {
2597         *outf = init_mdoutf(nfile,fnm,Flags,cr,ir,oenv);
2598
2599         *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : (*outf)->fp_ene,
2600                               mtop,ir, (*outf)->fp_dhdl);
2601     }
2602
2603     if (ir->bAdress)
2604     {
2605       please_cite(fplog,"Fritsch12");
2606       please_cite(fplog,"Junghans10");
2607     }
2608     /* Initiate variables */
2609     clear_mat(force_vir);
2610     clear_mat(shake_vir);
2611     clear_rvec(mu_tot);
2612
2613     debug_gmx();
2614 }
2615