src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSw_VdwLJSw_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 "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_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_ElecEwSw_VdwLJSw_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     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 101
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = fr->epsfac;
 114     charge           = mdatoms->chargeA;
 115     nvdwtype         = fr->ntype;
 116     vdwparam         = fr->nbfp;
 117     vdwtype          = mdatoms->typeA;
 118
 119     sh_ewald         = fr->ic->sh_ewald;
 120     ewtab            = fr->ic->tabq_coul_FDV0;
 121     ewtabscale       = fr->ic->tabq_scale;
 122     ewtabhalfspace   = 0.5/ewtabscale;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq0              = facel*charge[inr+0];
 127     iq1              = facel*charge[inr+1];
 128     iq2              = facel*charge[inr+2];
 129     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 130
 131     jq0              = charge[inr+0];
 132     jq1              = charge[inr+1];
 133     jq2              = charge[inr+2];
 134     vdwjidx0         = 2*vdwtype[inr+0];
 135     qq00             = iq0*jq0;
 136     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 137     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 138     qq01             = iq0*jq1;
 139     qq02             = iq0*jq2;
 140     qq10             = iq1*jq0;
 141     qq11             = iq1*jq1;
 142     qq12             = iq1*jq2;
 143     qq20             = iq2*jq0;
 144     qq21             = iq2*jq1;
 145     qq22             = iq2*jq2;
 146
 147     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 148     rcutoff          = fr->rcoulomb;
 149     rcutoff2         = rcutoff*rcutoff;
 150
 151     rswitch          = fr->rcoulomb_switch;
 152     /* Setup switch parameters */
 153     d                = rcutoff-rswitch;
 154     swV3             = -10.0/(d*d*d);
 155     swV4             =  15.0/(d*d*d*d);
 156     swV5             =  -6.0/(d*d*d*d*d);
 157     swF2             = -30.0/(d*d*d);
 158     swF3             =  60.0/(d*d*d*d);
 159     swF4             = -30.0/(d*d*d*d*d);
 160
 161     outeriter        = 0;
 162     inneriter        = 0;
 163
 164     /* Start outer loop over neighborlists */
 165     for(iidx=0; iidx<nri; iidx++)
 166     {
 167         /* Load shift vector for this list */
 168         i_shift_offset   = DIM*shiftidx[iidx];
 169         shX              = shiftvec[i_shift_offset+XX];
 170         shY              = shiftvec[i_shift_offset+YY];
 171         shZ              = shiftvec[i_shift_offset+ZZ];
 172
 173         /* Load limits for loop over neighbors */
 174         j_index_start    = jindex[iidx];
 175         j_index_end      = jindex[iidx+1];
 176
 177         /* Get outer coordinate index */
 178         inr              = iinr[iidx];
 179         i_coord_offset   = DIM*inr;
 180
 181         /* Load i particle coords and add shift vector */
 182         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 183         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 184         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 185         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 186         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 187         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 188         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 189         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 190         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 191
 192         fix0             = 0.0;
 193         fiy0             = 0.0;
 194         fiz0             = 0.0;
 195         fix1             = 0.0;
 196         fiy1             = 0.0;
 197         fiz1             = 0.0;
 198         fix2             = 0.0;
 199         fiy2             = 0.0;
 200         fiz2             = 0.0;
 201
 202         /* Reset potential sums */
 203         velecsum         = 0.0;
 204         vvdwsum          = 0.0;
 205
 206         /* Start inner kernel loop */
 207         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 208         {
 209             /* Get j neighbor index, and coordinate index */
 210             jnr              = jjnr[jidx];
 211             j_coord_offset   = DIM*jnr;
 212
 213             /* load j atom coordinates */
 214             jx0              = x[j_coord_offset+DIM*0+XX];
 215             jy0              = x[j_coord_offset+DIM*0+YY];
 216             jz0              = x[j_coord_offset+DIM*0+ZZ];
 217             jx1              = x[j_coord_offset+DIM*1+XX];
 218             jy1              = x[j_coord_offset+DIM*1+YY];
 219             jz1              = x[j_coord_offset+DIM*1+ZZ];
 220             jx2              = x[j_coord_offset+DIM*2+XX];
 221             jy2              = x[j_coord_offset+DIM*2+YY];
 222             jz2              = x[j_coord_offset+DIM*2+ZZ];
 223
 224             /* Calculate displacement vector */
 225             dx00             = ix0 - jx0;
 226             dy00             = iy0 - jy0;
 227             dz00             = iz0 - jz0;
 228             dx01             = ix0 - jx1;
 229             dy01             = iy0 - jy1;
 230             dz01             = iz0 - jz1;
 231             dx02             = ix0 - jx2;
 232             dy02             = iy0 - jy2;
 233             dz02             = iz0 - jz2;
 234             dx10             = ix1 - jx0;
 235             dy10             = iy1 - jy0;
 236             dz10             = iz1 - jz0;
 237             dx11             = ix1 - jx1;
 238             dy11             = iy1 - jy1;
 239             dz11             = iz1 - jz1;
 240             dx12             = ix1 - jx2;
 241             dy12             = iy1 - jy2;
 242             dz12             = iz1 - jz2;
 243             dx20             = ix2 - jx0;
 244             dy20             = iy2 - jy0;
 245             dz20             = iz2 - jz0;
 246             dx21             = ix2 - jx1;
 247             dy21             = iy2 - jy1;
 248             dz21             = iz2 - jz1;
 249             dx22             = ix2 - jx2;
 250             dy22             = iy2 - jy2;
 251             dz22             = iz2 - jz2;
 252
 253             /* Calculate squared distance and things based on it */
 254             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 255             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 256             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 257             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 258             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 259             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 260             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 261             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 262             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 263
 264             rinv00           = gmx_invsqrt(rsq00);
 265             rinv01           = gmx_invsqrt(rsq01);
 266             rinv02           = gmx_invsqrt(rsq02);
 267             rinv10           = gmx_invsqrt(rsq10);
 268             rinv11           = gmx_invsqrt(rsq11);
 269             rinv12           = gmx_invsqrt(rsq12);
 270             rinv20           = gmx_invsqrt(rsq20);
 271             rinv21           = gmx_invsqrt(rsq21);
 272             rinv22           = gmx_invsqrt(rsq22);
 273
 274             rinvsq00         = rinv00*rinv00;
 275             rinvsq01         = rinv01*rinv01;
 276             rinvsq02         = rinv02*rinv02;
 277             rinvsq10         = rinv10*rinv10;
 278             rinvsq11         = rinv11*rinv11;
 279             rinvsq12         = rinv12*rinv12;
 280             rinvsq20         = rinv20*rinv20;
 281             rinvsq21         = rinv21*rinv21;
 282             rinvsq22         = rinv22*rinv22;
 283
 284             /**************************
 285              * CALCULATE INTERACTIONS *
 286              **************************/
 287
 288             if (rsq00<rcutoff2)
 289             {
 290
 291             r00              = rsq00*rinv00;
 292
 293             /* EWALD ELECTROSTATICS */
 294
 295             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 296             ewrt             = r00*ewtabscale;
 297             ewitab           = ewrt;
 298             eweps            = ewrt-ewitab;
 299             ewitab           = 4*ewitab;
 300             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 301             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 302             felec            = qq00*rinv00*(rinvsq00-felec);
 303
 304             /* LENNARD-JONES DISPERSION/REPULSION */
 305
 306             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 307             vvdw6            = c6_00*rinvsix;
 308             vvdw12           = c12_00*rinvsix*rinvsix;
 309             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 310             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 311
 312             d                = r00-rswitch;
 313             d                = (d>0.0) ? d : 0.0;
 314             d2               = d*d;
 315             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 316
 317             dsw              = d2*(swF2+d*(swF3+d*swF4));
 318
 319             /* Evaluate switch function */
 320             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 321             felec            = felec*sw - rinv00*velec*dsw;
 322             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 323             velec           *= sw;
 324             vvdw            *= sw;
 325
 326             /* Update potential sums from outer loop */
 327             velecsum        += velec;
 328             vvdwsum         += vvdw;
 329
 330             fscal            = felec+fvdw;
 331
 332             /* Calculate temporary vectorial force */
 333             tx               = fscal*dx00;
 334             ty               = fscal*dy00;
 335             tz               = fscal*dz00;
 336
 337             /* Update vectorial force */
 338             fix0            += tx;
 339             fiy0            += ty;
 340             fiz0            += tz;
 341             f[j_coord_offset+DIM*0+XX] -= tx;
 342             f[j_coord_offset+DIM*0+YY] -= ty;
 343             f[j_coord_offset+DIM*0+ZZ] -= tz;
 344
 345             }
 346
 347             /**************************
 348              * CALCULATE INTERACTIONS *
 349              **************************/
 350
 351             if (rsq01<rcutoff2)
 352             {
 353
 354             r01              = rsq01*rinv01;
 355
 356             /* EWALD ELECTROSTATICS */
 357
 358             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 359             ewrt             = r01*ewtabscale;
 360             ewitab           = ewrt;
 361             eweps            = ewrt-ewitab;
 362             ewitab           = 4*ewitab;
 363             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 364             velec            = qq01*(rinv01-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 365             felec            = qq01*rinv01*(rinvsq01-felec);
 366
 367             d                = r01-rswitch;
 368             d                = (d>0.0) ? d : 0.0;
 369             d2               = d*d;
 370             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 371
 372             dsw              = d2*(swF2+d*(swF3+d*swF4));
 373
 374             /* Evaluate switch function */
 375             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 376             felec            = felec*sw - rinv01*velec*dsw;
 377             velec           *= sw;
 378
 379             /* Update potential sums from outer loop */
 380             velecsum        += velec;
 381
 382             fscal            = felec;
 383
 384             /* Calculate temporary vectorial force */
 385             tx               = fscal*dx01;
 386             ty               = fscal*dy01;
 387             tz               = fscal*dz01;
 388
 389             /* Update vectorial force */
 390             fix0            += tx;
 391             fiy0            += ty;
 392             fiz0            += tz;
 393             f[j_coord_offset+DIM*1+XX] -= tx;
 394             f[j_coord_offset+DIM*1+YY] -= ty;
 395             f[j_coord_offset+DIM*1+ZZ] -= tz;
 396
 397             }
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             if (rsq02<rcutoff2)
 404             {
 405
 406             r02              = rsq02*rinv02;
 407
 408             /* EWALD ELECTROSTATICS */
 409
 410             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 411             ewrt             = r02*ewtabscale;
 412             ewitab           = ewrt;
 413             eweps            = ewrt-ewitab;
 414             ewitab           = 4*ewitab;
 415             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 416             velec            = qq02*(rinv02-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 417             felec            = qq02*rinv02*(rinvsq02-felec);
 418
 419             d                = r02-rswitch;
 420             d                = (d>0.0) ? d : 0.0;
 421             d2               = d*d;
 422             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 423
 424             dsw              = d2*(swF2+d*(swF3+d*swF4));
 425
 426             /* Evaluate switch function */
 427             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 428             felec            = felec*sw - rinv02*velec*dsw;
 429             velec           *= sw;
 430
 431             /* Update potential sums from outer loop */
 432             velecsum        += velec;
 433
 434             fscal            = felec;
 435
 436             /* Calculate temporary vectorial force */
 437             tx               = fscal*dx02;
 438             ty               = fscal*dy02;
 439             tz               = fscal*dz02;
 440
 441             /* Update vectorial force */
 442             fix0            += tx;
 443             fiy0            += ty;
 444             fiz0            += tz;
 445             f[j_coord_offset+DIM*2+XX] -= tx;
 446             f[j_coord_offset+DIM*2+YY] -= ty;
 447             f[j_coord_offset+DIM*2+ZZ] -= tz;
 448
 449             }
 450
 451             /**************************
 452              * CALCULATE INTERACTIONS *
 453              **************************/
 454
 455             if (rsq10<rcutoff2)
 456             {
 457
 458             r10              = rsq10*rinv10;
 459
 460             /* EWALD ELECTROSTATICS */
 461
 462             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 463             ewrt             = r10*ewtabscale;
 464             ewitab           = ewrt;
 465             eweps            = ewrt-ewitab;
 466             ewitab           = 4*ewitab;
 467             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 468             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 469             felec            = qq10*rinv10*(rinvsq10-felec);
 470
 471             d                = r10-rswitch;
 472             d                = (d>0.0) ? d : 0.0;
 473             d2               = d*d;
 474             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 475
 476             dsw              = d2*(swF2+d*(swF3+d*swF4));
 477
 478             /* Evaluate switch function */
 479             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 480             felec            = felec*sw - rinv10*velec*dsw;
 481             velec           *= sw;
 482
 483             /* Update potential sums from outer loop */
 484             velecsum        += velec;
 485
 486             fscal            = felec;
 487
 488             /* Calculate temporary vectorial force */
 489             tx               = fscal*dx10;
 490             ty               = fscal*dy10;
 491             tz               = fscal*dz10;
 492
 493             /* Update vectorial force */
 494             fix1            += tx;
 495             fiy1            += ty;
 496             fiz1            += tz;
 497             f[j_coord_offset+DIM*0+XX] -= tx;
 498             f[j_coord_offset+DIM*0+YY] -= ty;
 499             f[j_coord_offset+DIM*0+ZZ] -= tz;
 500
 501             }
 502
 503             /**************************
 504              * CALCULATE INTERACTIONS *
 505              **************************/
 506
 507             if (rsq11<rcutoff2)
 508             {
 509
 510             r11              = rsq11*rinv11;
 511
 512             /* EWALD ELECTROSTATICS */
 513
 514             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 515             ewrt             = r11*ewtabscale;
 516             ewitab           = ewrt;
 517             eweps            = ewrt-ewitab;
 518             ewitab           = 4*ewitab;
 519             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 520             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 521             felec            = qq11*rinv11*(rinvsq11-felec);
 522
 523             d                = r11-rswitch;
 524             d                = (d>0.0) ? d : 0.0;
 525             d2               = d*d;
 526             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 527
 528             dsw              = d2*(swF2+d*(swF3+d*swF4));
 529
 530             /* Evaluate switch function */
 531             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 532             felec            = felec*sw - rinv11*velec*dsw;
 533             velec           *= sw;
 534
 535             /* Update potential sums from outer loop */
 536             velecsum        += velec;
 537
 538             fscal            = felec;
 539
 540             /* Calculate temporary vectorial force */
 541             tx               = fscal*dx11;
 542             ty               = fscal*dy11;
 543             tz               = fscal*dz11;
 544
 545             /* Update vectorial force */
 546             fix1            += tx;
 547             fiy1            += ty;
 548             fiz1            += tz;
 549             f[j_coord_offset+DIM*1+XX] -= tx;
 550             f[j_coord_offset+DIM*1+YY] -= ty;
 551             f[j_coord_offset+DIM*1+ZZ] -= tz;
 552
 553             }
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             if (rsq12<rcutoff2)
 560             {
 561
 562             r12              = rsq12*rinv12;
 563
 564             /* EWALD ELECTROSTATICS */
 565
 566             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 567             ewrt             = r12*ewtabscale;
 568             ewitab           = ewrt;
 569             eweps            = ewrt-ewitab;
 570             ewitab           = 4*ewitab;
 571             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 572             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 573             felec            = qq12*rinv12*(rinvsq12-felec);
 574
 575             d                = r12-rswitch;
 576             d                = (d>0.0) ? d : 0.0;
 577             d2               = d*d;
 578             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 579
 580             dsw              = d2*(swF2+d*(swF3+d*swF4));
 581
 582             /* Evaluate switch function */
 583             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 584             felec            = felec*sw - rinv12*velec*dsw;
 585             velec           *= sw;
 586
 587             /* Update potential sums from outer loop */
 588             velecsum        += velec;
 589
 590             fscal            = felec;
 591
 592             /* Calculate temporary vectorial force */
 593             tx               = fscal*dx12;
 594             ty               = fscal*dy12;
 595             tz               = fscal*dz12;
 596
 597             /* Update vectorial force */
 598             fix1            += tx;
 599             fiy1            += ty;
 600             fiz1            += tz;
 601             f[j_coord_offset+DIM*2+XX] -= tx;
 602             f[j_coord_offset+DIM*2+YY] -= ty;
 603             f[j_coord_offset+DIM*2+ZZ] -= tz;
 604
 605             }
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             if (rsq20<rcutoff2)
 612             {
 613
 614             r20              = rsq20*rinv20;
 615
 616             /* EWALD ELECTROSTATICS */
 617
 618             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 619             ewrt             = r20*ewtabscale;
 620             ewitab           = ewrt;
 621             eweps            = ewrt-ewitab;
 622             ewitab           = 4*ewitab;
 623             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 624             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 625             felec            = qq20*rinv20*(rinvsq20-felec);
 626
 627             d                = r20-rswitch;
 628             d                = (d>0.0) ? d : 0.0;
 629             d2               = d*d;
 630             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 631
 632             dsw              = d2*(swF2+d*(swF3+d*swF4));
 633
 634             /* Evaluate switch function */
 635             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 636             felec            = felec*sw - rinv20*velec*dsw;
 637             velec           *= sw;
 638
 639             /* Update potential sums from outer loop */
 640             velecsum        += velec;
 641
 642             fscal            = felec;
 643
 644             /* Calculate temporary vectorial force */
 645             tx               = fscal*dx20;
 646             ty               = fscal*dy20;
 647             tz               = fscal*dz20;
 648
 649             /* Update vectorial force */
 650             fix2            += tx;
 651             fiy2            += ty;
 652             fiz2            += tz;
 653             f[j_coord_offset+DIM*0+XX] -= tx;
 654             f[j_coord_offset+DIM*0+YY] -= ty;
 655             f[j_coord_offset+DIM*0+ZZ] -= tz;
 656
 657             }
 658
 659             /**************************
 660              * CALCULATE INTERACTIONS *
 661              **************************/
 662
 663             if (rsq21<rcutoff2)
 664             {
 665
 666             r21              = rsq21*rinv21;
 667
 668             /* EWALD ELECTROSTATICS */
 669
 670             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 671             ewrt             = r21*ewtabscale;
 672             ewitab           = ewrt;
 673             eweps            = ewrt-ewitab;
 674             ewitab           = 4*ewitab;
 675             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 676             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 677             felec            = qq21*rinv21*(rinvsq21-felec);
 678
 679             d                = r21-rswitch;
 680             d                = (d>0.0) ? d : 0.0;
 681             d2               = d*d;
 682             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 683
 684             dsw              = d2*(swF2+d*(swF3+d*swF4));
 685
 686             /* Evaluate switch function */
 687             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 688             felec            = felec*sw - rinv21*velec*dsw;
 689             velec           *= sw;
 690
 691             /* Update potential sums from outer loop */
 692             velecsum        += velec;
 693
 694             fscal            = felec;
 695
 696             /* Calculate temporary vectorial force */
 697             tx               = fscal*dx21;
 698             ty               = fscal*dy21;
 699             tz               = fscal*dz21;
 700
 701             /* Update vectorial force */
 702             fix2            += tx;
 703             fiy2            += ty;
 704             fiz2            += tz;
 705             f[j_coord_offset+DIM*1+XX] -= tx;
 706             f[j_coord_offset+DIM*1+YY] -= ty;
 707             f[j_coord_offset+DIM*1+ZZ] -= tz;
 708
 709             }
 710
 711             /**************************
 712              * CALCULATE INTERACTIONS *
 713              **************************/
 714
 715             if (rsq22<rcutoff2)
 716             {
 717
 718             r22              = rsq22*rinv22;
 719
 720             /* EWALD ELECTROSTATICS */
 721
 722             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 723             ewrt             = r22*ewtabscale;
 724             ewitab           = ewrt;
 725             eweps            = ewrt-ewitab;
 726             ewitab           = 4*ewitab;
 727             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 728             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 729             felec            = qq22*rinv22*(rinvsq22-felec);
 730
 731             d                = r22-rswitch;
 732             d                = (d>0.0) ? d : 0.0;
 733             d2               = d*d;
 734             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 735
 736             dsw              = d2*(swF2+d*(swF3+d*swF4));
 737
 738             /* Evaluate switch function */
 739             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 740             felec            = felec*sw - rinv22*velec*dsw;
 741             velec           *= sw;
 742
 743             /* Update potential sums from outer loop */
 744             velecsum        += velec;
 745
 746             fscal            = felec;
 747
 748             /* Calculate temporary vectorial force */
 749             tx               = fscal*dx22;
 750             ty               = fscal*dy22;
 751             tz               = fscal*dz22;
 752
 753             /* Update vectorial force */
 754             fix2            += tx;
 755             fiy2            += ty;
 756             fiz2            += tz;
 757             f[j_coord_offset+DIM*2+XX] -= tx;
 758             f[j_coord_offset+DIM*2+YY] -= ty;
 759             f[j_coord_offset+DIM*2+ZZ] -= tz;
 760
 761             }
 762
 763             /* Inner loop uses 538 flops */
 764         }
 765         /* End of innermost loop */
 766
 767         tx = ty = tz = 0;
 768         f[i_coord_offset+DIM*0+XX] += fix0;
 769         f[i_coord_offset+DIM*0+YY] += fiy0;
 770         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 771         tx                         += fix0;
 772         ty                         += fiy0;
 773         tz                         += fiz0;
 774         f[i_coord_offset+DIM*1+XX] += fix1;
 775         f[i_coord_offset+DIM*1+YY] += fiy1;
 776         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 777         tx                         += fix1;
 778         ty                         += fiy1;
 779         tz                         += fiz1;
 780         f[i_coord_offset+DIM*2+XX] += fix2;
 781         f[i_coord_offset+DIM*2+YY] += fiy2;
 782         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 783         tx                         += fix2;
 784         ty                         += fiy2;
 785         tz                         += fiz2;
 786         fshift[i_shift_offset+XX]  += tx;
 787         fshift[i_shift_offset+YY]  += ty;
 788         fshift[i_shift_offset+ZZ]  += tz;
 789
 790         ggid                        = gid[iidx];
 791         /* Update potential energies */
 792         kernel_data->energygrp_elec[ggid] += velecsum;
 793         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 794
 795         /* Increment number of inner iterations */
 796         inneriter                  += j_index_end - j_index_start;
 797
 798         /* Outer loop uses 32 flops */
 799     }
 800
 801     /* Increment number of outer iterations */
 802     outeriter        += nri;
 803
 804     /* Update outer/inner flops */
 805
 806     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*538);
 807 }
 808 /*
 809  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_F_c
 810  * Electrostatics interaction: Ewald
 811  * VdW interaction:            LennardJones
 812  * Geometry:                   Water3-Water3
 813  * Calculate force/pot:        Force
 814  */
 815 void
 816 nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_F_c
 817                     (t_nblist                    * gmx_restrict       nlist,
 818                      rvec                        * gmx_restrict          xx,
 819                      rvec                        * gmx_restrict          ff,
 820                      t_forcerec                  * gmx_restrict          fr,
 821                      t_mdatoms                   * gmx_restrict     mdatoms,
 822                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 823                      t_nrnb                      * gmx_restrict        nrnb)
 824 {
 825     int              i_shift_offset,i_coord_offset,j_coord_offset;
 826     int              j_index_start,j_index_end;
 827     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 828     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 829     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 830     real             *shiftvec,*fshift,*x,*f;
 831     int              vdwioffset0;
 832     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 833     int              vdwioffset1;
 834     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 835     int              vdwioffset2;
 836     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 837     int              vdwjidx0;
 838     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 839     int              vdwjidx1;
 840     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 841     int              vdwjidx2;
 842     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 843     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 844     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 845     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 846     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 847     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 848     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 849     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 850     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 851     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 852     real             velec,felec,velecsum,facel,crf,krf,krf2;
 853     real             *charge;
 854     int              nvdwtype;
 855     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 856     int              *vdwtype;
 857     real             *vdwparam;
 858     int              ewitab;
 859     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 860     real             *ewtab;
 861     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 862
 863     x                = xx[0];
 864     f                = ff[0];
 865
 866     nri              = nlist->nri;
 867     iinr             = nlist->iinr;
 868     jindex           = nlist->jindex;
 869     jjnr             = nlist->jjnr;
 870     shiftidx         = nlist->shift;
 871     gid              = nlist->gid;
 872     shiftvec         = fr->shift_vec[0];
 873     fshift           = fr->fshift[0];
 874     facel            = fr->epsfac;
 875     charge           = mdatoms->chargeA;
 876     nvdwtype         = fr->ntype;
 877     vdwparam         = fr->nbfp;
 878     vdwtype          = mdatoms->typeA;
 879
 880     sh_ewald         = fr->ic->sh_ewald;
 881     ewtab            = fr->ic->tabq_coul_FDV0;
 882     ewtabscale       = fr->ic->tabq_scale;
 883     ewtabhalfspace   = 0.5/ewtabscale;
 884
 885     /* Setup water-specific parameters */
 886     inr              = nlist->iinr[0];
 887     iq0              = facel*charge[inr+0];
 888     iq1              = facel*charge[inr+1];
 889     iq2              = facel*charge[inr+2];
 890     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 891
 892     jq0              = charge[inr+0];
 893     jq1              = charge[inr+1];
 894     jq2              = charge[inr+2];
 895     vdwjidx0         = 2*vdwtype[inr+0];
 896     qq00             = iq0*jq0;
 897     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 898     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 899     qq01             = iq0*jq1;
 900     qq02             = iq0*jq2;
 901     qq10             = iq1*jq0;
 902     qq11             = iq1*jq1;
 903     qq12             = iq1*jq2;
 904     qq20             = iq2*jq0;
 905     qq21             = iq2*jq1;
 906     qq22             = iq2*jq2;
 907
 908     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 909     rcutoff          = fr->rcoulomb;
 910     rcutoff2         = rcutoff*rcutoff;
 911
 912     rswitch          = fr->rcoulomb_switch;
 913     /* Setup switch parameters */
 914     d                = rcutoff-rswitch;
 915     swV3             = -10.0/(d*d*d);
 916     swV4             =  15.0/(d*d*d*d);
 917     swV5             =  -6.0/(d*d*d*d*d);
 918     swF2             = -30.0/(d*d*d);
 919     swF3             =  60.0/(d*d*d*d);
 920     swF4             = -30.0/(d*d*d*d*d);
 921
 922     outeriter        = 0;
 923     inneriter        = 0;
 924
 925     /* Start outer loop over neighborlists */
 926     for(iidx=0; iidx<nri; iidx++)
 927     {
 928         /* Load shift vector for this list */
 929         i_shift_offset   = DIM*shiftidx[iidx];
 930         shX              = shiftvec[i_shift_offset+XX];
 931         shY              = shiftvec[i_shift_offset+YY];
 932         shZ              = shiftvec[i_shift_offset+ZZ];
 933
 934         /* Load limits for loop over neighbors */
 935         j_index_start    = jindex[iidx];
 936         j_index_end      = jindex[iidx+1];
 937
 938         /* Get outer coordinate index */
 939         inr              = iinr[iidx];
 940         i_coord_offset   = DIM*inr;
 941
 942         /* Load i particle coords and add shift vector */
 943         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 944         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 945         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 946         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 947         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 948         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 949         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 950         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 951         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 952
 953         fix0             = 0.0;
 954         fiy0             = 0.0;
 955         fiz0             = 0.0;
 956         fix1             = 0.0;
 957         fiy1             = 0.0;
 958         fiz1             = 0.0;
 959         fix2             = 0.0;
 960         fiy2             = 0.0;
 961         fiz2             = 0.0;
 962
 963         /* Start inner kernel loop */
 964         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 965         {
 966             /* Get j neighbor index, and coordinate index */
 967             jnr              = jjnr[jidx];
 968             j_coord_offset   = DIM*jnr;
 969
 970             /* load j atom coordinates */
 971             jx0              = x[j_coord_offset+DIM*0+XX];
 972             jy0              = x[j_coord_offset+DIM*0+YY];
 973             jz0              = x[j_coord_offset+DIM*0+ZZ];
 974             jx1              = x[j_coord_offset+DIM*1+XX];
 975             jy1              = x[j_coord_offset+DIM*1+YY];
 976             jz1              = x[j_coord_offset+DIM*1+ZZ];
 977             jx2              = x[j_coord_offset+DIM*2+XX];
 978             jy2              = x[j_coord_offset+DIM*2+YY];
 979             jz2              = x[j_coord_offset+DIM*2+ZZ];
 980
 981             /* Calculate displacement vector */
 982             dx00             = ix0 - jx0;
 983             dy00             = iy0 - jy0;
 984             dz00             = iz0 - jz0;
 985             dx01             = ix0 - jx1;
 986             dy01             = iy0 - jy1;
 987             dz01             = iz0 - jz1;
 988             dx02             = ix0 - jx2;
 989             dy02             = iy0 - jy2;
 990             dz02             = iz0 - jz2;
 991             dx10             = ix1 - jx0;
 992             dy10             = iy1 - jy0;
 993             dz10             = iz1 - jz0;
 994             dx11             = ix1 - jx1;
 995             dy11             = iy1 - jy1;
 996             dz11             = iz1 - jz1;
 997             dx12             = ix1 - jx2;
 998             dy12             = iy1 - jy2;
 999             dz12             = iz1 - jz2;
1000             dx20             = ix2 - jx0;
1001             dy20             = iy2 - jy0;
1002             dz20             = iz2 - jz0;
1003             dx21             = ix2 - jx1;
1004             dy21             = iy2 - jy1;
1005             dz21             = iz2 - jz1;
1006             dx22             = ix2 - jx2;
1007             dy22             = iy2 - jy2;
1008             dz22             = iz2 - jz2;
1009
1010             /* Calculate squared distance and things based on it */
1011             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
1012             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
1013             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
1014             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
1015             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
1016             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
1017             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
1018             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
1019             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
1020
1021             rinv00           = gmx_invsqrt(rsq00);
1022             rinv01           = gmx_invsqrt(rsq01);
1023             rinv02           = gmx_invsqrt(rsq02);
1024             rinv10           = gmx_invsqrt(rsq10);
1025             rinv11           = gmx_invsqrt(rsq11);
1026             rinv12           = gmx_invsqrt(rsq12);
1027             rinv20           = gmx_invsqrt(rsq20);
1028             rinv21           = gmx_invsqrt(rsq21);
1029             rinv22           = gmx_invsqrt(rsq22);
1030
1031             rinvsq00         = rinv00*rinv00;
1032             rinvsq01         = rinv01*rinv01;
1033             rinvsq02         = rinv02*rinv02;
1034             rinvsq10         = rinv10*rinv10;
1035             rinvsq11         = rinv11*rinv11;
1036             rinvsq12         = rinv12*rinv12;
1037             rinvsq20         = rinv20*rinv20;
1038             rinvsq21         = rinv21*rinv21;
1039             rinvsq22         = rinv22*rinv22;
1040
1041             /**************************
1042              * CALCULATE INTERACTIONS *
1043              **************************/
1044
1045             if (rsq00<rcutoff2)
1046             {
1047
1048             r00              = rsq00*rinv00;
1049
1050             /* EWALD ELECTROSTATICS */
1051
1052             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1053             ewrt             = r00*ewtabscale;
1054             ewitab           = ewrt;
1055             eweps            = ewrt-ewitab;
1056             ewitab           = 4*ewitab;
1057             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1058             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1059             felec            = qq00*rinv00*(rinvsq00-felec);
1060
1061             /* LENNARD-JONES DISPERSION/REPULSION */
1062
1063             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
1064             vvdw6            = c6_00*rinvsix;
1065             vvdw12           = c12_00*rinvsix*rinvsix;
1066             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
1067             fvdw             = (vvdw12-vvdw6)*rinvsq00;
1068
1069             d                = r00-rswitch;
1070             d                = (d>0.0) ? d : 0.0;
1071             d2               = d*d;
1072             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1073
1074             dsw              = d2*(swF2+d*(swF3+d*swF4));
1075
1076             /* Evaluate switch function */
1077             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1078             felec            = felec*sw - rinv00*velec*dsw;
1079             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
1080
1081             fscal            = felec+fvdw;
1082
1083             /* Calculate temporary vectorial force */
1084             tx               = fscal*dx00;
1085             ty               = fscal*dy00;
1086             tz               = fscal*dz00;
1087
1088             /* Update vectorial force */
1089             fix0            += tx;
1090             fiy0            += ty;
1091             fiz0            += tz;
1092             f[j_coord_offset+DIM*0+XX] -= tx;
1093             f[j_coord_offset+DIM*0+YY] -= ty;
1094             f[j_coord_offset+DIM*0+ZZ] -= tz;
1095
1096             }
1097
1098             /**************************
1099              * CALCULATE INTERACTIONS *
1100              **************************/
1101
1102             if (rsq01<rcutoff2)
1103             {
1104
1105             r01              = rsq01*rinv01;
1106
1107             /* EWALD ELECTROSTATICS */
1108
1109             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1110             ewrt             = r01*ewtabscale;
1111             ewitab           = ewrt;
1112             eweps            = ewrt-ewitab;
1113             ewitab           = 4*ewitab;
1114             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1115             velec            = qq01*(rinv01-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1116             felec            = qq01*rinv01*(rinvsq01-felec);
1117
1118             d                = r01-rswitch;
1119             d                = (d>0.0) ? d : 0.0;
1120             d2               = d*d;
1121             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1122
1123             dsw              = d2*(swF2+d*(swF3+d*swF4));
1124
1125             /* Evaluate switch function */
1126             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1127             felec            = felec*sw - rinv01*velec*dsw;
1128
1129             fscal            = felec;
1130
1131             /* Calculate temporary vectorial force */
1132             tx               = fscal*dx01;
1133             ty               = fscal*dy01;
1134             tz               = fscal*dz01;
1135
1136             /* Update vectorial force */
1137             fix0            += tx;
1138             fiy0            += ty;
1139             fiz0            += tz;
1140             f[j_coord_offset+DIM*1+XX] -= tx;
1141             f[j_coord_offset+DIM*1+YY] -= ty;
1142             f[j_coord_offset+DIM*1+ZZ] -= tz;
1143
1144             }
1145
1146             /**************************
1147              * CALCULATE INTERACTIONS *
1148              **************************/
1149
1150             if (rsq02<rcutoff2)
1151             {
1152
1153             r02              = rsq02*rinv02;
1154
1155             /* EWALD ELECTROSTATICS */
1156
1157             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1158             ewrt             = r02*ewtabscale;
1159             ewitab           = ewrt;
1160             eweps            = ewrt-ewitab;
1161             ewitab           = 4*ewitab;
1162             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1163             velec            = qq02*(rinv02-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1164             felec            = qq02*rinv02*(rinvsq02-felec);
1165
1166             d                = r02-rswitch;
1167             d                = (d>0.0) ? d : 0.0;
1168             d2               = d*d;
1169             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1170
1171             dsw              = d2*(swF2+d*(swF3+d*swF4));
1172
1173             /* Evaluate switch function */
1174             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1175             felec            = felec*sw - rinv02*velec*dsw;
1176
1177             fscal            = felec;
1178
1179             /* Calculate temporary vectorial force */
1180             tx               = fscal*dx02;
1181             ty               = fscal*dy02;
1182             tz               = fscal*dz02;
1183
1184             /* Update vectorial force */
1185             fix0            += tx;
1186             fiy0            += ty;
1187             fiz0            += tz;
1188             f[j_coord_offset+DIM*2+XX] -= tx;
1189             f[j_coord_offset+DIM*2+YY] -= ty;
1190             f[j_coord_offset+DIM*2+ZZ] -= tz;
1191
1192             }
1193
1194             /**************************
1195              * CALCULATE INTERACTIONS *
1196              **************************/
1197
1198             if (rsq10<rcutoff2)
1199             {
1200
1201             r10              = rsq10*rinv10;
1202
1203             /* EWALD ELECTROSTATICS */
1204
1205             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1206             ewrt             = r10*ewtabscale;
1207             ewitab           = ewrt;
1208             eweps            = ewrt-ewitab;
1209             ewitab           = 4*ewitab;
1210             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1211             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1212             felec            = qq10*rinv10*(rinvsq10-felec);
1213
1214             d                = r10-rswitch;
1215             d                = (d>0.0) ? d : 0.0;
1216             d2               = d*d;
1217             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1218
1219             dsw              = d2*(swF2+d*(swF3+d*swF4));
1220
1221             /* Evaluate switch function */
1222             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1223             felec            = felec*sw - rinv10*velec*dsw;
1224
1225             fscal            = felec;
1226
1227             /* Calculate temporary vectorial force */
1228             tx               = fscal*dx10;
1229             ty               = fscal*dy10;
1230             tz               = fscal*dz10;
1231
1232             /* Update vectorial force */
1233             fix1            += tx;
1234             fiy1            += ty;
1235             fiz1            += tz;
1236             f[j_coord_offset+DIM*0+XX] -= tx;
1237             f[j_coord_offset+DIM*0+YY] -= ty;
1238             f[j_coord_offset+DIM*0+ZZ] -= tz;
1239
1240             }
1241
1242             /**************************
1243              * CALCULATE INTERACTIONS *
1244              **************************/
1245
1246             if (rsq11<rcutoff2)
1247             {
1248
1249             r11              = rsq11*rinv11;
1250
1251             /* EWALD ELECTROSTATICS */
1252
1253             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1254             ewrt             = r11*ewtabscale;
1255             ewitab           = ewrt;
1256             eweps            = ewrt-ewitab;
1257             ewitab           = 4*ewitab;
1258             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1259             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1260             felec            = qq11*rinv11*(rinvsq11-felec);
1261
1262             d                = r11-rswitch;
1263             d                = (d>0.0) ? d : 0.0;
1264             d2               = d*d;
1265             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1266
1267             dsw              = d2*(swF2+d*(swF3+d*swF4));
1268
1269             /* Evaluate switch function */
1270             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1271             felec            = felec*sw - rinv11*velec*dsw;
1272
1273             fscal            = felec;
1274
1275             /* Calculate temporary vectorial force */
1276             tx               = fscal*dx11;
1277             ty               = fscal*dy11;
1278             tz               = fscal*dz11;
1279
1280             /* Update vectorial force */
1281             fix1            += tx;
1282             fiy1            += ty;
1283             fiz1            += tz;
1284             f[j_coord_offset+DIM*1+XX] -= tx;
1285             f[j_coord_offset+DIM*1+YY] -= ty;
1286             f[j_coord_offset+DIM*1+ZZ] -= tz;
1287
1288             }
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             if (rsq12<rcutoff2)
1295             {
1296
1297             r12              = rsq12*rinv12;
1298
1299             /* EWALD ELECTROSTATICS */
1300
1301             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1302             ewrt             = r12*ewtabscale;
1303             ewitab           = ewrt;
1304             eweps            = ewrt-ewitab;
1305             ewitab           = 4*ewitab;
1306             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1307             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1308             felec            = qq12*rinv12*(rinvsq12-felec);
1309
1310             d                = r12-rswitch;
1311             d                = (d>0.0) ? d : 0.0;
1312             d2               = d*d;
1313             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1314
1315             dsw              = d2*(swF2+d*(swF3+d*swF4));
1316
1317             /* Evaluate switch function */
1318             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1319             felec            = felec*sw - rinv12*velec*dsw;
1320
1321             fscal            = felec;
1322
1323             /* Calculate temporary vectorial force */
1324             tx               = fscal*dx12;
1325             ty               = fscal*dy12;
1326             tz               = fscal*dz12;
1327
1328             /* Update vectorial force */
1329             fix1            += tx;
1330             fiy1            += ty;
1331             fiz1            += tz;
1332             f[j_coord_offset+DIM*2+XX] -= tx;
1333             f[j_coord_offset+DIM*2+YY] -= ty;
1334             f[j_coord_offset+DIM*2+ZZ] -= tz;
1335
1336             }
1337
1338             /**************************
1339              * CALCULATE INTERACTIONS *
1340              **************************/
1341
1342             if (rsq20<rcutoff2)
1343             {
1344
1345             r20              = rsq20*rinv20;
1346
1347             /* EWALD ELECTROSTATICS */
1348
1349             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1350             ewrt             = r20*ewtabscale;
1351             ewitab           = ewrt;
1352             eweps            = ewrt-ewitab;
1353             ewitab           = 4*ewitab;
1354             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1355             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1356             felec            = qq20*rinv20*(rinvsq20-felec);
1357
1358             d                = r20-rswitch;
1359             d                = (d>0.0) ? d : 0.0;
1360             d2               = d*d;
1361             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1362
1363             dsw              = d2*(swF2+d*(swF3+d*swF4));
1364
1365             /* Evaluate switch function */
1366             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1367             felec            = felec*sw - rinv20*velec*dsw;
1368
1369             fscal            = felec;
1370
1371             /* Calculate temporary vectorial force */
1372             tx               = fscal*dx20;
1373             ty               = fscal*dy20;
1374             tz               = fscal*dz20;
1375
1376             /* Update vectorial force */
1377             fix2            += tx;
1378             fiy2            += ty;
1379             fiz2            += tz;
1380             f[j_coord_offset+DIM*0+XX] -= tx;
1381             f[j_coord_offset+DIM*0+YY] -= ty;
1382             f[j_coord_offset+DIM*0+ZZ] -= tz;
1383
1384             }
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             if (rsq21<rcutoff2)
1391             {
1392
1393             r21              = rsq21*rinv21;
1394
1395             /* EWALD ELECTROSTATICS */
1396
1397             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1398             ewrt             = r21*ewtabscale;
1399             ewitab           = ewrt;
1400             eweps            = ewrt-ewitab;
1401             ewitab           = 4*ewitab;
1402             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1403             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1404             felec            = qq21*rinv21*(rinvsq21-felec);
1405
1406             d                = r21-rswitch;
1407             d                = (d>0.0) ? d : 0.0;
1408             d2               = d*d;
1409             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1410
1411             dsw              = d2*(swF2+d*(swF3+d*swF4));
1412
1413             /* Evaluate switch function */
1414             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1415             felec            = felec*sw - rinv21*velec*dsw;
1416
1417             fscal            = felec;
1418
1419             /* Calculate temporary vectorial force */
1420             tx               = fscal*dx21;
1421             ty               = fscal*dy21;
1422             tz               = fscal*dz21;
1423
1424             /* Update vectorial force */
1425             fix2            += tx;
1426             fiy2            += ty;
1427             fiz2            += tz;
1428             f[j_coord_offset+DIM*1+XX] -= tx;
1429             f[j_coord_offset+DIM*1+YY] -= ty;
1430             f[j_coord_offset+DIM*1+ZZ] -= tz;
1431
1432             }
1433
1434             /**************************
1435              * CALCULATE INTERACTIONS *
1436              **************************/
1437
1438             if (rsq22<rcutoff2)
1439             {
1440
1441             r22              = rsq22*rinv22;
1442
1443             /* EWALD ELECTROSTATICS */
1444
1445             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1446             ewrt             = r22*ewtabscale;
1447             ewitab           = ewrt;
1448             eweps            = ewrt-ewitab;
1449             ewitab           = 4*ewitab;
1450             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1451             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1452             felec            = qq22*rinv22*(rinvsq22-felec);
1453
1454             d                = r22-rswitch;
1455             d                = (d>0.0) ? d : 0.0;
1456             d2               = d*d;
1457             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1458
1459             dsw              = d2*(swF2+d*(swF3+d*swF4));
1460
1461             /* Evaluate switch function */
1462             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1463             felec            = felec*sw - rinv22*velec*dsw;
1464
1465             fscal            = felec;
1466
1467             /* Calculate temporary vectorial force */
1468             tx               = fscal*dx22;
1469             ty               = fscal*dy22;
1470             tz               = fscal*dz22;
1471
1472             /* Update vectorial force */
1473             fix2            += tx;
1474             fiy2            += ty;
1475             fiz2            += tz;
1476             f[j_coord_offset+DIM*2+XX] -= tx;
1477             f[j_coord_offset+DIM*2+YY] -= ty;
1478             f[j_coord_offset+DIM*2+ZZ] -= tz;
1479
1480             }
1481
1482             /* Inner loop uses 518 flops */
1483         }
1484         /* End of innermost loop */
1485
1486         tx = ty = tz = 0;
1487         f[i_coord_offset+DIM*0+XX] += fix0;
1488         f[i_coord_offset+DIM*0+YY] += fiy0;
1489         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1490         tx                         += fix0;
1491         ty                         += fiy0;
1492         tz                         += fiz0;
1493         f[i_coord_offset+DIM*1+XX] += fix1;
1494         f[i_coord_offset+DIM*1+YY] += fiy1;
1495         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1496         tx                         += fix1;
1497         ty                         += fiy1;
1498         tz                         += fiz1;
1499         f[i_coord_offset+DIM*2+XX] += fix2;
1500         f[i_coord_offset+DIM*2+YY] += fiy2;
1501         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1502         tx                         += fix2;
1503         ty                         += fiy2;
1504         tz                         += fiz2;
1505         fshift[i_shift_offset+XX]  += tx;
1506         fshift[i_shift_offset+YY]  += ty;
1507         fshift[i_shift_offset+ZZ]  += tz;
1508
1509         /* Increment number of inner iterations */
1510         inneriter                  += j_index_end - j_index_start;
1511
1512         /* Outer loop uses 30 flops */
1513     }
1514
1515     /* Increment number of outer iterations */
1516     outeriter        += nri;
1517
1518     /* Update outer/inner flops */
1519
1520     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*518);
1521 }