src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_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_GeomW3W3_VF_c
  49  * Electrostatics interaction: Ewald
  50  * VdW interaction:            LennardJones
  51  * Geometry:                   Water3-Water3
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_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              vdwjidx0;
  77     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  78     int              vdwjidx1;
  79     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  80     int              vdwjidx2;
  81     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  82     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  83     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
  84     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
  85     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  86     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  87     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  88     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  89     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  90     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  91     real             velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     int              ewitab;
  98     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
  99     real             *ewtab;
 100
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = fr->epsfac;
 113     charge           = mdatoms->chargeA;
 114     nvdwtype         = fr->ntype;
 115     vdwparam         = fr->nbfp;
 116     vdwtype          = mdatoms->typeA;
 117
 118     sh_ewald         = fr->ic->sh_ewald;
 119     ewtab            = fr->ic->tabq_coul_FDV0;
 120     ewtabscale       = fr->ic->tabq_scale;
 121     ewtabhalfspace   = 0.5/ewtabscale;
 122
 123     /* Setup water-specific parameters */
 124     inr              = nlist->iinr[0];
 125     iq0              = facel*charge[inr+0];
 126     iq1              = facel*charge[inr+1];
 127     iq2              = facel*charge[inr+2];
 128     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 129
 130     jq0              = charge[inr+0];
 131     jq1              = charge[inr+1];
 132     jq2              = charge[inr+2];
 133     vdwjidx0         = 2*vdwtype[inr+0];
 134     qq00             = iq0*jq0;
 135     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 136     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 137     qq01             = iq0*jq1;
 138     qq02             = iq0*jq2;
 139     qq10             = iq1*jq0;
 140     qq11             = iq1*jq1;
 141     qq12             = iq1*jq2;
 142     qq20             = iq2*jq0;
 143     qq21             = iq2*jq1;
 144     qq22             = iq2*jq2;
 145
 146     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 147     rcutoff          = fr->rcoulomb;
 148     rcutoff2         = rcutoff*rcutoff;
 149
 150     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 151     rvdw             = fr->rvdw;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161         shX              = shiftvec[i_shift_offset+XX];
 162         shY              = shiftvec[i_shift_offset+YY];
 163         shZ              = shiftvec[i_shift_offset+ZZ];
 164
 165         /* Load limits for loop over neighbors */
 166         j_index_start    = jindex[iidx];
 167         j_index_end      = jindex[iidx+1];
 168
 169         /* Get outer coordinate index */
 170         inr              = iinr[iidx];
 171         i_coord_offset   = DIM*inr;
 172
 173         /* Load i particle coords and add shift vector */
 174         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 175         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 176         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 177         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 178         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 179         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 180         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 181         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 182         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 183
 184         fix0             = 0.0;
 185         fiy0             = 0.0;
 186         fiz0             = 0.0;
 187         fix1             = 0.0;
 188         fiy1             = 0.0;
 189         fiz1             = 0.0;
 190         fix2             = 0.0;
 191         fiy2             = 0.0;
 192         fiz2             = 0.0;
 193
 194         /* Reset potential sums */
 195         velecsum         = 0.0;
 196         vvdwsum          = 0.0;
 197
 198         /* Start inner kernel loop */
 199         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 200         {
 201             /* Get j neighbor index, and coordinate index */
 202             jnr              = jjnr[jidx];
 203             j_coord_offset   = DIM*jnr;
 204
 205             /* load j atom coordinates */
 206             jx0              = x[j_coord_offset+DIM*0+XX];
 207             jy0              = x[j_coord_offset+DIM*0+YY];
 208             jz0              = x[j_coord_offset+DIM*0+ZZ];
 209             jx1              = x[j_coord_offset+DIM*1+XX];
 210             jy1              = x[j_coord_offset+DIM*1+YY];
 211             jz1              = x[j_coord_offset+DIM*1+ZZ];
 212             jx2              = x[j_coord_offset+DIM*2+XX];
 213             jy2              = x[j_coord_offset+DIM*2+YY];
 214             jz2              = x[j_coord_offset+DIM*2+ZZ];
 215
 216             /* Calculate displacement vector */
 217             dx00             = ix0 - jx0;
 218             dy00             = iy0 - jy0;
 219             dz00             = iz0 - jz0;
 220             dx01             = ix0 - jx1;
 221             dy01             = iy0 - jy1;
 222             dz01             = iz0 - jz1;
 223             dx02             = ix0 - jx2;
 224             dy02             = iy0 - jy2;
 225             dz02             = iz0 - jz2;
 226             dx10             = ix1 - jx0;
 227             dy10             = iy1 - jy0;
 228             dz10             = iz1 - jz0;
 229             dx11             = ix1 - jx1;
 230             dy11             = iy1 - jy1;
 231             dz11             = iz1 - jz1;
 232             dx12             = ix1 - jx2;
 233             dy12             = iy1 - jy2;
 234             dz12             = iz1 - jz2;
 235             dx20             = ix2 - jx0;
 236             dy20             = iy2 - jy0;
 237             dz20             = iz2 - jz0;
 238             dx21             = ix2 - jx1;
 239             dy21             = iy2 - jy1;
 240             dz21             = iz2 - jz1;
 241             dx22             = ix2 - jx2;
 242             dy22             = iy2 - jy2;
 243             dz22             = iz2 - jz2;
 244
 245             /* Calculate squared distance and things based on it */
 246             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 247             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 248             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 249             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 250             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 251             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 252             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 253             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 254             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 255
 256             rinv00           = gmx_invsqrt(rsq00);
 257             rinv01           = gmx_invsqrt(rsq01);
 258             rinv02           = gmx_invsqrt(rsq02);
 259             rinv10           = gmx_invsqrt(rsq10);
 260             rinv11           = gmx_invsqrt(rsq11);
 261             rinv12           = gmx_invsqrt(rsq12);
 262             rinv20           = gmx_invsqrt(rsq20);
 263             rinv21           = gmx_invsqrt(rsq21);
 264             rinv22           = gmx_invsqrt(rsq22);
 265
 266             rinvsq00         = rinv00*rinv00;
 267             rinvsq01         = rinv01*rinv01;
 268             rinvsq02         = rinv02*rinv02;
 269             rinvsq10         = rinv10*rinv10;
 270             rinvsq11         = rinv11*rinv11;
 271             rinvsq12         = rinv12*rinv12;
 272             rinvsq20         = rinv20*rinv20;
 273             rinvsq21         = rinv21*rinv21;
 274             rinvsq22         = rinv22*rinv22;
 275
 276             /**************************
 277              * CALCULATE INTERACTIONS *
 278              **************************/
 279
 280             if (rsq00<rcutoff2)
 281             {
 282
 283             r00              = rsq00*rinv00;
 284
 285             /* EWALD ELECTROSTATICS */
 286
 287             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 288             ewrt             = r00*ewtabscale;
 289             ewitab           = ewrt;
 290             eweps            = ewrt-ewitab;
 291             ewitab           = 4*ewitab;
 292             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 293             velec            = qq00*((rinv00-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 294             felec            = qq00*rinv00*(rinvsq00-felec);
 295
 296             /* LENNARD-JONES DISPERSION/REPULSION */
 297
 298             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 299             vvdw6            = c6_00*rinvsix;
 300             vvdw12           = c12_00*rinvsix*rinvsix;
 301             vvdw             = (vvdw12 - c12_00*sh_vdw_invrcut6*sh_vdw_invrcut6)*(1.0/12.0) - (vvdw6 - c6_00*sh_vdw_invrcut6)*(1.0/6.0);
 302             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 303
 304             /* Update potential sums from outer loop */
 305             velecsum        += velec;
 306             vvdwsum         += vvdw;
 307
 308             fscal            = felec+fvdw;
 309
 310             /* Calculate temporary vectorial force */
 311             tx               = fscal*dx00;
 312             ty               = fscal*dy00;
 313             tz               = fscal*dz00;
 314
 315             /* Update vectorial force */
 316             fix0            += tx;
 317             fiy0            += ty;
 318             fiz0            += tz;
 319             f[j_coord_offset+DIM*0+XX] -= tx;
 320             f[j_coord_offset+DIM*0+YY] -= ty;
 321             f[j_coord_offset+DIM*0+ZZ] -= tz;
 322
 323             }
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             if (rsq01<rcutoff2)
 330             {
 331
 332             r01              = rsq01*rinv01;
 333
 334             /* EWALD ELECTROSTATICS */
 335
 336             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 337             ewrt             = r01*ewtabscale;
 338             ewitab           = ewrt;
 339             eweps            = ewrt-ewitab;
 340             ewitab           = 4*ewitab;
 341             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 342             velec            = qq01*((rinv01-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 343             felec            = qq01*rinv01*(rinvsq01-felec);
 344
 345             /* Update potential sums from outer loop */
 346             velecsum        += velec;
 347
 348             fscal            = felec;
 349
 350             /* Calculate temporary vectorial force */
 351             tx               = fscal*dx01;
 352             ty               = fscal*dy01;
 353             tz               = fscal*dz01;
 354
 355             /* Update vectorial force */
 356             fix0            += tx;
 357             fiy0            += ty;
 358             fiz0            += tz;
 359             f[j_coord_offset+DIM*1+XX] -= tx;
 360             f[j_coord_offset+DIM*1+YY] -= ty;
 361             f[j_coord_offset+DIM*1+ZZ] -= tz;
 362
 363             }
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             if (rsq02<rcutoff2)
 370             {
 371
 372             r02              = rsq02*rinv02;
 373
 374             /* EWALD ELECTROSTATICS */
 375
 376             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 377             ewrt             = r02*ewtabscale;
 378             ewitab           = ewrt;
 379             eweps            = ewrt-ewitab;
 380             ewitab           = 4*ewitab;
 381             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 382             velec            = qq02*((rinv02-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 383             felec            = qq02*rinv02*(rinvsq02-felec);
 384
 385             /* Update potential sums from outer loop */
 386             velecsum        += velec;
 387
 388             fscal            = felec;
 389
 390             /* Calculate temporary vectorial force */
 391             tx               = fscal*dx02;
 392             ty               = fscal*dy02;
 393             tz               = fscal*dz02;
 394
 395             /* Update vectorial force */
 396             fix0            += tx;
 397             fiy0            += ty;
 398             fiz0            += tz;
 399             f[j_coord_offset+DIM*2+XX] -= tx;
 400             f[j_coord_offset+DIM*2+YY] -= ty;
 401             f[j_coord_offset+DIM*2+ZZ] -= tz;
 402
 403             }
 404
 405             /**************************
 406              * CALCULATE INTERACTIONS *
 407              **************************/
 408
 409             if (rsq10<rcutoff2)
 410             {
 411
 412             r10              = rsq10*rinv10;
 413
 414             /* EWALD ELECTROSTATICS */
 415
 416             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 417             ewrt             = r10*ewtabscale;
 418             ewitab           = ewrt;
 419             eweps            = ewrt-ewitab;
 420             ewitab           = 4*ewitab;
 421             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 422             velec            = qq10*((rinv10-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 423             felec            = qq10*rinv10*(rinvsq10-felec);
 424
 425             /* Update potential sums from outer loop */
 426             velecsum        += velec;
 427
 428             fscal            = felec;
 429
 430             /* Calculate temporary vectorial force */
 431             tx               = fscal*dx10;
 432             ty               = fscal*dy10;
 433             tz               = fscal*dz10;
 434
 435             /* Update vectorial force */
 436             fix1            += tx;
 437             fiy1            += ty;
 438             fiz1            += tz;
 439             f[j_coord_offset+DIM*0+XX] -= tx;
 440             f[j_coord_offset+DIM*0+YY] -= ty;
 441             f[j_coord_offset+DIM*0+ZZ] -= tz;
 442
 443             }
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             if (rsq11<rcutoff2)
 450             {
 451
 452             r11              = rsq11*rinv11;
 453
 454             /* EWALD ELECTROSTATICS */
 455
 456             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 457             ewrt             = r11*ewtabscale;
 458             ewitab           = ewrt;
 459             eweps            = ewrt-ewitab;
 460             ewitab           = 4*ewitab;
 461             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 462             velec            = qq11*((rinv11-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 463             felec            = qq11*rinv11*(rinvsq11-felec);
 464
 465             /* Update potential sums from outer loop */
 466             velecsum        += velec;
 467
 468             fscal            = felec;
 469
 470             /* Calculate temporary vectorial force */
 471             tx               = fscal*dx11;
 472             ty               = fscal*dy11;
 473             tz               = fscal*dz11;
 474
 475             /* Update vectorial force */
 476             fix1            += tx;
 477             fiy1            += ty;
 478             fiz1            += tz;
 479             f[j_coord_offset+DIM*1+XX] -= tx;
 480             f[j_coord_offset+DIM*1+YY] -= ty;
 481             f[j_coord_offset+DIM*1+ZZ] -= tz;
 482
 483             }
 484
 485             /**************************
 486              * CALCULATE INTERACTIONS *
 487              **************************/
 488
 489             if (rsq12<rcutoff2)
 490             {
 491
 492             r12              = rsq12*rinv12;
 493
 494             /* EWALD ELECTROSTATICS */
 495
 496             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 497             ewrt             = r12*ewtabscale;
 498             ewitab           = ewrt;
 499             eweps            = ewrt-ewitab;
 500             ewitab           = 4*ewitab;
 501             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 502             velec            = qq12*((rinv12-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 503             felec            = qq12*rinv12*(rinvsq12-felec);
 504
 505             /* Update potential sums from outer loop */
 506             velecsum        += velec;
 507
 508             fscal            = felec;
 509
 510             /* Calculate temporary vectorial force */
 511             tx               = fscal*dx12;
 512             ty               = fscal*dy12;
 513             tz               = fscal*dz12;
 514
 515             /* Update vectorial force */
 516             fix1            += tx;
 517             fiy1            += ty;
 518             fiz1            += tz;
 519             f[j_coord_offset+DIM*2+XX] -= tx;
 520             f[j_coord_offset+DIM*2+YY] -= ty;
 521             f[j_coord_offset+DIM*2+ZZ] -= tz;
 522
 523             }
 524
 525             /**************************
 526              * CALCULATE INTERACTIONS *
 527              **************************/
 528
 529             if (rsq20<rcutoff2)
 530             {
 531
 532             r20              = rsq20*rinv20;
 533
 534             /* EWALD ELECTROSTATICS */
 535
 536             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 537             ewrt             = r20*ewtabscale;
 538             ewitab           = ewrt;
 539             eweps            = ewrt-ewitab;
 540             ewitab           = 4*ewitab;
 541             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 542             velec            = qq20*((rinv20-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 543             felec            = qq20*rinv20*(rinvsq20-felec);
 544
 545             /* Update potential sums from outer loop */
 546             velecsum        += velec;
 547
 548             fscal            = felec;
 549
 550             /* Calculate temporary vectorial force */
 551             tx               = fscal*dx20;
 552             ty               = fscal*dy20;
 553             tz               = fscal*dz20;
 554
 555             /* Update vectorial force */
 556             fix2            += tx;
 557             fiy2            += ty;
 558             fiz2            += tz;
 559             f[j_coord_offset+DIM*0+XX] -= tx;
 560             f[j_coord_offset+DIM*0+YY] -= ty;
 561             f[j_coord_offset+DIM*0+ZZ] -= tz;
 562
 563             }
 564
 565             /**************************
 566              * CALCULATE INTERACTIONS *
 567              **************************/
 568
 569             if (rsq21<rcutoff2)
 570             {
 571
 572             r21              = rsq21*rinv21;
 573
 574             /* EWALD ELECTROSTATICS */
 575
 576             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 577             ewrt             = r21*ewtabscale;
 578             ewitab           = ewrt;
 579             eweps            = ewrt-ewitab;
 580             ewitab           = 4*ewitab;
 581             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 582             velec            = qq21*((rinv21-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 583             felec            = qq21*rinv21*(rinvsq21-felec);
 584
 585             /* Update potential sums from outer loop */
 586             velecsum        += velec;
 587
 588             fscal            = felec;
 589
 590             /* Calculate temporary vectorial force */
 591             tx               = fscal*dx21;
 592             ty               = fscal*dy21;
 593             tz               = fscal*dz21;
 594
 595             /* Update vectorial force */
 596             fix2            += tx;
 597             fiy2            += ty;
 598             fiz2            += tz;
 599             f[j_coord_offset+DIM*1+XX] -= tx;
 600             f[j_coord_offset+DIM*1+YY] -= ty;
 601             f[j_coord_offset+DIM*1+ZZ] -= tz;
 602
 603             }
 604
 605             /**************************
 606              * CALCULATE INTERACTIONS *
 607              **************************/
 608
 609             if (rsq22<rcutoff2)
 610             {
 611
 612             r22              = rsq22*rinv22;
 613
 614             /* EWALD ELECTROSTATICS */
 615
 616             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 617             ewrt             = r22*ewtabscale;
 618             ewitab           = ewrt;
 619             eweps            = ewrt-ewitab;
 620             ewitab           = 4*ewitab;
 621             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 622             velec            = qq22*((rinv22-sh_ewald)-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 623             felec            = qq22*rinv22*(rinvsq22-felec);
 624
 625             /* Update potential sums from outer loop */
 626             velecsum        += velec;
 627
 628             fscal            = felec;
 629
 630             /* Calculate temporary vectorial force */
 631             tx               = fscal*dx22;
 632             ty               = fscal*dy22;
 633             tz               = fscal*dz22;
 634
 635             /* Update vectorial force */
 636             fix2            += tx;
 637             fiy2            += ty;
 638             fiz2            += tz;
 639             f[j_coord_offset+DIM*2+XX] -= tx;
 640             f[j_coord_offset+DIM*2+YY] -= ty;
 641             f[j_coord_offset+DIM*2+ZZ] -= tz;
 642
 643             }
 644
 645             /* Inner loop uses 386 flops */
 646         }
 647         /* End of innermost loop */
 648
 649         tx = ty = tz = 0;
 650         f[i_coord_offset+DIM*0+XX] += fix0;
 651         f[i_coord_offset+DIM*0+YY] += fiy0;
 652         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 653         tx                         += fix0;
 654         ty                         += fiy0;
 655         tz                         += fiz0;
 656         f[i_coord_offset+DIM*1+XX] += fix1;
 657         f[i_coord_offset+DIM*1+YY] += fiy1;
 658         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 659         tx                         += fix1;
 660         ty                         += fiy1;
 661         tz                         += fiz1;
 662         f[i_coord_offset+DIM*2+XX] += fix2;
 663         f[i_coord_offset+DIM*2+YY] += fiy2;
 664         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 665         tx                         += fix2;
 666         ty                         += fiy2;
 667         tz                         += fiz2;
 668         fshift[i_shift_offset+XX]  += tx;
 669         fshift[i_shift_offset+YY]  += ty;
 670         fshift[i_shift_offset+ZZ]  += tz;
 671
 672         ggid                        = gid[iidx];
 673         /* Update potential energies */
 674         kernel_data->energygrp_elec[ggid] += velecsum;
 675         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 676
 677         /* Increment number of inner iterations */
 678         inneriter                  += j_index_end - j_index_start;
 679
 680         /* Outer loop uses 32 flops */
 681     }
 682
 683     /* Increment number of outer iterations */
 684     outeriter        += nri;
 685
 686     /* Update outer/inner flops */
 687
 688     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*386);
 689 }
 690 /*
 691  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_F_c
 692  * Electrostatics interaction: Ewald
 693  * VdW interaction:            LennardJones
 694  * Geometry:                   Water3-Water3
 695  * Calculate force/pot:        Force
 696  */
 697 void
 698 nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_F_c
 699                     (t_nblist                    * gmx_restrict       nlist,
 700                      rvec                        * gmx_restrict          xx,
 701                      rvec                        * gmx_restrict          ff,
 702                      t_forcerec                  * gmx_restrict          fr,
 703                      t_mdatoms                   * gmx_restrict     mdatoms,
 704                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 705                      t_nrnb                      * gmx_restrict        nrnb)
 706 {
 707     int              i_shift_offset,i_coord_offset,j_coord_offset;
 708     int              j_index_start,j_index_end;
 709     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 710     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 711     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 712     real             *shiftvec,*fshift,*x,*f;
 713     int              vdwioffset0;
 714     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 715     int              vdwioffset1;
 716     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 717     int              vdwioffset2;
 718     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 719     int              vdwjidx0;
 720     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 721     int              vdwjidx1;
 722     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 723     int              vdwjidx2;
 724     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 725     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 726     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 727     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 728     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 729     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 730     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 731     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 732     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 733     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 734     real             velec,felec,velecsum,facel,crf,krf,krf2;
 735     real             *charge;
 736     int              nvdwtype;
 737     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 738     int              *vdwtype;
 739     real             *vdwparam;
 740     int              ewitab;
 741     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 742     real             *ewtab;
 743
 744     x                = xx[0];
 745     f                = ff[0];
 746
 747     nri              = nlist->nri;
 748     iinr             = nlist->iinr;
 749     jindex           = nlist->jindex;
 750     jjnr             = nlist->jjnr;
 751     shiftidx         = nlist->shift;
 752     gid              = nlist->gid;
 753     shiftvec         = fr->shift_vec[0];
 754     fshift           = fr->fshift[0];
 755     facel            = fr->epsfac;
 756     charge           = mdatoms->chargeA;
 757     nvdwtype         = fr->ntype;
 758     vdwparam         = fr->nbfp;
 759     vdwtype          = mdatoms->typeA;
 760
 761     sh_ewald         = fr->ic->sh_ewald;
 762     ewtab            = fr->ic->tabq_coul_F;
 763     ewtabscale       = fr->ic->tabq_scale;
 764     ewtabhalfspace   = 0.5/ewtabscale;
 765
 766     /* Setup water-specific parameters */
 767     inr              = nlist->iinr[0];
 768     iq0              = facel*charge[inr+0];
 769     iq1              = facel*charge[inr+1];
 770     iq2              = facel*charge[inr+2];
 771     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 772
 773     jq0              = charge[inr+0];
 774     jq1              = charge[inr+1];
 775     jq2              = charge[inr+2];
 776     vdwjidx0         = 2*vdwtype[inr+0];
 777     qq00             = iq0*jq0;
 778     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 779     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 780     qq01             = iq0*jq1;
 781     qq02             = iq0*jq2;
 782     qq10             = iq1*jq0;
 783     qq11             = iq1*jq1;
 784     qq12             = iq1*jq2;
 785     qq20             = iq2*jq0;
 786     qq21             = iq2*jq1;
 787     qq22             = iq2*jq2;
 788
 789     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 790     rcutoff          = fr->rcoulomb;
 791     rcutoff2         = rcutoff*rcutoff;
 792
 793     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 794     rvdw             = fr->rvdw;
 795
 796     outeriter        = 0;
 797     inneriter        = 0;
 798
 799     /* Start outer loop over neighborlists */
 800     for(iidx=0; iidx<nri; iidx++)
 801     {
 802         /* Load shift vector for this list */
 803         i_shift_offset   = DIM*shiftidx[iidx];
 804         shX              = shiftvec[i_shift_offset+XX];
 805         shY              = shiftvec[i_shift_offset+YY];
 806         shZ              = shiftvec[i_shift_offset+ZZ];
 807
 808         /* Load limits for loop over neighbors */
 809         j_index_start    = jindex[iidx];
 810         j_index_end      = jindex[iidx+1];
 811
 812         /* Get outer coordinate index */
 813         inr              = iinr[iidx];
 814         i_coord_offset   = DIM*inr;
 815
 816         /* Load i particle coords and add shift vector */
 817         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 818         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 819         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 820         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 821         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 822         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 823         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 824         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 825         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 826
 827         fix0             = 0.0;
 828         fiy0             = 0.0;
 829         fiz0             = 0.0;
 830         fix1             = 0.0;
 831         fiy1             = 0.0;
 832         fiz1             = 0.0;
 833         fix2             = 0.0;
 834         fiy2             = 0.0;
 835         fiz2             = 0.0;
 836
 837         /* Start inner kernel loop */
 838         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 839         {
 840             /* Get j neighbor index, and coordinate index */
 841             jnr              = jjnr[jidx];
 842             j_coord_offset   = DIM*jnr;
 843
 844             /* load j atom coordinates */
 845             jx0              = x[j_coord_offset+DIM*0+XX];
 846             jy0              = x[j_coord_offset+DIM*0+YY];
 847             jz0              = x[j_coord_offset+DIM*0+ZZ];
 848             jx1              = x[j_coord_offset+DIM*1+XX];
 849             jy1              = x[j_coord_offset+DIM*1+YY];
 850             jz1              = x[j_coord_offset+DIM*1+ZZ];
 851             jx2              = x[j_coord_offset+DIM*2+XX];
 852             jy2              = x[j_coord_offset+DIM*2+YY];
 853             jz2              = x[j_coord_offset+DIM*2+ZZ];
 854
 855             /* Calculate displacement vector */
 856             dx00             = ix0 - jx0;
 857             dy00             = iy0 - jy0;
 858             dz00             = iz0 - jz0;
 859             dx01             = ix0 - jx1;
 860             dy01             = iy0 - jy1;
 861             dz01             = iz0 - jz1;
 862             dx02             = ix0 - jx2;
 863             dy02             = iy0 - jy2;
 864             dz02             = iz0 - jz2;
 865             dx10             = ix1 - jx0;
 866             dy10             = iy1 - jy0;
 867             dz10             = iz1 - jz0;
 868             dx11             = ix1 - jx1;
 869             dy11             = iy1 - jy1;
 870             dz11             = iz1 - jz1;
 871             dx12             = ix1 - jx2;
 872             dy12             = iy1 - jy2;
 873             dz12             = iz1 - jz2;
 874             dx20             = ix2 - jx0;
 875             dy20             = iy2 - jy0;
 876             dz20             = iz2 - jz0;
 877             dx21             = ix2 - jx1;
 878             dy21             = iy2 - jy1;
 879             dz21             = iz2 - jz1;
 880             dx22             = ix2 - jx2;
 881             dy22             = iy2 - jy2;
 882             dz22             = iz2 - jz2;
 883
 884             /* Calculate squared distance and things based on it */
 885             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 886             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 887             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 888             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 889             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 890             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 891             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 892             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 893             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 894
 895             rinv00           = gmx_invsqrt(rsq00);
 896             rinv01           = gmx_invsqrt(rsq01);
 897             rinv02           = gmx_invsqrt(rsq02);
 898             rinv10           = gmx_invsqrt(rsq10);
 899             rinv11           = gmx_invsqrt(rsq11);
 900             rinv12           = gmx_invsqrt(rsq12);
 901             rinv20           = gmx_invsqrt(rsq20);
 902             rinv21           = gmx_invsqrt(rsq21);
 903             rinv22           = gmx_invsqrt(rsq22);
 904
 905             rinvsq00         = rinv00*rinv00;
 906             rinvsq01         = rinv01*rinv01;
 907             rinvsq02         = rinv02*rinv02;
 908             rinvsq10         = rinv10*rinv10;
 909             rinvsq11         = rinv11*rinv11;
 910             rinvsq12         = rinv12*rinv12;
 911             rinvsq20         = rinv20*rinv20;
 912             rinvsq21         = rinv21*rinv21;
 913             rinvsq22         = rinv22*rinv22;
 914
 915             /**************************
 916              * CALCULATE INTERACTIONS *
 917              **************************/
 918
 919             if (rsq00<rcutoff2)
 920             {
 921
 922             r00              = rsq00*rinv00;
 923
 924             /* EWALD ELECTROSTATICS */
 925
 926             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 927             ewrt             = r00*ewtabscale;
 928             ewitab           = ewrt;
 929             eweps            = ewrt-ewitab;
 930             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
 931             felec            = qq00*rinv00*(rinvsq00-felec);
 932
 933             /* LENNARD-JONES DISPERSION/REPULSION */
 934
 935             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 936             fvdw             = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00;
 937
 938             fscal            = felec+fvdw;
 939
 940             /* Calculate temporary vectorial force */
 941             tx               = fscal*dx00;
 942             ty               = fscal*dy00;
 943             tz               = fscal*dz00;
 944
 945             /* Update vectorial force */
 946             fix0            += tx;
 947             fiy0            += ty;
 948             fiz0            += tz;
 949             f[j_coord_offset+DIM*0+XX] -= tx;
 950             f[j_coord_offset+DIM*0+YY] -= ty;
 951             f[j_coord_offset+DIM*0+ZZ] -= tz;
 952
 953             }
 954
 955             /**************************
 956              * CALCULATE INTERACTIONS *
 957              **************************/
 958
 959             if (rsq01<rcutoff2)
 960             {
 961
 962             r01              = rsq01*rinv01;
 963
 964             /* EWALD ELECTROSTATICS */
 965
 966             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 967             ewrt             = r01*ewtabscale;
 968             ewitab           = ewrt;
 969             eweps            = ewrt-ewitab;
 970             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
 971             felec            = qq01*rinv01*(rinvsq01-felec);
 972
 973             fscal            = felec;
 974
 975             /* Calculate temporary vectorial force */
 976             tx               = fscal*dx01;
 977             ty               = fscal*dy01;
 978             tz               = fscal*dz01;
 979
 980             /* Update vectorial force */
 981             fix0            += tx;
 982             fiy0            += ty;
 983             fiz0            += tz;
 984             f[j_coord_offset+DIM*1+XX] -= tx;
 985             f[j_coord_offset+DIM*1+YY] -= ty;
 986             f[j_coord_offset+DIM*1+ZZ] -= tz;
 987
 988             }
 989
 990             /**************************
 991              * CALCULATE INTERACTIONS *
 992              **************************/
 993
 994             if (rsq02<rcutoff2)
 995             {
 996
 997             r02              = rsq02*rinv02;
 998
 999             /* EWALD ELECTROSTATICS */
1000
1001             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1002             ewrt             = r02*ewtabscale;
1003             ewitab           = ewrt;
1004             eweps            = ewrt-ewitab;
1005             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1006             felec            = qq02*rinv02*(rinvsq02-felec);
1007
1008             fscal            = felec;
1009
1010             /* Calculate temporary vectorial force */
1011             tx               = fscal*dx02;
1012             ty               = fscal*dy02;
1013             tz               = fscal*dz02;
1014
1015             /* Update vectorial force */
1016             fix0            += tx;
1017             fiy0            += ty;
1018             fiz0            += tz;
1019             f[j_coord_offset+DIM*2+XX] -= tx;
1020             f[j_coord_offset+DIM*2+YY] -= ty;
1021             f[j_coord_offset+DIM*2+ZZ] -= tz;
1022
1023             }
1024
1025             /**************************
1026              * CALCULATE INTERACTIONS *
1027              **************************/
1028
1029             if (rsq10<rcutoff2)
1030             {
1031
1032             r10              = rsq10*rinv10;
1033
1034             /* EWALD ELECTROSTATICS */
1035
1036             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1037             ewrt             = r10*ewtabscale;
1038             ewitab           = ewrt;
1039             eweps            = ewrt-ewitab;
1040             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1041             felec            = qq10*rinv10*(rinvsq10-felec);
1042
1043             fscal            = felec;
1044
1045             /* Calculate temporary vectorial force */
1046             tx               = fscal*dx10;
1047             ty               = fscal*dy10;
1048             tz               = fscal*dz10;
1049
1050             /* Update vectorial force */
1051             fix1            += tx;
1052             fiy1            += ty;
1053             fiz1            += tz;
1054             f[j_coord_offset+DIM*0+XX] -= tx;
1055             f[j_coord_offset+DIM*0+YY] -= ty;
1056             f[j_coord_offset+DIM*0+ZZ] -= tz;
1057
1058             }
1059
1060             /**************************
1061              * CALCULATE INTERACTIONS *
1062              **************************/
1063
1064             if (rsq11<rcutoff2)
1065             {
1066
1067             r11              = rsq11*rinv11;
1068
1069             /* EWALD ELECTROSTATICS */
1070
1071             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1072             ewrt             = r11*ewtabscale;
1073             ewitab           = ewrt;
1074             eweps            = ewrt-ewitab;
1075             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1076             felec            = qq11*rinv11*(rinvsq11-felec);
1077
1078             fscal            = felec;
1079
1080             /* Calculate temporary vectorial force */
1081             tx               = fscal*dx11;
1082             ty               = fscal*dy11;
1083             tz               = fscal*dz11;
1084
1085             /* Update vectorial force */
1086             fix1            += tx;
1087             fiy1            += ty;
1088             fiz1            += tz;
1089             f[j_coord_offset+DIM*1+XX] -= tx;
1090             f[j_coord_offset+DIM*1+YY] -= ty;
1091             f[j_coord_offset+DIM*1+ZZ] -= tz;
1092
1093             }
1094
1095             /**************************
1096              * CALCULATE INTERACTIONS *
1097              **************************/
1098
1099             if (rsq12<rcutoff2)
1100             {
1101
1102             r12              = rsq12*rinv12;
1103
1104             /* EWALD ELECTROSTATICS */
1105
1106             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1107             ewrt             = r12*ewtabscale;
1108             ewitab           = ewrt;
1109             eweps            = ewrt-ewitab;
1110             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1111             felec            = qq12*rinv12*(rinvsq12-felec);
1112
1113             fscal            = felec;
1114
1115             /* Calculate temporary vectorial force */
1116             tx               = fscal*dx12;
1117             ty               = fscal*dy12;
1118             tz               = fscal*dz12;
1119
1120             /* Update vectorial force */
1121             fix1            += tx;
1122             fiy1            += ty;
1123             fiz1            += tz;
1124             f[j_coord_offset+DIM*2+XX] -= tx;
1125             f[j_coord_offset+DIM*2+YY] -= ty;
1126             f[j_coord_offset+DIM*2+ZZ] -= tz;
1127
1128             }
1129
1130             /**************************
1131              * CALCULATE INTERACTIONS *
1132              **************************/
1133
1134             if (rsq20<rcutoff2)
1135             {
1136
1137             r20              = rsq20*rinv20;
1138
1139             /* EWALD ELECTROSTATICS */
1140
1141             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1142             ewrt             = r20*ewtabscale;
1143             ewitab           = ewrt;
1144             eweps            = ewrt-ewitab;
1145             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1146             felec            = qq20*rinv20*(rinvsq20-felec);
1147
1148             fscal            = felec;
1149
1150             /* Calculate temporary vectorial force */
1151             tx               = fscal*dx20;
1152             ty               = fscal*dy20;
1153             tz               = fscal*dz20;
1154
1155             /* Update vectorial force */
1156             fix2            += tx;
1157             fiy2            += ty;
1158             fiz2            += tz;
1159             f[j_coord_offset+DIM*0+XX] -= tx;
1160             f[j_coord_offset+DIM*0+YY] -= ty;
1161             f[j_coord_offset+DIM*0+ZZ] -= tz;
1162
1163             }
1164
1165             /**************************
1166              * CALCULATE INTERACTIONS *
1167              **************************/
1168
1169             if (rsq21<rcutoff2)
1170             {
1171
1172             r21              = rsq21*rinv21;
1173
1174             /* EWALD ELECTROSTATICS */
1175
1176             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1177             ewrt             = r21*ewtabscale;
1178             ewitab           = ewrt;
1179             eweps            = ewrt-ewitab;
1180             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1181             felec            = qq21*rinv21*(rinvsq21-felec);
1182
1183             fscal            = felec;
1184
1185             /* Calculate temporary vectorial force */
1186             tx               = fscal*dx21;
1187             ty               = fscal*dy21;
1188             tz               = fscal*dz21;
1189
1190             /* Update vectorial force */
1191             fix2            += tx;
1192             fiy2            += ty;
1193             fiz2            += tz;
1194             f[j_coord_offset+DIM*1+XX] -= tx;
1195             f[j_coord_offset+DIM*1+YY] -= ty;
1196             f[j_coord_offset+DIM*1+ZZ] -= tz;
1197
1198             }
1199
1200             /**************************
1201              * CALCULATE INTERACTIONS *
1202              **************************/
1203
1204             if (rsq22<rcutoff2)
1205             {
1206
1207             r22              = rsq22*rinv22;
1208
1209             /* EWALD ELECTROSTATICS */
1210
1211             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1212             ewrt             = r22*ewtabscale;
1213             ewitab           = ewrt;
1214             eweps            = ewrt-ewitab;
1215             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1216             felec            = qq22*rinv22*(rinvsq22-felec);
1217
1218             fscal            = felec;
1219
1220             /* Calculate temporary vectorial force */
1221             tx               = fscal*dx22;
1222             ty               = fscal*dy22;
1223             tz               = fscal*dz22;
1224
1225             /* Update vectorial force */
1226             fix2            += tx;
1227             fiy2            += ty;
1228             fiz2            += tz;
1229             f[j_coord_offset+DIM*2+XX] -= tx;
1230             f[j_coord_offset+DIM*2+YY] -= ty;
1231             f[j_coord_offset+DIM*2+ZZ] -= tz;
1232
1233             }
1234
1235             /* Inner loop uses 304 flops */
1236         }
1237         /* End of innermost loop */
1238
1239         tx = ty = tz = 0;
1240         f[i_coord_offset+DIM*0+XX] += fix0;
1241         f[i_coord_offset+DIM*0+YY] += fiy0;
1242         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1243         tx                         += fix0;
1244         ty                         += fiy0;
1245         tz                         += fiz0;
1246         f[i_coord_offset+DIM*1+XX] += fix1;
1247         f[i_coord_offset+DIM*1+YY] += fiy1;
1248         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1249         tx                         += fix1;
1250         ty                         += fiy1;
1251         tz                         += fiz1;
1252         f[i_coord_offset+DIM*2+XX] += fix2;
1253         f[i_coord_offset+DIM*2+YY] += fiy2;
1254         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1255         tx                         += fix2;
1256         ty                         += fiy2;
1257         tz                         += fiz2;
1258         fshift[i_shift_offset+XX]  += tx;
1259         fshift[i_shift_offset+YY]  += ty;
1260         fshift[i_shift_offset+ZZ]  += tz;
1261
1262         /* Increment number of inner iterations */
1263         inneriter                  += j_index_end - j_index_start;
1264
1265         /* Outer loop uses 30 flops */
1266     }
1267
1268     /* Increment number of outer iterations */
1269     outeriter        += nri;
1270
1271     /* Update outer/inner flops */
1272
1273     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*304);
1274 }