src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSh_VdwBhamSh_GeomW4W4_c.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS c kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwBhamSh_GeomW4W4_VF_c
  49  * Electrostatics interaction: Ewald
  50  * VdW interaction:            Buckingham
  51  * Geometry:                   Water4-Water4
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecEwSh_VdwBhamSh_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      = 3*nvdwtype*vdwtype[inr+0];
 134
 135     jq1              = charge[inr+1];
 136     jq2              = charge[inr+2];
 137     jq3              = charge[inr+3];
 138     vdwjidx0         = 3*vdwtype[inr+0];
 139     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 140     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 141     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 142     qq11             = iq1*jq1;
 143     qq12             = iq1*jq2;
 144     qq13             = iq1*jq3;
 145     qq21             = iq2*jq1;
 146     qq22             = iq2*jq2;
 147     qq23             = iq2*jq3;
 148     qq31             = iq3*jq1;
 149     qq32             = iq3*jq2;
 150     qq33             = iq3*jq3;
 151
 152     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 153     rcutoff          = fr->rcoulomb;
 154     rcutoff2         = rcutoff*rcutoff;
 155
 156     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 157     rvdw             = fr->rvdw;
 158
 159     outeriter        = 0;
 160     inneriter        = 0;
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167         shX              = shiftvec[i_shift_offset+XX];
 168         shY              = shiftvec[i_shift_offset+YY];
 169         shZ              = shiftvec[i_shift_offset+ZZ];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 181         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 182         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 183         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 184         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 185         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 186         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 187         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 188         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 189         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 190         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 191         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 192
 193         fix0             = 0.0;
 194         fiy0             = 0.0;
 195         fiz0             = 0.0;
 196         fix1             = 0.0;
 197         fiy1             = 0.0;
 198         fiz1             = 0.0;
 199         fix2             = 0.0;
 200         fiy2             = 0.0;
 201         fiz2             = 0.0;
 202         fix3             = 0.0;
 203         fiy3             = 0.0;
 204         fiz3             = 0.0;
 205
 206         /* Reset potential sums */
 207         velecsum         = 0.0;
 208         vvdwsum          = 0.0;
 209
 210         /* Start inner kernel loop */
 211         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 212         {
 213             /* Get j neighbor index, and coordinate index */
 214             jnr              = jjnr[jidx];
 215             j_coord_offset   = DIM*jnr;
 216
 217             /* load j atom coordinates */
 218             jx0              = x[j_coord_offset+DIM*0+XX];
 219             jy0              = x[j_coord_offset+DIM*0+YY];
 220             jz0              = x[j_coord_offset+DIM*0+ZZ];
 221             jx1              = x[j_coord_offset+DIM*1+XX];
 222             jy1              = x[j_coord_offset+DIM*1+YY];
 223             jz1              = x[j_coord_offset+DIM*1+ZZ];
 224             jx2              = x[j_coord_offset+DIM*2+XX];
 225             jy2              = x[j_coord_offset+DIM*2+YY];
 226             jz2              = x[j_coord_offset+DIM*2+ZZ];
 227             jx3              = x[j_coord_offset+DIM*3+XX];
 228             jy3              = x[j_coord_offset+DIM*3+YY];
 229             jz3              = x[j_coord_offset+DIM*3+ZZ];
 230
 231             /* Calculate displacement vector */
 232             dx00             = ix0 - jx0;
 233             dy00             = iy0 - jy0;
 234             dz00             = iz0 - jz0;
 235             dx11             = ix1 - jx1;
 236             dy11             = iy1 - jy1;
 237             dz11             = iz1 - jz1;
 238             dx12             = ix1 - jx2;
 239             dy12             = iy1 - jy2;
 240             dz12             = iz1 - jz2;
 241             dx13             = ix1 - jx3;
 242             dy13             = iy1 - jy3;
 243             dz13             = iz1 - jz3;
 244             dx21             = ix2 - jx1;
 245             dy21             = iy2 - jy1;
 246             dz21             = iz2 - jz1;
 247             dx22             = ix2 - jx2;
 248             dy22             = iy2 - jy2;
 249             dz22             = iz2 - jz2;
 250             dx23             = ix2 - jx3;
 251             dy23             = iy2 - jy3;
 252             dz23             = iz2 - jz3;
 253             dx31             = ix3 - jx1;
 254             dy31             = iy3 - jy1;
 255             dz31             = iz3 - jz1;
 256             dx32             = ix3 - jx2;
 257             dy32             = iy3 - jy2;
 258             dz32             = iz3 - jz2;
 259             dx33             = ix3 - jx3;
 260             dy33             = iy3 - jy3;
 261             dz33             = iz3 - jz3;
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 265             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 266             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 267             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 268             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 269             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 270             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 271             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 272             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 273             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 274
 275             rinv00           = gmx_invsqrt(rsq00);
 276             rinv11           = gmx_invsqrt(rsq11);
 277             rinv12           = gmx_invsqrt(rsq12);
 278             rinv13           = gmx_invsqrt(rsq13);
 279             rinv21           = gmx_invsqrt(rsq21);
 280             rinv22           = gmx_invsqrt(rsq22);
 281             rinv23           = gmx_invsqrt(rsq23);
 282             rinv31           = gmx_invsqrt(rsq31);
 283             rinv32           = gmx_invsqrt(rsq32);
 284             rinv33           = gmx_invsqrt(rsq33);
 285
 286             rinvsq00         = rinv00*rinv00;
 287             rinvsq11         = rinv11*rinv11;
 288             rinvsq12         = rinv12*rinv12;
 289             rinvsq13         = rinv13*rinv13;
 290             rinvsq21         = rinv21*rinv21;
 291             rinvsq22         = rinv22*rinv22;
 292             rinvsq23         = rinv23*rinv23;
 293             rinvsq31         = rinv31*rinv31;
 294             rinvsq32         = rinv32*rinv32;
 295             rinvsq33         = rinv33*rinv33;
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             if (rsq00<rcutoff2)
 302             {
 303
 304             r00              = rsq00*rinv00;
 305
 306             /* BUCKINGHAM DISPERSION/REPULSION */
 307             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 308             vvdw6            = c6_00*rinvsix;
 309             br               = cexp2_00*r00;
 310             vvdwexp          = cexp1_00*exp(-br);
 311             vvdw             = (vvdwexp-cexp1_00*exp(-cexp2_00*rvdw)) - (vvdw6 - c6_00*sh_vdw_invrcut6)*(1.0/6.0);
 312             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 313
 314             /* Update potential sums from outer loop */
 315             vvdwsum         += vvdw;
 316
 317             fscal            = fvdw;
 318
 319             /* Calculate temporary vectorial force */
 320             tx               = fscal*dx00;
 321             ty               = fscal*dy00;
 322             tz               = fscal*dz00;
 323
 324             /* Update vectorial force */
 325             fix0            += tx;
 326             fiy0            += ty;
 327             fiz0            += tz;
 328             f[j_coord_offset+DIM*0+XX] -= tx;
 329             f[j_coord_offset+DIM*0+YY] -= ty;
 330             f[j_coord_offset+DIM*0+ZZ] -= tz;
 331
 332             }
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             if (rsq11<rcutoff2)
 339             {
 340
 341             r11              = rsq11*rinv11;
 342
 343             /* EWALD ELECTROSTATICS */
 344
 345             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 346             ewrt             = r11*ewtabscale;
 347             ewitab           = ewrt;
 348             eweps            = ewrt-ewitab;
 349             ewitab           = 4*ewitab;
 350             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 351             velec            = qq11*((rinv11-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 352             felec            = qq11*rinv11*(rinvsq11-felec);
 353
 354             /* Update potential sums from outer loop */
 355             velecsum        += velec;
 356
 357             fscal            = felec;
 358
 359             /* Calculate temporary vectorial force */
 360             tx               = fscal*dx11;
 361             ty               = fscal*dy11;
 362             tz               = fscal*dz11;
 363
 364             /* Update vectorial force */
 365             fix1            += tx;
 366             fiy1            += ty;
 367             fiz1            += tz;
 368             f[j_coord_offset+DIM*1+XX] -= tx;
 369             f[j_coord_offset+DIM*1+YY] -= ty;
 370             f[j_coord_offset+DIM*1+ZZ] -= tz;
 371
 372             }
 373
 374             /**************************
 375              * CALCULATE INTERACTIONS *
 376              **************************/
 377
 378             if (rsq12<rcutoff2)
 379             {
 380
 381             r12              = rsq12*rinv12;
 382
 383             /* EWALD ELECTROSTATICS */
 384
 385             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 386             ewrt             = r12*ewtabscale;
 387             ewitab           = ewrt;
 388             eweps            = ewrt-ewitab;
 389             ewitab           = 4*ewitab;
 390             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 391             velec            = qq12*((rinv12-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 392             felec            = qq12*rinv12*(rinvsq12-felec);
 393
 394             /* Update potential sums from outer loop */
 395             velecsum        += velec;
 396
 397             fscal            = felec;
 398
 399             /* Calculate temporary vectorial force */
 400             tx               = fscal*dx12;
 401             ty               = fscal*dy12;
 402             tz               = fscal*dz12;
 403
 404             /* Update vectorial force */
 405             fix1            += tx;
 406             fiy1            += ty;
 407             fiz1            += tz;
 408             f[j_coord_offset+DIM*2+XX] -= tx;
 409             f[j_coord_offset+DIM*2+YY] -= ty;
 410             f[j_coord_offset+DIM*2+ZZ] -= tz;
 411
 412             }
 413
 414             /**************************
 415              * CALCULATE INTERACTIONS *
 416              **************************/
 417
 418             if (rsq13<rcutoff2)
 419             {
 420
 421             r13              = rsq13*rinv13;
 422
 423             /* EWALD ELECTROSTATICS */
 424
 425             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 426             ewrt             = r13*ewtabscale;
 427             ewitab           = ewrt;
 428             eweps            = ewrt-ewitab;
 429             ewitab           = 4*ewitab;
 430             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 431             velec            = qq13*((rinv13-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 432             felec            = qq13*rinv13*(rinvsq13-felec);
 433
 434             /* Update potential sums from outer loop */
 435             velecsum        += velec;
 436
 437             fscal            = felec;
 438
 439             /* Calculate temporary vectorial force */
 440             tx               = fscal*dx13;
 441             ty               = fscal*dy13;
 442             tz               = fscal*dz13;
 443
 444             /* Update vectorial force */
 445             fix1            += tx;
 446             fiy1            += ty;
 447             fiz1            += tz;
 448             f[j_coord_offset+DIM*3+XX] -= tx;
 449             f[j_coord_offset+DIM*3+YY] -= ty;
 450             f[j_coord_offset+DIM*3+ZZ] -= tz;
 451
 452             }
 453
 454             /**************************
 455              * CALCULATE INTERACTIONS *
 456              **************************/
 457
 458             if (rsq21<rcutoff2)
 459             {
 460
 461             r21              = rsq21*rinv21;
 462
 463             /* EWALD ELECTROSTATICS */
 464
 465             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 466             ewrt             = r21*ewtabscale;
 467             ewitab           = ewrt;
 468             eweps            = ewrt-ewitab;
 469             ewitab           = 4*ewitab;
 470             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 471             velec            = qq21*((rinv21-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 472             felec            = qq21*rinv21*(rinvsq21-felec);
 473
 474             /* Update potential sums from outer loop */
 475             velecsum        += velec;
 476
 477             fscal            = felec;
 478
 479             /* Calculate temporary vectorial force */
 480             tx               = fscal*dx21;
 481             ty               = fscal*dy21;
 482             tz               = fscal*dz21;
 483
 484             /* Update vectorial force */
 485             fix2            += tx;
 486             fiy2            += ty;
 487             fiz2            += tz;
 488             f[j_coord_offset+DIM*1+XX] -= tx;
 489             f[j_coord_offset+DIM*1+YY] -= ty;
 490             f[j_coord_offset+DIM*1+ZZ] -= tz;
 491
 492             }
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             if (rsq22<rcutoff2)
 499             {
 500
 501             r22              = rsq22*rinv22;
 502
 503             /* EWALD ELECTROSTATICS */
 504
 505             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 506             ewrt             = r22*ewtabscale;
 507             ewitab           = ewrt;
 508             eweps            = ewrt-ewitab;
 509             ewitab           = 4*ewitab;
 510             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 511             velec            = qq22*((rinv22-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 512             felec            = qq22*rinv22*(rinvsq22-felec);
 513
 514             /* Update potential sums from outer loop */
 515             velecsum        += velec;
 516
 517             fscal            = felec;
 518
 519             /* Calculate temporary vectorial force */
 520             tx               = fscal*dx22;
 521             ty               = fscal*dy22;
 522             tz               = fscal*dz22;
 523
 524             /* Update vectorial force */
 525             fix2            += tx;
 526             fiy2            += ty;
 527             fiz2            += tz;
 528             f[j_coord_offset+DIM*2+XX] -= tx;
 529             f[j_coord_offset+DIM*2+YY] -= ty;
 530             f[j_coord_offset+DIM*2+ZZ] -= tz;
 531
 532             }
 533
 534             /**************************
 535              * CALCULATE INTERACTIONS *
 536              **************************/
 537
 538             if (rsq23<rcutoff2)
 539             {
 540
 541             r23              = rsq23*rinv23;
 542
 543             /* EWALD ELECTROSTATICS */
 544
 545             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 546             ewrt             = r23*ewtabscale;
 547             ewitab           = ewrt;
 548             eweps            = ewrt-ewitab;
 549             ewitab           = 4*ewitab;
 550             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 551             velec            = qq23*((rinv23-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 552             felec            = qq23*rinv23*(rinvsq23-felec);
 553
 554             /* Update potential sums from outer loop */
 555             velecsum        += velec;
 556
 557             fscal            = felec;
 558
 559             /* Calculate temporary vectorial force */
 560             tx               = fscal*dx23;
 561             ty               = fscal*dy23;
 562             tz               = fscal*dz23;
 563
 564             /* Update vectorial force */
 565             fix2            += tx;
 566             fiy2            += ty;
 567             fiz2            += tz;
 568             f[j_coord_offset+DIM*3+XX] -= tx;
 569             f[j_coord_offset+DIM*3+YY] -= ty;
 570             f[j_coord_offset+DIM*3+ZZ] -= tz;
 571
 572             }
 573
 574             /**************************
 575              * CALCULATE INTERACTIONS *
 576              **************************/
 577
 578             if (rsq31<rcutoff2)
 579             {
 580
 581             r31              = rsq31*rinv31;
 582
 583             /* EWALD ELECTROSTATICS */
 584
 585             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 586             ewrt             = r31*ewtabscale;
 587             ewitab           = ewrt;
 588             eweps            = ewrt-ewitab;
 589             ewitab           = 4*ewitab;
 590             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 591             velec            = qq31*((rinv31-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 592             felec            = qq31*rinv31*(rinvsq31-felec);
 593
 594             /* Update potential sums from outer loop */
 595             velecsum        += velec;
 596
 597             fscal            = felec;
 598
 599             /* Calculate temporary vectorial force */
 600             tx               = fscal*dx31;
 601             ty               = fscal*dy31;
 602             tz               = fscal*dz31;
 603
 604             /* Update vectorial force */
 605             fix3            += tx;
 606             fiy3            += ty;
 607             fiz3            += tz;
 608             f[j_coord_offset+DIM*1+XX] -= tx;
 609             f[j_coord_offset+DIM*1+YY] -= ty;
 610             f[j_coord_offset+DIM*1+ZZ] -= tz;
 611
 612             }
 613
 614             /**************************
 615              * CALCULATE INTERACTIONS *
 616              **************************/
 617
 618             if (rsq32<rcutoff2)
 619             {
 620
 621             r32              = rsq32*rinv32;
 622
 623             /* EWALD ELECTROSTATICS */
 624
 625             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 626             ewrt             = r32*ewtabscale;
 627             ewitab           = ewrt;
 628             eweps            = ewrt-ewitab;
 629             ewitab           = 4*ewitab;
 630             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 631             velec            = qq32*((rinv32-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 632             felec            = qq32*rinv32*(rinvsq32-felec);
 633
 634             /* Update potential sums from outer loop */
 635             velecsum        += velec;
 636
 637             fscal            = felec;
 638
 639             /* Calculate temporary vectorial force */
 640             tx               = fscal*dx32;
 641             ty               = fscal*dy32;
 642             tz               = fscal*dz32;
 643
 644             /* Update vectorial force */
 645             fix3            += tx;
 646             fiy3            += ty;
 647             fiz3            += tz;
 648             f[j_coord_offset+DIM*2+XX] -= tx;
 649             f[j_coord_offset+DIM*2+YY] -= ty;
 650             f[j_coord_offset+DIM*2+ZZ] -= tz;
 651
 652             }
 653
 654             /**************************
 655              * CALCULATE INTERACTIONS *
 656              **************************/
 657
 658             if (rsq33<rcutoff2)
 659             {
 660
 661             r33              = rsq33*rinv33;
 662
 663             /* EWALD ELECTROSTATICS */
 664
 665             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 666             ewrt             = r33*ewtabscale;
 667             ewitab           = ewrt;
 668             eweps            = ewrt-ewitab;
 669             ewitab           = 4*ewitab;
 670             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 671             velec            = qq33*((rinv33-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 672             felec            = qq33*rinv33*(rinvsq33-felec);
 673
 674             /* Update potential sums from outer loop */
 675             velecsum        += velec;
 676
 677             fscal            = felec;
 678
 679             /* Calculate temporary vectorial force */
 680             tx               = fscal*dx33;
 681             ty               = fscal*dy33;
 682             tz               = fscal*dz33;
 683
 684             /* Update vectorial force */
 685             fix3            += tx;
 686             fiy3            += ty;
 687             fiz3            += tz;
 688             f[j_coord_offset+DIM*3+XX] -= tx;
 689             f[j_coord_offset+DIM*3+YY] -= ty;
 690             f[j_coord_offset+DIM*3+ZZ] -= tz;
 691
 692             }
 693
 694             /* Inner loop uses 461 flops */
 695         }
 696         /* End of innermost loop */
 697
 698         tx = ty = tz = 0;
 699         f[i_coord_offset+DIM*0+XX] += fix0;
 700         f[i_coord_offset+DIM*0+YY] += fiy0;
 701         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 702         tx                         += fix0;
 703         ty                         += fiy0;
 704         tz                         += fiz0;
 705         f[i_coord_offset+DIM*1+XX] += fix1;
 706         f[i_coord_offset+DIM*1+YY] += fiy1;
 707         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 708         tx                         += fix1;
 709         ty                         += fiy1;
 710         tz                         += fiz1;
 711         f[i_coord_offset+DIM*2+XX] += fix2;
 712         f[i_coord_offset+DIM*2+YY] += fiy2;
 713         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 714         tx                         += fix2;
 715         ty                         += fiy2;
 716         tz                         += fiz2;
 717         f[i_coord_offset+DIM*3+XX] += fix3;
 718         f[i_coord_offset+DIM*3+YY] += fiy3;
 719         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 720         tx                         += fix3;
 721         ty                         += fiy3;
 722         tz                         += fiz3;
 723         fshift[i_shift_offset+XX]  += tx;
 724         fshift[i_shift_offset+YY]  += ty;
 725         fshift[i_shift_offset+ZZ]  += tz;
 726
 727         ggid                        = gid[iidx];
 728         /* Update potential energies */
 729         kernel_data->energygrp_elec[ggid] += velecsum;
 730         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 731
 732         /* Increment number of inner iterations */
 733         inneriter                  += j_index_end - j_index_start;
 734
 735         /* Outer loop uses 41 flops */
 736     }
 737
 738     /* Increment number of outer iterations */
 739     outeriter        += nri;
 740
 741     /* Update outer/inner flops */
 742
 743     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*461);
 744 }
 745 /*
 746  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwBhamSh_GeomW4W4_F_c
 747  * Electrostatics interaction: Ewald
 748  * VdW interaction:            Buckingham
 749  * Geometry:                   Water4-Water4
 750  * Calculate force/pot:        Force
 751  */
 752 void
 753 nb_kernel_ElecEwSh_VdwBhamSh_GeomW4W4_F_c
 754                     (t_nblist                    * gmx_restrict       nlist,
 755                      rvec                        * gmx_restrict          xx,
 756                      rvec                        * gmx_restrict          ff,
 757                      t_forcerec                  * gmx_restrict          fr,
 758                      t_mdatoms                   * gmx_restrict     mdatoms,
 759                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 760                      t_nrnb                      * gmx_restrict        nrnb)
 761 {
 762     int              i_shift_offset,i_coord_offset,j_coord_offset;
 763     int              j_index_start,j_index_end;
 764     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 765     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 766     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 767     real             *shiftvec,*fshift,*x,*f;
 768     int              vdwioffset0;
 769     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 770     int              vdwioffset1;
 771     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 772     int              vdwioffset2;
 773     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 774     int              vdwioffset3;
 775     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 776     int              vdwjidx0;
 777     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 778     int              vdwjidx1;
 779     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 780     int              vdwjidx2;
 781     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 782     int              vdwjidx3;
 783     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 784     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 785     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 786     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 787     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 788     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 789     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 790     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 791     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 792     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 793     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 794     real             velec,felec,velecsum,facel,crf,krf,krf2;
 795     real             *charge;
 796     int              nvdwtype;
 797     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 798     int              *vdwtype;
 799     real             *vdwparam;
 800     int              ewitab;
 801     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 802     real             *ewtab;
 803
 804     x                = xx[0];
 805     f                = ff[0];
 806
 807     nri              = nlist->nri;
 808     iinr             = nlist->iinr;
 809     jindex           = nlist->jindex;
 810     jjnr             = nlist->jjnr;
 811     shiftidx         = nlist->shift;
 812     gid              = nlist->gid;
 813     shiftvec         = fr->shift_vec[0];
 814     fshift           = fr->fshift[0];
 815     facel            = fr->epsfac;
 816     charge           = mdatoms->chargeA;
 817     nvdwtype         = fr->ntype;
 818     vdwparam         = fr->nbfp;
 819     vdwtype          = mdatoms->typeA;
 820
 821     sh_ewald         = fr->ic->sh_ewald;
 822     ewtab            = fr->ic->tabq_coul_F;
 823     ewtabscale       = fr->ic->tabq_scale;
 824     ewtabhalfspace   = 0.5/ewtabscale;
 825
 826     /* Setup water-specific parameters */
 827     inr              = nlist->iinr[0];
 828     iq1              = facel*charge[inr+1];
 829     iq2              = facel*charge[inr+2];
 830     iq3              = facel*charge[inr+3];
 831     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 832
 833     jq1              = charge[inr+1];
 834     jq2              = charge[inr+2];
 835     jq3              = charge[inr+3];
 836     vdwjidx0         = 3*vdwtype[inr+0];
 837     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 838     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 839     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 840     qq11             = iq1*jq1;
 841     qq12             = iq1*jq2;
 842     qq13             = iq1*jq3;
 843     qq21             = iq2*jq1;
 844     qq22             = iq2*jq2;
 845     qq23             = iq2*jq3;
 846     qq31             = iq3*jq1;
 847     qq32             = iq3*jq2;
 848     qq33             = iq3*jq3;
 849
 850     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 851     rcutoff          = fr->rcoulomb;
 852     rcutoff2         = rcutoff*rcutoff;
 853
 854     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 855     rvdw             = fr->rvdw;
 856
 857     outeriter        = 0;
 858     inneriter        = 0;
 859
 860     /* Start outer loop over neighborlists */
 861     for(iidx=0; iidx<nri; iidx++)
 862     {
 863         /* Load shift vector for this list */
 864         i_shift_offset   = DIM*shiftidx[iidx];
 865         shX              = shiftvec[i_shift_offset+XX];
 866         shY              = shiftvec[i_shift_offset+YY];
 867         shZ              = shiftvec[i_shift_offset+ZZ];
 868
 869         /* Load limits for loop over neighbors */
 870         j_index_start    = jindex[iidx];
 871         j_index_end      = jindex[iidx+1];
 872
 873         /* Get outer coordinate index */
 874         inr              = iinr[iidx];
 875         i_coord_offset   = DIM*inr;
 876
 877         /* Load i particle coords and add shift vector */
 878         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 879         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 880         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 881         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 882         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 883         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 884         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 885         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 886         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 887         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 888         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 889         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 890
 891         fix0             = 0.0;
 892         fiy0             = 0.0;
 893         fiz0             = 0.0;
 894         fix1             = 0.0;
 895         fiy1             = 0.0;
 896         fiz1             = 0.0;
 897         fix2             = 0.0;
 898         fiy2             = 0.0;
 899         fiz2             = 0.0;
 900         fix3             = 0.0;
 901         fiy3             = 0.0;
 902         fiz3             = 0.0;
 903
 904         /* Start inner kernel loop */
 905         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 906         {
 907             /* Get j neighbor index, and coordinate index */
 908             jnr              = jjnr[jidx];
 909             j_coord_offset   = DIM*jnr;
 910
 911             /* load j atom coordinates */
 912             jx0              = x[j_coord_offset+DIM*0+XX];
 913             jy0              = x[j_coord_offset+DIM*0+YY];
 914             jz0              = x[j_coord_offset+DIM*0+ZZ];
 915             jx1              = x[j_coord_offset+DIM*1+XX];
 916             jy1              = x[j_coord_offset+DIM*1+YY];
 917             jz1              = x[j_coord_offset+DIM*1+ZZ];
 918             jx2              = x[j_coord_offset+DIM*2+XX];
 919             jy2              = x[j_coord_offset+DIM*2+YY];
 920             jz2              = x[j_coord_offset+DIM*2+ZZ];
 921             jx3              = x[j_coord_offset+DIM*3+XX];
 922             jy3              = x[j_coord_offset+DIM*3+YY];
 923             jz3              = x[j_coord_offset+DIM*3+ZZ];
 924
 925             /* Calculate displacement vector */
 926             dx00             = ix0 - jx0;
 927             dy00             = iy0 - jy0;
 928             dz00             = iz0 - jz0;
 929             dx11             = ix1 - jx1;
 930             dy11             = iy1 - jy1;
 931             dz11             = iz1 - jz1;
 932             dx12             = ix1 - jx2;
 933             dy12             = iy1 - jy2;
 934             dz12             = iz1 - jz2;
 935             dx13             = ix1 - jx3;
 936             dy13             = iy1 - jy3;
 937             dz13             = iz1 - jz3;
 938             dx21             = ix2 - jx1;
 939             dy21             = iy2 - jy1;
 940             dz21             = iz2 - jz1;
 941             dx22             = ix2 - jx2;
 942             dy22             = iy2 - jy2;
 943             dz22             = iz2 - jz2;
 944             dx23             = ix2 - jx3;
 945             dy23             = iy2 - jy3;
 946             dz23             = iz2 - jz3;
 947             dx31             = ix3 - jx1;
 948             dy31             = iy3 - jy1;
 949             dz31             = iz3 - jz1;
 950             dx32             = ix3 - jx2;
 951             dy32             = iy3 - jy2;
 952             dz32             = iz3 - jz2;
 953             dx33             = ix3 - jx3;
 954             dy33             = iy3 - jy3;
 955             dz33             = iz3 - jz3;
 956
 957             /* Calculate squared distance and things based on it */
 958             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 959             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 960             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 961             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 962             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 963             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 964             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 965             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 966             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 967             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 968
 969             rinv00           = gmx_invsqrt(rsq00);
 970             rinv11           = gmx_invsqrt(rsq11);
 971             rinv12           = gmx_invsqrt(rsq12);
 972             rinv13           = gmx_invsqrt(rsq13);
 973             rinv21           = gmx_invsqrt(rsq21);
 974             rinv22           = gmx_invsqrt(rsq22);
 975             rinv23           = gmx_invsqrt(rsq23);
 976             rinv31           = gmx_invsqrt(rsq31);
 977             rinv32           = gmx_invsqrt(rsq32);
 978             rinv33           = gmx_invsqrt(rsq33);
 979
 980             rinvsq00         = rinv00*rinv00;
 981             rinvsq11         = rinv11*rinv11;
 982             rinvsq12         = rinv12*rinv12;
 983             rinvsq13         = rinv13*rinv13;
 984             rinvsq21         = rinv21*rinv21;
 985             rinvsq22         = rinv22*rinv22;
 986             rinvsq23         = rinv23*rinv23;
 987             rinvsq31         = rinv31*rinv31;
 988             rinvsq32         = rinv32*rinv32;
 989             rinvsq33         = rinv33*rinv33;
 990
 991             /**************************
 992              * CALCULATE INTERACTIONS *
 993              **************************/
 994
 995             if (rsq00<rcutoff2)
 996             {
 997
 998             r00              = rsq00*rinv00;
 999
1000             /* BUCKINGHAM DISPERSION/REPULSION */
1001             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
1002             vvdw6            = c6_00*rinvsix;
1003             br               = cexp2_00*r00;
1004             vvdwexp          = cexp1_00*exp(-br);
1005             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
1006
1007             fscal            = fvdw;
1008
1009             /* Calculate temporary vectorial force */
1010             tx               = fscal*dx00;
1011             ty               = fscal*dy00;
1012             tz               = fscal*dz00;
1013
1014             /* Update vectorial force */
1015             fix0            += tx;
1016             fiy0            += ty;
1017             fiz0            += tz;
1018             f[j_coord_offset+DIM*0+XX] -= tx;
1019             f[j_coord_offset+DIM*0+YY] -= ty;
1020             f[j_coord_offset+DIM*0+ZZ] -= tz;
1021
1022             }
1023
1024             /**************************
1025              * CALCULATE INTERACTIONS *
1026              **************************/
1027
1028             if (rsq11<rcutoff2)
1029             {
1030
1031             r11              = rsq11*rinv11;
1032
1033             /* EWALD ELECTROSTATICS */
1034
1035             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1036             ewrt             = r11*ewtabscale;
1037             ewitab           = ewrt;
1038             eweps            = ewrt-ewitab;
1039             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1040             felec            = qq11*rinv11*(rinvsq11-felec);
1041
1042             fscal            = felec;
1043
1044             /* Calculate temporary vectorial force */
1045             tx               = fscal*dx11;
1046             ty               = fscal*dy11;
1047             tz               = fscal*dz11;
1048
1049             /* Update vectorial force */
1050             fix1            += tx;
1051             fiy1            += ty;
1052             fiz1            += tz;
1053             f[j_coord_offset+DIM*1+XX] -= tx;
1054             f[j_coord_offset+DIM*1+YY] -= ty;
1055             f[j_coord_offset+DIM*1+ZZ] -= tz;
1056
1057             }
1058
1059             /**************************
1060              * CALCULATE INTERACTIONS *
1061              **************************/
1062
1063             if (rsq12<rcutoff2)
1064             {
1065
1066             r12              = rsq12*rinv12;
1067
1068             /* EWALD ELECTROSTATICS */
1069
1070             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1071             ewrt             = r12*ewtabscale;
1072             ewitab           = ewrt;
1073             eweps            = ewrt-ewitab;
1074             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1075             felec            = qq12*rinv12*(rinvsq12-felec);
1076
1077             fscal            = felec;
1078
1079             /* Calculate temporary vectorial force */
1080             tx               = fscal*dx12;
1081             ty               = fscal*dy12;
1082             tz               = fscal*dz12;
1083
1084             /* Update vectorial force */
1085             fix1            += tx;
1086             fiy1            += ty;
1087             fiz1            += tz;
1088             f[j_coord_offset+DIM*2+XX] -= tx;
1089             f[j_coord_offset+DIM*2+YY] -= ty;
1090             f[j_coord_offset+DIM*2+ZZ] -= tz;
1091
1092             }
1093
1094             /**************************
1095              * CALCULATE INTERACTIONS *
1096              **************************/
1097
1098             if (rsq13<rcutoff2)
1099             {
1100
1101             r13              = rsq13*rinv13;
1102
1103             /* EWALD ELECTROSTATICS */
1104
1105             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1106             ewrt             = r13*ewtabscale;
1107             ewitab           = ewrt;
1108             eweps            = ewrt-ewitab;
1109             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1110             felec            = qq13*rinv13*(rinvsq13-felec);
1111
1112             fscal            = felec;
1113
1114             /* Calculate temporary vectorial force */
1115             tx               = fscal*dx13;
1116             ty               = fscal*dy13;
1117             tz               = fscal*dz13;
1118
1119             /* Update vectorial force */
1120             fix1            += tx;
1121             fiy1            += ty;
1122             fiz1            += tz;
1123             f[j_coord_offset+DIM*3+XX] -= tx;
1124             f[j_coord_offset+DIM*3+YY] -= ty;
1125             f[j_coord_offset+DIM*3+ZZ] -= tz;
1126
1127             }
1128
1129             /**************************
1130              * CALCULATE INTERACTIONS *
1131              **************************/
1132
1133             if (rsq21<rcutoff2)
1134             {
1135
1136             r21              = rsq21*rinv21;
1137
1138             /* EWALD ELECTROSTATICS */
1139
1140             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1141             ewrt             = r21*ewtabscale;
1142             ewitab           = ewrt;
1143             eweps            = ewrt-ewitab;
1144             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1145             felec            = qq21*rinv21*(rinvsq21-felec);
1146
1147             fscal            = felec;
1148
1149             /* Calculate temporary vectorial force */
1150             tx               = fscal*dx21;
1151             ty               = fscal*dy21;
1152             tz               = fscal*dz21;
1153
1154             /* Update vectorial force */
1155             fix2            += tx;
1156             fiy2            += ty;
1157             fiz2            += tz;
1158             f[j_coord_offset+DIM*1+XX] -= tx;
1159             f[j_coord_offset+DIM*1+YY] -= ty;
1160             f[j_coord_offset+DIM*1+ZZ] -= tz;
1161
1162             }
1163
1164             /**************************
1165              * CALCULATE INTERACTIONS *
1166              **************************/
1167
1168             if (rsq22<rcutoff2)
1169             {
1170
1171             r22              = rsq22*rinv22;
1172
1173             /* EWALD ELECTROSTATICS */
1174
1175             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1176             ewrt             = r22*ewtabscale;
1177             ewitab           = ewrt;
1178             eweps            = ewrt-ewitab;
1179             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1180             felec            = qq22*rinv22*(rinvsq22-felec);
1181
1182             fscal            = felec;
1183
1184             /* Calculate temporary vectorial force */
1185             tx               = fscal*dx22;
1186             ty               = fscal*dy22;
1187             tz               = fscal*dz22;
1188
1189             /* Update vectorial force */
1190             fix2            += tx;
1191             fiy2            += ty;
1192             fiz2            += tz;
1193             f[j_coord_offset+DIM*2+XX] -= tx;
1194             f[j_coord_offset+DIM*2+YY] -= ty;
1195             f[j_coord_offset+DIM*2+ZZ] -= tz;
1196
1197             }
1198
1199             /**************************
1200              * CALCULATE INTERACTIONS *
1201              **************************/
1202
1203             if (rsq23<rcutoff2)
1204             {
1205
1206             r23              = rsq23*rinv23;
1207
1208             /* EWALD ELECTROSTATICS */
1209
1210             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1211             ewrt             = r23*ewtabscale;
1212             ewitab           = ewrt;
1213             eweps            = ewrt-ewitab;
1214             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1215             felec            = qq23*rinv23*(rinvsq23-felec);
1216
1217             fscal            = felec;
1218
1219             /* Calculate temporary vectorial force */
1220             tx               = fscal*dx23;
1221             ty               = fscal*dy23;
1222             tz               = fscal*dz23;
1223
1224             /* Update vectorial force */
1225             fix2            += tx;
1226             fiy2            += ty;
1227             fiz2            += tz;
1228             f[j_coord_offset+DIM*3+XX] -= tx;
1229             f[j_coord_offset+DIM*3+YY] -= ty;
1230             f[j_coord_offset+DIM*3+ZZ] -= tz;
1231
1232             }
1233
1234             /**************************
1235              * CALCULATE INTERACTIONS *
1236              **************************/
1237
1238             if (rsq31<rcutoff2)
1239             {
1240
1241             r31              = rsq31*rinv31;
1242
1243             /* EWALD ELECTROSTATICS */
1244
1245             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1246             ewrt             = r31*ewtabscale;
1247             ewitab           = ewrt;
1248             eweps            = ewrt-ewitab;
1249             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1250             felec            = qq31*rinv31*(rinvsq31-felec);
1251
1252             fscal            = felec;
1253
1254             /* Calculate temporary vectorial force */
1255             tx               = fscal*dx31;
1256             ty               = fscal*dy31;
1257             tz               = fscal*dz31;
1258
1259             /* Update vectorial force */
1260             fix3            += tx;
1261             fiy3            += ty;
1262             fiz3            += tz;
1263             f[j_coord_offset+DIM*1+XX] -= tx;
1264             f[j_coord_offset+DIM*1+YY] -= ty;
1265             f[j_coord_offset+DIM*1+ZZ] -= tz;
1266
1267             }
1268
1269             /**************************
1270              * CALCULATE INTERACTIONS *
1271              **************************/
1272
1273             if (rsq32<rcutoff2)
1274             {
1275
1276             r32              = rsq32*rinv32;
1277
1278             /* EWALD ELECTROSTATICS */
1279
1280             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1281             ewrt             = r32*ewtabscale;
1282             ewitab           = ewrt;
1283             eweps            = ewrt-ewitab;
1284             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1285             felec            = qq32*rinv32*(rinvsq32-felec);
1286
1287             fscal            = felec;
1288
1289             /* Calculate temporary vectorial force */
1290             tx               = fscal*dx32;
1291             ty               = fscal*dy32;
1292             tz               = fscal*dz32;
1293
1294             /* Update vectorial force */
1295             fix3            += tx;
1296             fiy3            += ty;
1297             fiz3            += tz;
1298             f[j_coord_offset+DIM*2+XX] -= tx;
1299             f[j_coord_offset+DIM*2+YY] -= ty;
1300             f[j_coord_offset+DIM*2+ZZ] -= tz;
1301
1302             }
1303
1304             /**************************
1305              * CALCULATE INTERACTIONS *
1306              **************************/
1307
1308             if (rsq33<rcutoff2)
1309             {
1310
1311             r33              = rsq33*rinv33;
1312
1313             /* EWALD ELECTROSTATICS */
1314
1315             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1316             ewrt             = r33*ewtabscale;
1317             ewitab           = ewrt;
1318             eweps            = ewrt-ewitab;
1319             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1320             felec            = qq33*rinv33*(rinvsq33-felec);
1321
1322             fscal            = felec;
1323
1324             /* Calculate temporary vectorial force */
1325             tx               = fscal*dx33;
1326             ty               = fscal*dy33;
1327             tz               = fscal*dz33;
1328
1329             /* Update vectorial force */
1330             fix3            += tx;
1331             fiy3            += ty;
1332             fiz3            += tz;
1333             f[j_coord_offset+DIM*3+XX] -= tx;
1334             f[j_coord_offset+DIM*3+YY] -= ty;
1335             f[j_coord_offset+DIM*3+ZZ] -= tz;
1336
1337             }
1338
1339             /* Inner loop uses 355 flops */
1340         }
1341         /* End of innermost loop */
1342
1343         tx = ty = tz = 0;
1344         f[i_coord_offset+DIM*0+XX] += fix0;
1345         f[i_coord_offset+DIM*0+YY] += fiy0;
1346         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1347         tx                         += fix0;
1348         ty                         += fiy0;
1349         tz                         += fiz0;
1350         f[i_coord_offset+DIM*1+XX] += fix1;
1351         f[i_coord_offset+DIM*1+YY] += fiy1;
1352         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1353         tx                         += fix1;
1354         ty                         += fiy1;
1355         tz                         += fiz1;
1356         f[i_coord_offset+DIM*2+XX] += fix2;
1357         f[i_coord_offset+DIM*2+YY] += fiy2;
1358         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1359         tx                         += fix2;
1360         ty                         += fiy2;
1361         tz                         += fiz2;
1362         f[i_coord_offset+DIM*3+XX] += fix3;
1363         f[i_coord_offset+DIM*3+YY] += fiy3;
1364         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1365         tx                         += fix3;
1366         ty                         += fiy3;
1367         tz                         += fiz3;
1368         fshift[i_shift_offset+XX]  += tx;
1369         fshift[i_shift_offset+YY]  += ty;
1370         fshift[i_shift_offset+ZZ]  += tz;
1371
1372         /* Increment number of inner iterations */
1373         inneriter                  += j_index_end - j_index_start;
1374
1375         /* Outer loop uses 39 flops */
1376     }
1377
1378     /* Increment number of outer iterations */
1379     outeriter        += nri;
1380
1381     /* Update outer/inner flops */
1382
1383     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*355);
1384 }