src/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSw_VdwLJSw_GeomW4W4_c.c

   1 /*
   2  * Note: this file was generated by the Gromacs c kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 /*
  34  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW4W4_VF_c
  35  * Electrostatics interaction: Ewald
  36  * VdW interaction:            LennardJones
  37  * Geometry:                   Water4-Water4
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecEwSw_VdwLJSw_GeomW4W4_VF_c
  42                     (t_nblist * gmx_restrict                nlist,
  43                      rvec * gmx_restrict                    xx,
  44                      rvec * gmx_restrict                    ff,
  45                      t_forcerec * gmx_restrict              fr,
  46                      t_mdatoms * gmx_restrict               mdatoms,
  47                      nb_kernel_data_t * gmx_restrict        kernel_data,
  48                      t_nrnb * gmx_restrict                  nrnb)
  49 {
  50     int              i_shift_offset,i_coord_offset,j_coord_offset;
  51     int              j_index_start,j_index_end;
  52     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  53     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  54     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  55     real             *shiftvec,*fshift,*x,*f;
  56     int              vdwioffset0;
  57     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  58     int              vdwioffset1;
  59     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  60     int              vdwioffset2;
  61     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  62     int              vdwioffset3;
  63     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  64     int              vdwjidx0;
  65     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  66     int              vdwjidx1;
  67     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  68     int              vdwjidx2;
  69     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  70     int              vdwjidx3;
  71     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  72     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  73     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  74     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  75     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
  76     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  77     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  78     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
  79     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
  80     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
  81     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
  82     real             velec,felec,velecsum,facel,crf,krf,krf2;
  83     real             *charge;
  84     int              nvdwtype;
  85     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  86     int              *vdwtype;
  87     real             *vdwparam;
  88     int              ewitab;
  89     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
  90     real             *ewtab;
  91     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  92
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = fr->epsfac;
 105     charge           = mdatoms->chargeA;
 106     nvdwtype         = fr->ntype;
 107     vdwparam         = fr->nbfp;
 108     vdwtype          = mdatoms->typeA;
 109
 110     sh_ewald         = fr->ic->sh_ewald;
 111     ewtab            = fr->ic->tabq_coul_FDV0;
 112     ewtabscale       = fr->ic->tabq_scale;
 113     ewtabhalfspace   = 0.5/ewtabscale;
 114
 115     /* Setup water-specific parameters */
 116     inr              = nlist->iinr[0];
 117     iq1              = facel*charge[inr+1];
 118     iq2              = facel*charge[inr+2];
 119     iq3              = facel*charge[inr+3];
 120     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 121
 122     jq1              = charge[inr+1];
 123     jq2              = charge[inr+2];
 124     jq3              = charge[inr+3];
 125     vdwjidx0         = 2*vdwtype[inr+0];
 126     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 127     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 128     qq11             = iq1*jq1;
 129     qq12             = iq1*jq2;
 130     qq13             = iq1*jq3;
 131     qq21             = iq2*jq1;
 132     qq22             = iq2*jq2;
 133     qq23             = iq2*jq3;
 134     qq31             = iq3*jq1;
 135     qq32             = iq3*jq2;
 136     qq33             = iq3*jq3;
 137
 138     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 139     rcutoff          = fr->rcoulomb;
 140     rcutoff2         = rcutoff*rcutoff;
 141
 142     rswitch          = fr->rcoulomb_switch;
 143     /* Setup switch parameters */
 144     d                = rcutoff-rswitch;
 145     swV3             = -10.0/(d*d*d);
 146     swV4             =  15.0/(d*d*d*d);
 147     swV5             =  -6.0/(d*d*d*d*d);
 148     swF2             = -30.0/(d*d*d);
 149     swF3             =  60.0/(d*d*d*d);
 150     swF4             = -30.0/(d*d*d*d*d);
 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         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 183         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 184         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 185
 186         fix0             = 0.0;
 187         fiy0             = 0.0;
 188         fiz0             = 0.0;
 189         fix1             = 0.0;
 190         fiy1             = 0.0;
 191         fiz1             = 0.0;
 192         fix2             = 0.0;
 193         fiy2             = 0.0;
 194         fiz2             = 0.0;
 195         fix3             = 0.0;
 196         fiy3             = 0.0;
 197         fiz3             = 0.0;
 198
 199         /* Reset potential sums */
 200         velecsum         = 0.0;
 201         vvdwsum          = 0.0;
 202
 203         /* Start inner kernel loop */
 204         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 205         {
 206             /* Get j neighbor index, and coordinate index */
 207             jnr              = jjnr[jidx];
 208             j_coord_offset   = DIM*jnr;
 209
 210             /* load j atom coordinates */
 211             jx0              = x[j_coord_offset+DIM*0+XX];
 212             jy0              = x[j_coord_offset+DIM*0+YY];
 213             jz0              = x[j_coord_offset+DIM*0+ZZ];
 214             jx1              = x[j_coord_offset+DIM*1+XX];
 215             jy1              = x[j_coord_offset+DIM*1+YY];
 216             jz1              = x[j_coord_offset+DIM*1+ZZ];
 217             jx2              = x[j_coord_offset+DIM*2+XX];
 218             jy2              = x[j_coord_offset+DIM*2+YY];
 219             jz2              = x[j_coord_offset+DIM*2+ZZ];
 220             jx3              = x[j_coord_offset+DIM*3+XX];
 221             jy3              = x[j_coord_offset+DIM*3+YY];
 222             jz3              = x[j_coord_offset+DIM*3+ZZ];
 223
 224             /* Calculate displacement vector */
 225             dx00             = ix0 - jx0;
 226             dy00             = iy0 - jy0;
 227             dz00             = iz0 - 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             dx13             = ix1 - jx3;
 235             dy13             = iy1 - jy3;
 236             dz13             = iz1 - jz3;
 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             dx23             = ix2 - jx3;
 244             dy23             = iy2 - jy3;
 245             dz23             = iz2 - jz3;
 246             dx31             = ix3 - jx1;
 247             dy31             = iy3 - jy1;
 248             dz31             = iz3 - jz1;
 249             dx32             = ix3 - jx2;
 250             dy32             = iy3 - jy2;
 251             dz32             = iz3 - jz2;
 252             dx33             = ix3 - jx3;
 253             dy33             = iy3 - jy3;
 254             dz33             = iz3 - jz3;
 255
 256             /* Calculate squared distance and things based on it */
 257             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 258             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 259             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 260             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 261             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 262             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 263             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 264             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 265             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 266             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 267
 268             rinv00           = gmx_invsqrt(rsq00);
 269             rinv11           = gmx_invsqrt(rsq11);
 270             rinv12           = gmx_invsqrt(rsq12);
 271             rinv13           = gmx_invsqrt(rsq13);
 272             rinv21           = gmx_invsqrt(rsq21);
 273             rinv22           = gmx_invsqrt(rsq22);
 274             rinv23           = gmx_invsqrt(rsq23);
 275             rinv31           = gmx_invsqrt(rsq31);
 276             rinv32           = gmx_invsqrt(rsq32);
 277             rinv33           = gmx_invsqrt(rsq33);
 278
 279             rinvsq00         = rinv00*rinv00;
 280             rinvsq11         = rinv11*rinv11;
 281             rinvsq12         = rinv12*rinv12;
 282             rinvsq13         = rinv13*rinv13;
 283             rinvsq21         = rinv21*rinv21;
 284             rinvsq22         = rinv22*rinv22;
 285             rinvsq23         = rinv23*rinv23;
 286             rinvsq31         = rinv31*rinv31;
 287             rinvsq32         = rinv32*rinv32;
 288             rinvsq33         = rinv33*rinv33;
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             if (rsq00<rcutoff2)
 295             {
 296
 297             r00              = rsq00*rinv00;
 298
 299             /* LENNARD-JONES DISPERSION/REPULSION */
 300
 301             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 302             vvdw6            = c6_00*rinvsix;
 303             vvdw12           = c12_00*rinvsix*rinvsix;
 304             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 305             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 306
 307             d                = r00-rswitch;
 308             d                = (d>0.0) ? d : 0.0;
 309             d2               = d*d;
 310             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 311
 312             dsw              = d2*(swF2+d*(swF3+d*swF4));
 313
 314             /* Evaluate switch function */
 315             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 316             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 317             vvdw            *= sw;
 318
 319             /* Update potential sums from outer loop */
 320             vvdwsum         += vvdw;
 321
 322             fscal            = fvdw;
 323
 324             /* Calculate temporary vectorial force */
 325             tx               = fscal*dx00;
 326             ty               = fscal*dy00;
 327             tz               = fscal*dz00;
 328
 329             /* Update vectorial force */
 330             fix0            += tx;
 331             fiy0            += ty;
 332             fiz0            += tz;
 333             f[j_coord_offset+DIM*0+XX] -= tx;
 334             f[j_coord_offset+DIM*0+YY] -= ty;
 335             f[j_coord_offset+DIM*0+ZZ] -= tz;
 336
 337             }
 338
 339             /**************************
 340              * CALCULATE INTERACTIONS *
 341              **************************/
 342
 343             if (rsq11<rcutoff2)
 344             {
 345
 346             r11              = rsq11*rinv11;
 347
 348             /* EWALD ELECTROSTATICS */
 349
 350             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 351             ewrt             = r11*ewtabscale;
 352             ewitab           = ewrt;
 353             eweps            = ewrt-ewitab;
 354             ewitab           = 4*ewitab;
 355             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 356             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 357             felec            = qq11*rinv11*(rinvsq11-felec);
 358
 359             d                = r11-rswitch;
 360             d                = (d>0.0) ? d : 0.0;
 361             d2               = d*d;
 362             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 363
 364             dsw              = d2*(swF2+d*(swF3+d*swF4));
 365
 366             /* Evaluate switch function */
 367             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 368             felec            = felec*sw - rinv11*velec*dsw;
 369             velec           *= sw;
 370
 371             /* Update potential sums from outer loop */
 372             velecsum        += velec;
 373
 374             fscal            = felec;
 375
 376             /* Calculate temporary vectorial force */
 377             tx               = fscal*dx11;
 378             ty               = fscal*dy11;
 379             tz               = fscal*dz11;
 380
 381             /* Update vectorial force */
 382             fix1            += tx;
 383             fiy1            += ty;
 384             fiz1            += tz;
 385             f[j_coord_offset+DIM*1+XX] -= tx;
 386             f[j_coord_offset+DIM*1+YY] -= ty;
 387             f[j_coord_offset+DIM*1+ZZ] -= tz;
 388
 389             }
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             if (rsq12<rcutoff2)
 396             {
 397
 398             r12              = rsq12*rinv12;
 399
 400             /* EWALD ELECTROSTATICS */
 401
 402             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 403             ewrt             = r12*ewtabscale;
 404             ewitab           = ewrt;
 405             eweps            = ewrt-ewitab;
 406             ewitab           = 4*ewitab;
 407             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 408             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 409             felec            = qq12*rinv12*(rinvsq12-felec);
 410
 411             d                = r12-rswitch;
 412             d                = (d>0.0) ? d : 0.0;
 413             d2               = d*d;
 414             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 415
 416             dsw              = d2*(swF2+d*(swF3+d*swF4));
 417
 418             /* Evaluate switch function */
 419             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 420             felec            = felec*sw - rinv12*velec*dsw;
 421             velec           *= sw;
 422
 423             /* Update potential sums from outer loop */
 424             velecsum        += velec;
 425
 426             fscal            = felec;
 427
 428             /* Calculate temporary vectorial force */
 429             tx               = fscal*dx12;
 430             ty               = fscal*dy12;
 431             tz               = fscal*dz12;
 432
 433             /* Update vectorial force */
 434             fix1            += tx;
 435             fiy1            += ty;
 436             fiz1            += tz;
 437             f[j_coord_offset+DIM*2+XX] -= tx;
 438             f[j_coord_offset+DIM*2+YY] -= ty;
 439             f[j_coord_offset+DIM*2+ZZ] -= tz;
 440
 441             }
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             if (rsq13<rcutoff2)
 448             {
 449
 450             r13              = rsq13*rinv13;
 451
 452             /* EWALD ELECTROSTATICS */
 453
 454             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 455             ewrt             = r13*ewtabscale;
 456             ewitab           = ewrt;
 457             eweps            = ewrt-ewitab;
 458             ewitab           = 4*ewitab;
 459             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 460             velec            = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 461             felec            = qq13*rinv13*(rinvsq13-felec);
 462
 463             d                = r13-rswitch;
 464             d                = (d>0.0) ? d : 0.0;
 465             d2               = d*d;
 466             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 467
 468             dsw              = d2*(swF2+d*(swF3+d*swF4));
 469
 470             /* Evaluate switch function */
 471             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 472             felec            = felec*sw - rinv13*velec*dsw;
 473             velec           *= sw;
 474
 475             /* Update potential sums from outer loop */
 476             velecsum        += velec;
 477
 478             fscal            = felec;
 479
 480             /* Calculate temporary vectorial force */
 481             tx               = fscal*dx13;
 482             ty               = fscal*dy13;
 483             tz               = fscal*dz13;
 484
 485             /* Update vectorial force */
 486             fix1            += tx;
 487             fiy1            += ty;
 488             fiz1            += tz;
 489             f[j_coord_offset+DIM*3+XX] -= tx;
 490             f[j_coord_offset+DIM*3+YY] -= ty;
 491             f[j_coord_offset+DIM*3+ZZ] -= tz;
 492
 493             }
 494
 495             /**************************
 496              * CALCULATE INTERACTIONS *
 497              **************************/
 498
 499             if (rsq21<rcutoff2)
 500             {
 501
 502             r21              = rsq21*rinv21;
 503
 504             /* EWALD ELECTROSTATICS */
 505
 506             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 507             ewrt             = r21*ewtabscale;
 508             ewitab           = ewrt;
 509             eweps            = ewrt-ewitab;
 510             ewitab           = 4*ewitab;
 511             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 512             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 513             felec            = qq21*rinv21*(rinvsq21-felec);
 514
 515             d                = r21-rswitch;
 516             d                = (d>0.0) ? d : 0.0;
 517             d2               = d*d;
 518             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 519
 520             dsw              = d2*(swF2+d*(swF3+d*swF4));
 521
 522             /* Evaluate switch function */
 523             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 524             felec            = felec*sw - rinv21*velec*dsw;
 525             velec           *= sw;
 526
 527             /* Update potential sums from outer loop */
 528             velecsum        += velec;
 529
 530             fscal            = felec;
 531
 532             /* Calculate temporary vectorial force */
 533             tx               = fscal*dx21;
 534             ty               = fscal*dy21;
 535             tz               = fscal*dz21;
 536
 537             /* Update vectorial force */
 538             fix2            += tx;
 539             fiy2            += ty;
 540             fiz2            += tz;
 541             f[j_coord_offset+DIM*1+XX] -= tx;
 542             f[j_coord_offset+DIM*1+YY] -= ty;
 543             f[j_coord_offset+DIM*1+ZZ] -= tz;
 544
 545             }
 546
 547             /**************************
 548              * CALCULATE INTERACTIONS *
 549              **************************/
 550
 551             if (rsq22<rcutoff2)
 552             {
 553
 554             r22              = rsq22*rinv22;
 555
 556             /* EWALD ELECTROSTATICS */
 557
 558             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 559             ewrt             = r22*ewtabscale;
 560             ewitab           = ewrt;
 561             eweps            = ewrt-ewitab;
 562             ewitab           = 4*ewitab;
 563             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 564             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 565             felec            = qq22*rinv22*(rinvsq22-felec);
 566
 567             d                = r22-rswitch;
 568             d                = (d>0.0) ? d : 0.0;
 569             d2               = d*d;
 570             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 571
 572             dsw              = d2*(swF2+d*(swF3+d*swF4));
 573
 574             /* Evaluate switch function */
 575             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 576             felec            = felec*sw - rinv22*velec*dsw;
 577             velec           *= sw;
 578
 579             /* Update potential sums from outer loop */
 580             velecsum        += velec;
 581
 582             fscal            = felec;
 583
 584             /* Calculate temporary vectorial force */
 585             tx               = fscal*dx22;
 586             ty               = fscal*dy22;
 587             tz               = fscal*dz22;
 588
 589             /* Update vectorial force */
 590             fix2            += tx;
 591             fiy2            += ty;
 592             fiz2            += tz;
 593             f[j_coord_offset+DIM*2+XX] -= tx;
 594             f[j_coord_offset+DIM*2+YY] -= ty;
 595             f[j_coord_offset+DIM*2+ZZ] -= tz;
 596
 597             }
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             if (rsq23<rcutoff2)
 604             {
 605
 606             r23              = rsq23*rinv23;
 607
 608             /* EWALD ELECTROSTATICS */
 609
 610             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 611             ewrt             = r23*ewtabscale;
 612             ewitab           = ewrt;
 613             eweps            = ewrt-ewitab;
 614             ewitab           = 4*ewitab;
 615             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 616             velec            = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 617             felec            = qq23*rinv23*(rinvsq23-felec);
 618
 619             d                = r23-rswitch;
 620             d                = (d>0.0) ? d : 0.0;
 621             d2               = d*d;
 622             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 623
 624             dsw              = d2*(swF2+d*(swF3+d*swF4));
 625
 626             /* Evaluate switch function */
 627             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 628             felec            = felec*sw - rinv23*velec*dsw;
 629             velec           *= sw;
 630
 631             /* Update potential sums from outer loop */
 632             velecsum        += velec;
 633
 634             fscal            = felec;
 635
 636             /* Calculate temporary vectorial force */
 637             tx               = fscal*dx23;
 638             ty               = fscal*dy23;
 639             tz               = fscal*dz23;
 640
 641             /* Update vectorial force */
 642             fix2            += tx;
 643             fiy2            += ty;
 644             fiz2            += tz;
 645             f[j_coord_offset+DIM*3+XX] -= tx;
 646             f[j_coord_offset+DIM*3+YY] -= ty;
 647             f[j_coord_offset+DIM*3+ZZ] -= tz;
 648
 649             }
 650
 651             /**************************
 652              * CALCULATE INTERACTIONS *
 653              **************************/
 654
 655             if (rsq31<rcutoff2)
 656             {
 657
 658             r31              = rsq31*rinv31;
 659
 660             /* EWALD ELECTROSTATICS */
 661
 662             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 663             ewrt             = r31*ewtabscale;
 664             ewitab           = ewrt;
 665             eweps            = ewrt-ewitab;
 666             ewitab           = 4*ewitab;
 667             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 668             velec            = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 669             felec            = qq31*rinv31*(rinvsq31-felec);
 670
 671             d                = r31-rswitch;
 672             d                = (d>0.0) ? d : 0.0;
 673             d2               = d*d;
 674             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 675
 676             dsw              = d2*(swF2+d*(swF3+d*swF4));
 677
 678             /* Evaluate switch function */
 679             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 680             felec            = felec*sw - rinv31*velec*dsw;
 681             velec           *= sw;
 682
 683             /* Update potential sums from outer loop */
 684             velecsum        += velec;
 685
 686             fscal            = felec;
 687
 688             /* Calculate temporary vectorial force */
 689             tx               = fscal*dx31;
 690             ty               = fscal*dy31;
 691             tz               = fscal*dz31;
 692
 693             /* Update vectorial force */
 694             fix3            += tx;
 695             fiy3            += ty;
 696             fiz3            += tz;
 697             f[j_coord_offset+DIM*1+XX] -= tx;
 698             f[j_coord_offset+DIM*1+YY] -= ty;
 699             f[j_coord_offset+DIM*1+ZZ] -= tz;
 700
 701             }
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             if (rsq32<rcutoff2)
 708             {
 709
 710             r32              = rsq32*rinv32;
 711
 712             /* EWALD ELECTROSTATICS */
 713
 714             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 715             ewrt             = r32*ewtabscale;
 716             ewitab           = ewrt;
 717             eweps            = ewrt-ewitab;
 718             ewitab           = 4*ewitab;
 719             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 720             velec            = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 721             felec            = qq32*rinv32*(rinvsq32-felec);
 722
 723             d                = r32-rswitch;
 724             d                = (d>0.0) ? d : 0.0;
 725             d2               = d*d;
 726             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 727
 728             dsw              = d2*(swF2+d*(swF3+d*swF4));
 729
 730             /* Evaluate switch function */
 731             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 732             felec            = felec*sw - rinv32*velec*dsw;
 733             velec           *= sw;
 734
 735             /* Update potential sums from outer loop */
 736             velecsum        += velec;
 737
 738             fscal            = felec;
 739
 740             /* Calculate temporary vectorial force */
 741             tx               = fscal*dx32;
 742             ty               = fscal*dy32;
 743             tz               = fscal*dz32;
 744
 745             /* Update vectorial force */
 746             fix3            += tx;
 747             fiy3            += ty;
 748             fiz3            += tz;
 749             f[j_coord_offset+DIM*2+XX] -= tx;
 750             f[j_coord_offset+DIM*2+YY] -= ty;
 751             f[j_coord_offset+DIM*2+ZZ] -= tz;
 752
 753             }
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             if (rsq33<rcutoff2)
 760             {
 761
 762             r33              = rsq33*rinv33;
 763
 764             /* EWALD ELECTROSTATICS */
 765
 766             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 767             ewrt             = r33*ewtabscale;
 768             ewitab           = ewrt;
 769             eweps            = ewrt-ewitab;
 770             ewitab           = 4*ewitab;
 771             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 772             velec            = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 773             felec            = qq33*rinv33*(rinvsq33-felec);
 774
 775             d                = r33-rswitch;
 776             d                = (d>0.0) ? d : 0.0;
 777             d2               = d*d;
 778             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 779
 780             dsw              = d2*(swF2+d*(swF3+d*swF4));
 781
 782             /* Evaluate switch function */
 783             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 784             felec            = felec*sw - rinv33*velec*dsw;
 785             velec           *= sw;
 786
 787             /* Update potential sums from outer loop */
 788             velecsum        += velec;
 789
 790             fscal            = felec;
 791
 792             /* Calculate temporary vectorial force */
 793             tx               = fscal*dx33;
 794             ty               = fscal*dy33;
 795             tz               = fscal*dz33;
 796
 797             /* Update vectorial force */
 798             fix3            += tx;
 799             fiy3            += ty;
 800             fiz3            += tz;
 801             f[j_coord_offset+DIM*3+XX] -= tx;
 802             f[j_coord_offset+DIM*3+YY] -= ty;
 803             f[j_coord_offset+DIM*3+ZZ] -= tz;
 804
 805             }
 806
 807             /* Inner loop uses 575 flops */
 808         }
 809         /* End of innermost loop */
 810
 811         tx = ty = tz = 0;
 812         f[i_coord_offset+DIM*0+XX] += fix0;
 813         f[i_coord_offset+DIM*0+YY] += fiy0;
 814         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 815         tx                         += fix0;
 816         ty                         += fiy0;
 817         tz                         += fiz0;
 818         f[i_coord_offset+DIM*1+XX] += fix1;
 819         f[i_coord_offset+DIM*1+YY] += fiy1;
 820         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 821         tx                         += fix1;
 822         ty                         += fiy1;
 823         tz                         += fiz1;
 824         f[i_coord_offset+DIM*2+XX] += fix2;
 825         f[i_coord_offset+DIM*2+YY] += fiy2;
 826         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 827         tx                         += fix2;
 828         ty                         += fiy2;
 829         tz                         += fiz2;
 830         f[i_coord_offset+DIM*3+XX] += fix3;
 831         f[i_coord_offset+DIM*3+YY] += fiy3;
 832         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 833         tx                         += fix3;
 834         ty                         += fiy3;
 835         tz                         += fiz3;
 836         fshift[i_shift_offset+XX]  += tx;
 837         fshift[i_shift_offset+YY]  += ty;
 838         fshift[i_shift_offset+ZZ]  += tz;
 839
 840         ggid                        = gid[iidx];
 841         /* Update potential energies */
 842         kernel_data->energygrp_elec[ggid] += velecsum;
 843         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 844
 845         /* Increment number of inner iterations */
 846         inneriter                  += j_index_end - j_index_start;
 847
 848         /* Outer loop uses 41 flops */
 849     }
 850
 851     /* Increment number of outer iterations */
 852     outeriter        += nri;
 853
 854     /* Update outer/inner flops */
 855
 856     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*575);
 857 }
 858 /*
 859  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW4W4_F_c
 860  * Electrostatics interaction: Ewald
 861  * VdW interaction:            LennardJones
 862  * Geometry:                   Water4-Water4
 863  * Calculate force/pot:        Force
 864  */
 865 void
 866 nb_kernel_ElecEwSw_VdwLJSw_GeomW4W4_F_c
 867                     (t_nblist * gmx_restrict                nlist,
 868                      rvec * gmx_restrict                    xx,
 869                      rvec * gmx_restrict                    ff,
 870                      t_forcerec * gmx_restrict              fr,
 871                      t_mdatoms * gmx_restrict               mdatoms,
 872                      nb_kernel_data_t * gmx_restrict        kernel_data,
 873                      t_nrnb * gmx_restrict                  nrnb)
 874 {
 875     int              i_shift_offset,i_coord_offset,j_coord_offset;
 876     int              j_index_start,j_index_end;
 877     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 878     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 879     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 880     real             *shiftvec,*fshift,*x,*f;
 881     int              vdwioffset0;
 882     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 883     int              vdwioffset1;
 884     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 885     int              vdwioffset2;
 886     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 887     int              vdwioffset3;
 888     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 889     int              vdwjidx0;
 890     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 891     int              vdwjidx1;
 892     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 893     int              vdwjidx2;
 894     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 895     int              vdwjidx3;
 896     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 897     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 898     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 899     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 900     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 901     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 902     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 903     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 904     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 905     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 906     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 907     real             velec,felec,velecsum,facel,crf,krf,krf2;
 908     real             *charge;
 909     int              nvdwtype;
 910     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 911     int              *vdwtype;
 912     real             *vdwparam;
 913     int              ewitab;
 914     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 915     real             *ewtab;
 916     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 917
 918     x                = xx[0];
 919     f                = ff[0];
 920
 921     nri              = nlist->nri;
 922     iinr             = nlist->iinr;
 923     jindex           = nlist->jindex;
 924     jjnr             = nlist->jjnr;
 925     shiftidx         = nlist->shift;
 926     gid              = nlist->gid;
 927     shiftvec         = fr->shift_vec[0];
 928     fshift           = fr->fshift[0];
 929     facel            = fr->epsfac;
 930     charge           = mdatoms->chargeA;
 931     nvdwtype         = fr->ntype;
 932     vdwparam         = fr->nbfp;
 933     vdwtype          = mdatoms->typeA;
 934
 935     sh_ewald         = fr->ic->sh_ewald;
 936     ewtab            = fr->ic->tabq_coul_FDV0;
 937     ewtabscale       = fr->ic->tabq_scale;
 938     ewtabhalfspace   = 0.5/ewtabscale;
 939
 940     /* Setup water-specific parameters */
 941     inr              = nlist->iinr[0];
 942     iq1              = facel*charge[inr+1];
 943     iq2              = facel*charge[inr+2];
 944     iq3              = facel*charge[inr+3];
 945     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 946
 947     jq1              = charge[inr+1];
 948     jq2              = charge[inr+2];
 949     jq3              = charge[inr+3];
 950     vdwjidx0         = 2*vdwtype[inr+0];
 951     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 952     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 953     qq11             = iq1*jq1;
 954     qq12             = iq1*jq2;
 955     qq13             = iq1*jq3;
 956     qq21             = iq2*jq1;
 957     qq22             = iq2*jq2;
 958     qq23             = iq2*jq3;
 959     qq31             = iq3*jq1;
 960     qq32             = iq3*jq2;
 961     qq33             = iq3*jq3;
 962
 963     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 964     rcutoff          = fr->rcoulomb;
 965     rcutoff2         = rcutoff*rcutoff;
 966
 967     rswitch          = fr->rcoulomb_switch;
 968     /* Setup switch parameters */
 969     d                = rcutoff-rswitch;
 970     swV3             = -10.0/(d*d*d);
 971     swV4             =  15.0/(d*d*d*d);
 972     swV5             =  -6.0/(d*d*d*d*d);
 973     swF2             = -30.0/(d*d*d);
 974     swF3             =  60.0/(d*d*d*d);
 975     swF4             = -30.0/(d*d*d*d*d);
 976
 977     outeriter        = 0;
 978     inneriter        = 0;
 979
 980     /* Start outer loop over neighborlists */
 981     for(iidx=0; iidx<nri; iidx++)
 982     {
 983         /* Load shift vector for this list */
 984         i_shift_offset   = DIM*shiftidx[iidx];
 985         shX              = shiftvec[i_shift_offset+XX];
 986         shY              = shiftvec[i_shift_offset+YY];
 987         shZ              = shiftvec[i_shift_offset+ZZ];
 988
 989         /* Load limits for loop over neighbors */
 990         j_index_start    = jindex[iidx];
 991         j_index_end      = jindex[iidx+1];
 992
 993         /* Get outer coordinate index */
 994         inr              = iinr[iidx];
 995         i_coord_offset   = DIM*inr;
 996
 997         /* Load i particle coords and add shift vector */
 998         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 999         iy0              = shY + x[i_coord_offset+DIM*0+YY];
1000         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
1001         ix1              = shX + x[i_coord_offset+DIM*1+XX];
1002         iy1              = shY + x[i_coord_offset+DIM*1+YY];
1003         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
1004         ix2              = shX + x[i_coord_offset+DIM*2+XX];
1005         iy2              = shY + x[i_coord_offset+DIM*2+YY];
1006         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
1007         ix3              = shX + x[i_coord_offset+DIM*3+XX];
1008         iy3              = shY + x[i_coord_offset+DIM*3+YY];
1009         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
1010
1011         fix0             = 0.0;
1012         fiy0             = 0.0;
1013         fiz0             = 0.0;
1014         fix1             = 0.0;
1015         fiy1             = 0.0;
1016         fiz1             = 0.0;
1017         fix2             = 0.0;
1018         fiy2             = 0.0;
1019         fiz2             = 0.0;
1020         fix3             = 0.0;
1021         fiy3             = 0.0;
1022         fiz3             = 0.0;
1023
1024         /* Start inner kernel loop */
1025         for(jidx=j_index_start; jidx<j_index_end; jidx++)
1026         {
1027             /* Get j neighbor index, and coordinate index */
1028             jnr              = jjnr[jidx];
1029             j_coord_offset   = DIM*jnr;
1030
1031             /* load j atom coordinates */
1032             jx0              = x[j_coord_offset+DIM*0+XX];
1033             jy0              = x[j_coord_offset+DIM*0+YY];
1034             jz0              = x[j_coord_offset+DIM*0+ZZ];
1035             jx1              = x[j_coord_offset+DIM*1+XX];
1036             jy1              = x[j_coord_offset+DIM*1+YY];
1037             jz1              = x[j_coord_offset+DIM*1+ZZ];
1038             jx2              = x[j_coord_offset+DIM*2+XX];
1039             jy2              = x[j_coord_offset+DIM*2+YY];
1040             jz2              = x[j_coord_offset+DIM*2+ZZ];
1041             jx3              = x[j_coord_offset+DIM*3+XX];
1042             jy3              = x[j_coord_offset+DIM*3+YY];
1043             jz3              = x[j_coord_offset+DIM*3+ZZ];
1044
1045             /* Calculate displacement vector */
1046             dx00             = ix0 - jx0;
1047             dy00             = iy0 - jy0;
1048             dz00             = iz0 - jz0;
1049             dx11             = ix1 - jx1;
1050             dy11             = iy1 - jy1;
1051             dz11             = iz1 - jz1;
1052             dx12             = ix1 - jx2;
1053             dy12             = iy1 - jy2;
1054             dz12             = iz1 - jz2;
1055             dx13             = ix1 - jx3;
1056             dy13             = iy1 - jy3;
1057             dz13             = iz1 - jz3;
1058             dx21             = ix2 - jx1;
1059             dy21             = iy2 - jy1;
1060             dz21             = iz2 - jz1;
1061             dx22             = ix2 - jx2;
1062             dy22             = iy2 - jy2;
1063             dz22             = iz2 - jz2;
1064             dx23             = ix2 - jx3;
1065             dy23             = iy2 - jy3;
1066             dz23             = iz2 - jz3;
1067             dx31             = ix3 - jx1;
1068             dy31             = iy3 - jy1;
1069             dz31             = iz3 - jz1;
1070             dx32             = ix3 - jx2;
1071             dy32             = iy3 - jy2;
1072             dz32             = iz3 - jz2;
1073             dx33             = ix3 - jx3;
1074             dy33             = iy3 - jy3;
1075             dz33             = iz3 - jz3;
1076
1077             /* Calculate squared distance and things based on it */
1078             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
1079             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
1080             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
1081             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
1082             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
1083             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
1084             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
1085             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
1086             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
1087             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
1088
1089             rinv00           = gmx_invsqrt(rsq00);
1090             rinv11           = gmx_invsqrt(rsq11);
1091             rinv12           = gmx_invsqrt(rsq12);
1092             rinv13           = gmx_invsqrt(rsq13);
1093             rinv21           = gmx_invsqrt(rsq21);
1094             rinv22           = gmx_invsqrt(rsq22);
1095             rinv23           = gmx_invsqrt(rsq23);
1096             rinv31           = gmx_invsqrt(rsq31);
1097             rinv32           = gmx_invsqrt(rsq32);
1098             rinv33           = gmx_invsqrt(rsq33);
1099
1100             rinvsq00         = rinv00*rinv00;
1101             rinvsq11         = rinv11*rinv11;
1102             rinvsq12         = rinv12*rinv12;
1103             rinvsq13         = rinv13*rinv13;
1104             rinvsq21         = rinv21*rinv21;
1105             rinvsq22         = rinv22*rinv22;
1106             rinvsq23         = rinv23*rinv23;
1107             rinvsq31         = rinv31*rinv31;
1108             rinvsq32         = rinv32*rinv32;
1109             rinvsq33         = rinv33*rinv33;
1110
1111             /**************************
1112              * CALCULATE INTERACTIONS *
1113              **************************/
1114
1115             if (rsq00<rcutoff2)
1116             {
1117
1118             r00              = rsq00*rinv00;
1119
1120             /* LENNARD-JONES DISPERSION/REPULSION */
1121
1122             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
1123             vvdw6            = c6_00*rinvsix;
1124             vvdw12           = c12_00*rinvsix*rinvsix;
1125             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
1126             fvdw             = (vvdw12-vvdw6)*rinvsq00;
1127
1128             d                = r00-rswitch;
1129             d                = (d>0.0) ? d : 0.0;
1130             d2               = d*d;
1131             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1132
1133             dsw              = d2*(swF2+d*(swF3+d*swF4));
1134
1135             /* Evaluate switch function */
1136             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1137             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
1138
1139             fscal            = fvdw;
1140
1141             /* Calculate temporary vectorial force */
1142             tx               = fscal*dx00;
1143             ty               = fscal*dy00;
1144             tz               = fscal*dz00;
1145
1146             /* Update vectorial force */
1147             fix0            += tx;
1148             fiy0            += ty;
1149             fiz0            += tz;
1150             f[j_coord_offset+DIM*0+XX] -= tx;
1151             f[j_coord_offset+DIM*0+YY] -= ty;
1152             f[j_coord_offset+DIM*0+ZZ] -= tz;
1153
1154             }
1155
1156             /**************************
1157              * CALCULATE INTERACTIONS *
1158              **************************/
1159
1160             if (rsq11<rcutoff2)
1161             {
1162
1163             r11              = rsq11*rinv11;
1164
1165             /* EWALD ELECTROSTATICS */
1166
1167             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1168             ewrt             = r11*ewtabscale;
1169             ewitab           = ewrt;
1170             eweps            = ewrt-ewitab;
1171             ewitab           = 4*ewitab;
1172             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1173             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1174             felec            = qq11*rinv11*(rinvsq11-felec);
1175
1176             d                = r11-rswitch;
1177             d                = (d>0.0) ? d : 0.0;
1178             d2               = d*d;
1179             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1180
1181             dsw              = d2*(swF2+d*(swF3+d*swF4));
1182
1183             /* Evaluate switch function */
1184             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1185             felec            = felec*sw - rinv11*velec*dsw;
1186
1187             fscal            = felec;
1188
1189             /* Calculate temporary vectorial force */
1190             tx               = fscal*dx11;
1191             ty               = fscal*dy11;
1192             tz               = fscal*dz11;
1193
1194             /* Update vectorial force */
1195             fix1            += tx;
1196             fiy1            += ty;
1197             fiz1            += tz;
1198             f[j_coord_offset+DIM*1+XX] -= tx;
1199             f[j_coord_offset+DIM*1+YY] -= ty;
1200             f[j_coord_offset+DIM*1+ZZ] -= tz;
1201
1202             }
1203
1204             /**************************
1205              * CALCULATE INTERACTIONS *
1206              **************************/
1207
1208             if (rsq12<rcutoff2)
1209             {
1210
1211             r12              = rsq12*rinv12;
1212
1213             /* EWALD ELECTROSTATICS */
1214
1215             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1216             ewrt             = r12*ewtabscale;
1217             ewitab           = ewrt;
1218             eweps            = ewrt-ewitab;
1219             ewitab           = 4*ewitab;
1220             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1221             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1222             felec            = qq12*rinv12*(rinvsq12-felec);
1223
1224             d                = r12-rswitch;
1225             d                = (d>0.0) ? d : 0.0;
1226             d2               = d*d;
1227             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1228
1229             dsw              = d2*(swF2+d*(swF3+d*swF4));
1230
1231             /* Evaluate switch function */
1232             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1233             felec            = felec*sw - rinv12*velec*dsw;
1234
1235             fscal            = felec;
1236
1237             /* Calculate temporary vectorial force */
1238             tx               = fscal*dx12;
1239             ty               = fscal*dy12;
1240             tz               = fscal*dz12;
1241
1242             /* Update vectorial force */
1243             fix1            += tx;
1244             fiy1            += ty;
1245             fiz1            += tz;
1246             f[j_coord_offset+DIM*2+XX] -= tx;
1247             f[j_coord_offset+DIM*2+YY] -= ty;
1248             f[j_coord_offset+DIM*2+ZZ] -= tz;
1249
1250             }
1251
1252             /**************************
1253              * CALCULATE INTERACTIONS *
1254              **************************/
1255
1256             if (rsq13<rcutoff2)
1257             {
1258
1259             r13              = rsq13*rinv13;
1260
1261             /* EWALD ELECTROSTATICS */
1262
1263             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1264             ewrt             = r13*ewtabscale;
1265             ewitab           = ewrt;
1266             eweps            = ewrt-ewitab;
1267             ewitab           = 4*ewitab;
1268             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1269             velec            = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1270             felec            = qq13*rinv13*(rinvsq13-felec);
1271
1272             d                = r13-rswitch;
1273             d                = (d>0.0) ? d : 0.0;
1274             d2               = d*d;
1275             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1276
1277             dsw              = d2*(swF2+d*(swF3+d*swF4));
1278
1279             /* Evaluate switch function */
1280             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1281             felec            = felec*sw - rinv13*velec*dsw;
1282
1283             fscal            = felec;
1284
1285             /* Calculate temporary vectorial force */
1286             tx               = fscal*dx13;
1287             ty               = fscal*dy13;
1288             tz               = fscal*dz13;
1289
1290             /* Update vectorial force */
1291             fix1            += tx;
1292             fiy1            += ty;
1293             fiz1            += tz;
1294             f[j_coord_offset+DIM*3+XX] -= tx;
1295             f[j_coord_offset+DIM*3+YY] -= ty;
1296             f[j_coord_offset+DIM*3+ZZ] -= tz;
1297
1298             }
1299
1300             /**************************
1301              * CALCULATE INTERACTIONS *
1302              **************************/
1303
1304             if (rsq21<rcutoff2)
1305             {
1306
1307             r21              = rsq21*rinv21;
1308
1309             /* EWALD ELECTROSTATICS */
1310
1311             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1312             ewrt             = r21*ewtabscale;
1313             ewitab           = ewrt;
1314             eweps            = ewrt-ewitab;
1315             ewitab           = 4*ewitab;
1316             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1317             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1318             felec            = qq21*rinv21*(rinvsq21-felec);
1319
1320             d                = r21-rswitch;
1321             d                = (d>0.0) ? d : 0.0;
1322             d2               = d*d;
1323             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1324
1325             dsw              = d2*(swF2+d*(swF3+d*swF4));
1326
1327             /* Evaluate switch function */
1328             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1329             felec            = felec*sw - rinv21*velec*dsw;
1330
1331             fscal            = felec;
1332
1333             /* Calculate temporary vectorial force */
1334             tx               = fscal*dx21;
1335             ty               = fscal*dy21;
1336             tz               = fscal*dz21;
1337
1338             /* Update vectorial force */
1339             fix2            += tx;
1340             fiy2            += ty;
1341             fiz2            += tz;
1342             f[j_coord_offset+DIM*1+XX] -= tx;
1343             f[j_coord_offset+DIM*1+YY] -= ty;
1344             f[j_coord_offset+DIM*1+ZZ] -= tz;
1345
1346             }
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             if (rsq22<rcutoff2)
1353             {
1354
1355             r22              = rsq22*rinv22;
1356
1357             /* EWALD ELECTROSTATICS */
1358
1359             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1360             ewrt             = r22*ewtabscale;
1361             ewitab           = ewrt;
1362             eweps            = ewrt-ewitab;
1363             ewitab           = 4*ewitab;
1364             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1365             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1366             felec            = qq22*rinv22*(rinvsq22-felec);
1367
1368             d                = r22-rswitch;
1369             d                = (d>0.0) ? d : 0.0;
1370             d2               = d*d;
1371             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1372
1373             dsw              = d2*(swF2+d*(swF3+d*swF4));
1374
1375             /* Evaluate switch function */
1376             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1377             felec            = felec*sw - rinv22*velec*dsw;
1378
1379             fscal            = felec;
1380
1381             /* Calculate temporary vectorial force */
1382             tx               = fscal*dx22;
1383             ty               = fscal*dy22;
1384             tz               = fscal*dz22;
1385
1386             /* Update vectorial force */
1387             fix2            += tx;
1388             fiy2            += ty;
1389             fiz2            += tz;
1390             f[j_coord_offset+DIM*2+XX] -= tx;
1391             f[j_coord_offset+DIM*2+YY] -= ty;
1392             f[j_coord_offset+DIM*2+ZZ] -= tz;
1393
1394             }
1395
1396             /**************************
1397              * CALCULATE INTERACTIONS *
1398              **************************/
1399
1400             if (rsq23<rcutoff2)
1401             {
1402
1403             r23              = rsq23*rinv23;
1404
1405             /* EWALD ELECTROSTATICS */
1406
1407             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1408             ewrt             = r23*ewtabscale;
1409             ewitab           = ewrt;
1410             eweps            = ewrt-ewitab;
1411             ewitab           = 4*ewitab;
1412             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1413             velec            = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1414             felec            = qq23*rinv23*(rinvsq23-felec);
1415
1416             d                = r23-rswitch;
1417             d                = (d>0.0) ? d : 0.0;
1418             d2               = d*d;
1419             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1420
1421             dsw              = d2*(swF2+d*(swF3+d*swF4));
1422
1423             /* Evaluate switch function */
1424             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1425             felec            = felec*sw - rinv23*velec*dsw;
1426
1427             fscal            = felec;
1428
1429             /* Calculate temporary vectorial force */
1430             tx               = fscal*dx23;
1431             ty               = fscal*dy23;
1432             tz               = fscal*dz23;
1433
1434             /* Update vectorial force */
1435             fix2            += tx;
1436             fiy2            += ty;
1437             fiz2            += tz;
1438             f[j_coord_offset+DIM*3+XX] -= tx;
1439             f[j_coord_offset+DIM*3+YY] -= ty;
1440             f[j_coord_offset+DIM*3+ZZ] -= tz;
1441
1442             }
1443
1444             /**************************
1445              * CALCULATE INTERACTIONS *
1446              **************************/
1447
1448             if (rsq31<rcutoff2)
1449             {
1450
1451             r31              = rsq31*rinv31;
1452
1453             /* EWALD ELECTROSTATICS */
1454
1455             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1456             ewrt             = r31*ewtabscale;
1457             ewitab           = ewrt;
1458             eweps            = ewrt-ewitab;
1459             ewitab           = 4*ewitab;
1460             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1461             velec            = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1462             felec            = qq31*rinv31*(rinvsq31-felec);
1463
1464             d                = r31-rswitch;
1465             d                = (d>0.0) ? d : 0.0;
1466             d2               = d*d;
1467             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1468
1469             dsw              = d2*(swF2+d*(swF3+d*swF4));
1470
1471             /* Evaluate switch function */
1472             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1473             felec            = felec*sw - rinv31*velec*dsw;
1474
1475             fscal            = felec;
1476
1477             /* Calculate temporary vectorial force */
1478             tx               = fscal*dx31;
1479             ty               = fscal*dy31;
1480             tz               = fscal*dz31;
1481
1482             /* Update vectorial force */
1483             fix3            += tx;
1484             fiy3            += ty;
1485             fiz3            += tz;
1486             f[j_coord_offset+DIM*1+XX] -= tx;
1487             f[j_coord_offset+DIM*1+YY] -= ty;
1488             f[j_coord_offset+DIM*1+ZZ] -= tz;
1489
1490             }
1491
1492             /**************************
1493              * CALCULATE INTERACTIONS *
1494              **************************/
1495
1496             if (rsq32<rcutoff2)
1497             {
1498
1499             r32              = rsq32*rinv32;
1500
1501             /* EWALD ELECTROSTATICS */
1502
1503             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1504             ewrt             = r32*ewtabscale;
1505             ewitab           = ewrt;
1506             eweps            = ewrt-ewitab;
1507             ewitab           = 4*ewitab;
1508             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1509             velec            = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1510             felec            = qq32*rinv32*(rinvsq32-felec);
1511
1512             d                = r32-rswitch;
1513             d                = (d>0.0) ? d : 0.0;
1514             d2               = d*d;
1515             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1516
1517             dsw              = d2*(swF2+d*(swF3+d*swF4));
1518
1519             /* Evaluate switch function */
1520             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1521             felec            = felec*sw - rinv32*velec*dsw;
1522
1523             fscal            = felec;
1524
1525             /* Calculate temporary vectorial force */
1526             tx               = fscal*dx32;
1527             ty               = fscal*dy32;
1528             tz               = fscal*dz32;
1529
1530             /* Update vectorial force */
1531             fix3            += tx;
1532             fiy3            += ty;
1533             fiz3            += tz;
1534             f[j_coord_offset+DIM*2+XX] -= tx;
1535             f[j_coord_offset+DIM*2+YY] -= ty;
1536             f[j_coord_offset+DIM*2+ZZ] -= tz;
1537
1538             }
1539
1540             /**************************
1541              * CALCULATE INTERACTIONS *
1542              **************************/
1543
1544             if (rsq33<rcutoff2)
1545             {
1546
1547             r33              = rsq33*rinv33;
1548
1549             /* EWALD ELECTROSTATICS */
1550
1551             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1552             ewrt             = r33*ewtabscale;
1553             ewitab           = ewrt;
1554             eweps            = ewrt-ewitab;
1555             ewitab           = 4*ewitab;
1556             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1557             velec            = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1558             felec            = qq33*rinv33*(rinvsq33-felec);
1559
1560             d                = r33-rswitch;
1561             d                = (d>0.0) ? d : 0.0;
1562             d2               = d*d;
1563             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1564
1565             dsw              = d2*(swF2+d*(swF3+d*swF4));
1566
1567             /* Evaluate switch function */
1568             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1569             felec            = felec*sw - rinv33*velec*dsw;
1570
1571             fscal            = felec;
1572
1573             /* Calculate temporary vectorial force */
1574             tx               = fscal*dx33;
1575             ty               = fscal*dy33;
1576             tz               = fscal*dz33;
1577
1578             /* Update vectorial force */
1579             fix3            += tx;
1580             fiy3            += ty;
1581             fiz3            += tz;
1582             f[j_coord_offset+DIM*3+XX] -= tx;
1583             f[j_coord_offset+DIM*3+YY] -= ty;
1584             f[j_coord_offset+DIM*3+ZZ] -= tz;
1585
1586             }
1587
1588             /* Inner loop uses 555 flops */
1589         }
1590         /* End of innermost loop */
1591
1592         tx = ty = tz = 0;
1593         f[i_coord_offset+DIM*0+XX] += fix0;
1594         f[i_coord_offset+DIM*0+YY] += fiy0;
1595         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1596         tx                         += fix0;
1597         ty                         += fiy0;
1598         tz                         += fiz0;
1599         f[i_coord_offset+DIM*1+XX] += fix1;
1600         f[i_coord_offset+DIM*1+YY] += fiy1;
1601         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1602         tx                         += fix1;
1603         ty                         += fiy1;
1604         tz                         += fiz1;
1605         f[i_coord_offset+DIM*2+XX] += fix2;
1606         f[i_coord_offset+DIM*2+YY] += fiy2;
1607         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1608         tx                         += fix2;
1609         ty                         += fiy2;
1610         tz                         += fiz2;
1611         f[i_coord_offset+DIM*3+XX] += fix3;
1612         f[i_coord_offset+DIM*3+YY] += fiy3;
1613         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1614         tx                         += fix3;
1615         ty                         += fiy3;
1616         tz                         += fiz3;
1617         fshift[i_shift_offset+XX]  += tx;
1618         fshift[i_shift_offset+YY]  += ty;
1619         fshift[i_shift_offset+ZZ]  += tz;
1620
1621         /* Increment number of inner iterations */
1622         inneriter                  += j_index_end - j_index_start;
1623
1624         /* Outer loop uses 39 flops */
1625     }
1626
1627     /* Increment number of outer iterations */
1628     outeriter        += nri;
1629
1630     /* Update outer/inner flops */
1631
1632     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*555);
1633 }