src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSh_VdwLJSh_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_ElecEwSh_VdwLJSh_GeomW4W4_VF_c
  49  * Electrostatics interaction: Ewald
  50  * VdW interaction:            LennardJones
  51  * Geometry:                   Water4-Water4
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecEwSh_VdwLJSh_GeomW4W4_VF_c
  56                     (t_nblist                    * gmx_restrict       nlist,
  57                      rvec                        * gmx_restrict          xx,
  58                      rvec                        * gmx_restrict          ff,
  59                      t_forcerec                  * gmx_restrict          fr,
  60                      t_mdatoms                   * gmx_restrict     mdatoms,
  61                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  62                      t_nrnb                      * gmx_restrict        nrnb)
  63 {
  64     int              i_shift_offset,i_coord_offset,j_coord_offset;
  65     int              j_index_start,j_index_end;
  66     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  67     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  68     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  69     real             *shiftvec,*fshift,*x,*f;
  70     int              vdwioffset0;
  71     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwioffset1;
  73     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  74     int              vdwioffset2;
  75     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  76     int              vdwioffset3;
  77     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  78     int              vdwjidx0;
  79     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  80     int              vdwjidx1;
  81     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  82     int              vdwjidx2;
  83     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  84     int              vdwjidx3;
  85     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  86     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  87     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  88     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  89     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
  90     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  91     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  92     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
  93     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
  94     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
  95     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
  96     real             velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     int              nvdwtype;
  99     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102     int              ewitab;
 103     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 104     real             *ewtab;
 105
 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     sh_ewald         = fr->ic->sh_ewald;
 124     ewtab            = fr->ic->tabq_coul_FDV0;
 125     ewtabscale       = fr->ic->tabq_scale;
 126     ewtabhalfspace   = 0.5/ewtabscale;
 127
 128     /* Setup water-specific parameters */
 129     inr              = nlist->iinr[0];
 130     iq1              = facel*charge[inr+1];
 131     iq2              = facel*charge[inr+2];
 132     iq3              = facel*charge[inr+3];
 133     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 134
 135     jq1              = charge[inr+1];
 136     jq2              = charge[inr+2];
 137     jq3              = charge[inr+3];
 138     vdwjidx0         = 2*vdwtype[inr+0];
 139     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 140     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 141     qq11             = iq1*jq1;
 142     qq12             = iq1*jq2;
 143     qq13             = iq1*jq3;
 144     qq21             = iq2*jq1;
 145     qq22             = iq2*jq2;
 146     qq23             = iq2*jq3;
 147     qq31             = iq3*jq1;
 148     qq32             = iq3*jq2;
 149     qq33             = iq3*jq3;
 150
 151     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 152     rcutoff          = fr->rcoulomb;
 153     rcutoff2         = rcutoff*rcutoff;
 154
 155     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 156     rvdw             = fr->rvdw;
 157
 158     outeriter        = 0;
 159     inneriter        = 0;
 160
 161     /* Start outer loop over neighborlists */
 162     for(iidx=0; iidx<nri; iidx++)
 163     {
 164         /* Load shift vector for this list */
 165         i_shift_offset   = DIM*shiftidx[iidx];
 166         shX              = shiftvec[i_shift_offset+XX];
 167         shY              = shiftvec[i_shift_offset+YY];
 168         shZ              = shiftvec[i_shift_offset+ZZ];
 169
 170         /* Load limits for loop over neighbors */
 171         j_index_start    = jindex[iidx];
 172         j_index_end      = jindex[iidx+1];
 173
 174         /* Get outer coordinate index */
 175         inr              = iinr[iidx];
 176         i_coord_offset   = DIM*inr;
 177
 178         /* Load i particle coords and add shift vector */
 179         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 180         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 181         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 182         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 183         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 184         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 185         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 186         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 187         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 188         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 189         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 190         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 191
 192         fix0             = 0.0;
 193         fiy0             = 0.0;
 194         fiz0             = 0.0;
 195         fix1             = 0.0;
 196         fiy1             = 0.0;
 197         fiz1             = 0.0;
 198         fix2             = 0.0;
 199         fiy2             = 0.0;
 200         fiz2             = 0.0;
 201         fix3             = 0.0;
 202         fiy3             = 0.0;
 203         fiz3             = 0.0;
 204
 205         /* Reset potential sums */
 206         velecsum         = 0.0;
 207         vvdwsum          = 0.0;
 208
 209         /* Start inner kernel loop */
 210         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 211         {
 212             /* Get j neighbor index, and coordinate index */
 213             jnr              = jjnr[jidx];
 214             j_coord_offset   = DIM*jnr;
 215
 216             /* load j atom coordinates */
 217             jx0              = x[j_coord_offset+DIM*0+XX];
 218             jy0              = x[j_coord_offset+DIM*0+YY];
 219             jz0              = x[j_coord_offset+DIM*0+ZZ];
 220             jx1              = x[j_coord_offset+DIM*1+XX];
 221             jy1              = x[j_coord_offset+DIM*1+YY];
 222             jz1              = x[j_coord_offset+DIM*1+ZZ];
 223             jx2              = x[j_coord_offset+DIM*2+XX];
 224             jy2              = x[j_coord_offset+DIM*2+YY];
 225             jz2              = x[j_coord_offset+DIM*2+ZZ];
 226             jx3              = x[j_coord_offset+DIM*3+XX];
 227             jy3              = x[j_coord_offset+DIM*3+YY];
 228             jz3              = x[j_coord_offset+DIM*3+ZZ];
 229
 230             /* Calculate displacement vector */
 231             dx00             = ix0 - jx0;
 232             dy00             = iy0 - jy0;
 233             dz00             = iz0 - jz0;
 234             dx11             = ix1 - jx1;
 235             dy11             = iy1 - jy1;
 236             dz11             = iz1 - jz1;
 237             dx12             = ix1 - jx2;
 238             dy12             = iy1 - jy2;
 239             dz12             = iz1 - jz2;
 240             dx13             = ix1 - jx3;
 241             dy13             = iy1 - jy3;
 242             dz13             = iz1 - jz3;
 243             dx21             = ix2 - jx1;
 244             dy21             = iy2 - jy1;
 245             dz21             = iz2 - jz1;
 246             dx22             = ix2 - jx2;
 247             dy22             = iy2 - jy2;
 248             dz22             = iz2 - jz2;
 249             dx23             = ix2 - jx3;
 250             dy23             = iy2 - jy3;
 251             dz23             = iz2 - jz3;
 252             dx31             = ix3 - jx1;
 253             dy31             = iy3 - jy1;
 254             dz31             = iz3 - jz1;
 255             dx32             = ix3 - jx2;
 256             dy32             = iy3 - jy2;
 257             dz32             = iz3 - jz2;
 258             dx33             = ix3 - jx3;
 259             dy33             = iy3 - jy3;
 260             dz33             = iz3 - jz3;
 261
 262             /* Calculate squared distance and things based on it */
 263             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 264             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 265             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 266             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 267             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 268             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 269             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 270             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 271             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 272             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 273
 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             rinvsq00         = 1.0/rsq00;
 285             rinvsq11         = rinv11*rinv11;
 286             rinvsq12         = rinv12*rinv12;
 287             rinvsq13         = rinv13*rinv13;
 288             rinvsq21         = rinv21*rinv21;
 289             rinvsq22         = rinv22*rinv22;
 290             rinvsq23         = rinv23*rinv23;
 291             rinvsq31         = rinv31*rinv31;
 292             rinvsq32         = rinv32*rinv32;
 293             rinvsq33         = rinv33*rinv33;
 294
 295             /**************************
 296              * CALCULATE INTERACTIONS *
 297              **************************/
 298
 299             if (rsq00<rcutoff2)
 300             {
 301
 302             /* LENNARD-JONES DISPERSION/REPULSION */
 303
 304             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 305             vvdw6            = c6_00*rinvsix;
 306             vvdw12           = c12_00*rinvsix*rinvsix;
 307             vvdw             = (vvdw12 - c12_00*sh_vdw_invrcut6*sh_vdw_invrcut6)*(1.0/12.0) - (vvdw6 - c6_00*sh_vdw_invrcut6)*(1.0/6.0);
 308             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 309
 310             /* Update potential sums from outer loop */
 311             vvdwsum         += vvdw;
 312
 313             fscal            = fvdw;
 314
 315             /* Calculate temporary vectorial force */
 316             tx               = fscal*dx00;
 317             ty               = fscal*dy00;
 318             tz               = fscal*dz00;
 319
 320             /* Update vectorial force */
 321             fix0            += tx;
 322             fiy0            += ty;
 323             fiz0            += tz;
 324             f[j_coord_offset+DIM*0+XX] -= tx;
 325             f[j_coord_offset+DIM*0+YY] -= ty;
 326             f[j_coord_offset+DIM*0+ZZ] -= tz;
 327
 328             }
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             if (rsq11<rcutoff2)
 335             {
 336
 337             r11              = rsq11*rinv11;
 338
 339             /* EWALD ELECTROSTATICS */
 340
 341             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 342             ewrt             = r11*ewtabscale;
 343             ewitab           = ewrt;
 344             eweps            = ewrt-ewitab;
 345             ewitab           = 4*ewitab;
 346             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 347             velec            = qq11*((rinv11-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 348             felec            = qq11*rinv11*(rinvsq11-felec);
 349
 350             /* Update potential sums from outer loop */
 351             velecsum        += velec;
 352
 353             fscal            = felec;
 354
 355             /* Calculate temporary vectorial force */
 356             tx               = fscal*dx11;
 357             ty               = fscal*dy11;
 358             tz               = fscal*dz11;
 359
 360             /* Update vectorial force */
 361             fix1            += tx;
 362             fiy1            += ty;
 363             fiz1            += tz;
 364             f[j_coord_offset+DIM*1+XX] -= tx;
 365             f[j_coord_offset+DIM*1+YY] -= ty;
 366             f[j_coord_offset+DIM*1+ZZ] -= tz;
 367
 368             }
 369
 370             /**************************
 371              * CALCULATE INTERACTIONS *
 372              **************************/
 373
 374             if (rsq12<rcutoff2)
 375             {
 376
 377             r12              = rsq12*rinv12;
 378
 379             /* EWALD ELECTROSTATICS */
 380
 381             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 382             ewrt             = r12*ewtabscale;
 383             ewitab           = ewrt;
 384             eweps            = ewrt-ewitab;
 385             ewitab           = 4*ewitab;
 386             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 387             velec            = qq12*((rinv12-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 388             felec            = qq12*rinv12*(rinvsq12-felec);
 389
 390             /* Update potential sums from outer loop */
 391             velecsum        += velec;
 392
 393             fscal            = felec;
 394
 395             /* Calculate temporary vectorial force */
 396             tx               = fscal*dx12;
 397             ty               = fscal*dy12;
 398             tz               = fscal*dz12;
 399
 400             /* Update vectorial force */
 401             fix1            += tx;
 402             fiy1            += ty;
 403             fiz1            += tz;
 404             f[j_coord_offset+DIM*2+XX] -= tx;
 405             f[j_coord_offset+DIM*2+YY] -= ty;
 406             f[j_coord_offset+DIM*2+ZZ] -= tz;
 407
 408             }
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             if (rsq13<rcutoff2)
 415             {
 416
 417             r13              = rsq13*rinv13;
 418
 419             /* EWALD ELECTROSTATICS */
 420
 421             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 422             ewrt             = r13*ewtabscale;
 423             ewitab           = ewrt;
 424             eweps            = ewrt-ewitab;
 425             ewitab           = 4*ewitab;
 426             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 427             velec            = qq13*((rinv13-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 428             felec            = qq13*rinv13*(rinvsq13-felec);
 429
 430             /* Update potential sums from outer loop */
 431             velecsum        += velec;
 432
 433             fscal            = felec;
 434
 435             /* Calculate temporary vectorial force */
 436             tx               = fscal*dx13;
 437             ty               = fscal*dy13;
 438             tz               = fscal*dz13;
 439
 440             /* Update vectorial force */
 441             fix1            += tx;
 442             fiy1            += ty;
 443             fiz1            += tz;
 444             f[j_coord_offset+DIM*3+XX] -= tx;
 445             f[j_coord_offset+DIM*3+YY] -= ty;
 446             f[j_coord_offset+DIM*3+ZZ] -= tz;
 447
 448             }
 449
 450             /**************************
 451              * CALCULATE INTERACTIONS *
 452              **************************/
 453
 454             if (rsq21<rcutoff2)
 455             {
 456
 457             r21              = rsq21*rinv21;
 458
 459             /* EWALD ELECTROSTATICS */
 460
 461             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 462             ewrt             = r21*ewtabscale;
 463             ewitab           = ewrt;
 464             eweps            = ewrt-ewitab;
 465             ewitab           = 4*ewitab;
 466             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 467             velec            = qq21*((rinv21-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 468             felec            = qq21*rinv21*(rinvsq21-felec);
 469
 470             /* Update potential sums from outer loop */
 471             velecsum        += velec;
 472
 473             fscal            = felec;
 474
 475             /* Calculate temporary vectorial force */
 476             tx               = fscal*dx21;
 477             ty               = fscal*dy21;
 478             tz               = fscal*dz21;
 479
 480             /* Update vectorial force */
 481             fix2            += tx;
 482             fiy2            += ty;
 483             fiz2            += tz;
 484             f[j_coord_offset+DIM*1+XX] -= tx;
 485             f[j_coord_offset+DIM*1+YY] -= ty;
 486             f[j_coord_offset+DIM*1+ZZ] -= tz;
 487
 488             }
 489
 490             /**************************
 491              * CALCULATE INTERACTIONS *
 492              **************************/
 493
 494             if (rsq22<rcutoff2)
 495             {
 496
 497             r22              = rsq22*rinv22;
 498
 499             /* EWALD ELECTROSTATICS */
 500
 501             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 502             ewrt             = r22*ewtabscale;
 503             ewitab           = ewrt;
 504             eweps            = ewrt-ewitab;
 505             ewitab           = 4*ewitab;
 506             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 507             velec            = qq22*((rinv22-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 508             felec            = qq22*rinv22*(rinvsq22-felec);
 509
 510             /* Update potential sums from outer loop */
 511             velecsum        += velec;
 512
 513             fscal            = felec;
 514
 515             /* Calculate temporary vectorial force */
 516             tx               = fscal*dx22;
 517             ty               = fscal*dy22;
 518             tz               = fscal*dz22;
 519
 520             /* Update vectorial force */
 521             fix2            += tx;
 522             fiy2            += ty;
 523             fiz2            += tz;
 524             f[j_coord_offset+DIM*2+XX] -= tx;
 525             f[j_coord_offset+DIM*2+YY] -= ty;
 526             f[j_coord_offset+DIM*2+ZZ] -= tz;
 527
 528             }
 529
 530             /**************************
 531              * CALCULATE INTERACTIONS *
 532              **************************/
 533
 534             if (rsq23<rcutoff2)
 535             {
 536
 537             r23              = rsq23*rinv23;
 538
 539             /* EWALD ELECTROSTATICS */
 540
 541             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 542             ewrt             = r23*ewtabscale;
 543             ewitab           = ewrt;
 544             eweps            = ewrt-ewitab;
 545             ewitab           = 4*ewitab;
 546             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 547             velec            = qq23*((rinv23-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 548             felec            = qq23*rinv23*(rinvsq23-felec);
 549
 550             /* Update potential sums from outer loop */
 551             velecsum        += velec;
 552
 553             fscal            = felec;
 554
 555             /* Calculate temporary vectorial force */
 556             tx               = fscal*dx23;
 557             ty               = fscal*dy23;
 558             tz               = fscal*dz23;
 559
 560             /* Update vectorial force */
 561             fix2            += tx;
 562             fiy2            += ty;
 563             fiz2            += tz;
 564             f[j_coord_offset+DIM*3+XX] -= tx;
 565             f[j_coord_offset+DIM*3+YY] -= ty;
 566             f[j_coord_offset+DIM*3+ZZ] -= tz;
 567
 568             }
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             if (rsq31<rcutoff2)
 575             {
 576
 577             r31              = rsq31*rinv31;
 578
 579             /* EWALD ELECTROSTATICS */
 580
 581             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 582             ewrt             = r31*ewtabscale;
 583             ewitab           = ewrt;
 584             eweps            = ewrt-ewitab;
 585             ewitab           = 4*ewitab;
 586             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 587             velec            = qq31*((rinv31-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 588             felec            = qq31*rinv31*(rinvsq31-felec);
 589
 590             /* Update potential sums from outer loop */
 591             velecsum        += velec;
 592
 593             fscal            = felec;
 594
 595             /* Calculate temporary vectorial force */
 596             tx               = fscal*dx31;
 597             ty               = fscal*dy31;
 598             tz               = fscal*dz31;
 599
 600             /* Update vectorial force */
 601             fix3            += tx;
 602             fiy3            += ty;
 603             fiz3            += tz;
 604             f[j_coord_offset+DIM*1+XX] -= tx;
 605             f[j_coord_offset+DIM*1+YY] -= ty;
 606             f[j_coord_offset+DIM*1+ZZ] -= tz;
 607
 608             }
 609
 610             /**************************
 611              * CALCULATE INTERACTIONS *
 612              **************************/
 613
 614             if (rsq32<rcutoff2)
 615             {
 616
 617             r32              = rsq32*rinv32;
 618
 619             /* EWALD ELECTROSTATICS */
 620
 621             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 622             ewrt             = r32*ewtabscale;
 623             ewitab           = ewrt;
 624             eweps            = ewrt-ewitab;
 625             ewitab           = 4*ewitab;
 626             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 627             velec            = qq32*((rinv32-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 628             felec            = qq32*rinv32*(rinvsq32-felec);
 629
 630             /* Update potential sums from outer loop */
 631             velecsum        += velec;
 632
 633             fscal            = felec;
 634
 635             /* Calculate temporary vectorial force */
 636             tx               = fscal*dx32;
 637             ty               = fscal*dy32;
 638             tz               = fscal*dz32;
 639
 640             /* Update vectorial force */
 641             fix3            += tx;
 642             fiy3            += ty;
 643             fiz3            += tz;
 644             f[j_coord_offset+DIM*2+XX] -= tx;
 645             f[j_coord_offset+DIM*2+YY] -= ty;
 646             f[j_coord_offset+DIM*2+ZZ] -= tz;
 647
 648             }
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             if (rsq33<rcutoff2)
 655             {
 656
 657             r33              = rsq33*rinv33;
 658
 659             /* EWALD ELECTROSTATICS */
 660
 661             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 662             ewrt             = r33*ewtabscale;
 663             ewitab           = ewrt;
 664             eweps            = ewrt-ewitab;
 665             ewitab           = 4*ewitab;
 666             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 667             velec            = qq33*((rinv33-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 668             felec            = qq33*rinv33*(rinvsq33-felec);
 669
 670             /* Update potential sums from outer loop */
 671             velecsum        += velec;
 672
 673             fscal            = felec;
 674
 675             /* Calculate temporary vectorial force */
 676             tx               = fscal*dx33;
 677             ty               = fscal*dy33;
 678             tz               = fscal*dz33;
 679
 680             /* Update vectorial force */
 681             fix3            += tx;
 682             fiy3            += ty;
 683             fiz3            += tz;
 684             f[j_coord_offset+DIM*3+XX] -= tx;
 685             f[j_coord_offset+DIM*3+YY] -= ty;
 686             f[j_coord_offset+DIM*3+ZZ] -= tz;
 687
 688             }
 689
 690             /* Inner loop uses 406 flops */
 691         }
 692         /* End of innermost loop */
 693
 694         tx = ty = tz = 0;
 695         f[i_coord_offset+DIM*0+XX] += fix0;
 696         f[i_coord_offset+DIM*0+YY] += fiy0;
 697         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 698         tx                         += fix0;
 699         ty                         += fiy0;
 700         tz                         += fiz0;
 701         f[i_coord_offset+DIM*1+XX] += fix1;
 702         f[i_coord_offset+DIM*1+YY] += fiy1;
 703         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 704         tx                         += fix1;
 705         ty                         += fiy1;
 706         tz                         += fiz1;
 707         f[i_coord_offset+DIM*2+XX] += fix2;
 708         f[i_coord_offset+DIM*2+YY] += fiy2;
 709         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 710         tx                         += fix2;
 711         ty                         += fiy2;
 712         tz                         += fiz2;
 713         f[i_coord_offset+DIM*3+XX] += fix3;
 714         f[i_coord_offset+DIM*3+YY] += fiy3;
 715         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 716         tx                         += fix3;
 717         ty                         += fiy3;
 718         tz                         += fiz3;
 719         fshift[i_shift_offset+XX]  += tx;
 720         fshift[i_shift_offset+YY]  += ty;
 721         fshift[i_shift_offset+ZZ]  += tz;
 722
 723         ggid                        = gid[iidx];
 724         /* Update potential energies */
 725         kernel_data->energygrp_elec[ggid] += velecsum;
 726         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 727
 728         /* Increment number of inner iterations */
 729         inneriter                  += j_index_end - j_index_start;
 730
 731         /* Outer loop uses 41 flops */
 732     }
 733
 734     /* Increment number of outer iterations */
 735     outeriter        += nri;
 736
 737     /* Update outer/inner flops */
 738
 739     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*406);
 740 }
 741 /*
 742  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomW4W4_F_c
 743  * Electrostatics interaction: Ewald
 744  * VdW interaction:            LennardJones
 745  * Geometry:                   Water4-Water4
 746  * Calculate force/pot:        Force
 747  */
 748 void
 749 nb_kernel_ElecEwSh_VdwLJSh_GeomW4W4_F_c
 750                     (t_nblist                    * gmx_restrict       nlist,
 751                      rvec                        * gmx_restrict          xx,
 752                      rvec                        * gmx_restrict          ff,
 753                      t_forcerec                  * gmx_restrict          fr,
 754                      t_mdatoms                   * gmx_restrict     mdatoms,
 755                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 756                      t_nrnb                      * gmx_restrict        nrnb)
 757 {
 758     int              i_shift_offset,i_coord_offset,j_coord_offset;
 759     int              j_index_start,j_index_end;
 760     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 761     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 762     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 763     real             *shiftvec,*fshift,*x,*f;
 764     int              vdwioffset0;
 765     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 766     int              vdwioffset1;
 767     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 768     int              vdwioffset2;
 769     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 770     int              vdwioffset3;
 771     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 772     int              vdwjidx0;
 773     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 774     int              vdwjidx1;
 775     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 776     int              vdwjidx2;
 777     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 778     int              vdwjidx3;
 779     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 780     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 781     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 782     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 783     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 784     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 785     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 786     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 787     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 788     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 789     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 790     real             velec,felec,velecsum,facel,crf,krf,krf2;
 791     real             *charge;
 792     int              nvdwtype;
 793     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 794     int              *vdwtype;
 795     real             *vdwparam;
 796     int              ewitab;
 797     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 798     real             *ewtab;
 799
 800     x                = xx[0];
 801     f                = ff[0];
 802
 803     nri              = nlist->nri;
 804     iinr             = nlist->iinr;
 805     jindex           = nlist->jindex;
 806     jjnr             = nlist->jjnr;
 807     shiftidx         = nlist->shift;
 808     gid              = nlist->gid;
 809     shiftvec         = fr->shift_vec[0];
 810     fshift           = fr->fshift[0];
 811     facel            = fr->epsfac;
 812     charge           = mdatoms->chargeA;
 813     nvdwtype         = fr->ntype;
 814     vdwparam         = fr->nbfp;
 815     vdwtype          = mdatoms->typeA;
 816
 817     sh_ewald         = fr->ic->sh_ewald;
 818     ewtab            = fr->ic->tabq_coul_F;
 819     ewtabscale       = fr->ic->tabq_scale;
 820     ewtabhalfspace   = 0.5/ewtabscale;
 821
 822     /* Setup water-specific parameters */
 823     inr              = nlist->iinr[0];
 824     iq1              = facel*charge[inr+1];
 825     iq2              = facel*charge[inr+2];
 826     iq3              = facel*charge[inr+3];
 827     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 828
 829     jq1              = charge[inr+1];
 830     jq2              = charge[inr+2];
 831     jq3              = charge[inr+3];
 832     vdwjidx0         = 2*vdwtype[inr+0];
 833     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 834     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 835     qq11             = iq1*jq1;
 836     qq12             = iq1*jq2;
 837     qq13             = iq1*jq3;
 838     qq21             = iq2*jq1;
 839     qq22             = iq2*jq2;
 840     qq23             = iq2*jq3;
 841     qq31             = iq3*jq1;
 842     qq32             = iq3*jq2;
 843     qq33             = iq3*jq3;
 844
 845     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 846     rcutoff          = fr->rcoulomb;
 847     rcutoff2         = rcutoff*rcutoff;
 848
 849     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 850     rvdw             = fr->rvdw;
 851
 852     outeriter        = 0;
 853     inneriter        = 0;
 854
 855     /* Start outer loop over neighborlists */
 856     for(iidx=0; iidx<nri; iidx++)
 857     {
 858         /* Load shift vector for this list */
 859         i_shift_offset   = DIM*shiftidx[iidx];
 860         shX              = shiftvec[i_shift_offset+XX];
 861         shY              = shiftvec[i_shift_offset+YY];
 862         shZ              = shiftvec[i_shift_offset+ZZ];
 863
 864         /* Load limits for loop over neighbors */
 865         j_index_start    = jindex[iidx];
 866         j_index_end      = jindex[iidx+1];
 867
 868         /* Get outer coordinate index */
 869         inr              = iinr[iidx];
 870         i_coord_offset   = DIM*inr;
 871
 872         /* Load i particle coords and add shift vector */
 873         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 874         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 875         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 876         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 877         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 878         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 879         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 880         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 881         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 882         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 883         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 884         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 885
 886         fix0             = 0.0;
 887         fiy0             = 0.0;
 888         fiz0             = 0.0;
 889         fix1             = 0.0;
 890         fiy1             = 0.0;
 891         fiz1             = 0.0;
 892         fix2             = 0.0;
 893         fiy2             = 0.0;
 894         fiz2             = 0.0;
 895         fix3             = 0.0;
 896         fiy3             = 0.0;
 897         fiz3             = 0.0;
 898
 899         /* Start inner kernel loop */
 900         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 901         {
 902             /* Get j neighbor index, and coordinate index */
 903             jnr              = jjnr[jidx];
 904             j_coord_offset   = DIM*jnr;
 905
 906             /* load j atom coordinates */
 907             jx0              = x[j_coord_offset+DIM*0+XX];
 908             jy0              = x[j_coord_offset+DIM*0+YY];
 909             jz0              = x[j_coord_offset+DIM*0+ZZ];
 910             jx1              = x[j_coord_offset+DIM*1+XX];
 911             jy1              = x[j_coord_offset+DIM*1+YY];
 912             jz1              = x[j_coord_offset+DIM*1+ZZ];
 913             jx2              = x[j_coord_offset+DIM*2+XX];
 914             jy2              = x[j_coord_offset+DIM*2+YY];
 915             jz2              = x[j_coord_offset+DIM*2+ZZ];
 916             jx3              = x[j_coord_offset+DIM*3+XX];
 917             jy3              = x[j_coord_offset+DIM*3+YY];
 918             jz3              = x[j_coord_offset+DIM*3+ZZ];
 919
 920             /* Calculate displacement vector */
 921             dx00             = ix0 - jx0;
 922             dy00             = iy0 - jy0;
 923             dz00             = iz0 - jz0;
 924             dx11             = ix1 - jx1;
 925             dy11             = iy1 - jy1;
 926             dz11             = iz1 - jz1;
 927             dx12             = ix1 - jx2;
 928             dy12             = iy1 - jy2;
 929             dz12             = iz1 - jz2;
 930             dx13             = ix1 - jx3;
 931             dy13             = iy1 - jy3;
 932             dz13             = iz1 - jz3;
 933             dx21             = ix2 - jx1;
 934             dy21             = iy2 - jy1;
 935             dz21             = iz2 - jz1;
 936             dx22             = ix2 - jx2;
 937             dy22             = iy2 - jy2;
 938             dz22             = iz2 - jz2;
 939             dx23             = ix2 - jx3;
 940             dy23             = iy2 - jy3;
 941             dz23             = iz2 - jz3;
 942             dx31             = ix3 - jx1;
 943             dy31             = iy3 - jy1;
 944             dz31             = iz3 - jz1;
 945             dx32             = ix3 - jx2;
 946             dy32             = iy3 - jy2;
 947             dz32             = iz3 - jz2;
 948             dx33             = ix3 - jx3;
 949             dy33             = iy3 - jy3;
 950             dz33             = iz3 - jz3;
 951
 952             /* Calculate squared distance and things based on it */
 953             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 954             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 955             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 956             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 957             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 958             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 959             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 960             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 961             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 962             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 963
 964             rinv11           = gmx_invsqrt(rsq11);
 965             rinv12           = gmx_invsqrt(rsq12);
 966             rinv13           = gmx_invsqrt(rsq13);
 967             rinv21           = gmx_invsqrt(rsq21);
 968             rinv22           = gmx_invsqrt(rsq22);
 969             rinv23           = gmx_invsqrt(rsq23);
 970             rinv31           = gmx_invsqrt(rsq31);
 971             rinv32           = gmx_invsqrt(rsq32);
 972             rinv33           = gmx_invsqrt(rsq33);
 973
 974             rinvsq00         = 1.0/rsq00;
 975             rinvsq11         = rinv11*rinv11;
 976             rinvsq12         = rinv12*rinv12;
 977             rinvsq13         = rinv13*rinv13;
 978             rinvsq21         = rinv21*rinv21;
 979             rinvsq22         = rinv22*rinv22;
 980             rinvsq23         = rinv23*rinv23;
 981             rinvsq31         = rinv31*rinv31;
 982             rinvsq32         = rinv32*rinv32;
 983             rinvsq33         = rinv33*rinv33;
 984
 985             /**************************
 986              * CALCULATE INTERACTIONS *
 987              **************************/
 988
 989             if (rsq00<rcutoff2)
 990             {
 991
 992             /* LENNARD-JONES DISPERSION/REPULSION */
 993
 994             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 995             fvdw             = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00;
 996
 997             fscal            = fvdw;
 998
 999             /* Calculate temporary vectorial force */
1000             tx               = fscal*dx00;
1001             ty               = fscal*dy00;
1002             tz               = fscal*dz00;
1003
1004             /* Update vectorial force */
1005             fix0            += tx;
1006             fiy0            += ty;
1007             fiz0            += tz;
1008             f[j_coord_offset+DIM*0+XX] -= tx;
1009             f[j_coord_offset+DIM*0+YY] -= ty;
1010             f[j_coord_offset+DIM*0+ZZ] -= tz;
1011
1012             }
1013
1014             /**************************
1015              * CALCULATE INTERACTIONS *
1016              **************************/
1017
1018             if (rsq11<rcutoff2)
1019             {
1020
1021             r11              = rsq11*rinv11;
1022
1023             /* EWALD ELECTROSTATICS */
1024
1025             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1026             ewrt             = r11*ewtabscale;
1027             ewitab           = ewrt;
1028             eweps            = ewrt-ewitab;
1029             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1030             felec            = qq11*rinv11*(rinvsq11-felec);
1031
1032             fscal            = felec;
1033
1034             /* Calculate temporary vectorial force */
1035             tx               = fscal*dx11;
1036             ty               = fscal*dy11;
1037             tz               = fscal*dz11;
1038
1039             /* Update vectorial force */
1040             fix1            += tx;
1041             fiy1            += ty;
1042             fiz1            += tz;
1043             f[j_coord_offset+DIM*1+XX] -= tx;
1044             f[j_coord_offset+DIM*1+YY] -= ty;
1045             f[j_coord_offset+DIM*1+ZZ] -= tz;
1046
1047             }
1048
1049             /**************************
1050              * CALCULATE INTERACTIONS *
1051              **************************/
1052
1053             if (rsq12<rcutoff2)
1054             {
1055
1056             r12              = rsq12*rinv12;
1057
1058             /* EWALD ELECTROSTATICS */
1059
1060             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1061             ewrt             = r12*ewtabscale;
1062             ewitab           = ewrt;
1063             eweps            = ewrt-ewitab;
1064             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1065             felec            = qq12*rinv12*(rinvsq12-felec);
1066
1067             fscal            = felec;
1068
1069             /* Calculate temporary vectorial force */
1070             tx               = fscal*dx12;
1071             ty               = fscal*dy12;
1072             tz               = fscal*dz12;
1073
1074             /* Update vectorial force */
1075             fix1            += tx;
1076             fiy1            += ty;
1077             fiz1            += tz;
1078             f[j_coord_offset+DIM*2+XX] -= tx;
1079             f[j_coord_offset+DIM*2+YY] -= ty;
1080             f[j_coord_offset+DIM*2+ZZ] -= tz;
1081
1082             }
1083
1084             /**************************
1085              * CALCULATE INTERACTIONS *
1086              **************************/
1087
1088             if (rsq13<rcutoff2)
1089             {
1090
1091             r13              = rsq13*rinv13;
1092
1093             /* EWALD ELECTROSTATICS */
1094
1095             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1096             ewrt             = r13*ewtabscale;
1097             ewitab           = ewrt;
1098             eweps            = ewrt-ewitab;
1099             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1100             felec            = qq13*rinv13*(rinvsq13-felec);
1101
1102             fscal            = felec;
1103
1104             /* Calculate temporary vectorial force */
1105             tx               = fscal*dx13;
1106             ty               = fscal*dy13;
1107             tz               = fscal*dz13;
1108
1109             /* Update vectorial force */
1110             fix1            += tx;
1111             fiy1            += ty;
1112             fiz1            += tz;
1113             f[j_coord_offset+DIM*3+XX] -= tx;
1114             f[j_coord_offset+DIM*3+YY] -= ty;
1115             f[j_coord_offset+DIM*3+ZZ] -= tz;
1116
1117             }
1118
1119             /**************************
1120              * CALCULATE INTERACTIONS *
1121              **************************/
1122
1123             if (rsq21<rcutoff2)
1124             {
1125
1126             r21              = rsq21*rinv21;
1127
1128             /* EWALD ELECTROSTATICS */
1129
1130             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1131             ewrt             = r21*ewtabscale;
1132             ewitab           = ewrt;
1133             eweps            = ewrt-ewitab;
1134             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1135             felec            = qq21*rinv21*(rinvsq21-felec);
1136
1137             fscal            = felec;
1138
1139             /* Calculate temporary vectorial force */
1140             tx               = fscal*dx21;
1141             ty               = fscal*dy21;
1142             tz               = fscal*dz21;
1143
1144             /* Update vectorial force */
1145             fix2            += tx;
1146             fiy2            += ty;
1147             fiz2            += tz;
1148             f[j_coord_offset+DIM*1+XX] -= tx;
1149             f[j_coord_offset+DIM*1+YY] -= ty;
1150             f[j_coord_offset+DIM*1+ZZ] -= tz;
1151
1152             }
1153
1154             /**************************
1155              * CALCULATE INTERACTIONS *
1156              **************************/
1157
1158             if (rsq22<rcutoff2)
1159             {
1160
1161             r22              = rsq22*rinv22;
1162
1163             /* EWALD ELECTROSTATICS */
1164
1165             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1166             ewrt             = r22*ewtabscale;
1167             ewitab           = ewrt;
1168             eweps            = ewrt-ewitab;
1169             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1170             felec            = qq22*rinv22*(rinvsq22-felec);
1171
1172             fscal            = felec;
1173
1174             /* Calculate temporary vectorial force */
1175             tx               = fscal*dx22;
1176             ty               = fscal*dy22;
1177             tz               = fscal*dz22;
1178
1179             /* Update vectorial force */
1180             fix2            += tx;
1181             fiy2            += ty;
1182             fiz2            += tz;
1183             f[j_coord_offset+DIM*2+XX] -= tx;
1184             f[j_coord_offset+DIM*2+YY] -= ty;
1185             f[j_coord_offset+DIM*2+ZZ] -= tz;
1186
1187             }
1188
1189             /**************************
1190              * CALCULATE INTERACTIONS *
1191              **************************/
1192
1193             if (rsq23<rcutoff2)
1194             {
1195
1196             r23              = rsq23*rinv23;
1197
1198             /* EWALD ELECTROSTATICS */
1199
1200             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1201             ewrt             = r23*ewtabscale;
1202             ewitab           = ewrt;
1203             eweps            = ewrt-ewitab;
1204             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1205             felec            = qq23*rinv23*(rinvsq23-felec);
1206
1207             fscal            = felec;
1208
1209             /* Calculate temporary vectorial force */
1210             tx               = fscal*dx23;
1211             ty               = fscal*dy23;
1212             tz               = fscal*dz23;
1213
1214             /* Update vectorial force */
1215             fix2            += tx;
1216             fiy2            += ty;
1217             fiz2            += tz;
1218             f[j_coord_offset+DIM*3+XX] -= tx;
1219             f[j_coord_offset+DIM*3+YY] -= ty;
1220             f[j_coord_offset+DIM*3+ZZ] -= tz;
1221
1222             }
1223
1224             /**************************
1225              * CALCULATE INTERACTIONS *
1226              **************************/
1227
1228             if (rsq31<rcutoff2)
1229             {
1230
1231             r31              = rsq31*rinv31;
1232
1233             /* EWALD ELECTROSTATICS */
1234
1235             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1236             ewrt             = r31*ewtabscale;
1237             ewitab           = ewrt;
1238             eweps            = ewrt-ewitab;
1239             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1240             felec            = qq31*rinv31*(rinvsq31-felec);
1241
1242             fscal            = felec;
1243
1244             /* Calculate temporary vectorial force */
1245             tx               = fscal*dx31;
1246             ty               = fscal*dy31;
1247             tz               = fscal*dz31;
1248
1249             /* Update vectorial force */
1250             fix3            += tx;
1251             fiy3            += ty;
1252             fiz3            += tz;
1253             f[j_coord_offset+DIM*1+XX] -= tx;
1254             f[j_coord_offset+DIM*1+YY] -= ty;
1255             f[j_coord_offset+DIM*1+ZZ] -= tz;
1256
1257             }
1258
1259             /**************************
1260              * CALCULATE INTERACTIONS *
1261              **************************/
1262
1263             if (rsq32<rcutoff2)
1264             {
1265
1266             r32              = rsq32*rinv32;
1267
1268             /* EWALD ELECTROSTATICS */
1269
1270             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1271             ewrt             = r32*ewtabscale;
1272             ewitab           = ewrt;
1273             eweps            = ewrt-ewitab;
1274             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1275             felec            = qq32*rinv32*(rinvsq32-felec);
1276
1277             fscal            = felec;
1278
1279             /* Calculate temporary vectorial force */
1280             tx               = fscal*dx32;
1281             ty               = fscal*dy32;
1282             tz               = fscal*dz32;
1283
1284             /* Update vectorial force */
1285             fix3            += tx;
1286             fiy3            += ty;
1287             fiz3            += tz;
1288             f[j_coord_offset+DIM*2+XX] -= tx;
1289             f[j_coord_offset+DIM*2+YY] -= ty;
1290             f[j_coord_offset+DIM*2+ZZ] -= tz;
1291
1292             }
1293
1294             /**************************
1295              * CALCULATE INTERACTIONS *
1296              **************************/
1297
1298             if (rsq33<rcutoff2)
1299             {
1300
1301             r33              = rsq33*rinv33;
1302
1303             /* EWALD ELECTROSTATICS */
1304
1305             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1306             ewrt             = r33*ewtabscale;
1307             ewitab           = ewrt;
1308             eweps            = ewrt-ewitab;
1309             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1310             felec            = qq33*rinv33*(rinvsq33-felec);
1311
1312             fscal            = felec;
1313
1314             /* Calculate temporary vectorial force */
1315             tx               = fscal*dx33;
1316             ty               = fscal*dy33;
1317             tz               = fscal*dz33;
1318
1319             /* Update vectorial force */
1320             fix3            += tx;
1321             fiy3            += ty;
1322             fiz3            += tz;
1323             f[j_coord_offset+DIM*3+XX] -= tx;
1324             f[j_coord_offset+DIM*3+YY] -= ty;
1325             f[j_coord_offset+DIM*3+ZZ] -= tz;
1326
1327             }
1328
1329             /* Inner loop uses 324 flops */
1330         }
1331         /* End of innermost loop */
1332
1333         tx = ty = tz = 0;
1334         f[i_coord_offset+DIM*0+XX] += fix0;
1335         f[i_coord_offset+DIM*0+YY] += fiy0;
1336         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1337         tx                         += fix0;
1338         ty                         += fiy0;
1339         tz                         += fiz0;
1340         f[i_coord_offset+DIM*1+XX] += fix1;
1341         f[i_coord_offset+DIM*1+YY] += fiy1;
1342         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1343         tx                         += fix1;
1344         ty                         += fiy1;
1345         tz                         += fiz1;
1346         f[i_coord_offset+DIM*2+XX] += fix2;
1347         f[i_coord_offset+DIM*2+YY] += fiy2;
1348         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1349         tx                         += fix2;
1350         ty                         += fiy2;
1351         tz                         += fiz2;
1352         f[i_coord_offset+DIM*3+XX] += fix3;
1353         f[i_coord_offset+DIM*3+YY] += fiy3;
1354         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1355         tx                         += fix3;
1356         ty                         += fiy3;
1357         tz                         += fiz3;
1358         fshift[i_shift_offset+XX]  += tx;
1359         fshift[i_shift_offset+YY]  += ty;
1360         fshift[i_shift_offset+ZZ]  += tz;
1361
1362         /* Increment number of inner iterations */
1363         inneriter                  += j_index_end - j_index_start;
1364
1365         /* Outer loop uses 39 flops */
1366     }
1367
1368     /* Increment number of outer iterations */
1369     outeriter        += nri;
1370
1371     /* Update outer/inner flops */
1372
1373     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*324);
1374 }