src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecCoul_VdwCSTab_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 "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_VF_c
  49  * Electrostatics interaction: Coulomb
  50  * VdW interaction:            CubicSplineTable
  51  * Geometry:                   Water4-Water4
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecCoul_VdwCSTab_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              vfitab;
 103     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 104     real             *vftab;
 105
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = fr->epsfac;
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_vdw->data;
 124     vftabscale       = kernel_data->table_vdw->scale;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq1              = facel*charge[inr+1];
 129     iq2              = facel*charge[inr+2];
 130     iq3              = facel*charge[inr+3];
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     jq1              = charge[inr+1];
 134     jq2              = charge[inr+2];
 135     jq3              = charge[inr+3];
 136     vdwjidx0         = 2*vdwtype[inr+0];
 137     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 138     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 139     qq11             = iq1*jq1;
 140     qq12             = iq1*jq2;
 141     qq13             = iq1*jq3;
 142     qq21             = iq2*jq1;
 143     qq22             = iq2*jq2;
 144     qq23             = iq2*jq3;
 145     qq31             = iq3*jq1;
 146     qq32             = iq3*jq2;
 147     qq33             = iq3*jq3;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157         shX              = shiftvec[i_shift_offset+XX];
 158         shY              = shiftvec[i_shift_offset+YY];
 159         shZ              = shiftvec[i_shift_offset+ZZ];
 160
 161         /* Load limits for loop over neighbors */
 162         j_index_start    = jindex[iidx];
 163         j_index_end      = jindex[iidx+1];
 164
 165         /* Get outer coordinate index */
 166         inr              = iinr[iidx];
 167         i_coord_offset   = DIM*inr;
 168
 169         /* Load i particle coords and add shift vector */
 170         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 171         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 172         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 173         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 174         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 175         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 176         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 177         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 178         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 179         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 180         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 181         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 182
 183         fix0             = 0.0;
 184         fiy0             = 0.0;
 185         fiz0             = 0.0;
 186         fix1             = 0.0;
 187         fiy1             = 0.0;
 188         fiz1             = 0.0;
 189         fix2             = 0.0;
 190         fiy2             = 0.0;
 191         fiz2             = 0.0;
 192         fix3             = 0.0;
 193         fiy3             = 0.0;
 194         fiz3             = 0.0;
 195
 196         /* Reset potential sums */
 197         velecsum         = 0.0;
 198         vvdwsum          = 0.0;
 199
 200         /* Start inner kernel loop */
 201         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 202         {
 203             /* Get j neighbor index, and coordinate index */
 204             jnr              = jjnr[jidx];
 205             j_coord_offset   = DIM*jnr;
 206
 207             /* load j atom coordinates */
 208             jx0              = x[j_coord_offset+DIM*0+XX];
 209             jy0              = x[j_coord_offset+DIM*0+YY];
 210             jz0              = x[j_coord_offset+DIM*0+ZZ];
 211             jx1              = x[j_coord_offset+DIM*1+XX];
 212             jy1              = x[j_coord_offset+DIM*1+YY];
 213             jz1              = x[j_coord_offset+DIM*1+ZZ];
 214             jx2              = x[j_coord_offset+DIM*2+XX];
 215             jy2              = x[j_coord_offset+DIM*2+YY];
 216             jz2              = x[j_coord_offset+DIM*2+ZZ];
 217             jx3              = x[j_coord_offset+DIM*3+XX];
 218             jy3              = x[j_coord_offset+DIM*3+YY];
 219             jz3              = x[j_coord_offset+DIM*3+ZZ];
 220
 221             /* Calculate displacement vector */
 222             dx00             = ix0 - jx0;
 223             dy00             = iy0 - jy0;
 224             dz00             = iz0 - jz0;
 225             dx11             = ix1 - jx1;
 226             dy11             = iy1 - jy1;
 227             dz11             = iz1 - jz1;
 228             dx12             = ix1 - jx2;
 229             dy12             = iy1 - jy2;
 230             dz12             = iz1 - jz2;
 231             dx13             = ix1 - jx3;
 232             dy13             = iy1 - jy3;
 233             dz13             = iz1 - jz3;
 234             dx21             = ix2 - jx1;
 235             dy21             = iy2 - jy1;
 236             dz21             = iz2 - jz1;
 237             dx22             = ix2 - jx2;
 238             dy22             = iy2 - jy2;
 239             dz22             = iz2 - jz2;
 240             dx23             = ix2 - jx3;
 241             dy23             = iy2 - jy3;
 242             dz23             = iz2 - jz3;
 243             dx31             = ix3 - jx1;
 244             dy31             = iy3 - jy1;
 245             dz31             = iz3 - jz1;
 246             dx32             = ix3 - jx2;
 247             dy32             = iy3 - jy2;
 248             dz32             = iz3 - jz2;
 249             dx33             = ix3 - jx3;
 250             dy33             = iy3 - jy3;
 251             dz33             = iz3 - jz3;
 252
 253             /* Calculate squared distance and things based on it */
 254             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 255             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 256             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 257             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 258             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 259             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 260             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 261             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 262             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 263             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 264
 265             rinv00           = gmx_invsqrt(rsq00);
 266             rinv11           = gmx_invsqrt(rsq11);
 267             rinv12           = gmx_invsqrt(rsq12);
 268             rinv13           = gmx_invsqrt(rsq13);
 269             rinv21           = gmx_invsqrt(rsq21);
 270             rinv22           = gmx_invsqrt(rsq22);
 271             rinv23           = gmx_invsqrt(rsq23);
 272             rinv31           = gmx_invsqrt(rsq31);
 273             rinv32           = gmx_invsqrt(rsq32);
 274             rinv33           = gmx_invsqrt(rsq33);
 275
 276             rinvsq11         = rinv11*rinv11;
 277             rinvsq12         = rinv12*rinv12;
 278             rinvsq13         = rinv13*rinv13;
 279             rinvsq21         = rinv21*rinv21;
 280             rinvsq22         = rinv22*rinv22;
 281             rinvsq23         = rinv23*rinv23;
 282             rinvsq31         = rinv31*rinv31;
 283             rinvsq32         = rinv32*rinv32;
 284             rinvsq33         = rinv33*rinv33;
 285
 286             /**************************
 287              * CALCULATE INTERACTIONS *
 288              **************************/
 289
 290             r00              = rsq00*rinv00;
 291
 292             /* Calculate table index by multiplying r with table scale and truncate to integer */
 293             rt               = r00*vftabscale;
 294             vfitab           = rt;
 295             vfeps            = rt-vfitab;
 296             vfitab           = 2*4*vfitab;
 297
 298             /* CUBIC SPLINE TABLE DISPERSION */
 299             vfitab          += 0;
 300             Y                = vftab[vfitab];
 301             F                = vftab[vfitab+1];
 302             Geps             = vfeps*vftab[vfitab+2];
 303             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 304             Fp               = F+Geps+Heps2;
 305             VV               = Y+vfeps*Fp;
 306             vvdw6            = c6_00*VV;
 307             FF               = Fp+Geps+2.0*Heps2;
 308             fvdw6            = c6_00*FF;
 309
 310             /* CUBIC SPLINE TABLE REPULSION */
 311             Y                = vftab[vfitab+4];
 312             F                = vftab[vfitab+5];
 313             Geps             = vfeps*vftab[vfitab+6];
 314             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 315             Fp               = F+Geps+Heps2;
 316             VV               = Y+vfeps*Fp;
 317             vvdw12           = c12_00*VV;
 318             FF               = Fp+Geps+2.0*Heps2;
 319             fvdw12           = c12_00*FF;
 320             vvdw             = vvdw12+vvdw6;
 321             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 322
 323             /* Update potential sums from outer loop */
 324             vvdwsum         += vvdw;
 325
 326             fscal            = fvdw;
 327
 328             /* Calculate temporary vectorial force */
 329             tx               = fscal*dx00;
 330             ty               = fscal*dy00;
 331             tz               = fscal*dz00;
 332
 333             /* Update vectorial force */
 334             fix0            += tx;
 335             fiy0            += ty;
 336             fiz0            += tz;
 337             f[j_coord_offset+DIM*0+XX] -= tx;
 338             f[j_coord_offset+DIM*0+YY] -= ty;
 339             f[j_coord_offset+DIM*0+ZZ] -= tz;
 340
 341             /**************************
 342              * CALCULATE INTERACTIONS *
 343              **************************/
 344
 345             /* COULOMB ELECTROSTATICS */
 346             velec            = qq11*rinv11;
 347             felec            = velec*rinvsq11;
 348
 349             /* Update potential sums from outer loop */
 350             velecsum        += velec;
 351
 352             fscal            = felec;
 353
 354             /* Calculate temporary vectorial force */
 355             tx               = fscal*dx11;
 356             ty               = fscal*dy11;
 357             tz               = fscal*dz11;
 358
 359             /* Update vectorial force */
 360             fix1            += tx;
 361             fiy1            += ty;
 362             fiz1            += tz;
 363             f[j_coord_offset+DIM*1+XX] -= tx;
 364             f[j_coord_offset+DIM*1+YY] -= ty;
 365             f[j_coord_offset+DIM*1+ZZ] -= tz;
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             /* COULOMB ELECTROSTATICS */
 372             velec            = qq12*rinv12;
 373             felec            = velec*rinvsq12;
 374
 375             /* Update potential sums from outer loop */
 376             velecsum        += velec;
 377
 378             fscal            = felec;
 379
 380             /* Calculate temporary vectorial force */
 381             tx               = fscal*dx12;
 382             ty               = fscal*dy12;
 383             tz               = fscal*dz12;
 384
 385             /* Update vectorial force */
 386             fix1            += tx;
 387             fiy1            += ty;
 388             fiz1            += tz;
 389             f[j_coord_offset+DIM*2+XX] -= tx;
 390             f[j_coord_offset+DIM*2+YY] -= ty;
 391             f[j_coord_offset+DIM*2+ZZ] -= tz;
 392
 393             /**************************
 394              * CALCULATE INTERACTIONS *
 395              **************************/
 396
 397             /* COULOMB ELECTROSTATICS */
 398             velec            = qq13*rinv13;
 399             felec            = velec*rinvsq13;
 400
 401             /* Update potential sums from outer loop */
 402             velecsum        += velec;
 403
 404             fscal            = felec;
 405
 406             /* Calculate temporary vectorial force */
 407             tx               = fscal*dx13;
 408             ty               = fscal*dy13;
 409             tz               = fscal*dz13;
 410
 411             /* Update vectorial force */
 412             fix1            += tx;
 413             fiy1            += ty;
 414             fiz1            += tz;
 415             f[j_coord_offset+DIM*3+XX] -= tx;
 416             f[j_coord_offset+DIM*3+YY] -= ty;
 417             f[j_coord_offset+DIM*3+ZZ] -= tz;
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             /* COULOMB ELECTROSTATICS */
 424             velec            = qq21*rinv21;
 425             felec            = velec*rinvsq21;
 426
 427             /* Update potential sums from outer loop */
 428             velecsum        += velec;
 429
 430             fscal            = felec;
 431
 432             /* Calculate temporary vectorial force */
 433             tx               = fscal*dx21;
 434             ty               = fscal*dy21;
 435             tz               = fscal*dz21;
 436
 437             /* Update vectorial force */
 438             fix2            += tx;
 439             fiy2            += ty;
 440             fiz2            += tz;
 441             f[j_coord_offset+DIM*1+XX] -= tx;
 442             f[j_coord_offset+DIM*1+YY] -= ty;
 443             f[j_coord_offset+DIM*1+ZZ] -= tz;
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* COULOMB ELECTROSTATICS */
 450             velec            = qq22*rinv22;
 451             felec            = velec*rinvsq22;
 452
 453             /* Update potential sums from outer loop */
 454             velecsum        += velec;
 455
 456             fscal            = felec;
 457
 458             /* Calculate temporary vectorial force */
 459             tx               = fscal*dx22;
 460             ty               = fscal*dy22;
 461             tz               = fscal*dz22;
 462
 463             /* Update vectorial force */
 464             fix2            += tx;
 465             fiy2            += ty;
 466             fiz2            += tz;
 467             f[j_coord_offset+DIM*2+XX] -= tx;
 468             f[j_coord_offset+DIM*2+YY] -= ty;
 469             f[j_coord_offset+DIM*2+ZZ] -= tz;
 470
 471             /**************************
 472              * CALCULATE INTERACTIONS *
 473              **************************/
 474
 475             /* COULOMB ELECTROSTATICS */
 476             velec            = qq23*rinv23;
 477             felec            = velec*rinvsq23;
 478
 479             /* Update potential sums from outer loop */
 480             velecsum        += velec;
 481
 482             fscal            = felec;
 483
 484             /* Calculate temporary vectorial force */
 485             tx               = fscal*dx23;
 486             ty               = fscal*dy23;
 487             tz               = fscal*dz23;
 488
 489             /* Update vectorial force */
 490             fix2            += tx;
 491             fiy2            += ty;
 492             fiz2            += tz;
 493             f[j_coord_offset+DIM*3+XX] -= tx;
 494             f[j_coord_offset+DIM*3+YY] -= ty;
 495             f[j_coord_offset+DIM*3+ZZ] -= tz;
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             /* COULOMB ELECTROSTATICS */
 502             velec            = qq31*rinv31;
 503             felec            = velec*rinvsq31;
 504
 505             /* Update potential sums from outer loop */
 506             velecsum        += velec;
 507
 508             fscal            = felec;
 509
 510             /* Calculate temporary vectorial force */
 511             tx               = fscal*dx31;
 512             ty               = fscal*dy31;
 513             tz               = fscal*dz31;
 514
 515             /* Update vectorial force */
 516             fix3            += tx;
 517             fiy3            += ty;
 518             fiz3            += tz;
 519             f[j_coord_offset+DIM*1+XX] -= tx;
 520             f[j_coord_offset+DIM*1+YY] -= ty;
 521             f[j_coord_offset+DIM*1+ZZ] -= tz;
 522
 523             /**************************
 524              * CALCULATE INTERACTIONS *
 525              **************************/
 526
 527             /* COULOMB ELECTROSTATICS */
 528             velec            = qq32*rinv32;
 529             felec            = velec*rinvsq32;
 530
 531             /* Update potential sums from outer loop */
 532             velecsum        += velec;
 533
 534             fscal            = felec;
 535
 536             /* Calculate temporary vectorial force */
 537             tx               = fscal*dx32;
 538             ty               = fscal*dy32;
 539             tz               = fscal*dz32;
 540
 541             /* Update vectorial force */
 542             fix3            += tx;
 543             fiy3            += ty;
 544             fiz3            += tz;
 545             f[j_coord_offset+DIM*2+XX] -= tx;
 546             f[j_coord_offset+DIM*2+YY] -= ty;
 547             f[j_coord_offset+DIM*2+ZZ] -= tz;
 548
 549             /**************************
 550              * CALCULATE INTERACTIONS *
 551              **************************/
 552
 553             /* COULOMB ELECTROSTATICS */
 554             velec            = qq33*rinv33;
 555             felec            = velec*rinvsq33;
 556
 557             /* Update potential sums from outer loop */
 558             velecsum        += velec;
 559
 560             fscal            = felec;
 561
 562             /* Calculate temporary vectorial force */
 563             tx               = fscal*dx33;
 564             ty               = fscal*dy33;
 565             tz               = fscal*dz33;
 566
 567             /* Update vectorial force */
 568             fix3            += tx;
 569             fiy3            += ty;
 570             fiz3            += tz;
 571             f[j_coord_offset+DIM*3+XX] -= tx;
 572             f[j_coord_offset+DIM*3+YY] -= ty;
 573             f[j_coord_offset+DIM*3+ZZ] -= tz;
 574
 575             /* Inner loop uses 298 flops */
 576         }
 577         /* End of innermost loop */
 578
 579         tx = ty = tz = 0;
 580         f[i_coord_offset+DIM*0+XX] += fix0;
 581         f[i_coord_offset+DIM*0+YY] += fiy0;
 582         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 583         tx                         += fix0;
 584         ty                         += fiy0;
 585         tz                         += fiz0;
 586         f[i_coord_offset+DIM*1+XX] += fix1;
 587         f[i_coord_offset+DIM*1+YY] += fiy1;
 588         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 589         tx                         += fix1;
 590         ty                         += fiy1;
 591         tz                         += fiz1;
 592         f[i_coord_offset+DIM*2+XX] += fix2;
 593         f[i_coord_offset+DIM*2+YY] += fiy2;
 594         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 595         tx                         += fix2;
 596         ty                         += fiy2;
 597         tz                         += fiz2;
 598         f[i_coord_offset+DIM*3+XX] += fix3;
 599         f[i_coord_offset+DIM*3+YY] += fiy3;
 600         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 601         tx                         += fix3;
 602         ty                         += fiy3;
 603         tz                         += fiz3;
 604         fshift[i_shift_offset+XX]  += tx;
 605         fshift[i_shift_offset+YY]  += ty;
 606         fshift[i_shift_offset+ZZ]  += tz;
 607
 608         ggid                        = gid[iidx];
 609         /* Update potential energies */
 610         kernel_data->energygrp_elec[ggid] += velecsum;
 611         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 612
 613         /* Increment number of inner iterations */
 614         inneriter                  += j_index_end - j_index_start;
 615
 616         /* Outer loop uses 41 flops */
 617     }
 618
 619     /* Increment number of outer iterations */
 620     outeriter        += nri;
 621
 622     /* Update outer/inner flops */
 623
 624     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*298);
 625 }
 626 /*
 627  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_c
 628  * Electrostatics interaction: Coulomb
 629  * VdW interaction:            CubicSplineTable
 630  * Geometry:                   Water4-Water4
 631  * Calculate force/pot:        Force
 632  */
 633 void
 634 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_c
 635                     (t_nblist                    * gmx_restrict       nlist,
 636                      rvec                        * gmx_restrict          xx,
 637                      rvec                        * gmx_restrict          ff,
 638                      t_forcerec                  * gmx_restrict          fr,
 639                      t_mdatoms                   * gmx_restrict     mdatoms,
 640                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 641                      t_nrnb                      * gmx_restrict        nrnb)
 642 {
 643     int              i_shift_offset,i_coord_offset,j_coord_offset;
 644     int              j_index_start,j_index_end;
 645     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 646     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 647     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 648     real             *shiftvec,*fshift,*x,*f;
 649     int              vdwioffset0;
 650     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 651     int              vdwioffset1;
 652     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 653     int              vdwioffset2;
 654     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 655     int              vdwioffset3;
 656     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 657     int              vdwjidx0;
 658     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 659     int              vdwjidx1;
 660     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 661     int              vdwjidx2;
 662     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 663     int              vdwjidx3;
 664     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 665     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 666     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 667     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 668     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 669     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 670     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 671     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 672     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 673     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 674     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 675     real             velec,felec,velecsum,facel,crf,krf,krf2;
 676     real             *charge;
 677     int              nvdwtype;
 678     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 679     int              *vdwtype;
 680     real             *vdwparam;
 681     int              vfitab;
 682     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 683     real             *vftab;
 684
 685     x                = xx[0];
 686     f                = ff[0];
 687
 688     nri              = nlist->nri;
 689     iinr             = nlist->iinr;
 690     jindex           = nlist->jindex;
 691     jjnr             = nlist->jjnr;
 692     shiftidx         = nlist->shift;
 693     gid              = nlist->gid;
 694     shiftvec         = fr->shift_vec[0];
 695     fshift           = fr->fshift[0];
 696     facel            = fr->epsfac;
 697     charge           = mdatoms->chargeA;
 698     nvdwtype         = fr->ntype;
 699     vdwparam         = fr->nbfp;
 700     vdwtype          = mdatoms->typeA;
 701
 702     vftab            = kernel_data->table_vdw->data;
 703     vftabscale       = kernel_data->table_vdw->scale;
 704
 705     /* Setup water-specific parameters */
 706     inr              = nlist->iinr[0];
 707     iq1              = facel*charge[inr+1];
 708     iq2              = facel*charge[inr+2];
 709     iq3              = facel*charge[inr+3];
 710     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 711
 712     jq1              = charge[inr+1];
 713     jq2              = charge[inr+2];
 714     jq3              = charge[inr+3];
 715     vdwjidx0         = 2*vdwtype[inr+0];
 716     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 717     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 718     qq11             = iq1*jq1;
 719     qq12             = iq1*jq2;
 720     qq13             = iq1*jq3;
 721     qq21             = iq2*jq1;
 722     qq22             = iq2*jq2;
 723     qq23             = iq2*jq3;
 724     qq31             = iq3*jq1;
 725     qq32             = iq3*jq2;
 726     qq33             = iq3*jq3;
 727
 728     outeriter        = 0;
 729     inneriter        = 0;
 730
 731     /* Start outer loop over neighborlists */
 732     for(iidx=0; iidx<nri; iidx++)
 733     {
 734         /* Load shift vector for this list */
 735         i_shift_offset   = DIM*shiftidx[iidx];
 736         shX              = shiftvec[i_shift_offset+XX];
 737         shY              = shiftvec[i_shift_offset+YY];
 738         shZ              = shiftvec[i_shift_offset+ZZ];
 739
 740         /* Load limits for loop over neighbors */
 741         j_index_start    = jindex[iidx];
 742         j_index_end      = jindex[iidx+1];
 743
 744         /* Get outer coordinate index */
 745         inr              = iinr[iidx];
 746         i_coord_offset   = DIM*inr;
 747
 748         /* Load i particle coords and add shift vector */
 749         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 750         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 751         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 752         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 753         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 754         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 755         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 756         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 757         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 758         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 759         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 760         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 761
 762         fix0             = 0.0;
 763         fiy0             = 0.0;
 764         fiz0             = 0.0;
 765         fix1             = 0.0;
 766         fiy1             = 0.0;
 767         fiz1             = 0.0;
 768         fix2             = 0.0;
 769         fiy2             = 0.0;
 770         fiz2             = 0.0;
 771         fix3             = 0.0;
 772         fiy3             = 0.0;
 773         fiz3             = 0.0;
 774
 775         /* Start inner kernel loop */
 776         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 777         {
 778             /* Get j neighbor index, and coordinate index */
 779             jnr              = jjnr[jidx];
 780             j_coord_offset   = DIM*jnr;
 781
 782             /* load j atom coordinates */
 783             jx0              = x[j_coord_offset+DIM*0+XX];
 784             jy0              = x[j_coord_offset+DIM*0+YY];
 785             jz0              = x[j_coord_offset+DIM*0+ZZ];
 786             jx1              = x[j_coord_offset+DIM*1+XX];
 787             jy1              = x[j_coord_offset+DIM*1+YY];
 788             jz1              = x[j_coord_offset+DIM*1+ZZ];
 789             jx2              = x[j_coord_offset+DIM*2+XX];
 790             jy2              = x[j_coord_offset+DIM*2+YY];
 791             jz2              = x[j_coord_offset+DIM*2+ZZ];
 792             jx3              = x[j_coord_offset+DIM*3+XX];
 793             jy3              = x[j_coord_offset+DIM*3+YY];
 794             jz3              = x[j_coord_offset+DIM*3+ZZ];
 795
 796             /* Calculate displacement vector */
 797             dx00             = ix0 - jx0;
 798             dy00             = iy0 - jy0;
 799             dz00             = iz0 - jz0;
 800             dx11             = ix1 - jx1;
 801             dy11             = iy1 - jy1;
 802             dz11             = iz1 - jz1;
 803             dx12             = ix1 - jx2;
 804             dy12             = iy1 - jy2;
 805             dz12             = iz1 - jz2;
 806             dx13             = ix1 - jx3;
 807             dy13             = iy1 - jy3;
 808             dz13             = iz1 - jz3;
 809             dx21             = ix2 - jx1;
 810             dy21             = iy2 - jy1;
 811             dz21             = iz2 - jz1;
 812             dx22             = ix2 - jx2;
 813             dy22             = iy2 - jy2;
 814             dz22             = iz2 - jz2;
 815             dx23             = ix2 - jx3;
 816             dy23             = iy2 - jy3;
 817             dz23             = iz2 - jz3;
 818             dx31             = ix3 - jx1;
 819             dy31             = iy3 - jy1;
 820             dz31             = iz3 - jz1;
 821             dx32             = ix3 - jx2;
 822             dy32             = iy3 - jy2;
 823             dz32             = iz3 - jz2;
 824             dx33             = ix3 - jx3;
 825             dy33             = iy3 - jy3;
 826             dz33             = iz3 - jz3;
 827
 828             /* Calculate squared distance and things based on it */
 829             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 830             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 831             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 832             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 833             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 834             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 835             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 836             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 837             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 838             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 839
 840             rinv00           = gmx_invsqrt(rsq00);
 841             rinv11           = gmx_invsqrt(rsq11);
 842             rinv12           = gmx_invsqrt(rsq12);
 843             rinv13           = gmx_invsqrt(rsq13);
 844             rinv21           = gmx_invsqrt(rsq21);
 845             rinv22           = gmx_invsqrt(rsq22);
 846             rinv23           = gmx_invsqrt(rsq23);
 847             rinv31           = gmx_invsqrt(rsq31);
 848             rinv32           = gmx_invsqrt(rsq32);
 849             rinv33           = gmx_invsqrt(rsq33);
 850
 851             rinvsq11         = rinv11*rinv11;
 852             rinvsq12         = rinv12*rinv12;
 853             rinvsq13         = rinv13*rinv13;
 854             rinvsq21         = rinv21*rinv21;
 855             rinvsq22         = rinv22*rinv22;
 856             rinvsq23         = rinv23*rinv23;
 857             rinvsq31         = rinv31*rinv31;
 858             rinvsq32         = rinv32*rinv32;
 859             rinvsq33         = rinv33*rinv33;
 860
 861             /**************************
 862              * CALCULATE INTERACTIONS *
 863              **************************/
 864
 865             r00              = rsq00*rinv00;
 866
 867             /* Calculate table index by multiplying r with table scale and truncate to integer */
 868             rt               = r00*vftabscale;
 869             vfitab           = rt;
 870             vfeps            = rt-vfitab;
 871             vfitab           = 2*4*vfitab;
 872
 873             /* CUBIC SPLINE TABLE DISPERSION */
 874             vfitab          += 0;
 875             F                = vftab[vfitab+1];
 876             Geps             = vfeps*vftab[vfitab+2];
 877             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 878             Fp               = F+Geps+Heps2;
 879             FF               = Fp+Geps+2.0*Heps2;
 880             fvdw6            = c6_00*FF;
 881
 882             /* CUBIC SPLINE TABLE REPULSION */
 883             F                = vftab[vfitab+5];
 884             Geps             = vfeps*vftab[vfitab+6];
 885             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 886             Fp               = F+Geps+Heps2;
 887             FF               = Fp+Geps+2.0*Heps2;
 888             fvdw12           = c12_00*FF;
 889             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 890
 891             fscal            = fvdw;
 892
 893             /* Calculate temporary vectorial force */
 894             tx               = fscal*dx00;
 895             ty               = fscal*dy00;
 896             tz               = fscal*dz00;
 897
 898             /* Update vectorial force */
 899             fix0            += tx;
 900             fiy0            += ty;
 901             fiz0            += tz;
 902             f[j_coord_offset+DIM*0+XX] -= tx;
 903             f[j_coord_offset+DIM*0+YY] -= ty;
 904             f[j_coord_offset+DIM*0+ZZ] -= tz;
 905
 906             /**************************
 907              * CALCULATE INTERACTIONS *
 908              **************************/
 909
 910             /* COULOMB ELECTROSTATICS */
 911             velec            = qq11*rinv11;
 912             felec            = velec*rinvsq11;
 913
 914             fscal            = felec;
 915
 916             /* Calculate temporary vectorial force */
 917             tx               = fscal*dx11;
 918             ty               = fscal*dy11;
 919             tz               = fscal*dz11;
 920
 921             /* Update vectorial force */
 922             fix1            += tx;
 923             fiy1            += ty;
 924             fiz1            += tz;
 925             f[j_coord_offset+DIM*1+XX] -= tx;
 926             f[j_coord_offset+DIM*1+YY] -= ty;
 927             f[j_coord_offset+DIM*1+ZZ] -= tz;
 928
 929             /**************************
 930              * CALCULATE INTERACTIONS *
 931              **************************/
 932
 933             /* COULOMB ELECTROSTATICS */
 934             velec            = qq12*rinv12;
 935             felec            = velec*rinvsq12;
 936
 937             fscal            = felec;
 938
 939             /* Calculate temporary vectorial force */
 940             tx               = fscal*dx12;
 941             ty               = fscal*dy12;
 942             tz               = fscal*dz12;
 943
 944             /* Update vectorial force */
 945             fix1            += tx;
 946             fiy1            += ty;
 947             fiz1            += tz;
 948             f[j_coord_offset+DIM*2+XX] -= tx;
 949             f[j_coord_offset+DIM*2+YY] -= ty;
 950             f[j_coord_offset+DIM*2+ZZ] -= tz;
 951
 952             /**************************
 953              * CALCULATE INTERACTIONS *
 954              **************************/
 955
 956             /* COULOMB ELECTROSTATICS */
 957             velec            = qq13*rinv13;
 958             felec            = velec*rinvsq13;
 959
 960             fscal            = felec;
 961
 962             /* Calculate temporary vectorial force */
 963             tx               = fscal*dx13;
 964             ty               = fscal*dy13;
 965             tz               = fscal*dz13;
 966
 967             /* Update vectorial force */
 968             fix1            += tx;
 969             fiy1            += ty;
 970             fiz1            += tz;
 971             f[j_coord_offset+DIM*3+XX] -= tx;
 972             f[j_coord_offset+DIM*3+YY] -= ty;
 973             f[j_coord_offset+DIM*3+ZZ] -= tz;
 974
 975             /**************************
 976              * CALCULATE INTERACTIONS *
 977              **************************/
 978
 979             /* COULOMB ELECTROSTATICS */
 980             velec            = qq21*rinv21;
 981             felec            = velec*rinvsq21;
 982
 983             fscal            = felec;
 984
 985             /* Calculate temporary vectorial force */
 986             tx               = fscal*dx21;
 987             ty               = fscal*dy21;
 988             tz               = fscal*dz21;
 989
 990             /* Update vectorial force */
 991             fix2            += tx;
 992             fiy2            += ty;
 993             fiz2            += tz;
 994             f[j_coord_offset+DIM*1+XX] -= tx;
 995             f[j_coord_offset+DIM*1+YY] -= ty;
 996             f[j_coord_offset+DIM*1+ZZ] -= tz;
 997
 998             /**************************
 999              * CALCULATE INTERACTIONS *
1000              **************************/
1001
1002             /* COULOMB ELECTROSTATICS */
1003             velec            = qq22*rinv22;
1004             felec            = velec*rinvsq22;
1005
1006             fscal            = felec;
1007
1008             /* Calculate temporary vectorial force */
1009             tx               = fscal*dx22;
1010             ty               = fscal*dy22;
1011             tz               = fscal*dz22;
1012
1013             /* Update vectorial force */
1014             fix2            += tx;
1015             fiy2            += ty;
1016             fiz2            += tz;
1017             f[j_coord_offset+DIM*2+XX] -= tx;
1018             f[j_coord_offset+DIM*2+YY] -= ty;
1019             f[j_coord_offset+DIM*2+ZZ] -= tz;
1020
1021             /**************************
1022              * CALCULATE INTERACTIONS *
1023              **************************/
1024
1025             /* COULOMB ELECTROSTATICS */
1026             velec            = qq23*rinv23;
1027             felec            = velec*rinvsq23;
1028
1029             fscal            = felec;
1030
1031             /* Calculate temporary vectorial force */
1032             tx               = fscal*dx23;
1033             ty               = fscal*dy23;
1034             tz               = fscal*dz23;
1035
1036             /* Update vectorial force */
1037             fix2            += tx;
1038             fiy2            += ty;
1039             fiz2            += tz;
1040             f[j_coord_offset+DIM*3+XX] -= tx;
1041             f[j_coord_offset+DIM*3+YY] -= ty;
1042             f[j_coord_offset+DIM*3+ZZ] -= tz;
1043
1044             /**************************
1045              * CALCULATE INTERACTIONS *
1046              **************************/
1047
1048             /* COULOMB ELECTROSTATICS */
1049             velec            = qq31*rinv31;
1050             felec            = velec*rinvsq31;
1051
1052             fscal            = felec;
1053
1054             /* Calculate temporary vectorial force */
1055             tx               = fscal*dx31;
1056             ty               = fscal*dy31;
1057             tz               = fscal*dz31;
1058
1059             /* Update vectorial force */
1060             fix3            += tx;
1061             fiy3            += ty;
1062             fiz3            += tz;
1063             f[j_coord_offset+DIM*1+XX] -= tx;
1064             f[j_coord_offset+DIM*1+YY] -= ty;
1065             f[j_coord_offset+DIM*1+ZZ] -= tz;
1066
1067             /**************************
1068              * CALCULATE INTERACTIONS *
1069              **************************/
1070
1071             /* COULOMB ELECTROSTATICS */
1072             velec            = qq32*rinv32;
1073             felec            = velec*rinvsq32;
1074
1075             fscal            = felec;
1076
1077             /* Calculate temporary vectorial force */
1078             tx               = fscal*dx32;
1079             ty               = fscal*dy32;
1080             tz               = fscal*dz32;
1081
1082             /* Update vectorial force */
1083             fix3            += tx;
1084             fiy3            += ty;
1085             fiz3            += tz;
1086             f[j_coord_offset+DIM*2+XX] -= tx;
1087             f[j_coord_offset+DIM*2+YY] -= ty;
1088             f[j_coord_offset+DIM*2+ZZ] -= tz;
1089
1090             /**************************
1091              * CALCULATE INTERACTIONS *
1092              **************************/
1093
1094             /* COULOMB ELECTROSTATICS */
1095             velec            = qq33*rinv33;
1096             felec            = velec*rinvsq33;
1097
1098             fscal            = felec;
1099
1100             /* Calculate temporary vectorial force */
1101             tx               = fscal*dx33;
1102             ty               = fscal*dy33;
1103             tz               = fscal*dz33;
1104
1105             /* Update vectorial force */
1106             fix3            += tx;
1107             fiy3            += ty;
1108             fiz3            += tz;
1109             f[j_coord_offset+DIM*3+XX] -= tx;
1110             f[j_coord_offset+DIM*3+YY] -= ty;
1111             f[j_coord_offset+DIM*3+ZZ] -= tz;
1112
1113             /* Inner loop uses 281 flops */
1114         }
1115         /* End of innermost loop */
1116
1117         tx = ty = tz = 0;
1118         f[i_coord_offset+DIM*0+XX] += fix0;
1119         f[i_coord_offset+DIM*0+YY] += fiy0;
1120         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1121         tx                         += fix0;
1122         ty                         += fiy0;
1123         tz                         += fiz0;
1124         f[i_coord_offset+DIM*1+XX] += fix1;
1125         f[i_coord_offset+DIM*1+YY] += fiy1;
1126         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1127         tx                         += fix1;
1128         ty                         += fiy1;
1129         tz                         += fiz1;
1130         f[i_coord_offset+DIM*2+XX] += fix2;
1131         f[i_coord_offset+DIM*2+YY] += fiy2;
1132         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1133         tx                         += fix2;
1134         ty                         += fiy2;
1135         tz                         += fiz2;
1136         f[i_coord_offset+DIM*3+XX] += fix3;
1137         f[i_coord_offset+DIM*3+YY] += fiy3;
1138         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1139         tx                         += fix3;
1140         ty                         += fiy3;
1141         tz                         += fiz3;
1142         fshift[i_shift_offset+XX]  += tx;
1143         fshift[i_shift_offset+YY]  += ty;
1144         fshift[i_shift_offset+ZZ]  += tz;
1145
1146         /* Increment number of inner iterations */
1147         inneriter                  += j_index_end - j_index_start;
1148
1149         /* Outer loop uses 39 flops */
1150     }
1151
1152     /* Increment number of outer iterations */
1153     outeriter        += nri;
1154
1155     /* Update outer/inner flops */
1156
1157     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*281);
1158 }