src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEw_VdwCSTab_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_ElecEw_VdwCSTab_GeomW3W3_VF_c
  49  * Electrostatics interaction: Ewald
  50  * VdW interaction:            CubicSplineTable
  51  * Geometry:                   Water3-Water3
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecEw_VdwCSTab_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              vfitab;
  98     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
  99     real             *vftab;
 100     int              ewitab;
 101     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 102     real             *ewtab;
 103
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = fr->epsfac;
 116     charge           = mdatoms->chargeA;
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     vftab            = kernel_data->table_vdw->data;
 122     vftabscale       = kernel_data->table_vdw->scale;
 123
 124     sh_ewald         = fr->ic->sh_ewald;
 125     ewtab            = fr->ic->tabq_coul_FDV0;
 126     ewtabscale       = fr->ic->tabq_scale;
 127     ewtabhalfspace   = 0.5/ewtabscale;
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq0              = facel*charge[inr+0];
 132     iq1              = facel*charge[inr+1];
 133     iq2              = facel*charge[inr+2];
 134     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 135
 136     jq0              = charge[inr+0];
 137     jq1              = charge[inr+1];
 138     jq2              = charge[inr+2];
 139     vdwjidx0         = 2*vdwtype[inr+0];
 140     qq00             = iq0*jq0;
 141     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 142     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 143     qq01             = iq0*jq1;
 144     qq02             = iq0*jq2;
 145     qq10             = iq1*jq0;
 146     qq11             = iq1*jq1;
 147     qq12             = iq1*jq2;
 148     qq20             = iq2*jq0;
 149     qq21             = iq2*jq1;
 150     qq22             = iq2*jq2;
 151
 152     outeriter        = 0;
 153     inneriter        = 0;
 154
 155     /* Start outer loop over neighborlists */
 156     for(iidx=0; iidx<nri; iidx++)
 157     {
 158         /* Load shift vector for this list */
 159         i_shift_offset   = DIM*shiftidx[iidx];
 160         shX              = shiftvec[i_shift_offset+XX];
 161         shY              = shiftvec[i_shift_offset+YY];
 162         shZ              = shiftvec[i_shift_offset+ZZ];
 163
 164         /* Load limits for loop over neighbors */
 165         j_index_start    = jindex[iidx];
 166         j_index_end      = jindex[iidx+1];
 167
 168         /* Get outer coordinate index */
 169         inr              = iinr[iidx];
 170         i_coord_offset   = DIM*inr;
 171
 172         /* Load i particle coords and add shift vector */
 173         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 174         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 175         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 176         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 177         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 178         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 179         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 180         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 181         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 182
 183         fix0             = 0.0;
 184         fiy0             = 0.0;
 185         fiz0             = 0.0;
 186         fix1             = 0.0;
 187         fiy1             = 0.0;
 188         fiz1             = 0.0;
 189         fix2             = 0.0;
 190         fiy2             = 0.0;
 191         fiz2             = 0.0;
 192
 193         /* Reset potential sums */
 194         velecsum         = 0.0;
 195         vvdwsum          = 0.0;
 196
 197         /* Start inner kernel loop */
 198         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 199         {
 200             /* Get j neighbor index, and coordinate index */
 201             jnr              = jjnr[jidx];
 202             j_coord_offset   = DIM*jnr;
 203
 204             /* load j atom coordinates */
 205             jx0              = x[j_coord_offset+DIM*0+XX];
 206             jy0              = x[j_coord_offset+DIM*0+YY];
 207             jz0              = x[j_coord_offset+DIM*0+ZZ];
 208             jx1              = x[j_coord_offset+DIM*1+XX];
 209             jy1              = x[j_coord_offset+DIM*1+YY];
 210             jz1              = x[j_coord_offset+DIM*1+ZZ];
 211             jx2              = x[j_coord_offset+DIM*2+XX];
 212             jy2              = x[j_coord_offset+DIM*2+YY];
 213             jz2              = x[j_coord_offset+DIM*2+ZZ];
 214
 215             /* Calculate displacement vector */
 216             dx00             = ix0 - jx0;
 217             dy00             = iy0 - jy0;
 218             dz00             = iz0 - jz0;
 219             dx01             = ix0 - jx1;
 220             dy01             = iy0 - jy1;
 221             dz01             = iz0 - jz1;
 222             dx02             = ix0 - jx2;
 223             dy02             = iy0 - jy2;
 224             dz02             = iz0 - jz2;
 225             dx10             = ix1 - jx0;
 226             dy10             = iy1 - jy0;
 227             dz10             = iz1 - jz0;
 228             dx11             = ix1 - jx1;
 229             dy11             = iy1 - jy1;
 230             dz11             = iz1 - jz1;
 231             dx12             = ix1 - jx2;
 232             dy12             = iy1 - jy2;
 233             dz12             = iz1 - jz2;
 234             dx20             = ix2 - jx0;
 235             dy20             = iy2 - jy0;
 236             dz20             = iz2 - jz0;
 237             dx21             = ix2 - jx1;
 238             dy21             = iy2 - jy1;
 239             dz21             = iz2 - jz1;
 240             dx22             = ix2 - jx2;
 241             dy22             = iy2 - jy2;
 242             dz22             = iz2 - jz2;
 243
 244             /* Calculate squared distance and things based on it */
 245             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 246             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 247             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 248             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 249             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 250             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 251             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 252             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 253             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 254
 255             rinv00           = gmx_invsqrt(rsq00);
 256             rinv01           = gmx_invsqrt(rsq01);
 257             rinv02           = gmx_invsqrt(rsq02);
 258             rinv10           = gmx_invsqrt(rsq10);
 259             rinv11           = gmx_invsqrt(rsq11);
 260             rinv12           = gmx_invsqrt(rsq12);
 261             rinv20           = gmx_invsqrt(rsq20);
 262             rinv21           = gmx_invsqrt(rsq21);
 263             rinv22           = gmx_invsqrt(rsq22);
 264
 265             rinvsq00         = rinv00*rinv00;
 266             rinvsq01         = rinv01*rinv01;
 267             rinvsq02         = rinv02*rinv02;
 268             rinvsq10         = rinv10*rinv10;
 269             rinvsq11         = rinv11*rinv11;
 270             rinvsq12         = rinv12*rinv12;
 271             rinvsq20         = rinv20*rinv20;
 272             rinvsq21         = rinv21*rinv21;
 273             rinvsq22         = rinv22*rinv22;
 274
 275             /**************************
 276              * CALCULATE INTERACTIONS *
 277              **************************/
 278
 279             r00              = rsq00*rinv00;
 280
 281             /* Calculate table index by multiplying r with table scale and truncate to integer */
 282             rt               = r00*vftabscale;
 283             vfitab           = rt;
 284             vfeps            = rt-vfitab;
 285             vfitab           = 2*4*vfitab;
 286
 287             /* EWALD ELECTROSTATICS */
 288
 289             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 290             ewrt             = r00*ewtabscale;
 291             ewitab           = ewrt;
 292             eweps            = ewrt-ewitab;
 293             ewitab           = 4*ewitab;
 294             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 295             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 296             felec            = qq00*rinv00*(rinvsq00-felec);
 297
 298             /* CUBIC SPLINE TABLE DISPERSION */
 299             vfitab          += 0;
 300             Y                = vftab[vfitab];
 301             F                = vftab[vfitab+1];
 302             Geps             = vfeps*vftab[vfitab+2];
 303             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 304             Fp               = F+Geps+Heps2;
 305             VV               = Y+vfeps*Fp;
 306             vvdw6            = c6_00*VV;
 307             FF               = Fp+Geps+2.0*Heps2;
 308             fvdw6            = c6_00*FF;
 309
 310             /* CUBIC SPLINE TABLE REPULSION */
 311             Y                = vftab[vfitab+4];
 312             F                = vftab[vfitab+5];
 313             Geps             = vfeps*vftab[vfitab+6];
 314             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 315             Fp               = F+Geps+Heps2;
 316             VV               = Y+vfeps*Fp;
 317             vvdw12           = c12_00*VV;
 318             FF               = Fp+Geps+2.0*Heps2;
 319             fvdw12           = c12_00*FF;
 320             vvdw             = vvdw12+vvdw6;
 321             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 322
 323             /* Update potential sums from outer loop */
 324             velecsum        += velec;
 325             vvdwsum         += vvdw;
 326
 327             fscal            = felec+fvdw;
 328
 329             /* Calculate temporary vectorial force */
 330             tx               = fscal*dx00;
 331             ty               = fscal*dy00;
 332             tz               = fscal*dz00;
 333
 334             /* Update vectorial force */
 335             fix0            += tx;
 336             fiy0            += ty;
 337             fiz0            += tz;
 338             f[j_coord_offset+DIM*0+XX] -= tx;
 339             f[j_coord_offset+DIM*0+YY] -= ty;
 340             f[j_coord_offset+DIM*0+ZZ] -= tz;
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             r01              = rsq01*rinv01;
 347
 348             /* EWALD ELECTROSTATICS */
 349
 350             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 351             ewrt             = r01*ewtabscale;
 352             ewitab           = ewrt;
 353             eweps            = ewrt-ewitab;
 354             ewitab           = 4*ewitab;
 355             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 356             velec            = qq01*(rinv01-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 357             felec            = qq01*rinv01*(rinvsq01-felec);
 358
 359             /* Update potential sums from outer loop */
 360             velecsum        += velec;
 361
 362             fscal            = felec;
 363
 364             /* Calculate temporary vectorial force */
 365             tx               = fscal*dx01;
 366             ty               = fscal*dy01;
 367             tz               = fscal*dz01;
 368
 369             /* Update vectorial force */
 370             fix0            += tx;
 371             fiy0            += ty;
 372             fiz0            += tz;
 373             f[j_coord_offset+DIM*1+XX] -= tx;
 374             f[j_coord_offset+DIM*1+YY] -= ty;
 375             f[j_coord_offset+DIM*1+ZZ] -= tz;
 376
 377             /**************************
 378              * CALCULATE INTERACTIONS *
 379              **************************/
 380
 381             r02              = rsq02*rinv02;
 382
 383             /* EWALD ELECTROSTATICS */
 384
 385             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 386             ewrt             = r02*ewtabscale;
 387             ewitab           = ewrt;
 388             eweps            = ewrt-ewitab;
 389             ewitab           = 4*ewitab;
 390             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 391             velec            = qq02*(rinv02-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 392             felec            = qq02*rinv02*(rinvsq02-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*dx02;
 401             ty               = fscal*dy02;
 402             tz               = fscal*dz02;
 403
 404             /* Update vectorial force */
 405             fix0            += tx;
 406             fiy0            += ty;
 407             fiz0            += 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              * CALCULATE INTERACTIONS *
 414              **************************/
 415
 416             r10              = rsq10*rinv10;
 417
 418             /* EWALD ELECTROSTATICS */
 419
 420             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 421             ewrt             = r10*ewtabscale;
 422             ewitab           = ewrt;
 423             eweps            = ewrt-ewitab;
 424             ewitab           = 4*ewitab;
 425             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 426             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 427             felec            = qq10*rinv10*(rinvsq10-felec);
 428
 429             /* Update potential sums from outer loop */
 430             velecsum        += velec;
 431
 432             fscal            = felec;
 433
 434             /* Calculate temporary vectorial force */
 435             tx               = fscal*dx10;
 436             ty               = fscal*dy10;
 437             tz               = fscal*dz10;
 438
 439             /* Update vectorial force */
 440             fix1            += tx;
 441             fiy1            += ty;
 442             fiz1            += tz;
 443             f[j_coord_offset+DIM*0+XX] -= tx;
 444             f[j_coord_offset+DIM*0+YY] -= ty;
 445             f[j_coord_offset+DIM*0+ZZ] -= tz;
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             r11              = rsq11*rinv11;
 452
 453             /* EWALD ELECTROSTATICS */
 454
 455             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 456             ewrt             = r11*ewtabscale;
 457             ewitab           = ewrt;
 458             eweps            = ewrt-ewitab;
 459             ewitab           = 4*ewitab;
 460             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 461             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 462             felec            = qq11*rinv11*(rinvsq11-felec);
 463
 464             /* Update potential sums from outer loop */
 465             velecsum        += velec;
 466
 467             fscal            = felec;
 468
 469             /* Calculate temporary vectorial force */
 470             tx               = fscal*dx11;
 471             ty               = fscal*dy11;
 472             tz               = fscal*dz11;
 473
 474             /* Update vectorial force */
 475             fix1            += tx;
 476             fiy1            += ty;
 477             fiz1            += tz;
 478             f[j_coord_offset+DIM*1+XX] -= tx;
 479             f[j_coord_offset+DIM*1+YY] -= ty;
 480             f[j_coord_offset+DIM*1+ZZ] -= tz;
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             r12              = rsq12*rinv12;
 487
 488             /* EWALD ELECTROSTATICS */
 489
 490             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 491             ewrt             = r12*ewtabscale;
 492             ewitab           = ewrt;
 493             eweps            = ewrt-ewitab;
 494             ewitab           = 4*ewitab;
 495             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 496             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 497             felec            = qq12*rinv12*(rinvsq12-felec);
 498
 499             /* Update potential sums from outer loop */
 500             velecsum        += velec;
 501
 502             fscal            = felec;
 503
 504             /* Calculate temporary vectorial force */
 505             tx               = fscal*dx12;
 506             ty               = fscal*dy12;
 507             tz               = fscal*dz12;
 508
 509             /* Update vectorial force */
 510             fix1            += tx;
 511             fiy1            += ty;
 512             fiz1            += tz;
 513             f[j_coord_offset+DIM*2+XX] -= tx;
 514             f[j_coord_offset+DIM*2+YY] -= ty;
 515             f[j_coord_offset+DIM*2+ZZ] -= tz;
 516
 517             /**************************
 518              * CALCULATE INTERACTIONS *
 519              **************************/
 520
 521             r20              = rsq20*rinv20;
 522
 523             /* EWALD ELECTROSTATICS */
 524
 525             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 526             ewrt             = r20*ewtabscale;
 527             ewitab           = ewrt;
 528             eweps            = ewrt-ewitab;
 529             ewitab           = 4*ewitab;
 530             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 531             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 532             felec            = qq20*rinv20*(rinvsq20-felec);
 533
 534             /* Update potential sums from outer loop */
 535             velecsum        += velec;
 536
 537             fscal            = felec;
 538
 539             /* Calculate temporary vectorial force */
 540             tx               = fscal*dx20;
 541             ty               = fscal*dy20;
 542             tz               = fscal*dz20;
 543
 544             /* Update vectorial force */
 545             fix2            += tx;
 546             fiy2            += ty;
 547             fiz2            += tz;
 548             f[j_coord_offset+DIM*0+XX] -= tx;
 549             f[j_coord_offset+DIM*0+YY] -= ty;
 550             f[j_coord_offset+DIM*0+ZZ] -= tz;
 551
 552             /**************************
 553              * CALCULATE INTERACTIONS *
 554              **************************/
 555
 556             r21              = rsq21*rinv21;
 557
 558             /* EWALD ELECTROSTATICS */
 559
 560             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 561             ewrt             = r21*ewtabscale;
 562             ewitab           = ewrt;
 563             eweps            = ewrt-ewitab;
 564             ewitab           = 4*ewitab;
 565             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 566             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 567             felec            = qq21*rinv21*(rinvsq21-felec);
 568
 569             /* Update potential sums from outer loop */
 570             velecsum        += velec;
 571
 572             fscal            = felec;
 573
 574             /* Calculate temporary vectorial force */
 575             tx               = fscal*dx21;
 576             ty               = fscal*dy21;
 577             tz               = fscal*dz21;
 578
 579             /* Update vectorial force */
 580             fix2            += tx;
 581             fiy2            += ty;
 582             fiz2            += tz;
 583             f[j_coord_offset+DIM*1+XX] -= tx;
 584             f[j_coord_offset+DIM*1+YY] -= ty;
 585             f[j_coord_offset+DIM*1+ZZ] -= tz;
 586
 587             /**************************
 588              * CALCULATE INTERACTIONS *
 589              **************************/
 590
 591             r22              = rsq22*rinv22;
 592
 593             /* EWALD ELECTROSTATICS */
 594
 595             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 596             ewrt             = r22*ewtabscale;
 597             ewitab           = ewrt;
 598             eweps            = ewrt-ewitab;
 599             ewitab           = 4*ewitab;
 600             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 601             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 602             felec            = qq22*rinv22*(rinvsq22-felec);
 603
 604             /* Update potential sums from outer loop */
 605             velecsum        += velec;
 606
 607             fscal            = felec;
 608
 609             /* Calculate temporary vectorial force */
 610             tx               = fscal*dx22;
 611             ty               = fscal*dy22;
 612             tz               = fscal*dz22;
 613
 614             /* Update vectorial force */
 615             fix2            += tx;
 616             fiy2            += ty;
 617             fiz2            += tz;
 618             f[j_coord_offset+DIM*2+XX] -= tx;
 619             f[j_coord_offset+DIM*2+YY] -= ty;
 620             f[j_coord_offset+DIM*2+ZZ] -= tz;
 621
 622             /* Inner loop uses 393 flops */
 623         }
 624         /* End of innermost loop */
 625
 626         tx = ty = tz = 0;
 627         f[i_coord_offset+DIM*0+XX] += fix0;
 628         f[i_coord_offset+DIM*0+YY] += fiy0;
 629         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 630         tx                         += fix0;
 631         ty                         += fiy0;
 632         tz                         += fiz0;
 633         f[i_coord_offset+DIM*1+XX] += fix1;
 634         f[i_coord_offset+DIM*1+YY] += fiy1;
 635         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 636         tx                         += fix1;
 637         ty                         += fiy1;
 638         tz                         += fiz1;
 639         f[i_coord_offset+DIM*2+XX] += fix2;
 640         f[i_coord_offset+DIM*2+YY] += fiy2;
 641         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 642         tx                         += fix2;
 643         ty                         += fiy2;
 644         tz                         += fiz2;
 645         fshift[i_shift_offset+XX]  += tx;
 646         fshift[i_shift_offset+YY]  += ty;
 647         fshift[i_shift_offset+ZZ]  += tz;
 648
 649         ggid                        = gid[iidx];
 650         /* Update potential energies */
 651         kernel_data->energygrp_elec[ggid] += velecsum;
 652         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 653
 654         /* Increment number of inner iterations */
 655         inneriter                  += j_index_end - j_index_start;
 656
 657         /* Outer loop uses 32 flops */
 658     }
 659
 660     /* Increment number of outer iterations */
 661     outeriter        += nri;
 662
 663     /* Update outer/inner flops */
 664
 665     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*393);
 666 }
 667 /*
 668  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW3W3_F_c
 669  * Electrostatics interaction: Ewald
 670  * VdW interaction:            CubicSplineTable
 671  * Geometry:                   Water3-Water3
 672  * Calculate force/pot:        Force
 673  */
 674 void
 675 nb_kernel_ElecEw_VdwCSTab_GeomW3W3_F_c
 676                     (t_nblist                    * gmx_restrict       nlist,
 677                      rvec                        * gmx_restrict          xx,
 678                      rvec                        * gmx_restrict          ff,
 679                      t_forcerec                  * gmx_restrict          fr,
 680                      t_mdatoms                   * gmx_restrict     mdatoms,
 681                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 682                      t_nrnb                      * gmx_restrict        nrnb)
 683 {
 684     int              i_shift_offset,i_coord_offset,j_coord_offset;
 685     int              j_index_start,j_index_end;
 686     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 687     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 688     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 689     real             *shiftvec,*fshift,*x,*f;
 690     int              vdwioffset0;
 691     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 692     int              vdwioffset1;
 693     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 694     int              vdwioffset2;
 695     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 696     int              vdwjidx0;
 697     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 698     int              vdwjidx1;
 699     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 700     int              vdwjidx2;
 701     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 702     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 703     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 704     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 705     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 706     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 707     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 708     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 709     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 710     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 711     real             velec,felec,velecsum,facel,crf,krf,krf2;
 712     real             *charge;
 713     int              nvdwtype;
 714     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 715     int              *vdwtype;
 716     real             *vdwparam;
 717     int              vfitab;
 718     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 719     real             *vftab;
 720     int              ewitab;
 721     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 722     real             *ewtab;
 723
 724     x                = xx[0];
 725     f                = ff[0];
 726
 727     nri              = nlist->nri;
 728     iinr             = nlist->iinr;
 729     jindex           = nlist->jindex;
 730     jjnr             = nlist->jjnr;
 731     shiftidx         = nlist->shift;
 732     gid              = nlist->gid;
 733     shiftvec         = fr->shift_vec[0];
 734     fshift           = fr->fshift[0];
 735     facel            = fr->epsfac;
 736     charge           = mdatoms->chargeA;
 737     nvdwtype         = fr->ntype;
 738     vdwparam         = fr->nbfp;
 739     vdwtype          = mdatoms->typeA;
 740
 741     vftab            = kernel_data->table_vdw->data;
 742     vftabscale       = kernel_data->table_vdw->scale;
 743
 744     sh_ewald         = fr->ic->sh_ewald;
 745     ewtab            = fr->ic->tabq_coul_F;
 746     ewtabscale       = fr->ic->tabq_scale;
 747     ewtabhalfspace   = 0.5/ewtabscale;
 748
 749     /* Setup water-specific parameters */
 750     inr              = nlist->iinr[0];
 751     iq0              = facel*charge[inr+0];
 752     iq1              = facel*charge[inr+1];
 753     iq2              = facel*charge[inr+2];
 754     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 755
 756     jq0              = charge[inr+0];
 757     jq1              = charge[inr+1];
 758     jq2              = charge[inr+2];
 759     vdwjidx0         = 2*vdwtype[inr+0];
 760     qq00             = iq0*jq0;
 761     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 762     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 763     qq01             = iq0*jq1;
 764     qq02             = iq0*jq2;
 765     qq10             = iq1*jq0;
 766     qq11             = iq1*jq1;
 767     qq12             = iq1*jq2;
 768     qq20             = iq2*jq0;
 769     qq21             = iq2*jq1;
 770     qq22             = iq2*jq2;
 771
 772     outeriter        = 0;
 773     inneriter        = 0;
 774
 775     /* Start outer loop over neighborlists */
 776     for(iidx=0; iidx<nri; iidx++)
 777     {
 778         /* Load shift vector for this list */
 779         i_shift_offset   = DIM*shiftidx[iidx];
 780         shX              = shiftvec[i_shift_offset+XX];
 781         shY              = shiftvec[i_shift_offset+YY];
 782         shZ              = shiftvec[i_shift_offset+ZZ];
 783
 784         /* Load limits for loop over neighbors */
 785         j_index_start    = jindex[iidx];
 786         j_index_end      = jindex[iidx+1];
 787
 788         /* Get outer coordinate index */
 789         inr              = iinr[iidx];
 790         i_coord_offset   = DIM*inr;
 791
 792         /* Load i particle coords and add shift vector */
 793         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 794         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 795         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 796         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 797         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 798         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 799         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 800         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 801         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 802
 803         fix0             = 0.0;
 804         fiy0             = 0.0;
 805         fiz0             = 0.0;
 806         fix1             = 0.0;
 807         fiy1             = 0.0;
 808         fiz1             = 0.0;
 809         fix2             = 0.0;
 810         fiy2             = 0.0;
 811         fiz2             = 0.0;
 812
 813         /* Start inner kernel loop */
 814         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 815         {
 816             /* Get j neighbor index, and coordinate index */
 817             jnr              = jjnr[jidx];
 818             j_coord_offset   = DIM*jnr;
 819
 820             /* load j atom coordinates */
 821             jx0              = x[j_coord_offset+DIM*0+XX];
 822             jy0              = x[j_coord_offset+DIM*0+YY];
 823             jz0              = x[j_coord_offset+DIM*0+ZZ];
 824             jx1              = x[j_coord_offset+DIM*1+XX];
 825             jy1              = x[j_coord_offset+DIM*1+YY];
 826             jz1              = x[j_coord_offset+DIM*1+ZZ];
 827             jx2              = x[j_coord_offset+DIM*2+XX];
 828             jy2              = x[j_coord_offset+DIM*2+YY];
 829             jz2              = x[j_coord_offset+DIM*2+ZZ];
 830
 831             /* Calculate displacement vector */
 832             dx00             = ix0 - jx0;
 833             dy00             = iy0 - jy0;
 834             dz00             = iz0 - jz0;
 835             dx01             = ix0 - jx1;
 836             dy01             = iy0 - jy1;
 837             dz01             = iz0 - jz1;
 838             dx02             = ix0 - jx2;
 839             dy02             = iy0 - jy2;
 840             dz02             = iz0 - jz2;
 841             dx10             = ix1 - jx0;
 842             dy10             = iy1 - jy0;
 843             dz10             = iz1 - jz0;
 844             dx11             = ix1 - jx1;
 845             dy11             = iy1 - jy1;
 846             dz11             = iz1 - jz1;
 847             dx12             = ix1 - jx2;
 848             dy12             = iy1 - jy2;
 849             dz12             = iz1 - jz2;
 850             dx20             = ix2 - jx0;
 851             dy20             = iy2 - jy0;
 852             dz20             = iz2 - jz0;
 853             dx21             = ix2 - jx1;
 854             dy21             = iy2 - jy1;
 855             dz21             = iz2 - jz1;
 856             dx22             = ix2 - jx2;
 857             dy22             = iy2 - jy2;
 858             dz22             = iz2 - jz2;
 859
 860             /* Calculate squared distance and things based on it */
 861             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 862             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 863             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 864             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 865             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 866             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 867             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 868             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 869             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 870
 871             rinv00           = gmx_invsqrt(rsq00);
 872             rinv01           = gmx_invsqrt(rsq01);
 873             rinv02           = gmx_invsqrt(rsq02);
 874             rinv10           = gmx_invsqrt(rsq10);
 875             rinv11           = gmx_invsqrt(rsq11);
 876             rinv12           = gmx_invsqrt(rsq12);
 877             rinv20           = gmx_invsqrt(rsq20);
 878             rinv21           = gmx_invsqrt(rsq21);
 879             rinv22           = gmx_invsqrt(rsq22);
 880
 881             rinvsq00         = rinv00*rinv00;
 882             rinvsq01         = rinv01*rinv01;
 883             rinvsq02         = rinv02*rinv02;
 884             rinvsq10         = rinv10*rinv10;
 885             rinvsq11         = rinv11*rinv11;
 886             rinvsq12         = rinv12*rinv12;
 887             rinvsq20         = rinv20*rinv20;
 888             rinvsq21         = rinv21*rinv21;
 889             rinvsq22         = rinv22*rinv22;
 890
 891             /**************************
 892              * CALCULATE INTERACTIONS *
 893              **************************/
 894
 895             r00              = rsq00*rinv00;
 896
 897             /* Calculate table index by multiplying r with table scale and truncate to integer */
 898             rt               = r00*vftabscale;
 899             vfitab           = rt;
 900             vfeps            = rt-vfitab;
 901             vfitab           = 2*4*vfitab;
 902
 903             /* EWALD ELECTROSTATICS */
 904
 905             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 906             ewrt             = r00*ewtabscale;
 907             ewitab           = ewrt;
 908             eweps            = ewrt-ewitab;
 909             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
 910             felec            = qq00*rinv00*(rinvsq00-felec);
 911
 912             /* CUBIC SPLINE TABLE DISPERSION */
 913             vfitab          += 0;
 914             F                = vftab[vfitab+1];
 915             Geps             = vfeps*vftab[vfitab+2];
 916             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 917             Fp               = F+Geps+Heps2;
 918             FF               = Fp+Geps+2.0*Heps2;
 919             fvdw6            = c6_00*FF;
 920
 921             /* CUBIC SPLINE TABLE REPULSION */
 922             F                = vftab[vfitab+5];
 923             Geps             = vfeps*vftab[vfitab+6];
 924             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 925             Fp               = F+Geps+Heps2;
 926             FF               = Fp+Geps+2.0*Heps2;
 927             fvdw12           = c12_00*FF;
 928             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 929
 930             fscal            = felec+fvdw;
 931
 932             /* Calculate temporary vectorial force */
 933             tx               = fscal*dx00;
 934             ty               = fscal*dy00;
 935             tz               = fscal*dz00;
 936
 937             /* Update vectorial force */
 938             fix0            += tx;
 939             fiy0            += ty;
 940             fiz0            += tz;
 941             f[j_coord_offset+DIM*0+XX] -= tx;
 942             f[j_coord_offset+DIM*0+YY] -= ty;
 943             f[j_coord_offset+DIM*0+ZZ] -= tz;
 944
 945             /**************************
 946              * CALCULATE INTERACTIONS *
 947              **************************/
 948
 949             r01              = rsq01*rinv01;
 950
 951             /* EWALD ELECTROSTATICS */
 952
 953             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 954             ewrt             = r01*ewtabscale;
 955             ewitab           = ewrt;
 956             eweps            = ewrt-ewitab;
 957             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
 958             felec            = qq01*rinv01*(rinvsq01-felec);
 959
 960             fscal            = felec;
 961
 962             /* Calculate temporary vectorial force */
 963             tx               = fscal*dx01;
 964             ty               = fscal*dy01;
 965             tz               = fscal*dz01;
 966
 967             /* Update vectorial force */
 968             fix0            += tx;
 969             fiy0            += ty;
 970             fiz0            += tz;
 971             f[j_coord_offset+DIM*1+XX] -= tx;
 972             f[j_coord_offset+DIM*1+YY] -= ty;
 973             f[j_coord_offset+DIM*1+ZZ] -= tz;
 974
 975             /**************************
 976              * CALCULATE INTERACTIONS *
 977              **************************/
 978
 979             r02              = rsq02*rinv02;
 980
 981             /* EWALD ELECTROSTATICS */
 982
 983             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 984             ewrt             = r02*ewtabscale;
 985             ewitab           = ewrt;
 986             eweps            = ewrt-ewitab;
 987             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
 988             felec            = qq02*rinv02*(rinvsq02-felec);
 989
 990             fscal            = felec;
 991
 992             /* Calculate temporary vectorial force */
 993             tx               = fscal*dx02;
 994             ty               = fscal*dy02;
 995             tz               = fscal*dz02;
 996
 997             /* Update vectorial force */
 998             fix0            += tx;
 999             fiy0            += ty;
1000             fiz0            += tz;
1001             f[j_coord_offset+DIM*2+XX] -= tx;
1002             f[j_coord_offset+DIM*2+YY] -= ty;
1003             f[j_coord_offset+DIM*2+ZZ] -= tz;
1004
1005             /**************************
1006              * CALCULATE INTERACTIONS *
1007              **************************/
1008
1009             r10              = rsq10*rinv10;
1010
1011             /* EWALD ELECTROSTATICS */
1012
1013             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1014             ewrt             = r10*ewtabscale;
1015             ewitab           = ewrt;
1016             eweps            = ewrt-ewitab;
1017             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1018             felec            = qq10*rinv10*(rinvsq10-felec);
1019
1020             fscal            = felec;
1021
1022             /* Calculate temporary vectorial force */
1023             tx               = fscal*dx10;
1024             ty               = fscal*dy10;
1025             tz               = fscal*dz10;
1026
1027             /* Update vectorial force */
1028             fix1            += tx;
1029             fiy1            += ty;
1030             fiz1            += tz;
1031             f[j_coord_offset+DIM*0+XX] -= tx;
1032             f[j_coord_offset+DIM*0+YY] -= ty;
1033             f[j_coord_offset+DIM*0+ZZ] -= tz;
1034
1035             /**************************
1036              * CALCULATE INTERACTIONS *
1037              **************************/
1038
1039             r11              = rsq11*rinv11;
1040
1041             /* EWALD ELECTROSTATICS */
1042
1043             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1044             ewrt             = r11*ewtabscale;
1045             ewitab           = ewrt;
1046             eweps            = ewrt-ewitab;
1047             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1048             felec            = qq11*rinv11*(rinvsq11-felec);
1049
1050             fscal            = felec;
1051
1052             /* Calculate temporary vectorial force */
1053             tx               = fscal*dx11;
1054             ty               = fscal*dy11;
1055             tz               = fscal*dz11;
1056
1057             /* Update vectorial force */
1058             fix1            += tx;
1059             fiy1            += ty;
1060             fiz1            += tz;
1061             f[j_coord_offset+DIM*1+XX] -= tx;
1062             f[j_coord_offset+DIM*1+YY] -= ty;
1063             f[j_coord_offset+DIM*1+ZZ] -= tz;
1064
1065             /**************************
1066              * CALCULATE INTERACTIONS *
1067              **************************/
1068
1069             r12              = rsq12*rinv12;
1070
1071             /* EWALD ELECTROSTATICS */
1072
1073             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1074             ewrt             = r12*ewtabscale;
1075             ewitab           = ewrt;
1076             eweps            = ewrt-ewitab;
1077             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1078             felec            = qq12*rinv12*(rinvsq12-felec);
1079
1080             fscal            = felec;
1081
1082             /* Calculate temporary vectorial force */
1083             tx               = fscal*dx12;
1084             ty               = fscal*dy12;
1085             tz               = fscal*dz12;
1086
1087             /* Update vectorial force */
1088             fix1            += tx;
1089             fiy1            += ty;
1090             fiz1            += tz;
1091             f[j_coord_offset+DIM*2+XX] -= tx;
1092             f[j_coord_offset+DIM*2+YY] -= ty;
1093             f[j_coord_offset+DIM*2+ZZ] -= tz;
1094
1095             /**************************
1096              * CALCULATE INTERACTIONS *
1097              **************************/
1098
1099             r20              = rsq20*rinv20;
1100
1101             /* EWALD ELECTROSTATICS */
1102
1103             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1104             ewrt             = r20*ewtabscale;
1105             ewitab           = ewrt;
1106             eweps            = ewrt-ewitab;
1107             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1108             felec            = qq20*rinv20*(rinvsq20-felec);
1109
1110             fscal            = felec;
1111
1112             /* Calculate temporary vectorial force */
1113             tx               = fscal*dx20;
1114             ty               = fscal*dy20;
1115             tz               = fscal*dz20;
1116
1117             /* Update vectorial force */
1118             fix2            += tx;
1119             fiy2            += ty;
1120             fiz2            += tz;
1121             f[j_coord_offset+DIM*0+XX] -= tx;
1122             f[j_coord_offset+DIM*0+YY] -= ty;
1123             f[j_coord_offset+DIM*0+ZZ] -= tz;
1124
1125             /**************************
1126              * CALCULATE INTERACTIONS *
1127              **************************/
1128
1129             r21              = rsq21*rinv21;
1130
1131             /* EWALD ELECTROSTATICS */
1132
1133             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1134             ewrt             = r21*ewtabscale;
1135             ewitab           = ewrt;
1136             eweps            = ewrt-ewitab;
1137             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1138             felec            = qq21*rinv21*(rinvsq21-felec);
1139
1140             fscal            = felec;
1141
1142             /* Calculate temporary vectorial force */
1143             tx               = fscal*dx21;
1144             ty               = fscal*dy21;
1145             tz               = fscal*dz21;
1146
1147             /* Update vectorial force */
1148             fix2            += tx;
1149             fiy2            += ty;
1150             fiz2            += tz;
1151             f[j_coord_offset+DIM*1+XX] -= tx;
1152             f[j_coord_offset+DIM*1+YY] -= ty;
1153             f[j_coord_offset+DIM*1+ZZ] -= tz;
1154
1155             /**************************
1156              * CALCULATE INTERACTIONS *
1157              **************************/
1158
1159             r22              = rsq22*rinv22;
1160
1161             /* EWALD ELECTROSTATICS */
1162
1163             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1164             ewrt             = r22*ewtabscale;
1165             ewitab           = ewrt;
1166             eweps            = ewrt-ewitab;
1167             felec            = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1];
1168             felec            = qq22*rinv22*(rinvsq22-felec);
1169
1170             fscal            = felec;
1171
1172             /* Calculate temporary vectorial force */
1173             tx               = fscal*dx22;
1174             ty               = fscal*dy22;
1175             tz               = fscal*dz22;
1176
1177             /* Update vectorial force */
1178             fix2            += tx;
1179             fiy2            += ty;
1180             fiz2            += tz;
1181             f[j_coord_offset+DIM*2+XX] -= tx;
1182             f[j_coord_offset+DIM*2+YY] -= ty;
1183             f[j_coord_offset+DIM*2+ZZ] -= tz;
1184
1185             /* Inner loop uses 322 flops */
1186         }
1187         /* End of innermost loop */
1188
1189         tx = ty = tz = 0;
1190         f[i_coord_offset+DIM*0+XX] += fix0;
1191         f[i_coord_offset+DIM*0+YY] += fiy0;
1192         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1193         tx                         += fix0;
1194         ty                         += fiy0;
1195         tz                         += fiz0;
1196         f[i_coord_offset+DIM*1+XX] += fix1;
1197         f[i_coord_offset+DIM*1+YY] += fiy1;
1198         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1199         tx                         += fix1;
1200         ty                         += fiy1;
1201         tz                         += fiz1;
1202         f[i_coord_offset+DIM*2+XX] += fix2;
1203         f[i_coord_offset+DIM*2+YY] += fiy2;
1204         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1205         tx                         += fix2;
1206         ty                         += fiy2;
1207         tz                         += fiz2;
1208         fshift[i_shift_offset+XX]  += tx;
1209         fshift[i_shift_offset+YY]  += ty;
1210         fshift[i_shift_offset+ZZ]  += tz;
1211
1212         /* Increment number of inner iterations */
1213         inneriter                  += j_index_end - j_index_start;
1214
1215         /* Outer loop uses 30 flops */
1216     }
1217
1218     /* Increment number of outer iterations */
1219     outeriter        += nri;
1220
1221     /* Update outer/inner flops */
1222
1223     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*322);
1224 }