src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSw_VdwBhamSw_GeomW4W4_c.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS c kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwBhamSw_GeomW4W4_VF_c
  49  * Electrostatics interaction: Ewald
  50  * VdW interaction:            Buckingham
  51  * Geometry:                   Water4-Water4
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecEwSw_VdwBhamSw_GeomW4W4_VF_c
  56                     (t_nblist                    * gmx_restrict       nlist,
  57                      rvec                        * gmx_restrict          xx,
  58                      rvec                        * gmx_restrict          ff,
  59                      t_forcerec                  * gmx_restrict          fr,
  60                      t_mdatoms                   * gmx_restrict     mdatoms,
  61                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  62                      t_nrnb                      * gmx_restrict        nrnb)
  63 {
  64     int              i_shift_offset,i_coord_offset,j_coord_offset;
  65     int              j_index_start,j_index_end;
  66     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  67     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  68     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  69     real             *shiftvec,*fshift,*x,*f;
  70     int              vdwioffset0;
  71     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwioffset1;
  73     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  74     int              vdwioffset2;
  75     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  76     int              vdwioffset3;
  77     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  78     int              vdwjidx0;
  79     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  80     int              vdwjidx1;
  81     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  82     int              vdwjidx2;
  83     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  84     int              vdwjidx3;
  85     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  86     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  87     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  88     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  89     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
  90     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  91     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  92     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
  93     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
  94     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
  95     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
  96     real             velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     int              nvdwtype;
  99     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102     int              ewitab;
 103     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 104     real             *ewtab;
 105     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 106
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = fr->epsfac;
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     sh_ewald         = fr->ic->sh_ewald;
 125     ewtab            = fr->ic->tabq_coul_FDV0;
 126     ewtabscale       = fr->ic->tabq_scale;
 127     ewtabhalfspace   = 0.5/ewtabscale;
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = facel*charge[inr+1];
 132     iq2              = facel*charge[inr+2];
 133     iq3              = facel*charge[inr+3];
 134     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 135
 136     jq1              = charge[inr+1];
 137     jq2              = charge[inr+2];
 138     jq3              = charge[inr+3];
 139     vdwjidx0         = 3*vdwtype[inr+0];
 140     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 141     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 142     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 143     qq11             = iq1*jq1;
 144     qq12             = iq1*jq2;
 145     qq13             = iq1*jq3;
 146     qq21             = iq2*jq1;
 147     qq22             = iq2*jq2;
 148     qq23             = iq2*jq3;
 149     qq31             = iq3*jq1;
 150     qq32             = iq3*jq2;
 151     qq33             = iq3*jq3;
 152
 153     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 154     rcutoff          = fr->rcoulomb;
 155     rcutoff2         = rcutoff*rcutoff;
 156
 157     rswitch          = fr->rcoulomb_switch;
 158     /* Setup switch parameters */
 159     d                = rcutoff-rswitch;
 160     swV3             = -10.0/(d*d*d);
 161     swV4             =  15.0/(d*d*d*d);
 162     swV5             =  -6.0/(d*d*d*d*d);
 163     swF2             = -30.0/(d*d*d);
 164     swF3             =  60.0/(d*d*d*d);
 165     swF4             = -30.0/(d*d*d*d*d);
 166
 167     outeriter        = 0;
 168     inneriter        = 0;
 169
 170     /* Start outer loop over neighborlists */
 171     for(iidx=0; iidx<nri; iidx++)
 172     {
 173         /* Load shift vector for this list */
 174         i_shift_offset   = DIM*shiftidx[iidx];
 175         shX              = shiftvec[i_shift_offset+XX];
 176         shY              = shiftvec[i_shift_offset+YY];
 177         shZ              = shiftvec[i_shift_offset+ZZ];
 178
 179         /* Load limits for loop over neighbors */
 180         j_index_start    = jindex[iidx];
 181         j_index_end      = jindex[iidx+1];
 182
 183         /* Get outer coordinate index */
 184         inr              = iinr[iidx];
 185         i_coord_offset   = DIM*inr;
 186
 187         /* Load i particle coords and add shift vector */
 188         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 189         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 190         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 191         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 192         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 193         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 194         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 195         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 196         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 197         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 198         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 199         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 200
 201         fix0             = 0.0;
 202         fiy0             = 0.0;
 203         fiz0             = 0.0;
 204         fix1             = 0.0;
 205         fiy1             = 0.0;
 206         fiz1             = 0.0;
 207         fix2             = 0.0;
 208         fiy2             = 0.0;
 209         fiz2             = 0.0;
 210         fix3             = 0.0;
 211         fiy3             = 0.0;
 212         fiz3             = 0.0;
 213
 214         /* Reset potential sums */
 215         velecsum         = 0.0;
 216         vvdwsum          = 0.0;
 217
 218         /* Start inner kernel loop */
 219         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 220         {
 221             /* Get j neighbor index, and coordinate index */
 222             jnr              = jjnr[jidx];
 223             j_coord_offset   = DIM*jnr;
 224
 225             /* load j atom coordinates */
 226             jx0              = x[j_coord_offset+DIM*0+XX];
 227             jy0              = x[j_coord_offset+DIM*0+YY];
 228             jz0              = x[j_coord_offset+DIM*0+ZZ];
 229             jx1              = x[j_coord_offset+DIM*1+XX];
 230             jy1              = x[j_coord_offset+DIM*1+YY];
 231             jz1              = x[j_coord_offset+DIM*1+ZZ];
 232             jx2              = x[j_coord_offset+DIM*2+XX];
 233             jy2              = x[j_coord_offset+DIM*2+YY];
 234             jz2              = x[j_coord_offset+DIM*2+ZZ];
 235             jx3              = x[j_coord_offset+DIM*3+XX];
 236             jy3              = x[j_coord_offset+DIM*3+YY];
 237             jz3              = x[j_coord_offset+DIM*3+ZZ];
 238
 239             /* Calculate displacement vector */
 240             dx00             = ix0 - jx0;
 241             dy00             = iy0 - jy0;
 242             dz00             = iz0 - jz0;
 243             dx11             = ix1 - jx1;
 244             dy11             = iy1 - jy1;
 245             dz11             = iz1 - jz1;
 246             dx12             = ix1 - jx2;
 247             dy12             = iy1 - jy2;
 248             dz12             = iz1 - jz2;
 249             dx13             = ix1 - jx3;
 250             dy13             = iy1 - jy3;
 251             dz13             = iz1 - jz3;
 252             dx21             = ix2 - jx1;
 253             dy21             = iy2 - jy1;
 254             dz21             = iz2 - jz1;
 255             dx22             = ix2 - jx2;
 256             dy22             = iy2 - jy2;
 257             dz22             = iz2 - jz2;
 258             dx23             = ix2 - jx3;
 259             dy23             = iy2 - jy3;
 260             dz23             = iz2 - jz3;
 261             dx31             = ix3 - jx1;
 262             dy31             = iy3 - jy1;
 263             dz31             = iz3 - jz1;
 264             dx32             = ix3 - jx2;
 265             dy32             = iy3 - jy2;
 266             dz32             = iz3 - jz2;
 267             dx33             = ix3 - jx3;
 268             dy33             = iy3 - jy3;
 269             dz33             = iz3 - jz3;
 270
 271             /* Calculate squared distance and things based on it */
 272             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 273             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 274             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 275             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 276             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 277             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 278             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 279             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 280             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 281             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 282
 283             rinv00           = gmx_invsqrt(rsq00);
 284             rinv11           = gmx_invsqrt(rsq11);
 285             rinv12           = gmx_invsqrt(rsq12);
 286             rinv13           = gmx_invsqrt(rsq13);
 287             rinv21           = gmx_invsqrt(rsq21);
 288             rinv22           = gmx_invsqrt(rsq22);
 289             rinv23           = gmx_invsqrt(rsq23);
 290             rinv31           = gmx_invsqrt(rsq31);
 291             rinv32           = gmx_invsqrt(rsq32);
 292             rinv33           = gmx_invsqrt(rsq33);
 293
 294             rinvsq00         = rinv00*rinv00;
 295             rinvsq11         = rinv11*rinv11;
 296             rinvsq12         = rinv12*rinv12;
 297             rinvsq13         = rinv13*rinv13;
 298             rinvsq21         = rinv21*rinv21;
 299             rinvsq22         = rinv22*rinv22;
 300             rinvsq23         = rinv23*rinv23;
 301             rinvsq31         = rinv31*rinv31;
 302             rinvsq32         = rinv32*rinv32;
 303             rinvsq33         = rinv33*rinv33;
 304
 305             /**************************
 306              * CALCULATE INTERACTIONS *
 307              **************************/
 308
 309             if (rsq00<rcutoff2)
 310             {
 311
 312             r00              = rsq00*rinv00;
 313
 314             /* BUCKINGHAM DISPERSION/REPULSION */
 315             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 316             vvdw6            = c6_00*rinvsix;
 317             br               = cexp2_00*r00;
 318             vvdwexp          = cexp1_00*exp(-br);
 319             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
 320             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 321
 322             d                = r00-rswitch;
 323             d                = (d>0.0) ? d : 0.0;
 324             d2               = d*d;
 325             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 326
 327             dsw              = d2*(swF2+d*(swF3+d*swF4));
 328
 329             /* Evaluate switch function */
 330             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 331             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 332             vvdw            *= sw;
 333
 334             /* Update potential sums from outer loop */
 335             vvdwsum         += vvdw;
 336
 337             fscal            = fvdw;
 338
 339             /* Calculate temporary vectorial force */
 340             tx               = fscal*dx00;
 341             ty               = fscal*dy00;
 342             tz               = fscal*dz00;
 343
 344             /* Update vectorial force */
 345             fix0            += tx;
 346             fiy0            += ty;
 347             fiz0            += tz;
 348             f[j_coord_offset+DIM*0+XX] -= tx;
 349             f[j_coord_offset+DIM*0+YY] -= ty;
 350             f[j_coord_offset+DIM*0+ZZ] -= tz;
 351
 352             }
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             if (rsq11<rcutoff2)
 359             {
 360
 361             r11              = rsq11*rinv11;
 362
 363             /* EWALD ELECTROSTATICS */
 364
 365             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 366             ewrt             = r11*ewtabscale;
 367             ewitab           = ewrt;
 368             eweps            = ewrt-ewitab;
 369             ewitab           = 4*ewitab;
 370             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 371             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 372             felec            = qq11*rinv11*(rinvsq11-felec);
 373
 374             d                = r11-rswitch;
 375             d                = (d>0.0) ? d : 0.0;
 376             d2               = d*d;
 377             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 378
 379             dsw              = d2*(swF2+d*(swF3+d*swF4));
 380
 381             /* Evaluate switch function */
 382             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 383             felec            = felec*sw - rinv11*velec*dsw;
 384             velec           *= sw;
 385
 386             /* Update potential sums from outer loop */
 387             velecsum        += velec;
 388
 389             fscal            = felec;
 390
 391             /* Calculate temporary vectorial force */
 392             tx               = fscal*dx11;
 393             ty               = fscal*dy11;
 394             tz               = fscal*dz11;
 395
 396             /* Update vectorial force */
 397             fix1            += tx;
 398             fiy1            += ty;
 399             fiz1            += tz;
 400             f[j_coord_offset+DIM*1+XX] -= tx;
 401             f[j_coord_offset+DIM*1+YY] -= ty;
 402             f[j_coord_offset+DIM*1+ZZ] -= tz;
 403
 404             }
 405
 406             /**************************
 407              * CALCULATE INTERACTIONS *
 408              **************************/
 409
 410             if (rsq12<rcutoff2)
 411             {
 412
 413             r12              = rsq12*rinv12;
 414
 415             /* EWALD ELECTROSTATICS */
 416
 417             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 418             ewrt             = r12*ewtabscale;
 419             ewitab           = ewrt;
 420             eweps            = ewrt-ewitab;
 421             ewitab           = 4*ewitab;
 422             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 423             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 424             felec            = qq12*rinv12*(rinvsq12-felec);
 425
 426             d                = r12-rswitch;
 427             d                = (d>0.0) ? d : 0.0;
 428             d2               = d*d;
 429             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 430
 431             dsw              = d2*(swF2+d*(swF3+d*swF4));
 432
 433             /* Evaluate switch function */
 434             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 435             felec            = felec*sw - rinv12*velec*dsw;
 436             velec           *= sw;
 437
 438             /* Update potential sums from outer loop */
 439             velecsum        += velec;
 440
 441             fscal            = felec;
 442
 443             /* Calculate temporary vectorial force */
 444             tx               = fscal*dx12;
 445             ty               = fscal*dy12;
 446             tz               = fscal*dz12;
 447
 448             /* Update vectorial force */
 449             fix1            += tx;
 450             fiy1            += ty;
 451             fiz1            += tz;
 452             f[j_coord_offset+DIM*2+XX] -= tx;
 453             f[j_coord_offset+DIM*2+YY] -= ty;
 454             f[j_coord_offset+DIM*2+ZZ] -= tz;
 455
 456             }
 457
 458             /**************************
 459              * CALCULATE INTERACTIONS *
 460              **************************/
 461
 462             if (rsq13<rcutoff2)
 463             {
 464
 465             r13              = rsq13*rinv13;
 466
 467             /* EWALD ELECTROSTATICS */
 468
 469             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 470             ewrt             = r13*ewtabscale;
 471             ewitab           = ewrt;
 472             eweps            = ewrt-ewitab;
 473             ewitab           = 4*ewitab;
 474             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 475             velec            = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 476             felec            = qq13*rinv13*(rinvsq13-felec);
 477
 478             d                = r13-rswitch;
 479             d                = (d>0.0) ? d : 0.0;
 480             d2               = d*d;
 481             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 482
 483             dsw              = d2*(swF2+d*(swF3+d*swF4));
 484
 485             /* Evaluate switch function */
 486             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 487             felec            = felec*sw - rinv13*velec*dsw;
 488             velec           *= sw;
 489
 490             /* Update potential sums from outer loop */
 491             velecsum        += velec;
 492
 493             fscal            = felec;
 494
 495             /* Calculate temporary vectorial force */
 496             tx               = fscal*dx13;
 497             ty               = fscal*dy13;
 498             tz               = fscal*dz13;
 499
 500             /* Update vectorial force */
 501             fix1            += tx;
 502             fiy1            += ty;
 503             fiz1            += tz;
 504             f[j_coord_offset+DIM*3+XX] -= tx;
 505             f[j_coord_offset+DIM*3+YY] -= ty;
 506             f[j_coord_offset+DIM*3+ZZ] -= tz;
 507
 508             }
 509
 510             /**************************
 511              * CALCULATE INTERACTIONS *
 512              **************************/
 513
 514             if (rsq21<rcutoff2)
 515             {
 516
 517             r21              = rsq21*rinv21;
 518
 519             /* EWALD ELECTROSTATICS */
 520
 521             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 522             ewrt             = r21*ewtabscale;
 523             ewitab           = ewrt;
 524             eweps            = ewrt-ewitab;
 525             ewitab           = 4*ewitab;
 526             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 527             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 528             felec            = qq21*rinv21*(rinvsq21-felec);
 529
 530             d                = r21-rswitch;
 531             d                = (d>0.0) ? d : 0.0;
 532             d2               = d*d;
 533             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 534
 535             dsw              = d2*(swF2+d*(swF3+d*swF4));
 536
 537             /* Evaluate switch function */
 538             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 539             felec            = felec*sw - rinv21*velec*dsw;
 540             velec           *= sw;
 541
 542             /* Update potential sums from outer loop */
 543             velecsum        += velec;
 544
 545             fscal            = felec;
 546
 547             /* Calculate temporary vectorial force */
 548             tx               = fscal*dx21;
 549             ty               = fscal*dy21;
 550             tz               = fscal*dz21;
 551
 552             /* Update vectorial force */
 553             fix2            += tx;
 554             fiy2            += ty;
 555             fiz2            += tz;
 556             f[j_coord_offset+DIM*1+XX] -= tx;
 557             f[j_coord_offset+DIM*1+YY] -= ty;
 558             f[j_coord_offset+DIM*1+ZZ] -= tz;
 559
 560             }
 561
 562             /**************************
 563              * CALCULATE INTERACTIONS *
 564              **************************/
 565
 566             if (rsq22<rcutoff2)
 567             {
 568
 569             r22              = rsq22*rinv22;
 570
 571             /* EWALD ELECTROSTATICS */
 572
 573             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 574             ewrt             = r22*ewtabscale;
 575             ewitab           = ewrt;
 576             eweps            = ewrt-ewitab;
 577             ewitab           = 4*ewitab;
 578             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 579             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 580             felec            = qq22*rinv22*(rinvsq22-felec);
 581
 582             d                = r22-rswitch;
 583             d                = (d>0.0) ? d : 0.0;
 584             d2               = d*d;
 585             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 586
 587             dsw              = d2*(swF2+d*(swF3+d*swF4));
 588
 589             /* Evaluate switch function */
 590             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 591             felec            = felec*sw - rinv22*velec*dsw;
 592             velec           *= sw;
 593
 594             /* Update potential sums from outer loop */
 595             velecsum        += velec;
 596
 597             fscal            = felec;
 598
 599             /* Calculate temporary vectorial force */
 600             tx               = fscal*dx22;
 601             ty               = fscal*dy22;
 602             tz               = fscal*dz22;
 603
 604             /* Update vectorial force */
 605             fix2            += tx;
 606             fiy2            += ty;
 607             fiz2            += tz;
 608             f[j_coord_offset+DIM*2+XX] -= tx;
 609             f[j_coord_offset+DIM*2+YY] -= ty;
 610             f[j_coord_offset+DIM*2+ZZ] -= tz;
 611
 612             }
 613
 614             /**************************
 615              * CALCULATE INTERACTIONS *
 616              **************************/
 617
 618             if (rsq23<rcutoff2)
 619             {
 620
 621             r23              = rsq23*rinv23;
 622
 623             /* EWALD ELECTROSTATICS */
 624
 625             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 626             ewrt             = r23*ewtabscale;
 627             ewitab           = ewrt;
 628             eweps            = ewrt-ewitab;
 629             ewitab           = 4*ewitab;
 630             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 631             velec            = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 632             felec            = qq23*rinv23*(rinvsq23-felec);
 633
 634             d                = r23-rswitch;
 635             d                = (d>0.0) ? d : 0.0;
 636             d2               = d*d;
 637             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 638
 639             dsw              = d2*(swF2+d*(swF3+d*swF4));
 640
 641             /* Evaluate switch function */
 642             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 643             felec            = felec*sw - rinv23*velec*dsw;
 644             velec           *= sw;
 645
 646             /* Update potential sums from outer loop */
 647             velecsum        += velec;
 648
 649             fscal            = felec;
 650
 651             /* Calculate temporary vectorial force */
 652             tx               = fscal*dx23;
 653             ty               = fscal*dy23;
 654             tz               = fscal*dz23;
 655
 656             /* Update vectorial force */
 657             fix2            += tx;
 658             fiy2            += ty;
 659             fiz2            += tz;
 660             f[j_coord_offset+DIM*3+XX] -= tx;
 661             f[j_coord_offset+DIM*3+YY] -= ty;
 662             f[j_coord_offset+DIM*3+ZZ] -= tz;
 663
 664             }
 665
 666             /**************************
 667              * CALCULATE INTERACTIONS *
 668              **************************/
 669
 670             if (rsq31<rcutoff2)
 671             {
 672
 673             r31              = rsq31*rinv31;
 674
 675             /* EWALD ELECTROSTATICS */
 676
 677             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 678             ewrt             = r31*ewtabscale;
 679             ewitab           = ewrt;
 680             eweps            = ewrt-ewitab;
 681             ewitab           = 4*ewitab;
 682             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 683             velec            = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 684             felec            = qq31*rinv31*(rinvsq31-felec);
 685
 686             d                = r31-rswitch;
 687             d                = (d>0.0) ? d : 0.0;
 688             d2               = d*d;
 689             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 690
 691             dsw              = d2*(swF2+d*(swF3+d*swF4));
 692
 693             /* Evaluate switch function */
 694             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 695             felec            = felec*sw - rinv31*velec*dsw;
 696             velec           *= sw;
 697
 698             /* Update potential sums from outer loop */
 699             velecsum        += velec;
 700
 701             fscal            = felec;
 702
 703             /* Calculate temporary vectorial force */
 704             tx               = fscal*dx31;
 705             ty               = fscal*dy31;
 706             tz               = fscal*dz31;
 707
 708             /* Update vectorial force */
 709             fix3            += tx;
 710             fiy3            += ty;
 711             fiz3            += tz;
 712             f[j_coord_offset+DIM*1+XX] -= tx;
 713             f[j_coord_offset+DIM*1+YY] -= ty;
 714             f[j_coord_offset+DIM*1+ZZ] -= tz;
 715
 716             }
 717
 718             /**************************
 719              * CALCULATE INTERACTIONS *
 720              **************************/
 721
 722             if (rsq32<rcutoff2)
 723             {
 724
 725             r32              = rsq32*rinv32;
 726
 727             /* EWALD ELECTROSTATICS */
 728
 729             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 730             ewrt             = r32*ewtabscale;
 731             ewitab           = ewrt;
 732             eweps            = ewrt-ewitab;
 733             ewitab           = 4*ewitab;
 734             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 735             velec            = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 736             felec            = qq32*rinv32*(rinvsq32-felec);
 737
 738             d                = r32-rswitch;
 739             d                = (d>0.0) ? d : 0.0;
 740             d2               = d*d;
 741             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 742
 743             dsw              = d2*(swF2+d*(swF3+d*swF4));
 744
 745             /* Evaluate switch function */
 746             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 747             felec            = felec*sw - rinv32*velec*dsw;
 748             velec           *= sw;
 749
 750             /* Update potential sums from outer loop */
 751             velecsum        += velec;
 752
 753             fscal            = felec;
 754
 755             /* Calculate temporary vectorial force */
 756             tx               = fscal*dx32;
 757             ty               = fscal*dy32;
 758             tz               = fscal*dz32;
 759
 760             /* Update vectorial force */
 761             fix3            += tx;
 762             fiy3            += ty;
 763             fiz3            += tz;
 764             f[j_coord_offset+DIM*2+XX] -= tx;
 765             f[j_coord_offset+DIM*2+YY] -= ty;
 766             f[j_coord_offset+DIM*2+ZZ] -= tz;
 767
 768             }
 769
 770             /**************************
 771              * CALCULATE INTERACTIONS *
 772              **************************/
 773
 774             if (rsq33<rcutoff2)
 775             {
 776
 777             r33              = rsq33*rinv33;
 778
 779             /* EWALD ELECTROSTATICS */
 780
 781             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 782             ewrt             = r33*ewtabscale;
 783             ewitab           = ewrt;
 784             eweps            = ewrt-ewitab;
 785             ewitab           = 4*ewitab;
 786             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 787             velec            = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 788             felec            = qq33*rinv33*(rinvsq33-felec);
 789
 790             d                = r33-rswitch;
 791             d                = (d>0.0) ? d : 0.0;
 792             d2               = d*d;
 793             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 794
 795             dsw              = d2*(swF2+d*(swF3+d*swF4));
 796
 797             /* Evaluate switch function */
 798             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 799             felec            = felec*sw - rinv33*velec*dsw;
 800             velec           *= sw;
 801
 802             /* Update potential sums from outer loop */
 803             velecsum        += velec;
 804
 805             fscal            = felec;
 806
 807             /* Calculate temporary vectorial force */
 808             tx               = fscal*dx33;
 809             ty               = fscal*dy33;
 810             tz               = fscal*dz33;
 811
 812             /* Update vectorial force */
 813             fix3            += tx;
 814             fiy3            += ty;
 815             fiz3            += tz;
 816             f[j_coord_offset+DIM*3+XX] -= tx;
 817             f[j_coord_offset+DIM*3+YY] -= ty;
 818             f[j_coord_offset+DIM*3+ZZ] -= tz;
 819
 820             }
 821
 822             /* Inner loop uses 601 flops */
 823         }
 824         /* End of innermost loop */
 825
 826         tx = ty = tz = 0;
 827         f[i_coord_offset+DIM*0+XX] += fix0;
 828         f[i_coord_offset+DIM*0+YY] += fiy0;
 829         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 830         tx                         += fix0;
 831         ty                         += fiy0;
 832         tz                         += fiz0;
 833         f[i_coord_offset+DIM*1+XX] += fix1;
 834         f[i_coord_offset+DIM*1+YY] += fiy1;
 835         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 836         tx                         += fix1;
 837         ty                         += fiy1;
 838         tz                         += fiz1;
 839         f[i_coord_offset+DIM*2+XX] += fix2;
 840         f[i_coord_offset+DIM*2+YY] += fiy2;
 841         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 842         tx                         += fix2;
 843         ty                         += fiy2;
 844         tz                         += fiz2;
 845         f[i_coord_offset+DIM*3+XX] += fix3;
 846         f[i_coord_offset+DIM*3+YY] += fiy3;
 847         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 848         tx                         += fix3;
 849         ty                         += fiy3;
 850         tz                         += fiz3;
 851         fshift[i_shift_offset+XX]  += tx;
 852         fshift[i_shift_offset+YY]  += ty;
 853         fshift[i_shift_offset+ZZ]  += tz;
 854
 855         ggid                        = gid[iidx];
 856         /* Update potential energies */
 857         kernel_data->energygrp_elec[ggid] += velecsum;
 858         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 859
 860         /* Increment number of inner iterations */
 861         inneriter                  += j_index_end - j_index_start;
 862
 863         /* Outer loop uses 41 flops */
 864     }
 865
 866     /* Increment number of outer iterations */
 867     outeriter        += nri;
 868
 869     /* Update outer/inner flops */
 870
 871     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*601);
 872 }
 873 /*
 874  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwBhamSw_GeomW4W4_F_c
 875  * Electrostatics interaction: Ewald
 876  * VdW interaction:            Buckingham
 877  * Geometry:                   Water4-Water4
 878  * Calculate force/pot:        Force
 879  */
 880 void
 881 nb_kernel_ElecEwSw_VdwBhamSw_GeomW4W4_F_c
 882                     (t_nblist                    * gmx_restrict       nlist,
 883                      rvec                        * gmx_restrict          xx,
 884                      rvec                        * gmx_restrict          ff,
 885                      t_forcerec                  * gmx_restrict          fr,
 886                      t_mdatoms                   * gmx_restrict     mdatoms,
 887                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 888                      t_nrnb                      * gmx_restrict        nrnb)
 889 {
 890     int              i_shift_offset,i_coord_offset,j_coord_offset;
 891     int              j_index_start,j_index_end;
 892     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 893     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 894     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 895     real             *shiftvec,*fshift,*x,*f;
 896     int              vdwioffset0;
 897     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 898     int              vdwioffset1;
 899     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 900     int              vdwioffset2;
 901     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 902     int              vdwioffset3;
 903     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 904     int              vdwjidx0;
 905     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 906     int              vdwjidx1;
 907     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 908     int              vdwjidx2;
 909     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 910     int              vdwjidx3;
 911     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 912     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 913     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 914     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 915     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 916     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 917     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 918     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 919     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 920     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 921     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 922     real             velec,felec,velecsum,facel,crf,krf,krf2;
 923     real             *charge;
 924     int              nvdwtype;
 925     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 926     int              *vdwtype;
 927     real             *vdwparam;
 928     int              ewitab;
 929     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 930     real             *ewtab;
 931     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 932
 933     x                = xx[0];
 934     f                = ff[0];
 935
 936     nri              = nlist->nri;
 937     iinr             = nlist->iinr;
 938     jindex           = nlist->jindex;
 939     jjnr             = nlist->jjnr;
 940     shiftidx         = nlist->shift;
 941     gid              = nlist->gid;
 942     shiftvec         = fr->shift_vec[0];
 943     fshift           = fr->fshift[0];
 944     facel            = fr->epsfac;
 945     charge           = mdatoms->chargeA;
 946     nvdwtype         = fr->ntype;
 947     vdwparam         = fr->nbfp;
 948     vdwtype          = mdatoms->typeA;
 949
 950     sh_ewald         = fr->ic->sh_ewald;
 951     ewtab            = fr->ic->tabq_coul_FDV0;
 952     ewtabscale       = fr->ic->tabq_scale;
 953     ewtabhalfspace   = 0.5/ewtabscale;
 954
 955     /* Setup water-specific parameters */
 956     inr              = nlist->iinr[0];
 957     iq1              = facel*charge[inr+1];
 958     iq2              = facel*charge[inr+2];
 959     iq3              = facel*charge[inr+3];
 960     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 961
 962     jq1              = charge[inr+1];
 963     jq2              = charge[inr+2];
 964     jq3              = charge[inr+3];
 965     vdwjidx0         = 3*vdwtype[inr+0];
 966     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 967     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 968     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 969     qq11             = iq1*jq1;
 970     qq12             = iq1*jq2;
 971     qq13             = iq1*jq3;
 972     qq21             = iq2*jq1;
 973     qq22             = iq2*jq2;
 974     qq23             = iq2*jq3;
 975     qq31             = iq3*jq1;
 976     qq32             = iq3*jq2;
 977     qq33             = iq3*jq3;
 978
 979     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 980     rcutoff          = fr->rcoulomb;
 981     rcutoff2         = rcutoff*rcutoff;
 982
 983     rswitch          = fr->rcoulomb_switch;
 984     /* Setup switch parameters */
 985     d                = rcutoff-rswitch;
 986     swV3             = -10.0/(d*d*d);
 987     swV4             =  15.0/(d*d*d*d);
 988     swV5             =  -6.0/(d*d*d*d*d);
 989     swF2             = -30.0/(d*d*d);
 990     swF3             =  60.0/(d*d*d*d);
 991     swF4             = -30.0/(d*d*d*d*d);
 992
 993     outeriter        = 0;
 994     inneriter        = 0;
 995
 996     /* Start outer loop over neighborlists */
 997     for(iidx=0; iidx<nri; iidx++)
 998     {
 999         /* Load shift vector for this list */
1000         i_shift_offset   = DIM*shiftidx[iidx];
1001         shX              = shiftvec[i_shift_offset+XX];
1002         shY              = shiftvec[i_shift_offset+YY];
1003         shZ              = shiftvec[i_shift_offset+ZZ];
1004
1005         /* Load limits for loop over neighbors */
1006         j_index_start    = jindex[iidx];
1007         j_index_end      = jindex[iidx+1];
1008
1009         /* Get outer coordinate index */
1010         inr              = iinr[iidx];
1011         i_coord_offset   = DIM*inr;
1012
1013         /* Load i particle coords and add shift vector */
1014         ix0              = shX + x[i_coord_offset+DIM*0+XX];
1015         iy0              = shY + x[i_coord_offset+DIM*0+YY];
1016         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
1017         ix1              = shX + x[i_coord_offset+DIM*1+XX];
1018         iy1              = shY + x[i_coord_offset+DIM*1+YY];
1019         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
1020         ix2              = shX + x[i_coord_offset+DIM*2+XX];
1021         iy2              = shY + x[i_coord_offset+DIM*2+YY];
1022         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
1023         ix3              = shX + x[i_coord_offset+DIM*3+XX];
1024         iy3              = shY + x[i_coord_offset+DIM*3+YY];
1025         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
1026
1027         fix0             = 0.0;
1028         fiy0             = 0.0;
1029         fiz0             = 0.0;
1030         fix1             = 0.0;
1031         fiy1             = 0.0;
1032         fiz1             = 0.0;
1033         fix2             = 0.0;
1034         fiy2             = 0.0;
1035         fiz2             = 0.0;
1036         fix3             = 0.0;
1037         fiy3             = 0.0;
1038         fiz3             = 0.0;
1039
1040         /* Start inner kernel loop */
1041         for(jidx=j_index_start; jidx<j_index_end; jidx++)
1042         {
1043             /* Get j neighbor index, and coordinate index */
1044             jnr              = jjnr[jidx];
1045             j_coord_offset   = DIM*jnr;
1046
1047             /* load j atom coordinates */
1048             jx0              = x[j_coord_offset+DIM*0+XX];
1049             jy0              = x[j_coord_offset+DIM*0+YY];
1050             jz0              = x[j_coord_offset+DIM*0+ZZ];
1051             jx1              = x[j_coord_offset+DIM*1+XX];
1052             jy1              = x[j_coord_offset+DIM*1+YY];
1053             jz1              = x[j_coord_offset+DIM*1+ZZ];
1054             jx2              = x[j_coord_offset+DIM*2+XX];
1055             jy2              = x[j_coord_offset+DIM*2+YY];
1056             jz2              = x[j_coord_offset+DIM*2+ZZ];
1057             jx3              = x[j_coord_offset+DIM*3+XX];
1058             jy3              = x[j_coord_offset+DIM*3+YY];
1059             jz3              = x[j_coord_offset+DIM*3+ZZ];
1060
1061             /* Calculate displacement vector */
1062             dx00             = ix0 - jx0;
1063             dy00             = iy0 - jy0;
1064             dz00             = iz0 - jz0;
1065             dx11             = ix1 - jx1;
1066             dy11             = iy1 - jy1;
1067             dz11             = iz1 - jz1;
1068             dx12             = ix1 - jx2;
1069             dy12             = iy1 - jy2;
1070             dz12             = iz1 - jz2;
1071             dx13             = ix1 - jx3;
1072             dy13             = iy1 - jy3;
1073             dz13             = iz1 - jz3;
1074             dx21             = ix2 - jx1;
1075             dy21             = iy2 - jy1;
1076             dz21             = iz2 - jz1;
1077             dx22             = ix2 - jx2;
1078             dy22             = iy2 - jy2;
1079             dz22             = iz2 - jz2;
1080             dx23             = ix2 - jx3;
1081             dy23             = iy2 - jy3;
1082             dz23             = iz2 - jz3;
1083             dx31             = ix3 - jx1;
1084             dy31             = iy3 - jy1;
1085             dz31             = iz3 - jz1;
1086             dx32             = ix3 - jx2;
1087             dy32             = iy3 - jy2;
1088             dz32             = iz3 - jz2;
1089             dx33             = ix3 - jx3;
1090             dy33             = iy3 - jy3;
1091             dz33             = iz3 - jz3;
1092
1093             /* Calculate squared distance and things based on it */
1094             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
1095             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
1096             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
1097             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
1098             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
1099             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
1100             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
1101             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
1102             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
1103             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
1104
1105             rinv00           = gmx_invsqrt(rsq00);
1106             rinv11           = gmx_invsqrt(rsq11);
1107             rinv12           = gmx_invsqrt(rsq12);
1108             rinv13           = gmx_invsqrt(rsq13);
1109             rinv21           = gmx_invsqrt(rsq21);
1110             rinv22           = gmx_invsqrt(rsq22);
1111             rinv23           = gmx_invsqrt(rsq23);
1112             rinv31           = gmx_invsqrt(rsq31);
1113             rinv32           = gmx_invsqrt(rsq32);
1114             rinv33           = gmx_invsqrt(rsq33);
1115
1116             rinvsq00         = rinv00*rinv00;
1117             rinvsq11         = rinv11*rinv11;
1118             rinvsq12         = rinv12*rinv12;
1119             rinvsq13         = rinv13*rinv13;
1120             rinvsq21         = rinv21*rinv21;
1121             rinvsq22         = rinv22*rinv22;
1122             rinvsq23         = rinv23*rinv23;
1123             rinvsq31         = rinv31*rinv31;
1124             rinvsq32         = rinv32*rinv32;
1125             rinvsq33         = rinv33*rinv33;
1126
1127             /**************************
1128              * CALCULATE INTERACTIONS *
1129              **************************/
1130
1131             if (rsq00<rcutoff2)
1132             {
1133
1134             r00              = rsq00*rinv00;
1135
1136             /* BUCKINGHAM DISPERSION/REPULSION */
1137             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
1138             vvdw6            = c6_00*rinvsix;
1139             br               = cexp2_00*r00;
1140             vvdwexp          = cexp1_00*exp(-br);
1141             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
1142             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
1143
1144             d                = r00-rswitch;
1145             d                = (d>0.0) ? d : 0.0;
1146             d2               = d*d;
1147             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1148
1149             dsw              = d2*(swF2+d*(swF3+d*swF4));
1150
1151             /* Evaluate switch function */
1152             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1153             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
1154
1155             fscal            = fvdw;
1156
1157             /* Calculate temporary vectorial force */
1158             tx               = fscal*dx00;
1159             ty               = fscal*dy00;
1160             tz               = fscal*dz00;
1161
1162             /* Update vectorial force */
1163             fix0            += tx;
1164             fiy0            += ty;
1165             fiz0            += tz;
1166             f[j_coord_offset+DIM*0+XX] -= tx;
1167             f[j_coord_offset+DIM*0+YY] -= ty;
1168             f[j_coord_offset+DIM*0+ZZ] -= tz;
1169
1170             }
1171
1172             /**************************
1173              * CALCULATE INTERACTIONS *
1174              **************************/
1175
1176             if (rsq11<rcutoff2)
1177             {
1178
1179             r11              = rsq11*rinv11;
1180
1181             /* EWALD ELECTROSTATICS */
1182
1183             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1184             ewrt             = r11*ewtabscale;
1185             ewitab           = ewrt;
1186             eweps            = ewrt-ewitab;
1187             ewitab           = 4*ewitab;
1188             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1189             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1190             felec            = qq11*rinv11*(rinvsq11-felec);
1191
1192             d                = r11-rswitch;
1193             d                = (d>0.0) ? d : 0.0;
1194             d2               = d*d;
1195             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1196
1197             dsw              = d2*(swF2+d*(swF3+d*swF4));
1198
1199             /* Evaluate switch function */
1200             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1201             felec            = felec*sw - rinv11*velec*dsw;
1202
1203             fscal            = felec;
1204
1205             /* Calculate temporary vectorial force */
1206             tx               = fscal*dx11;
1207             ty               = fscal*dy11;
1208             tz               = fscal*dz11;
1209
1210             /* Update vectorial force */
1211             fix1            += tx;
1212             fiy1            += ty;
1213             fiz1            += tz;
1214             f[j_coord_offset+DIM*1+XX] -= tx;
1215             f[j_coord_offset+DIM*1+YY] -= ty;
1216             f[j_coord_offset+DIM*1+ZZ] -= tz;
1217
1218             }
1219
1220             /**************************
1221              * CALCULATE INTERACTIONS *
1222              **************************/
1223
1224             if (rsq12<rcutoff2)
1225             {
1226
1227             r12              = rsq12*rinv12;
1228
1229             /* EWALD ELECTROSTATICS */
1230
1231             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1232             ewrt             = r12*ewtabscale;
1233             ewitab           = ewrt;
1234             eweps            = ewrt-ewitab;
1235             ewitab           = 4*ewitab;
1236             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1237             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1238             felec            = qq12*rinv12*(rinvsq12-felec);
1239
1240             d                = r12-rswitch;
1241             d                = (d>0.0) ? d : 0.0;
1242             d2               = d*d;
1243             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1244
1245             dsw              = d2*(swF2+d*(swF3+d*swF4));
1246
1247             /* Evaluate switch function */
1248             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1249             felec            = felec*sw - rinv12*velec*dsw;
1250
1251             fscal            = felec;
1252
1253             /* Calculate temporary vectorial force */
1254             tx               = fscal*dx12;
1255             ty               = fscal*dy12;
1256             tz               = fscal*dz12;
1257
1258             /* Update vectorial force */
1259             fix1            += tx;
1260             fiy1            += ty;
1261             fiz1            += tz;
1262             f[j_coord_offset+DIM*2+XX] -= tx;
1263             f[j_coord_offset+DIM*2+YY] -= ty;
1264             f[j_coord_offset+DIM*2+ZZ] -= tz;
1265
1266             }
1267
1268             /**************************
1269              * CALCULATE INTERACTIONS *
1270              **************************/
1271
1272             if (rsq13<rcutoff2)
1273             {
1274
1275             r13              = rsq13*rinv13;
1276
1277             /* EWALD ELECTROSTATICS */
1278
1279             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1280             ewrt             = r13*ewtabscale;
1281             ewitab           = ewrt;
1282             eweps            = ewrt-ewitab;
1283             ewitab           = 4*ewitab;
1284             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1285             velec            = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1286             felec            = qq13*rinv13*(rinvsq13-felec);
1287
1288             d                = r13-rswitch;
1289             d                = (d>0.0) ? d : 0.0;
1290             d2               = d*d;
1291             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1292
1293             dsw              = d2*(swF2+d*(swF3+d*swF4));
1294
1295             /* Evaluate switch function */
1296             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1297             felec            = felec*sw - rinv13*velec*dsw;
1298
1299             fscal            = felec;
1300
1301             /* Calculate temporary vectorial force */
1302             tx               = fscal*dx13;
1303             ty               = fscal*dy13;
1304             tz               = fscal*dz13;
1305
1306             /* Update vectorial force */
1307             fix1            += tx;
1308             fiy1            += ty;
1309             fiz1            += tz;
1310             f[j_coord_offset+DIM*3+XX] -= tx;
1311             f[j_coord_offset+DIM*3+YY] -= ty;
1312             f[j_coord_offset+DIM*3+ZZ] -= tz;
1313
1314             }
1315
1316             /**************************
1317              * CALCULATE INTERACTIONS *
1318              **************************/
1319
1320             if (rsq21<rcutoff2)
1321             {
1322
1323             r21              = rsq21*rinv21;
1324
1325             /* EWALD ELECTROSTATICS */
1326
1327             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1328             ewrt             = r21*ewtabscale;
1329             ewitab           = ewrt;
1330             eweps            = ewrt-ewitab;
1331             ewitab           = 4*ewitab;
1332             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1333             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1334             felec            = qq21*rinv21*(rinvsq21-felec);
1335
1336             d                = r21-rswitch;
1337             d                = (d>0.0) ? d : 0.0;
1338             d2               = d*d;
1339             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1340
1341             dsw              = d2*(swF2+d*(swF3+d*swF4));
1342
1343             /* Evaluate switch function */
1344             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1345             felec            = felec*sw - rinv21*velec*dsw;
1346
1347             fscal            = felec;
1348
1349             /* Calculate temporary vectorial force */
1350             tx               = fscal*dx21;
1351             ty               = fscal*dy21;
1352             tz               = fscal*dz21;
1353
1354             /* Update vectorial force */
1355             fix2            += tx;
1356             fiy2            += ty;
1357             fiz2            += tz;
1358             f[j_coord_offset+DIM*1+XX] -= tx;
1359             f[j_coord_offset+DIM*1+YY] -= ty;
1360             f[j_coord_offset+DIM*1+ZZ] -= tz;
1361
1362             }
1363
1364             /**************************
1365              * CALCULATE INTERACTIONS *
1366              **************************/
1367
1368             if (rsq22<rcutoff2)
1369             {
1370
1371             r22              = rsq22*rinv22;
1372
1373             /* EWALD ELECTROSTATICS */
1374
1375             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1376             ewrt             = r22*ewtabscale;
1377             ewitab           = ewrt;
1378             eweps            = ewrt-ewitab;
1379             ewitab           = 4*ewitab;
1380             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1381             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1382             felec            = qq22*rinv22*(rinvsq22-felec);
1383
1384             d                = r22-rswitch;
1385             d                = (d>0.0) ? d : 0.0;
1386             d2               = d*d;
1387             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1388
1389             dsw              = d2*(swF2+d*(swF3+d*swF4));
1390
1391             /* Evaluate switch function */
1392             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1393             felec            = felec*sw - rinv22*velec*dsw;
1394
1395             fscal            = felec;
1396
1397             /* Calculate temporary vectorial force */
1398             tx               = fscal*dx22;
1399             ty               = fscal*dy22;
1400             tz               = fscal*dz22;
1401
1402             /* Update vectorial force */
1403             fix2            += tx;
1404             fiy2            += ty;
1405             fiz2            += tz;
1406             f[j_coord_offset+DIM*2+XX] -= tx;
1407             f[j_coord_offset+DIM*2+YY] -= ty;
1408             f[j_coord_offset+DIM*2+ZZ] -= tz;
1409
1410             }
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             if (rsq23<rcutoff2)
1417             {
1418
1419             r23              = rsq23*rinv23;
1420
1421             /* EWALD ELECTROSTATICS */
1422
1423             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1424             ewrt             = r23*ewtabscale;
1425             ewitab           = ewrt;
1426             eweps            = ewrt-ewitab;
1427             ewitab           = 4*ewitab;
1428             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1429             velec            = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1430             felec            = qq23*rinv23*(rinvsq23-felec);
1431
1432             d                = r23-rswitch;
1433             d                = (d>0.0) ? d : 0.0;
1434             d2               = d*d;
1435             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1436
1437             dsw              = d2*(swF2+d*(swF3+d*swF4));
1438
1439             /* Evaluate switch function */
1440             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1441             felec            = felec*sw - rinv23*velec*dsw;
1442
1443             fscal            = felec;
1444
1445             /* Calculate temporary vectorial force */
1446             tx               = fscal*dx23;
1447             ty               = fscal*dy23;
1448             tz               = fscal*dz23;
1449
1450             /* Update vectorial force */
1451             fix2            += tx;
1452             fiy2            += ty;
1453             fiz2            += tz;
1454             f[j_coord_offset+DIM*3+XX] -= tx;
1455             f[j_coord_offset+DIM*3+YY] -= ty;
1456             f[j_coord_offset+DIM*3+ZZ] -= tz;
1457
1458             }
1459
1460             /**************************
1461              * CALCULATE INTERACTIONS *
1462              **************************/
1463
1464             if (rsq31<rcutoff2)
1465             {
1466
1467             r31              = rsq31*rinv31;
1468
1469             /* EWALD ELECTROSTATICS */
1470
1471             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1472             ewrt             = r31*ewtabscale;
1473             ewitab           = ewrt;
1474             eweps            = ewrt-ewitab;
1475             ewitab           = 4*ewitab;
1476             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1477             velec            = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1478             felec            = qq31*rinv31*(rinvsq31-felec);
1479
1480             d                = r31-rswitch;
1481             d                = (d>0.0) ? d : 0.0;
1482             d2               = d*d;
1483             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1484
1485             dsw              = d2*(swF2+d*(swF3+d*swF4));
1486
1487             /* Evaluate switch function */
1488             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1489             felec            = felec*sw - rinv31*velec*dsw;
1490
1491             fscal            = felec;
1492
1493             /* Calculate temporary vectorial force */
1494             tx               = fscal*dx31;
1495             ty               = fscal*dy31;
1496             tz               = fscal*dz31;
1497
1498             /* Update vectorial force */
1499             fix3            += tx;
1500             fiy3            += ty;
1501             fiz3            += tz;
1502             f[j_coord_offset+DIM*1+XX] -= tx;
1503             f[j_coord_offset+DIM*1+YY] -= ty;
1504             f[j_coord_offset+DIM*1+ZZ] -= tz;
1505
1506             }
1507
1508             /**************************
1509              * CALCULATE INTERACTIONS *
1510              **************************/
1511
1512             if (rsq32<rcutoff2)
1513             {
1514
1515             r32              = rsq32*rinv32;
1516
1517             /* EWALD ELECTROSTATICS */
1518
1519             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1520             ewrt             = r32*ewtabscale;
1521             ewitab           = ewrt;
1522             eweps            = ewrt-ewitab;
1523             ewitab           = 4*ewitab;
1524             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1525             velec            = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1526             felec            = qq32*rinv32*(rinvsq32-felec);
1527
1528             d                = r32-rswitch;
1529             d                = (d>0.0) ? d : 0.0;
1530             d2               = d*d;
1531             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1532
1533             dsw              = d2*(swF2+d*(swF3+d*swF4));
1534
1535             /* Evaluate switch function */
1536             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1537             felec            = felec*sw - rinv32*velec*dsw;
1538
1539             fscal            = felec;
1540
1541             /* Calculate temporary vectorial force */
1542             tx               = fscal*dx32;
1543             ty               = fscal*dy32;
1544             tz               = fscal*dz32;
1545
1546             /* Update vectorial force */
1547             fix3            += tx;
1548             fiy3            += ty;
1549             fiz3            += tz;
1550             f[j_coord_offset+DIM*2+XX] -= tx;
1551             f[j_coord_offset+DIM*2+YY] -= ty;
1552             f[j_coord_offset+DIM*2+ZZ] -= tz;
1553
1554             }
1555
1556             /**************************
1557              * CALCULATE INTERACTIONS *
1558              **************************/
1559
1560             if (rsq33<rcutoff2)
1561             {
1562
1563             r33              = rsq33*rinv33;
1564
1565             /* EWALD ELECTROSTATICS */
1566
1567             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1568             ewrt             = r33*ewtabscale;
1569             ewitab           = ewrt;
1570             eweps            = ewrt-ewitab;
1571             ewitab           = 4*ewitab;
1572             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1573             velec            = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1574             felec            = qq33*rinv33*(rinvsq33-felec);
1575
1576             d                = r33-rswitch;
1577             d                = (d>0.0) ? d : 0.0;
1578             d2               = d*d;
1579             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1580
1581             dsw              = d2*(swF2+d*(swF3+d*swF4));
1582
1583             /* Evaluate switch function */
1584             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1585             felec            = felec*sw - rinv33*velec*dsw;
1586
1587             fscal            = felec;
1588
1589             /* Calculate temporary vectorial force */
1590             tx               = fscal*dx33;
1591             ty               = fscal*dy33;
1592             tz               = fscal*dz33;
1593
1594             /* Update vectorial force */
1595             fix3            += tx;
1596             fiy3            += ty;
1597             fiz3            += tz;
1598             f[j_coord_offset+DIM*3+XX] -= tx;
1599             f[j_coord_offset+DIM*3+YY] -= ty;
1600             f[j_coord_offset+DIM*3+ZZ] -= tz;
1601
1602             }
1603
1604             /* Inner loop uses 581 flops */
1605         }
1606         /* End of innermost loop */
1607
1608         tx = ty = tz = 0;
1609         f[i_coord_offset+DIM*0+XX] += fix0;
1610         f[i_coord_offset+DIM*0+YY] += fiy0;
1611         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1612         tx                         += fix0;
1613         ty                         += fiy0;
1614         tz                         += fiz0;
1615         f[i_coord_offset+DIM*1+XX] += fix1;
1616         f[i_coord_offset+DIM*1+YY] += fiy1;
1617         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1618         tx                         += fix1;
1619         ty                         += fiy1;
1620         tz                         += fiz1;
1621         f[i_coord_offset+DIM*2+XX] += fix2;
1622         f[i_coord_offset+DIM*2+YY] += fiy2;
1623         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1624         tx                         += fix2;
1625         ty                         += fiy2;
1626         tz                         += fiz2;
1627         f[i_coord_offset+DIM*3+XX] += fix3;
1628         f[i_coord_offset+DIM*3+YY] += fiy3;
1629         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1630         tx                         += fix3;
1631         ty                         += fiy3;
1632         tz                         += fiz3;
1633         fshift[i_shift_offset+XX]  += tx;
1634         fshift[i_shift_offset+YY]  += ty;
1635         fshift[i_shift_offset+ZZ]  += tz;
1636
1637         /* Increment number of inner iterations */
1638         inneriter                  += j_index_end - j_index_start;
1639
1640         /* Outer loop uses 39 flops */
1641     }
1642
1643     /* Increment number of outer iterations */
1644     outeriter        += nri;
1645
1646     /* Update outer/inner flops */
1647
1648     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*581);
1649 }