src/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecRFCut_VdwBhamSw_GeomW4P1_c.c

   1 /*
   2  * Note: this file was generated by the Gromacs c kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 /*
  34  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwBhamSw_GeomW4P1_VF_c
  35  * Electrostatics interaction: ReactionField
  36  * VdW interaction:            Buckingham
  37  * Geometry:                   Water4-Particle
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecRFCut_VdwBhamSw_GeomW4P1_VF_c
  42                     (t_nblist * gmx_restrict                nlist,
  43                      rvec * gmx_restrict                    xx,
  44                      rvec * gmx_restrict                    ff,
  45                      t_forcerec * gmx_restrict              fr,
  46                      t_mdatoms * gmx_restrict               mdatoms,
  47                      nb_kernel_data_t * gmx_restrict        kernel_data,
  48                      t_nrnb * gmx_restrict                  nrnb)
  49 {
  50     int              i_shift_offset,i_coord_offset,j_coord_offset;
  51     int              j_index_start,j_index_end;
  52     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  53     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  54     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  55     real             *shiftvec,*fshift,*x,*f;
  56     int              vdwioffset0;
  57     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  58     int              vdwioffset1;
  59     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  60     int              vdwioffset2;
  61     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  62     int              vdwioffset3;
  63     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  64     int              vdwjidx0;
  65     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  66     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  67     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  68     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  69     real             dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30,cexp1_30,cexp2_30;
  70     real             velec,felec,velecsum,facel,crf,krf,krf2;
  71     real             *charge;
  72     int              nvdwtype;
  73     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  74     int              *vdwtype;
  75     real             *vdwparam;
  76     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  77
  78     x                = xx[0];
  79     f                = ff[0];
  80
  81     nri              = nlist->nri;
  82     iinr             = nlist->iinr;
  83     jindex           = nlist->jindex;
  84     jjnr             = nlist->jjnr;
  85     shiftidx         = nlist->shift;
  86     gid              = nlist->gid;
  87     shiftvec         = fr->shift_vec[0];
  88     fshift           = fr->fshift[0];
  89     facel            = fr->epsfac;
  90     charge           = mdatoms->chargeA;
  91     krf              = fr->ic->k_rf;
  92     krf2             = krf*2.0;
  93     crf              = fr->ic->c_rf;
  94     nvdwtype         = fr->ntype;
  95     vdwparam         = fr->nbfp;
  96     vdwtype          = mdatoms->typeA;
  97
  98     /* Setup water-specific parameters */
  99     inr              = nlist->iinr[0];
 100     iq1              = facel*charge[inr+1];
 101     iq2              = facel*charge[inr+2];
 102     iq3              = facel*charge[inr+3];
 103     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 104
 105     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 106     rcutoff          = fr->rcoulomb;
 107     rcutoff2         = rcutoff*rcutoff;
 108
 109     rswitch          = fr->rvdw_switch;
 110     /* Setup switch parameters */
 111     d                = rcutoff-rswitch;
 112     swV3             = -10.0/(d*d*d);
 113     swV4             =  15.0/(d*d*d*d);
 114     swV5             =  -6.0/(d*d*d*d*d);
 115     swF2             = -30.0/(d*d*d);
 116     swF3             =  60.0/(d*d*d*d);
 117     swF4             = -30.0/(d*d*d*d*d);
 118
 119     outeriter        = 0;
 120     inneriter        = 0;
 121
 122     /* Start outer loop over neighborlists */
 123     for(iidx=0; iidx<nri; iidx++)
 124     {
 125         /* Load shift vector for this list */
 126         i_shift_offset   = DIM*shiftidx[iidx];
 127         shX              = shiftvec[i_shift_offset+XX];
 128         shY              = shiftvec[i_shift_offset+YY];
 129         shZ              = shiftvec[i_shift_offset+ZZ];
 130
 131         /* Load limits for loop over neighbors */
 132         j_index_start    = jindex[iidx];
 133         j_index_end      = jindex[iidx+1];
 134
 135         /* Get outer coordinate index */
 136         inr              = iinr[iidx];
 137         i_coord_offset   = DIM*inr;
 138
 139         /* Load i particle coords and add shift vector */
 140         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 141         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 142         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 143         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 144         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 145         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 146         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 147         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 148         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 149         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 150         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 151         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 152
 153         fix0             = 0.0;
 154         fiy0             = 0.0;
 155         fiz0             = 0.0;
 156         fix1             = 0.0;
 157         fiy1             = 0.0;
 158         fiz1             = 0.0;
 159         fix2             = 0.0;
 160         fiy2             = 0.0;
 161         fiz2             = 0.0;
 162         fix3             = 0.0;
 163         fiy3             = 0.0;
 164         fiz3             = 0.0;
 165
 166         /* Reset potential sums */
 167         velecsum         = 0.0;
 168         vvdwsum          = 0.0;
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 172         {
 173             /* Get j neighbor index, and coordinate index */
 174             jnr              = jjnr[jidx];
 175             j_coord_offset   = DIM*jnr;
 176
 177             /* load j atom coordinates */
 178             jx0              = x[j_coord_offset+DIM*0+XX];
 179             jy0              = x[j_coord_offset+DIM*0+YY];
 180             jz0              = x[j_coord_offset+DIM*0+ZZ];
 181
 182             /* Calculate displacement vector */
 183             dx00             = ix0 - jx0;
 184             dy00             = iy0 - jy0;
 185             dz00             = iz0 - jz0;
 186             dx10             = ix1 - jx0;
 187             dy10             = iy1 - jy0;
 188             dz10             = iz1 - jz0;
 189             dx20             = ix2 - jx0;
 190             dy20             = iy2 - jy0;
 191             dz20             = iz2 - jz0;
 192             dx30             = ix3 - jx0;
 193             dy30             = iy3 - jy0;
 194             dz30             = iz3 - jz0;
 195
 196             /* Calculate squared distance and things based on it */
 197             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 198             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 199             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 200             rsq30            = dx30*dx30+dy30*dy30+dz30*dz30;
 201
 202             rinv00           = gmx_invsqrt(rsq00);
 203             rinv10           = gmx_invsqrt(rsq10);
 204             rinv20           = gmx_invsqrt(rsq20);
 205             rinv30           = gmx_invsqrt(rsq30);
 206
 207             rinvsq00         = rinv00*rinv00;
 208             rinvsq10         = rinv10*rinv10;
 209             rinvsq20         = rinv20*rinv20;
 210             rinvsq30         = rinv30*rinv30;
 211
 212             /* Load parameters for j particles */
 213             jq0              = charge[jnr+0];
 214             vdwjidx0         = 3*vdwtype[jnr+0];
 215
 216             /**************************
 217              * CALCULATE INTERACTIONS *
 218              **************************/
 219
 220             if (rsq00<rcutoff2)
 221             {
 222
 223             r00              = rsq00*rinv00;
 224
 225             c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 226             cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 227             cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 228
 229             /* BUCKINGHAM DISPERSION/REPULSION */
 230             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 231             vvdw6            = c6_00*rinvsix;
 232             br               = cexp2_00*r00;
 233             vvdwexp          = cexp1_00*exp(-br);
 234             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
 235             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 236
 237             d                = r00-rswitch;
 238             d                = (d>0.0) ? d : 0.0;
 239             d2               = d*d;
 240             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 241
 242             dsw              = d2*(swF2+d*(swF3+d*swF4));
 243
 244             /* Evaluate switch function */
 245             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 246             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 247             vvdw            *= sw;
 248
 249             /* Update potential sums from outer loop */
 250             vvdwsum         += vvdw;
 251
 252             fscal            = fvdw;
 253
 254             /* Calculate temporary vectorial force */
 255             tx               = fscal*dx00;
 256             ty               = fscal*dy00;
 257             tz               = fscal*dz00;
 258
 259             /* Update vectorial force */
 260             fix0            += tx;
 261             fiy0            += ty;
 262             fiz0            += tz;
 263             f[j_coord_offset+DIM*0+XX] -= tx;
 264             f[j_coord_offset+DIM*0+YY] -= ty;
 265             f[j_coord_offset+DIM*0+ZZ] -= tz;
 266
 267             }
 268
 269             /**************************
 270              * CALCULATE INTERACTIONS *
 271              **************************/
 272
 273             if (rsq10<rcutoff2)
 274             {
 275
 276             qq10             = iq1*jq0;
 277
 278             /* REACTION-FIELD ELECTROSTATICS */
 279             velec            = qq10*(rinv10+krf*rsq10-crf);
 280             felec            = qq10*(rinv10*rinvsq10-krf2);
 281
 282             /* Update potential sums from outer loop */
 283             velecsum        += velec;
 284
 285             fscal            = felec;
 286
 287             /* Calculate temporary vectorial force */
 288             tx               = fscal*dx10;
 289             ty               = fscal*dy10;
 290             tz               = fscal*dz10;
 291
 292             /* Update vectorial force */
 293             fix1            += tx;
 294             fiy1            += ty;
 295             fiz1            += tz;
 296             f[j_coord_offset+DIM*0+XX] -= tx;
 297             f[j_coord_offset+DIM*0+YY] -= ty;
 298             f[j_coord_offset+DIM*0+ZZ] -= tz;
 299
 300             }
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             if (rsq20<rcutoff2)
 307             {
 308
 309             qq20             = iq2*jq0;
 310
 311             /* REACTION-FIELD ELECTROSTATICS */
 312             velec            = qq20*(rinv20+krf*rsq20-crf);
 313             felec            = qq20*(rinv20*rinvsq20-krf2);
 314
 315             /* Update potential sums from outer loop */
 316             velecsum        += velec;
 317
 318             fscal            = felec;
 319
 320             /* Calculate temporary vectorial force */
 321             tx               = fscal*dx20;
 322             ty               = fscal*dy20;
 323             tz               = fscal*dz20;
 324
 325             /* Update vectorial force */
 326             fix2            += tx;
 327             fiy2            += ty;
 328             fiz2            += tz;
 329             f[j_coord_offset+DIM*0+XX] -= tx;
 330             f[j_coord_offset+DIM*0+YY] -= ty;
 331             f[j_coord_offset+DIM*0+ZZ] -= tz;
 332
 333             }
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             if (rsq30<rcutoff2)
 340             {
 341
 342             qq30             = iq3*jq0;
 343
 344             /* REACTION-FIELD ELECTROSTATICS */
 345             velec            = qq30*(rinv30+krf*rsq30-crf);
 346             felec            = qq30*(rinv30*rinvsq30-krf2);
 347
 348             /* Update potential sums from outer loop */
 349             velecsum        += velec;
 350
 351             fscal            = felec;
 352
 353             /* Calculate temporary vectorial force */
 354             tx               = fscal*dx30;
 355             ty               = fscal*dy30;
 356             tz               = fscal*dz30;
 357
 358             /* Update vectorial force */
 359             fix3            += tx;
 360             fiy3            += ty;
 361             fiz3            += tz;
 362             f[j_coord_offset+DIM*0+XX] -= tx;
 363             f[j_coord_offset+DIM*0+YY] -= ty;
 364             f[j_coord_offset+DIM*0+ZZ] -= tz;
 365
 366             }
 367
 368             /* Inner loop uses 175 flops */
 369         }
 370         /* End of innermost loop */
 371
 372         tx = ty = tz = 0;
 373         f[i_coord_offset+DIM*0+XX] += fix0;
 374         f[i_coord_offset+DIM*0+YY] += fiy0;
 375         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 376         tx                         += fix0;
 377         ty                         += fiy0;
 378         tz                         += fiz0;
 379         f[i_coord_offset+DIM*1+XX] += fix1;
 380         f[i_coord_offset+DIM*1+YY] += fiy1;
 381         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 382         tx                         += fix1;
 383         ty                         += fiy1;
 384         tz                         += fiz1;
 385         f[i_coord_offset+DIM*2+XX] += fix2;
 386         f[i_coord_offset+DIM*2+YY] += fiy2;
 387         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 388         tx                         += fix2;
 389         ty                         += fiy2;
 390         tz                         += fiz2;
 391         f[i_coord_offset+DIM*3+XX] += fix3;
 392         f[i_coord_offset+DIM*3+YY] += fiy3;
 393         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 394         tx                         += fix3;
 395         ty                         += fiy3;
 396         tz                         += fiz3;
 397         fshift[i_shift_offset+XX]  += tx;
 398         fshift[i_shift_offset+YY]  += ty;
 399         fshift[i_shift_offset+ZZ]  += tz;
 400
 401         ggid                        = gid[iidx];
 402         /* Update potential energies */
 403         kernel_data->energygrp_elec[ggid] += velecsum;
 404         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 405
 406         /* Increment number of inner iterations */
 407         inneriter                  += j_index_end - j_index_start;
 408
 409         /* Outer loop uses 41 flops */
 410     }
 411
 412     /* Increment number of outer iterations */
 413     outeriter        += nri;
 414
 415     /* Update outer/inner flops */
 416
 417     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*41 + inneriter*175);
 418 }
 419 /*
 420  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwBhamSw_GeomW4P1_F_c
 421  * Electrostatics interaction: ReactionField
 422  * VdW interaction:            Buckingham
 423  * Geometry:                   Water4-Particle
 424  * Calculate force/pot:        Force
 425  */
 426 void
 427 nb_kernel_ElecRFCut_VdwBhamSw_GeomW4P1_F_c
 428                     (t_nblist * gmx_restrict                nlist,
 429                      rvec * gmx_restrict                    xx,
 430                      rvec * gmx_restrict                    ff,
 431                      t_forcerec * gmx_restrict              fr,
 432                      t_mdatoms * gmx_restrict               mdatoms,
 433                      nb_kernel_data_t * gmx_restrict        kernel_data,
 434                      t_nrnb * gmx_restrict                  nrnb)
 435 {
 436     int              i_shift_offset,i_coord_offset,j_coord_offset;
 437     int              j_index_start,j_index_end;
 438     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 439     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 440     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 441     real             *shiftvec,*fshift,*x,*f;
 442     int              vdwioffset0;
 443     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 444     int              vdwioffset1;
 445     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 446     int              vdwioffset2;
 447     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 448     int              vdwioffset3;
 449     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 450     int              vdwjidx0;
 451     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 452     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 453     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 454     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 455     real             dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30,cexp1_30,cexp2_30;
 456     real             velec,felec,velecsum,facel,crf,krf,krf2;
 457     real             *charge;
 458     int              nvdwtype;
 459     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 460     int              *vdwtype;
 461     real             *vdwparam;
 462     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 463
 464     x                = xx[0];
 465     f                = ff[0];
 466
 467     nri              = nlist->nri;
 468     iinr             = nlist->iinr;
 469     jindex           = nlist->jindex;
 470     jjnr             = nlist->jjnr;
 471     shiftidx         = nlist->shift;
 472     gid              = nlist->gid;
 473     shiftvec         = fr->shift_vec[0];
 474     fshift           = fr->fshift[0];
 475     facel            = fr->epsfac;
 476     charge           = mdatoms->chargeA;
 477     krf              = fr->ic->k_rf;
 478     krf2             = krf*2.0;
 479     crf              = fr->ic->c_rf;
 480     nvdwtype         = fr->ntype;
 481     vdwparam         = fr->nbfp;
 482     vdwtype          = mdatoms->typeA;
 483
 484     /* Setup water-specific parameters */
 485     inr              = nlist->iinr[0];
 486     iq1              = facel*charge[inr+1];
 487     iq2              = facel*charge[inr+2];
 488     iq3              = facel*charge[inr+3];
 489     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 490
 491     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 492     rcutoff          = fr->rcoulomb;
 493     rcutoff2         = rcutoff*rcutoff;
 494
 495     rswitch          = fr->rvdw_switch;
 496     /* Setup switch parameters */
 497     d                = rcutoff-rswitch;
 498     swV3             = -10.0/(d*d*d);
 499     swV4             =  15.0/(d*d*d*d);
 500     swV5             =  -6.0/(d*d*d*d*d);
 501     swF2             = -30.0/(d*d*d);
 502     swF3             =  60.0/(d*d*d*d);
 503     swF4             = -30.0/(d*d*d*d*d);
 504
 505     outeriter        = 0;
 506     inneriter        = 0;
 507
 508     /* Start outer loop over neighborlists */
 509     for(iidx=0; iidx<nri; iidx++)
 510     {
 511         /* Load shift vector for this list */
 512         i_shift_offset   = DIM*shiftidx[iidx];
 513         shX              = shiftvec[i_shift_offset+XX];
 514         shY              = shiftvec[i_shift_offset+YY];
 515         shZ              = shiftvec[i_shift_offset+ZZ];
 516
 517         /* Load limits for loop over neighbors */
 518         j_index_start    = jindex[iidx];
 519         j_index_end      = jindex[iidx+1];
 520
 521         /* Get outer coordinate index */
 522         inr              = iinr[iidx];
 523         i_coord_offset   = DIM*inr;
 524
 525         /* Load i particle coords and add shift vector */
 526         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 527         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 528         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 529         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 530         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 531         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 532         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 533         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 534         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 535         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 536         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 537         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 538
 539         fix0             = 0.0;
 540         fiy0             = 0.0;
 541         fiz0             = 0.0;
 542         fix1             = 0.0;
 543         fiy1             = 0.0;
 544         fiz1             = 0.0;
 545         fix2             = 0.0;
 546         fiy2             = 0.0;
 547         fiz2             = 0.0;
 548         fix3             = 0.0;
 549         fiy3             = 0.0;
 550         fiz3             = 0.0;
 551
 552         /* Start inner kernel loop */
 553         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 554         {
 555             /* Get j neighbor index, and coordinate index */
 556             jnr              = jjnr[jidx];
 557             j_coord_offset   = DIM*jnr;
 558
 559             /* load j atom coordinates */
 560             jx0              = x[j_coord_offset+DIM*0+XX];
 561             jy0              = x[j_coord_offset+DIM*0+YY];
 562             jz0              = x[j_coord_offset+DIM*0+ZZ];
 563
 564             /* Calculate displacement vector */
 565             dx00             = ix0 - jx0;
 566             dy00             = iy0 - jy0;
 567             dz00             = iz0 - jz0;
 568             dx10             = ix1 - jx0;
 569             dy10             = iy1 - jy0;
 570             dz10             = iz1 - jz0;
 571             dx20             = ix2 - jx0;
 572             dy20             = iy2 - jy0;
 573             dz20             = iz2 - jz0;
 574             dx30             = ix3 - jx0;
 575             dy30             = iy3 - jy0;
 576             dz30             = iz3 - jz0;
 577
 578             /* Calculate squared distance and things based on it */
 579             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 580             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 581             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 582             rsq30            = dx30*dx30+dy30*dy30+dz30*dz30;
 583
 584             rinv00           = gmx_invsqrt(rsq00);
 585             rinv10           = gmx_invsqrt(rsq10);
 586             rinv20           = gmx_invsqrt(rsq20);
 587             rinv30           = gmx_invsqrt(rsq30);
 588
 589             rinvsq00         = rinv00*rinv00;
 590             rinvsq10         = rinv10*rinv10;
 591             rinvsq20         = rinv20*rinv20;
 592             rinvsq30         = rinv30*rinv30;
 593
 594             /* Load parameters for j particles */
 595             jq0              = charge[jnr+0];
 596             vdwjidx0         = 3*vdwtype[jnr+0];
 597
 598             /**************************
 599              * CALCULATE INTERACTIONS *
 600              **************************/
 601
 602             if (rsq00<rcutoff2)
 603             {
 604
 605             r00              = rsq00*rinv00;
 606
 607             c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 608             cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 609             cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 610
 611             /* BUCKINGHAM DISPERSION/REPULSION */
 612             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 613             vvdw6            = c6_00*rinvsix;
 614             br               = cexp2_00*r00;
 615             vvdwexp          = cexp1_00*exp(-br);
 616             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
 617             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 618
 619             d                = r00-rswitch;
 620             d                = (d>0.0) ? d : 0.0;
 621             d2               = d*d;
 622             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 623
 624             dsw              = d2*(swF2+d*(swF3+d*swF4));
 625
 626             /* Evaluate switch function */
 627             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 628             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 629
 630             fscal            = fvdw;
 631
 632             /* Calculate temporary vectorial force */
 633             tx               = fscal*dx00;
 634             ty               = fscal*dy00;
 635             tz               = fscal*dz00;
 636
 637             /* Update vectorial force */
 638             fix0            += tx;
 639             fiy0            += ty;
 640             fiz0            += tz;
 641             f[j_coord_offset+DIM*0+XX] -= tx;
 642             f[j_coord_offset+DIM*0+YY] -= ty;
 643             f[j_coord_offset+DIM*0+ZZ] -= tz;
 644
 645             }
 646
 647             /**************************
 648              * CALCULATE INTERACTIONS *
 649              **************************/
 650
 651             if (rsq10<rcutoff2)
 652             {
 653
 654             qq10             = iq1*jq0;
 655
 656             /* REACTION-FIELD ELECTROSTATICS */
 657             felec            = qq10*(rinv10*rinvsq10-krf2);
 658
 659             fscal            = felec;
 660
 661             /* Calculate temporary vectorial force */
 662             tx               = fscal*dx10;
 663             ty               = fscal*dy10;
 664             tz               = fscal*dz10;
 665
 666             /* Update vectorial force */
 667             fix1            += tx;
 668             fiy1            += ty;
 669             fiz1            += tz;
 670             f[j_coord_offset+DIM*0+XX] -= tx;
 671             f[j_coord_offset+DIM*0+YY] -= ty;
 672             f[j_coord_offset+DIM*0+ZZ] -= tz;
 673
 674             }
 675
 676             /**************************
 677              * CALCULATE INTERACTIONS *
 678              **************************/
 679
 680             if (rsq20<rcutoff2)
 681             {
 682
 683             qq20             = iq2*jq0;
 684
 685             /* REACTION-FIELD ELECTROSTATICS */
 686             felec            = qq20*(rinv20*rinvsq20-krf2);
 687
 688             fscal            = felec;
 689
 690             /* Calculate temporary vectorial force */
 691             tx               = fscal*dx20;
 692             ty               = fscal*dy20;
 693             tz               = fscal*dz20;
 694
 695             /* Update vectorial force */
 696             fix2            += tx;
 697             fiy2            += ty;
 698             fiz2            += tz;
 699             f[j_coord_offset+DIM*0+XX] -= tx;
 700             f[j_coord_offset+DIM*0+YY] -= ty;
 701             f[j_coord_offset+DIM*0+ZZ] -= tz;
 702
 703             }
 704
 705             /**************************
 706              * CALCULATE INTERACTIONS *
 707              **************************/
 708
 709             if (rsq30<rcutoff2)
 710             {
 711
 712             qq30             = iq3*jq0;
 713
 714             /* REACTION-FIELD ELECTROSTATICS */
 715             felec            = qq30*(rinv30*rinvsq30-krf2);
 716
 717             fscal            = felec;
 718
 719             /* Calculate temporary vectorial force */
 720             tx               = fscal*dx30;
 721             ty               = fscal*dy30;
 722             tz               = fscal*dz30;
 723
 724             /* Update vectorial force */
 725             fix3            += tx;
 726             fiy3            += ty;
 727             fiz3            += tz;
 728             f[j_coord_offset+DIM*0+XX] -= tx;
 729             f[j_coord_offset+DIM*0+YY] -= ty;
 730             f[j_coord_offset+DIM*0+ZZ] -= tz;
 731
 732             }
 733
 734             /* Inner loop uses 158 flops */
 735         }
 736         /* End of innermost loop */
 737
 738         tx = ty = tz = 0;
 739         f[i_coord_offset+DIM*0+XX] += fix0;
 740         f[i_coord_offset+DIM*0+YY] += fiy0;
 741         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 742         tx                         += fix0;
 743         ty                         += fiy0;
 744         tz                         += fiz0;
 745         f[i_coord_offset+DIM*1+XX] += fix1;
 746         f[i_coord_offset+DIM*1+YY] += fiy1;
 747         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 748         tx                         += fix1;
 749         ty                         += fiy1;
 750         tz                         += fiz1;
 751         f[i_coord_offset+DIM*2+XX] += fix2;
 752         f[i_coord_offset+DIM*2+YY] += fiy2;
 753         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 754         tx                         += fix2;
 755         ty                         += fiy2;
 756         tz                         += fiz2;
 757         f[i_coord_offset+DIM*3+XX] += fix3;
 758         f[i_coord_offset+DIM*3+YY] += fiy3;
 759         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 760         tx                         += fix3;
 761         ty                         += fiy3;
 762         tz                         += fiz3;
 763         fshift[i_shift_offset+XX]  += tx;
 764         fshift[i_shift_offset+YY]  += ty;
 765         fshift[i_shift_offset+ZZ]  += tz;
 766
 767         /* Increment number of inner iterations */
 768         inneriter                  += j_index_end - j_index_start;
 769
 770         /* Outer loop uses 39 flops */
 771     }
 772
 773     /* Increment number of outer iterations */
 774     outeriter        += nri;
 775
 776     /* Update outer/inner flops */
 777
 778     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*39 + inneriter*158);
 779 }