src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEw_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_ElecEw_VdwCSTab_GeomW4W4_VF_c
  49  * Electrostatics interaction: Ewald
  50  * VdW interaction:            CubicSplineTable
  51  * Geometry:                   Water4-Water4
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecEw_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     int              ewitab;
 106     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 107     real             *ewtab;
 108
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = fr->epsfac;
 121     charge           = mdatoms->chargeA;
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     vftab            = kernel_data->table_vdw->data;
 127     vftabscale       = kernel_data->table_vdw->scale;
 128
 129     sh_ewald         = fr->ic->sh_ewald;
 130     ewtab            = fr->ic->tabq_coul_FDV0;
 131     ewtabscale       = fr->ic->tabq_scale;
 132     ewtabhalfspace   = 0.5/ewtabscale;
 133
 134     /* Setup water-specific parameters */
 135     inr              = nlist->iinr[0];
 136     iq1              = facel*charge[inr+1];
 137     iq2              = facel*charge[inr+2];
 138     iq3              = facel*charge[inr+3];
 139     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 140
 141     jq1              = charge[inr+1];
 142     jq2              = charge[inr+2];
 143     jq3              = charge[inr+3];
 144     vdwjidx0         = 2*vdwtype[inr+0];
 145     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 146     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 147     qq11             = iq1*jq1;
 148     qq12             = iq1*jq2;
 149     qq13             = iq1*jq3;
 150     qq21             = iq2*jq1;
 151     qq22             = iq2*jq2;
 152     qq23             = iq2*jq3;
 153     qq31             = iq3*jq1;
 154     qq32             = iq3*jq2;
 155     qq33             = iq3*jq3;
 156
 157     outeriter        = 0;
 158     inneriter        = 0;
 159
 160     /* Start outer loop over neighborlists */
 161     for(iidx=0; iidx<nri; iidx++)
 162     {
 163         /* Load shift vector for this list */
 164         i_shift_offset   = DIM*shiftidx[iidx];
 165         shX              = shiftvec[i_shift_offset+XX];
 166         shY              = shiftvec[i_shift_offset+YY];
 167         shZ              = shiftvec[i_shift_offset+ZZ];
 168
 169         /* Load limits for loop over neighbors */
 170         j_index_start    = jindex[iidx];
 171         j_index_end      = jindex[iidx+1];
 172
 173         /* Get outer coordinate index */
 174         inr              = iinr[iidx];
 175         i_coord_offset   = DIM*inr;
 176
 177         /* Load i particle coords and add shift vector */
 178         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 179         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 180         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 181         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 182         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 183         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 184         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 185         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 186         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 187         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 188         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 189         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 190
 191         fix0             = 0.0;
 192         fiy0             = 0.0;
 193         fiz0             = 0.0;
 194         fix1             = 0.0;
 195         fiy1             = 0.0;
 196         fiz1             = 0.0;
 197         fix2             = 0.0;
 198         fiy2             = 0.0;
 199         fiz2             = 0.0;
 200         fix3             = 0.0;
 201         fiy3             = 0.0;
 202         fiz3             = 0.0;
 203
 204         /* Reset potential sums */
 205         velecsum         = 0.0;
 206         vvdwsum          = 0.0;
 207
 208         /* Start inner kernel loop */
 209         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 210         {
 211             /* Get j neighbor index, and coordinate index */
 212             jnr              = jjnr[jidx];
 213             j_coord_offset   = DIM*jnr;
 214
 215             /* load j atom coordinates */
 216             jx0              = x[j_coord_offset+DIM*0+XX];
 217             jy0              = x[j_coord_offset+DIM*0+YY];
 218             jz0              = x[j_coord_offset+DIM*0+ZZ];
 219             jx1              = x[j_coord_offset+DIM*1+XX];
 220             jy1              = x[j_coord_offset+DIM*1+YY];
 221             jz1              = x[j_coord_offset+DIM*1+ZZ];
 222             jx2              = x[j_coord_offset+DIM*2+XX];
 223             jy2              = x[j_coord_offset+DIM*2+YY];
 224             jz2              = x[j_coord_offset+DIM*2+ZZ];
 225             jx3              = x[j_coord_offset+DIM*3+XX];
 226             jy3              = x[j_coord_offset+DIM*3+YY];
 227             jz3              = x[j_coord_offset+DIM*3+ZZ];
 228
 229             /* Calculate displacement vector */
 230             dx00             = ix0 - jx0;
 231             dy00             = iy0 - jy0;
 232             dz00             = iz0 - jz0;
 233             dx11             = ix1 - jx1;
 234             dy11             = iy1 - jy1;
 235             dz11             = iz1 - jz1;
 236             dx12             = ix1 - jx2;
 237             dy12             = iy1 - jy2;
 238             dz12             = iz1 - jz2;
 239             dx13             = ix1 - jx3;
 240             dy13             = iy1 - jy3;
 241             dz13             = iz1 - jz3;
 242             dx21             = ix2 - jx1;
 243             dy21             = iy2 - jy1;
 244             dz21             = iz2 - jz1;
 245             dx22             = ix2 - jx2;
 246             dy22             = iy2 - jy2;
 247             dz22             = iz2 - jz2;
 248             dx23             = ix2 - jx3;
 249             dy23             = iy2 - jy3;
 250             dz23             = iz2 - jz3;
 251             dx31             = ix3 - jx1;
 252             dy31             = iy3 - jy1;
 253             dz31             = iz3 - jz1;
 254             dx32             = ix3 - jx2;
 255             dy32             = iy3 - jy2;
 256             dz32             = iz3 - jz2;
 257             dx33             = ix3 - jx3;
 258             dy33             = iy3 - jy3;
 259             dz33             = iz3 - jz3;
 260
 261             /* Calculate squared distance and things based on it */
 262             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 263             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 264             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 265             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 266             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 267             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 268             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 269             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 270             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 271             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 272
 273             rinv00           = gmx_invsqrt(rsq00);
 274             rinv11           = gmx_invsqrt(rsq11);
 275             rinv12           = gmx_invsqrt(rsq12);
 276             rinv13           = gmx_invsqrt(rsq13);
 277             rinv21           = gmx_invsqrt(rsq21);
 278             rinv22           = gmx_invsqrt(rsq22);
 279             rinv23           = gmx_invsqrt(rsq23);
 280             rinv31           = gmx_invsqrt(rsq31);
 281             rinv32           = gmx_invsqrt(rsq32);
 282             rinv33           = gmx_invsqrt(rsq33);
 283
 284             rinvsq11         = rinv11*rinv11;
 285             rinvsq12         = rinv12*rinv12;
 286             rinvsq13         = rinv13*rinv13;
 287             rinvsq21         = rinv21*rinv21;
 288             rinvsq22         = rinv22*rinv22;
 289             rinvsq23         = rinv23*rinv23;
 290             rinvsq31         = rinv31*rinv31;
 291             rinvsq32         = rinv32*rinv32;
 292             rinvsq33         = rinv33*rinv33;
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             r00              = rsq00*rinv00;
 299
 300             /* Calculate table index by multiplying r with table scale and truncate to integer */
 301             rt               = r00*vftabscale;
 302             vfitab           = rt;
 303             vfeps            = rt-vfitab;
 304             vfitab           = 2*4*vfitab;
 305
 306             /* CUBIC SPLINE TABLE DISPERSION */
 307             vfitab          += 0;
 308             Y                = vftab[vfitab];
 309             F                = vftab[vfitab+1];
 310             Geps             = vfeps*vftab[vfitab+2];
 311             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 312             Fp               = F+Geps+Heps2;
 313             VV               = Y+vfeps*Fp;
 314             vvdw6            = c6_00*VV;
 315             FF               = Fp+Geps+2.0*Heps2;
 316             fvdw6            = c6_00*FF;
 317
 318             /* CUBIC SPLINE TABLE REPULSION */
 319             Y                = vftab[vfitab+4];
 320             F                = vftab[vfitab+5];
 321             Geps             = vfeps*vftab[vfitab+6];
 322             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 323             Fp               = F+Geps+Heps2;
 324             VV               = Y+vfeps*Fp;
 325             vvdw12           = c12_00*VV;
 326             FF               = Fp+Geps+2.0*Heps2;
 327             fvdw12           = c12_00*FF;
 328             vvdw             = vvdw12+vvdw6;
 329             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 330
 331             /* Update potential sums from outer loop */
 332             vvdwsum         += vvdw;
 333
 334             fscal            = fvdw;
 335
 336             /* Calculate temporary vectorial force */
 337             tx               = fscal*dx00;
 338             ty               = fscal*dy00;
 339             tz               = fscal*dz00;
 340
 341             /* Update vectorial force */
 342             fix0            += tx;
 343             fiy0            += ty;
 344             fiz0            += tz;
 345             f[j_coord_offset+DIM*0+XX] -= tx;
 346             f[j_coord_offset+DIM*0+YY] -= ty;
 347             f[j_coord_offset+DIM*0+ZZ] -= tz;
 348
 349             /**************************
 350              * CALCULATE INTERACTIONS *
 351              **************************/
 352
 353             r11              = rsq11*rinv11;
 354
 355             /* EWALD ELECTROSTATICS */
 356
 357             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 358             ewrt             = r11*ewtabscale;
 359             ewitab           = ewrt;
 360             eweps            = ewrt-ewitab;
 361             ewitab           = 4*ewitab;
 362             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 363             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 364             felec            = qq11*rinv11*(rinvsq11-felec);
 365
 366             /* Update potential sums from outer loop */
 367             velecsum        += velec;
 368
 369             fscal            = felec;
 370
 371             /* Calculate temporary vectorial force */
 372             tx               = fscal*dx11;
 373             ty               = fscal*dy11;
 374             tz               = fscal*dz11;
 375
 376             /* Update vectorial force */
 377             fix1            += tx;
 378             fiy1            += ty;
 379             fiz1            += tz;
 380             f[j_coord_offset+DIM*1+XX] -= tx;
 381             f[j_coord_offset+DIM*1+YY] -= ty;
 382             f[j_coord_offset+DIM*1+ZZ] -= tz;
 383
 384             /**************************
 385              * CALCULATE INTERACTIONS *
 386              **************************/
 387
 388             r12              = rsq12*rinv12;
 389
 390             /* EWALD ELECTROSTATICS */
 391
 392             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 393             ewrt             = r12*ewtabscale;
 394             ewitab           = ewrt;
 395             eweps            = ewrt-ewitab;
 396             ewitab           = 4*ewitab;
 397             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 398             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 399             felec            = qq12*rinv12*(rinvsq12-felec);
 400
 401             /* Update potential sums from outer loop */
 402             velecsum        += velec;
 403
 404             fscal            = felec;
 405
 406             /* Calculate temporary vectorial force */
 407             tx               = fscal*dx12;
 408             ty               = fscal*dy12;
 409             tz               = fscal*dz12;
 410
 411             /* Update vectorial force */
 412             fix1            += tx;
 413             fiy1            += ty;
 414             fiz1            += tz;
 415             f[j_coord_offset+DIM*2+XX] -= tx;
 416             f[j_coord_offset+DIM*2+YY] -= ty;
 417             f[j_coord_offset+DIM*2+ZZ] -= tz;
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             r13              = rsq13*rinv13;
 424
 425             /* EWALD ELECTROSTATICS */
 426
 427             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 428             ewrt             = r13*ewtabscale;
 429             ewitab           = ewrt;
 430             eweps            = ewrt-ewitab;
 431             ewitab           = 4*ewitab;
 432             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 433             velec            = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 434             felec            = qq13*rinv13*(rinvsq13-felec);
 435
 436             /* Update potential sums from outer loop */
 437             velecsum        += velec;
 438
 439             fscal            = felec;
 440
 441             /* Calculate temporary vectorial force */
 442             tx               = fscal*dx13;
 443             ty               = fscal*dy13;
 444             tz               = fscal*dz13;
 445
 446             /* Update vectorial force */
 447             fix1            += tx;
 448             fiy1            += ty;
 449             fiz1            += tz;
 450             f[j_coord_offset+DIM*3+XX] -= tx;
 451             f[j_coord_offset+DIM*3+YY] -= ty;
 452             f[j_coord_offset+DIM*3+ZZ] -= tz;
 453
 454             /**************************
 455              * CALCULATE INTERACTIONS *
 456              **************************/
 457
 458             r21              = rsq21*rinv21;
 459
 460             /* EWALD ELECTROSTATICS */
 461
 462             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 463             ewrt             = r21*ewtabscale;
 464             ewitab           = ewrt;
 465             eweps            = ewrt-ewitab;
 466             ewitab           = 4*ewitab;
 467             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 468             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 469             felec            = qq21*rinv21*(rinvsq21-felec);
 470
 471             /* Update potential sums from outer loop */
 472             velecsum        += velec;
 473
 474             fscal            = felec;
 475
 476             /* Calculate temporary vectorial force */
 477             tx               = fscal*dx21;
 478             ty               = fscal*dy21;
 479             tz               = fscal*dz21;
 480
 481             /* Update vectorial force */
 482             fix2            += tx;
 483             fiy2            += ty;
 484             fiz2            += tz;
 485             f[j_coord_offset+DIM*1+XX] -= tx;
 486             f[j_coord_offset+DIM*1+YY] -= ty;
 487             f[j_coord_offset+DIM*1+ZZ] -= tz;
 488
 489             /**************************
 490              * CALCULATE INTERACTIONS *
 491              **************************/
 492
 493             r22              = rsq22*rinv22;
 494
 495             /* EWALD ELECTROSTATICS */
 496
 497             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 498             ewrt             = r22*ewtabscale;
 499             ewitab           = ewrt;
 500             eweps            = ewrt-ewitab;
 501             ewitab           = 4*ewitab;
 502             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 503             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 504             felec            = qq22*rinv22*(rinvsq22-felec);
 505
 506             /* Update potential sums from outer loop */
 507             velecsum        += velec;
 508
 509             fscal            = felec;
 510
 511             /* Calculate temporary vectorial force */
 512             tx               = fscal*dx22;
 513             ty               = fscal*dy22;
 514             tz               = fscal*dz22;
 515
 516             /* Update vectorial force */
 517             fix2            += tx;
 518             fiy2            += ty;
 519             fiz2            += tz;
 520             f[j_coord_offset+DIM*2+XX] -= tx;
 521             f[j_coord_offset+DIM*2+YY] -= ty;
 522             f[j_coord_offset+DIM*2+ZZ] -= tz;
 523
 524             /**************************
 525              * CALCULATE INTERACTIONS *
 526              **************************/
 527
 528             r23              = rsq23*rinv23;
 529
 530             /* EWALD ELECTROSTATICS */
 531
 532             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 533             ewrt             = r23*ewtabscale;
 534             ewitab           = ewrt;
 535             eweps            = ewrt-ewitab;
 536             ewitab           = 4*ewitab;
 537             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 538             velec            = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 539             felec            = qq23*rinv23*(rinvsq23-felec);
 540
 541             /* Update potential sums from outer loop */
 542             velecsum        += velec;
 543
 544             fscal            = felec;
 545
 546             /* Calculate temporary vectorial force */
 547             tx               = fscal*dx23;
 548             ty               = fscal*dy23;
 549             tz               = fscal*dz23;
 550
 551             /* Update vectorial force */
 552             fix2            += tx;
 553             fiy2            += ty;
 554             fiz2            += tz;
 555             f[j_coord_offset+DIM*3+XX] -= tx;
 556             f[j_coord_offset+DIM*3+YY] -= ty;
 557             f[j_coord_offset+DIM*3+ZZ] -= tz;
 558
 559             /**************************
 560              * CALCULATE INTERACTIONS *
 561              **************************/
 562
 563             r31              = rsq31*rinv31;
 564
 565             /* EWALD ELECTROSTATICS */
 566
 567             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 568             ewrt             = r31*ewtabscale;
 569             ewitab           = ewrt;
 570             eweps            = ewrt-ewitab;
 571             ewitab           = 4*ewitab;
 572             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 573             velec            = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 574             felec            = qq31*rinv31*(rinvsq31-felec);
 575
 576             /* Update potential sums from outer loop */
 577             velecsum        += velec;
 578
 579             fscal            = felec;
 580
 581             /* Calculate temporary vectorial force */
 582             tx               = fscal*dx31;
 583             ty               = fscal*dy31;
 584             tz               = fscal*dz31;
 585
 586             /* Update vectorial force */
 587             fix3            += tx;
 588             fiy3            += ty;
 589             fiz3            += tz;
 590             f[j_coord_offset+DIM*1+XX] -= tx;
 591             f[j_coord_offset+DIM*1+YY] -= ty;
 592             f[j_coord_offset+DIM*1+ZZ] -= tz;
 593
 594             /**************************
 595              * CALCULATE INTERACTIONS *
 596              **************************/
 597
 598             r32              = rsq32*rinv32;
 599
 600             /* EWALD ELECTROSTATICS */
 601
 602             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 603             ewrt             = r32*ewtabscale;
 604             ewitab           = ewrt;
 605             eweps            = ewrt-ewitab;
 606             ewitab           = 4*ewitab;
 607             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 608             velec            = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 609             felec            = qq32*rinv32*(rinvsq32-felec);
 610
 611             /* Update potential sums from outer loop */
 612             velecsum        += velec;
 613
 614             fscal            = felec;
 615
 616             /* Calculate temporary vectorial force */
 617             tx               = fscal*dx32;
 618             ty               = fscal*dy32;
 619             tz               = fscal*dz32;
 620
 621             /* Update vectorial force */
 622             fix3            += tx;
 623             fiy3            += ty;
 624             fiz3            += tz;
 625             f[j_coord_offset+DIM*2+XX] -= tx;
 626             f[j_coord_offset+DIM*2+YY] -= ty;
 627             f[j_coord_offset+DIM*2+ZZ] -= tz;
 628
 629             /**************************
 630              * CALCULATE INTERACTIONS *
 631              **************************/
 632
 633             r33              = rsq33*rinv33;
 634
 635             /* EWALD ELECTROSTATICS */
 636
 637             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 638             ewrt             = r33*ewtabscale;
 639             ewitab           = ewrt;
 640             eweps            = ewrt-ewitab;
 641             ewitab           = 4*ewitab;
 642             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 643             velec            = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 644             felec            = qq33*rinv33*(rinvsq33-felec);
 645
 646             /* Update potential sums from outer loop */
 647             velecsum        += velec;
 648
 649             fscal            = felec;
 650
 651             /* Calculate temporary vectorial force */
 652             tx               = fscal*dx33;
 653             ty               = fscal*dy33;
 654             tz               = fscal*dz33;
 655
 656             /* Update vectorial force */
 657             fix3            += tx;
 658             fiy3            += ty;
 659             fiz3            += 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             /* Inner loop uses 415 flops */
 665         }
 666         /* End of innermost loop */
 667
 668         tx = ty = tz = 0;
 669         f[i_coord_offset+DIM*0+XX] += fix0;
 670         f[i_coord_offset+DIM*0+YY] += fiy0;
 671         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 672         tx                         += fix0;
 673         ty                         += fiy0;
 674         tz                         += fiz0;
 675         f[i_coord_offset+DIM*1+XX] += fix1;
 676         f[i_coord_offset+DIM*1+YY] += fiy1;
 677         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 678         tx                         += fix1;
 679         ty                         += fiy1;
 680         tz                         += fiz1;
 681         f[i_coord_offset+DIM*2+XX] += fix2;
 682         f[i_coord_offset+DIM*2+YY] += fiy2;
 683         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 684         tx                         += fix2;
 685         ty                         += fiy2;
 686         tz                         += fiz2;
 687         f[i_coord_offset+DIM*3+XX] += fix3;
 688         f[i_coord_offset+DIM*3+YY] += fiy3;
 689         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 690         tx                         += fix3;
 691         ty                         += fiy3;
 692         tz                         += fiz3;
 693         fshift[i_shift_offset+XX]  += tx;
 694         fshift[i_shift_offset+YY]  += ty;
 695         fshift[i_shift_offset+ZZ]  += tz;
 696
 697         ggid                        = gid[iidx];
 698         /* Update potential energies */
 699         kernel_data->energygrp_elec[ggid] += velecsum;
 700         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 701
 702         /* Increment number of inner iterations */
 703         inneriter                  += j_index_end - j_index_start;
 704
 705         /* Outer loop uses 41 flops */
 706     }
 707
 708     /* Increment number of outer iterations */
 709     outeriter        += nri;
 710
 711     /* Update outer/inner flops */
 712
 713     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*415);
 714 }
 715 /*
 716  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW4W4_F_c
 717  * Electrostatics interaction: Ewald
 718  * VdW interaction:            CubicSplineTable
 719  * Geometry:                   Water4-Water4
 720  * Calculate force/pot:        Force
 721  */
 722 void
 723 nb_kernel_ElecEw_VdwCSTab_GeomW4W4_F_c
 724                     (t_nblist                    * gmx_restrict       nlist,
 725                      rvec                        * gmx_restrict          xx,
 726                      rvec                        * gmx_restrict          ff,
 727                      t_forcerec                  * gmx_restrict          fr,
 728                      t_mdatoms                   * gmx_restrict     mdatoms,
 729                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 730                      t_nrnb                      * gmx_restrict        nrnb)
 731 {
 732     int              i_shift_offset,i_coord_offset,j_coord_offset;
 733     int              j_index_start,j_index_end;
 734     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 735     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 736     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 737     real             *shiftvec,*fshift,*x,*f;
 738     int              vdwioffset0;
 739     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 740     int              vdwioffset1;
 741     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 742     int              vdwioffset2;
 743     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 744     int              vdwioffset3;
 745     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 746     int              vdwjidx0;
 747     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 748     int              vdwjidx1;
 749     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 750     int              vdwjidx2;
 751     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 752     int              vdwjidx3;
 753     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 754     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 755     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 756     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 757     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 758     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 759     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 760     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 761     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 762     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 763     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 764     real             velec,felec,velecsum,facel,crf,krf,krf2;
 765     real             *charge;
 766     int              nvdwtype;
 767     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 768     int              *vdwtype;
 769     real             *vdwparam;
 770     int              vfitab;
 771     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 772     real             *vftab;
 773     int              ewitab;
 774     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 775     real             *ewtab;
 776
 777     x                = xx[0];
 778     f                = ff[0];
 779
 780     nri              = nlist->nri;
 781     iinr             = nlist->iinr;
 782     jindex           = nlist->jindex;
 783     jjnr             = nlist->jjnr;
 784     shiftidx         = nlist->shift;
 785     gid              = nlist->gid;
 786     shiftvec         = fr->shift_vec[0];
 787     fshift           = fr->fshift[0];
 788     facel            = fr->epsfac;
 789     charge           = mdatoms->chargeA;
 790     nvdwtype         = fr->ntype;
 791     vdwparam         = fr->nbfp;
 792     vdwtype          = mdatoms->typeA;
 793
 794     vftab            = kernel_data->table_vdw->data;
 795     vftabscale       = kernel_data->table_vdw->scale;
 796
 797     sh_ewald         = fr->ic->sh_ewald;
 798     ewtab            = fr->ic->tabq_coul_F;
 799     ewtabscale       = fr->ic->tabq_scale;
 800     ewtabhalfspace   = 0.5/ewtabscale;
 801
 802     /* Setup water-specific parameters */
 803     inr              = nlist->iinr[0];
 804     iq1              = facel*charge[inr+1];
 805     iq2              = facel*charge[inr+2];
 806     iq3              = facel*charge[inr+3];
 807     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 808
 809     jq1              = charge[inr+1];
 810     jq2              = charge[inr+2];
 811     jq3              = charge[inr+3];
 812     vdwjidx0         = 2*vdwtype[inr+0];
 813     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 814     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 815     qq11             = iq1*jq1;
 816     qq12             = iq1*jq2;
 817     qq13             = iq1*jq3;
 818     qq21             = iq2*jq1;
 819     qq22             = iq2*jq2;
 820     qq23             = iq2*jq3;
 821     qq31             = iq3*jq1;
 822     qq32             = iq3*jq2;
 823     qq33             = iq3*jq3;
 824
 825     outeriter        = 0;
 826     inneriter        = 0;
 827
 828     /* Start outer loop over neighborlists */
 829     for(iidx=0; iidx<nri; iidx++)
 830     {
 831         /* Load shift vector for this list */
 832         i_shift_offset   = DIM*shiftidx[iidx];
 833         shX              = shiftvec[i_shift_offset+XX];
 834         shY              = shiftvec[i_shift_offset+YY];
 835         shZ              = shiftvec[i_shift_offset+ZZ];
 836
 837         /* Load limits for loop over neighbors */
 838         j_index_start    = jindex[iidx];
 839         j_index_end      = jindex[iidx+1];
 840
 841         /* Get outer coordinate index */
 842         inr              = iinr[iidx];
 843         i_coord_offset   = DIM*inr;
 844
 845         /* Load i particle coords and add shift vector */
 846         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 847         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 848         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 849         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 850         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 851         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 852         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 853         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 854         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 855         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 856         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 857         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 858
 859         fix0             = 0.0;
 860         fiy0             = 0.0;
 861         fiz0             = 0.0;
 862         fix1             = 0.0;
 863         fiy1             = 0.0;
 864         fiz1             = 0.0;
 865         fix2             = 0.0;
 866         fiy2             = 0.0;
 867         fiz2             = 0.0;
 868         fix3             = 0.0;
 869         fiy3             = 0.0;
 870         fiz3             = 0.0;
 871
 872         /* Start inner kernel loop */
 873         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 874         {
 875             /* Get j neighbor index, and coordinate index */
 876             jnr              = jjnr[jidx];
 877             j_coord_offset   = DIM*jnr;
 878
 879             /* load j atom coordinates */
 880             jx0              = x[j_coord_offset+DIM*0+XX];
 881             jy0              = x[j_coord_offset+DIM*0+YY];
 882             jz0              = x[j_coord_offset+DIM*0+ZZ];
 883             jx1              = x[j_coord_offset+DIM*1+XX];
 884             jy1              = x[j_coord_offset+DIM*1+YY];
 885             jz1              = x[j_coord_offset+DIM*1+ZZ];
 886             jx2              = x[j_coord_offset+DIM*2+XX];
 887             jy2              = x[j_coord_offset+DIM*2+YY];
 888             jz2              = x[j_coord_offset+DIM*2+ZZ];
 889             jx3              = x[j_coord_offset+DIM*3+XX];
 890             jy3              = x[j_coord_offset+DIM*3+YY];
 891             jz3              = x[j_coord_offset+DIM*3+ZZ];
 892
 893             /* Calculate displacement vector */
 894             dx00             = ix0 - jx0;
 895             dy00             = iy0 - jy0;
 896             dz00             = iz0 - jz0;
 897             dx11             = ix1 - jx1;
 898             dy11             = iy1 - jy1;
 899             dz11             = iz1 - jz1;
 900             dx12             = ix1 - jx2;
 901             dy12             = iy1 - jy2;
 902             dz12             = iz1 - jz2;
 903             dx13             = ix1 - jx3;
 904             dy13             = iy1 - jy3;
 905             dz13             = iz1 - jz3;
 906             dx21             = ix2 - jx1;
 907             dy21             = iy2 - jy1;
 908             dz21             = iz2 - jz1;
 909             dx22             = ix2 - jx2;
 910             dy22             = iy2 - jy2;
 911             dz22             = iz2 - jz2;
 912             dx23             = ix2 - jx3;
 913             dy23             = iy2 - jy3;
 914             dz23             = iz2 - jz3;
 915             dx31             = ix3 - jx1;
 916             dy31             = iy3 - jy1;
 917             dz31             = iz3 - jz1;
 918             dx32             = ix3 - jx2;
 919             dy32             = iy3 - jy2;
 920             dz32             = iz3 - jz2;
 921             dx33             = ix3 - jx3;
 922             dy33             = iy3 - jy3;
 923             dz33             = iz3 - jz3;
 924
 925             /* Calculate squared distance and things based on it */
 926             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 927             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 928             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 929             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 930             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 931             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 932             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 933             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 934             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 935             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 936
 937             rinv00           = gmx_invsqrt(rsq00);
 938             rinv11           = gmx_invsqrt(rsq11);
 939             rinv12           = gmx_invsqrt(rsq12);
 940             rinv13           = gmx_invsqrt(rsq13);
 941             rinv21           = gmx_invsqrt(rsq21);
 942             rinv22           = gmx_invsqrt(rsq22);
 943             rinv23           = gmx_invsqrt(rsq23);
 944             rinv31           = gmx_invsqrt(rsq31);
 945             rinv32           = gmx_invsqrt(rsq32);
 946             rinv33           = gmx_invsqrt(rsq33);
 947
 948             rinvsq11         = rinv11*rinv11;
 949             rinvsq12         = rinv12*rinv12;
 950             rinvsq13         = rinv13*rinv13;
 951             rinvsq21         = rinv21*rinv21;
 952             rinvsq22         = rinv22*rinv22;
 953             rinvsq23         = rinv23*rinv23;
 954             rinvsq31         = rinv31*rinv31;
 955             rinvsq32         = rinv32*rinv32;
 956             rinvsq33         = rinv33*rinv33;
 957
 958             /**************************
 959              * CALCULATE INTERACTIONS *
 960              **************************/
 961
 962             r00              = rsq00*rinv00;
 963
 964             /* Calculate table index by multiplying r with table scale and truncate to integer */
 965             rt               = r00*vftabscale;
 966             vfitab           = rt;
 967             vfeps            = rt-vfitab;
 968             vfitab           = 2*4*vfitab;
 969
 970             /* CUBIC SPLINE TABLE DISPERSION */
 971             vfitab          += 0;
 972             F                = vftab[vfitab+1];
 973             Geps             = vfeps*vftab[vfitab+2];
 974             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 975             Fp               = F+Geps+Heps2;
 976             FF               = Fp+Geps+2.0*Heps2;
 977             fvdw6            = c6_00*FF;
 978
 979             /* CUBIC SPLINE TABLE REPULSION */
 980             F                = vftab[vfitab+5];
 981             Geps             = vfeps*vftab[vfitab+6];
 982             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 983             Fp               = F+Geps+Heps2;
 984             FF               = Fp+Geps+2.0*Heps2;
 985             fvdw12           = c12_00*FF;
 986             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 987
 988             fscal            = fvdw;
 989
 990             /* Calculate temporary vectorial force */
 991             tx               = fscal*dx00;
 992             ty               = fscal*dy00;
 993             tz               = fscal*dz00;
 994
 995             /* Update vectorial force */
 996             fix0            += tx;
 997             fiy0            += ty;
 998             fiz0            += tz;
 999             f[j_coord_offset+DIM*0+XX] -= tx;
1000             f[j_coord_offset+DIM*0+YY] -= ty;
1001             f[j_coord_offset+DIM*0+ZZ] -= tz;
1002
1003             /**************************
1004              * CALCULATE INTERACTIONS *
1005              **************************/
1006
1007             r11              = rsq11*rinv11;
1008
1009             /* EWALD ELECTROSTATICS */
1010
1011             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1012             ewrt             = r11*ewtabscale;
1013             ewitab           = ewrt;
1014             eweps            = ewrt-ewitab;
1015             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1016             felec            = qq11*rinv11*(rinvsq11-felec);
1017
1018             fscal            = felec;
1019
1020             /* Calculate temporary vectorial force */
1021             tx               = fscal*dx11;
1022             ty               = fscal*dy11;
1023             tz               = fscal*dz11;
1024
1025             /* Update vectorial force */
1026             fix1            += tx;
1027             fiy1            += ty;
1028             fiz1            += tz;
1029             f[j_coord_offset+DIM*1+XX] -= tx;
1030             f[j_coord_offset+DIM*1+YY] -= ty;
1031             f[j_coord_offset+DIM*1+ZZ] -= tz;
1032
1033             /**************************
1034              * CALCULATE INTERACTIONS *
1035              **************************/
1036
1037             r12              = rsq12*rinv12;
1038
1039             /* EWALD ELECTROSTATICS */
1040
1041             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1042             ewrt             = r12*ewtabscale;
1043             ewitab           = ewrt;
1044             eweps            = ewrt-ewitab;
1045             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1046             felec            = qq12*rinv12*(rinvsq12-felec);
1047
1048             fscal            = felec;
1049
1050             /* Calculate temporary vectorial force */
1051             tx               = fscal*dx12;
1052             ty               = fscal*dy12;
1053             tz               = fscal*dz12;
1054
1055             /* Update vectorial force */
1056             fix1            += tx;
1057             fiy1            += ty;
1058             fiz1            += tz;
1059             f[j_coord_offset+DIM*2+XX] -= tx;
1060             f[j_coord_offset+DIM*2+YY] -= ty;
1061             f[j_coord_offset+DIM*2+ZZ] -= tz;
1062
1063             /**************************
1064              * CALCULATE INTERACTIONS *
1065              **************************/
1066
1067             r13              = rsq13*rinv13;
1068
1069             /* EWALD ELECTROSTATICS */
1070
1071             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1072             ewrt             = r13*ewtabscale;
1073             ewitab           = ewrt;
1074             eweps            = ewrt-ewitab;
1075             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1076             felec            = qq13*rinv13*(rinvsq13-felec);
1077
1078             fscal            = felec;
1079
1080             /* Calculate temporary vectorial force */
1081             tx               = fscal*dx13;
1082             ty               = fscal*dy13;
1083             tz               = fscal*dz13;
1084
1085             /* Update vectorial force */
1086             fix1            += tx;
1087             fiy1            += ty;
1088             fiz1            += tz;
1089             f[j_coord_offset+DIM*3+XX] -= tx;
1090             f[j_coord_offset+DIM*3+YY] -= ty;
1091             f[j_coord_offset+DIM*3+ZZ] -= tz;
1092
1093             /**************************
1094              * CALCULATE INTERACTIONS *
1095              **************************/
1096
1097             r21              = rsq21*rinv21;
1098
1099             /* EWALD ELECTROSTATICS */
1100
1101             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1102             ewrt             = r21*ewtabscale;
1103             ewitab           = ewrt;
1104             eweps            = ewrt-ewitab;
1105             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1106             felec            = qq21*rinv21*(rinvsq21-felec);
1107
1108             fscal            = felec;
1109
1110             /* Calculate temporary vectorial force */
1111             tx               = fscal*dx21;
1112             ty               = fscal*dy21;
1113             tz               = fscal*dz21;
1114
1115             /* Update vectorial force */
1116             fix2            += tx;
1117             fiy2            += ty;
1118             fiz2            += tz;
1119             f[j_coord_offset+DIM*1+XX] -= tx;
1120             f[j_coord_offset+DIM*1+YY] -= ty;
1121             f[j_coord_offset+DIM*1+ZZ] -= tz;
1122
1123             /**************************
1124              * CALCULATE INTERACTIONS *
1125              **************************/
1126
1127             r22              = rsq22*rinv22;
1128
1129             /* EWALD ELECTROSTATICS */
1130
1131             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1132             ewrt             = r22*ewtabscale;
1133             ewitab           = ewrt;
1134             eweps            = ewrt-ewitab;
1135             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1136             felec            = qq22*rinv22*(rinvsq22-felec);
1137
1138             fscal            = felec;
1139
1140             /* Calculate temporary vectorial force */
1141             tx               = fscal*dx22;
1142             ty               = fscal*dy22;
1143             tz               = fscal*dz22;
1144
1145             /* Update vectorial force */
1146             fix2            += tx;
1147             fiy2            += ty;
1148             fiz2            += tz;
1149             f[j_coord_offset+DIM*2+XX] -= tx;
1150             f[j_coord_offset+DIM*2+YY] -= ty;
1151             f[j_coord_offset+DIM*2+ZZ] -= tz;
1152
1153             /**************************
1154              * CALCULATE INTERACTIONS *
1155              **************************/
1156
1157             r23              = rsq23*rinv23;
1158
1159             /* EWALD ELECTROSTATICS */
1160
1161             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1162             ewrt             = r23*ewtabscale;
1163             ewitab           = ewrt;
1164             eweps            = ewrt-ewitab;
1165             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1166             felec            = qq23*rinv23*(rinvsq23-felec);
1167
1168             fscal            = felec;
1169
1170             /* Calculate temporary vectorial force */
1171             tx               = fscal*dx23;
1172             ty               = fscal*dy23;
1173             tz               = fscal*dz23;
1174
1175             /* Update vectorial force */
1176             fix2            += tx;
1177             fiy2            += ty;
1178             fiz2            += tz;
1179             f[j_coord_offset+DIM*3+XX] -= tx;
1180             f[j_coord_offset+DIM*3+YY] -= ty;
1181             f[j_coord_offset+DIM*3+ZZ] -= tz;
1182
1183             /**************************
1184              * CALCULATE INTERACTIONS *
1185              **************************/
1186
1187             r31              = rsq31*rinv31;
1188
1189             /* EWALD ELECTROSTATICS */
1190
1191             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1192             ewrt             = r31*ewtabscale;
1193             ewitab           = ewrt;
1194             eweps            = ewrt-ewitab;
1195             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1196             felec            = qq31*rinv31*(rinvsq31-felec);
1197
1198             fscal            = felec;
1199
1200             /* Calculate temporary vectorial force */
1201             tx               = fscal*dx31;
1202             ty               = fscal*dy31;
1203             tz               = fscal*dz31;
1204
1205             /* Update vectorial force */
1206             fix3            += tx;
1207             fiy3            += ty;
1208             fiz3            += tz;
1209             f[j_coord_offset+DIM*1+XX] -= tx;
1210             f[j_coord_offset+DIM*1+YY] -= ty;
1211             f[j_coord_offset+DIM*1+ZZ] -= tz;
1212
1213             /**************************
1214              * CALCULATE INTERACTIONS *
1215              **************************/
1216
1217             r32              = rsq32*rinv32;
1218
1219             /* EWALD ELECTROSTATICS */
1220
1221             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1222             ewrt             = r32*ewtabscale;
1223             ewitab           = ewrt;
1224             eweps            = ewrt-ewitab;
1225             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1226             felec            = qq32*rinv32*(rinvsq32-felec);
1227
1228             fscal            = felec;
1229
1230             /* Calculate temporary vectorial force */
1231             tx               = fscal*dx32;
1232             ty               = fscal*dy32;
1233             tz               = fscal*dz32;
1234
1235             /* Update vectorial force */
1236             fix3            += tx;
1237             fiy3            += ty;
1238             fiz3            += tz;
1239             f[j_coord_offset+DIM*2+XX] -= tx;
1240             f[j_coord_offset+DIM*2+YY] -= ty;
1241             f[j_coord_offset+DIM*2+ZZ] -= tz;
1242
1243             /**************************
1244              * CALCULATE INTERACTIONS *
1245              **************************/
1246
1247             r33              = rsq33*rinv33;
1248
1249             /* EWALD ELECTROSTATICS */
1250
1251             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1252             ewrt             = r33*ewtabscale;
1253             ewitab           = ewrt;
1254             eweps            = ewrt-ewitab;
1255             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1256             felec            = qq33*rinv33*(rinvsq33-felec);
1257
1258             fscal            = felec;
1259
1260             /* Calculate temporary vectorial force */
1261             tx               = fscal*dx33;
1262             ty               = fscal*dy33;
1263             tz               = fscal*dz33;
1264
1265             /* Update vectorial force */
1266             fix3            += tx;
1267             fiy3            += ty;
1268             fiz3            += tz;
1269             f[j_coord_offset+DIM*3+XX] -= tx;
1270             f[j_coord_offset+DIM*3+YY] -= ty;
1271             f[j_coord_offset+DIM*3+ZZ] -= tz;
1272
1273             /* Inner loop uses 344 flops */
1274         }
1275         /* End of innermost loop */
1276
1277         tx = ty = tz = 0;
1278         f[i_coord_offset+DIM*0+XX] += fix0;
1279         f[i_coord_offset+DIM*0+YY] += fiy0;
1280         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1281         tx                         += fix0;
1282         ty                         += fiy0;
1283         tz                         += fiz0;
1284         f[i_coord_offset+DIM*1+XX] += fix1;
1285         f[i_coord_offset+DIM*1+YY] += fiy1;
1286         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1287         tx                         += fix1;
1288         ty                         += fiy1;
1289         tz                         += fiz1;
1290         f[i_coord_offset+DIM*2+XX] += fix2;
1291         f[i_coord_offset+DIM*2+YY] += fiy2;
1292         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1293         tx                         += fix2;
1294         ty                         += fiy2;
1295         tz                         += fiz2;
1296         f[i_coord_offset+DIM*3+XX] += fix3;
1297         f[i_coord_offset+DIM*3+YY] += fiy3;
1298         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1299         tx                         += fix3;
1300         ty                         += fiy3;
1301         tz                         += fiz3;
1302         fshift[i_shift_offset+XX]  += tx;
1303         fshift[i_shift_offset+YY]  += ty;
1304         fshift[i_shift_offset+ZZ]  += tz;
1305
1306         /* Increment number of inner iterations */
1307         inneriter                  += j_index_end - j_index_start;
1308
1309         /* Outer loop uses 39 flops */
1310     }
1311
1312     /* Increment number of outer iterations */
1313     outeriter        += nri;
1314
1315     /* Update outer/inner flops */
1316
1317     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*344);
1318 }