src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSw_VdwNone_GeomW4W4_c.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS c kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomW4W4_VF_c
  49  * Electrostatics interaction: Ewald
  50  * VdW interaction:            None
  51  * Geometry:                   Water4-Water4
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecEwSw_VdwNone_GeomW4W4_VF_c
  56                     (t_nblist                    * gmx_restrict       nlist,
  57                      rvec                        * gmx_restrict          xx,
  58                      rvec                        * gmx_restrict          ff,
  59                      t_forcerec                  * gmx_restrict          fr,
  60                      t_mdatoms                   * gmx_restrict     mdatoms,
  61                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  62                      t_nrnb                      * gmx_restrict        nrnb)
  63 {
  64     int              i_shift_offset,i_coord_offset,j_coord_offset;
  65     int              j_index_start,j_index_end;
  66     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  67     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  68     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  69     real             *shiftvec,*fshift,*x,*f;
  70     int              vdwioffset1;
  71     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  72     int              vdwioffset2;
  73     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  74     int              vdwioffset3;
  75     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  76     int              vdwjidx1;
  77     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2;
  79     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3;
  81     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  83     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  84     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
  85     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  86     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  87     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
  88     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
  89     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
  90     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
  91     real             velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              ewitab;
  94     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
  95     real             *ewtab;
  96     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  97
  98     x                = xx[0];
  99     f                = ff[0];
 100
 101     nri              = nlist->nri;
 102     iinr             = nlist->iinr;
 103     jindex           = nlist->jindex;
 104     jjnr             = nlist->jjnr;
 105     shiftidx         = nlist->shift;
 106     gid              = nlist->gid;
 107     shiftvec         = fr->shift_vec[0];
 108     fshift           = fr->fshift[0];
 109     facel            = fr->epsfac;
 110     charge           = mdatoms->chargeA;
 111
 112     sh_ewald         = fr->ic->sh_ewald;
 113     ewtab            = fr->ic->tabq_coul_FDV0;
 114     ewtabscale       = fr->ic->tabq_scale;
 115     ewtabhalfspace   = 0.5/ewtabscale;
 116
 117     /* Setup water-specific parameters */
 118     inr              = nlist->iinr[0];
 119     iq1              = facel*charge[inr+1];
 120     iq2              = facel*charge[inr+2];
 121     iq3              = facel*charge[inr+3];
 122
 123     jq1              = charge[inr+1];
 124     jq2              = charge[inr+2];
 125     jq3              = charge[inr+3];
 126     qq11             = iq1*jq1;
 127     qq12             = iq1*jq2;
 128     qq13             = iq1*jq3;
 129     qq21             = iq2*jq1;
 130     qq22             = iq2*jq2;
 131     qq23             = iq2*jq3;
 132     qq31             = iq3*jq1;
 133     qq32             = iq3*jq2;
 134     qq33             = iq3*jq3;
 135
 136     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 137     rcutoff          = fr->rcoulomb;
 138     rcutoff2         = rcutoff*rcutoff;
 139
 140     rswitch          = fr->rcoulomb_switch;
 141     /* Setup switch parameters */
 142     d                = rcutoff-rswitch;
 143     swV3             = -10.0/(d*d*d);
 144     swV4             =  15.0/(d*d*d*d);
 145     swV5             =  -6.0/(d*d*d*d*d);
 146     swF2             = -30.0/(d*d*d);
 147     swF3             =  60.0/(d*d*d*d);
 148     swF4             = -30.0/(d*d*d*d*d);
 149
 150     outeriter        = 0;
 151     inneriter        = 0;
 152
 153     /* Start outer loop over neighborlists */
 154     for(iidx=0; iidx<nri; iidx++)
 155     {
 156         /* Load shift vector for this list */
 157         i_shift_offset   = DIM*shiftidx[iidx];
 158         shX              = shiftvec[i_shift_offset+XX];
 159         shY              = shiftvec[i_shift_offset+YY];
 160         shZ              = shiftvec[i_shift_offset+ZZ];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 172         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 173         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 174         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 175         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 176         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 177         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 178         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 179         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 180
 181         fix1             = 0.0;
 182         fiy1             = 0.0;
 183         fiz1             = 0.0;
 184         fix2             = 0.0;
 185         fiy2             = 0.0;
 186         fiz2             = 0.0;
 187         fix3             = 0.0;
 188         fiy3             = 0.0;
 189         fiz3             = 0.0;
 190
 191         /* Reset potential sums */
 192         velecsum         = 0.0;
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 196         {
 197             /* Get j neighbor index, and coordinate index */
 198             jnr              = jjnr[jidx];
 199             j_coord_offset   = DIM*jnr;
 200
 201             /* load j atom coordinates */
 202             jx1              = x[j_coord_offset+DIM*1+XX];
 203             jy1              = x[j_coord_offset+DIM*1+YY];
 204             jz1              = x[j_coord_offset+DIM*1+ZZ];
 205             jx2              = x[j_coord_offset+DIM*2+XX];
 206             jy2              = x[j_coord_offset+DIM*2+YY];
 207             jz2              = x[j_coord_offset+DIM*2+ZZ];
 208             jx3              = x[j_coord_offset+DIM*3+XX];
 209             jy3              = x[j_coord_offset+DIM*3+YY];
 210             jz3              = x[j_coord_offset+DIM*3+ZZ];
 211
 212             /* Calculate displacement vector */
 213             dx11             = ix1 - jx1;
 214             dy11             = iy1 - jy1;
 215             dz11             = iz1 - jz1;
 216             dx12             = ix1 - jx2;
 217             dy12             = iy1 - jy2;
 218             dz12             = iz1 - jz2;
 219             dx13             = ix1 - jx3;
 220             dy13             = iy1 - jy3;
 221             dz13             = iz1 - jz3;
 222             dx21             = ix2 - jx1;
 223             dy21             = iy2 - jy1;
 224             dz21             = iz2 - jz1;
 225             dx22             = ix2 - jx2;
 226             dy22             = iy2 - jy2;
 227             dz22             = iz2 - jz2;
 228             dx23             = ix2 - jx3;
 229             dy23             = iy2 - jy3;
 230             dz23             = iz2 - jz3;
 231             dx31             = ix3 - jx1;
 232             dy31             = iy3 - jy1;
 233             dz31             = iz3 - jz1;
 234             dx32             = ix3 - jx2;
 235             dy32             = iy3 - jy2;
 236             dz32             = iz3 - jz2;
 237             dx33             = ix3 - jx3;
 238             dy33             = iy3 - jy3;
 239             dz33             = iz3 - jz3;
 240
 241             /* Calculate squared distance and things based on it */
 242             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 243             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 244             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 245             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 246             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 247             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 248             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 249             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 250             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 251
 252             rinv11           = gmx_invsqrt(rsq11);
 253             rinv12           = gmx_invsqrt(rsq12);
 254             rinv13           = gmx_invsqrt(rsq13);
 255             rinv21           = gmx_invsqrt(rsq21);
 256             rinv22           = gmx_invsqrt(rsq22);
 257             rinv23           = gmx_invsqrt(rsq23);
 258             rinv31           = gmx_invsqrt(rsq31);
 259             rinv32           = gmx_invsqrt(rsq32);
 260             rinv33           = gmx_invsqrt(rsq33);
 261
 262             rinvsq11         = rinv11*rinv11;
 263             rinvsq12         = rinv12*rinv12;
 264             rinvsq13         = rinv13*rinv13;
 265             rinvsq21         = rinv21*rinv21;
 266             rinvsq22         = rinv22*rinv22;
 267             rinvsq23         = rinv23*rinv23;
 268             rinvsq31         = rinv31*rinv31;
 269             rinvsq32         = rinv32*rinv32;
 270             rinvsq33         = rinv33*rinv33;
 271
 272             /**************************
 273              * CALCULATE INTERACTIONS *
 274              **************************/
 275
 276             if (rsq11<rcutoff2)
 277             {
 278
 279             r11              = rsq11*rinv11;
 280
 281             /* EWALD ELECTROSTATICS */
 282
 283             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 284             ewrt             = r11*ewtabscale;
 285             ewitab           = ewrt;
 286             eweps            = ewrt-ewitab;
 287             ewitab           = 4*ewitab;
 288             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 289             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 290             felec            = qq11*rinv11*(rinvsq11-felec);
 291
 292             d                = r11-rswitch;
 293             d                = (d>0.0) ? d : 0.0;
 294             d2               = d*d;
 295             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 296
 297             dsw              = d2*(swF2+d*(swF3+d*swF4));
 298
 299             /* Evaluate switch function */
 300             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 301             felec            = felec*sw - rinv11*velec*dsw;
 302             velec           *= sw;
 303
 304             /* Update potential sums from outer loop */
 305             velecsum        += velec;
 306
 307             fscal            = felec;
 308
 309             /* Calculate temporary vectorial force */
 310             tx               = fscal*dx11;
 311             ty               = fscal*dy11;
 312             tz               = fscal*dz11;
 313
 314             /* Update vectorial force */
 315             fix1            += tx;
 316             fiy1            += ty;
 317             fiz1            += tz;
 318             f[j_coord_offset+DIM*1+XX] -= tx;
 319             f[j_coord_offset+DIM*1+YY] -= ty;
 320             f[j_coord_offset+DIM*1+ZZ] -= tz;
 321
 322             }
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             if (rsq12<rcutoff2)
 329             {
 330
 331             r12              = rsq12*rinv12;
 332
 333             /* EWALD ELECTROSTATICS */
 334
 335             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 336             ewrt             = r12*ewtabscale;
 337             ewitab           = ewrt;
 338             eweps            = ewrt-ewitab;
 339             ewitab           = 4*ewitab;
 340             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 341             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 342             felec            = qq12*rinv12*(rinvsq12-felec);
 343
 344             d                = r12-rswitch;
 345             d                = (d>0.0) ? d : 0.0;
 346             d2               = d*d;
 347             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 348
 349             dsw              = d2*(swF2+d*(swF3+d*swF4));
 350
 351             /* Evaluate switch function */
 352             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 353             felec            = felec*sw - rinv12*velec*dsw;
 354             velec           *= sw;
 355
 356             /* Update potential sums from outer loop */
 357             velecsum        += velec;
 358
 359             fscal            = felec;
 360
 361             /* Calculate temporary vectorial force */
 362             tx               = fscal*dx12;
 363             ty               = fscal*dy12;
 364             tz               = fscal*dz12;
 365
 366             /* Update vectorial force */
 367             fix1            += tx;
 368             fiy1            += ty;
 369             fiz1            += tz;
 370             f[j_coord_offset+DIM*2+XX] -= tx;
 371             f[j_coord_offset+DIM*2+YY] -= ty;
 372             f[j_coord_offset+DIM*2+ZZ] -= tz;
 373
 374             }
 375
 376             /**************************
 377              * CALCULATE INTERACTIONS *
 378              **************************/
 379
 380             if (rsq13<rcutoff2)
 381             {
 382
 383             r13              = rsq13*rinv13;
 384
 385             /* EWALD ELECTROSTATICS */
 386
 387             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 388             ewrt             = r13*ewtabscale;
 389             ewitab           = ewrt;
 390             eweps            = ewrt-ewitab;
 391             ewitab           = 4*ewitab;
 392             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 393             velec            = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 394             felec            = qq13*rinv13*(rinvsq13-felec);
 395
 396             d                = r13-rswitch;
 397             d                = (d>0.0) ? d : 0.0;
 398             d2               = d*d;
 399             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 400
 401             dsw              = d2*(swF2+d*(swF3+d*swF4));
 402
 403             /* Evaluate switch function */
 404             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 405             felec            = felec*sw - rinv13*velec*dsw;
 406             velec           *= sw;
 407
 408             /* Update potential sums from outer loop */
 409             velecsum        += velec;
 410
 411             fscal            = felec;
 412
 413             /* Calculate temporary vectorial force */
 414             tx               = fscal*dx13;
 415             ty               = fscal*dy13;
 416             tz               = fscal*dz13;
 417
 418             /* Update vectorial force */
 419             fix1            += tx;
 420             fiy1            += ty;
 421             fiz1            += tz;
 422             f[j_coord_offset+DIM*3+XX] -= tx;
 423             f[j_coord_offset+DIM*3+YY] -= ty;
 424             f[j_coord_offset+DIM*3+ZZ] -= tz;
 425
 426             }
 427
 428             /**************************
 429              * CALCULATE INTERACTIONS *
 430              **************************/
 431
 432             if (rsq21<rcutoff2)
 433             {
 434
 435             r21              = rsq21*rinv21;
 436
 437             /* EWALD ELECTROSTATICS */
 438
 439             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 440             ewrt             = r21*ewtabscale;
 441             ewitab           = ewrt;
 442             eweps            = ewrt-ewitab;
 443             ewitab           = 4*ewitab;
 444             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 445             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 446             felec            = qq21*rinv21*(rinvsq21-felec);
 447
 448             d                = r21-rswitch;
 449             d                = (d>0.0) ? d : 0.0;
 450             d2               = d*d;
 451             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 452
 453             dsw              = d2*(swF2+d*(swF3+d*swF4));
 454
 455             /* Evaluate switch function */
 456             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 457             felec            = felec*sw - rinv21*velec*dsw;
 458             velec           *= sw;
 459
 460             /* Update potential sums from outer loop */
 461             velecsum        += velec;
 462
 463             fscal            = felec;
 464
 465             /* Calculate temporary vectorial force */
 466             tx               = fscal*dx21;
 467             ty               = fscal*dy21;
 468             tz               = fscal*dz21;
 469
 470             /* Update vectorial force */
 471             fix2            += tx;
 472             fiy2            += ty;
 473             fiz2            += tz;
 474             f[j_coord_offset+DIM*1+XX] -= tx;
 475             f[j_coord_offset+DIM*1+YY] -= ty;
 476             f[j_coord_offset+DIM*1+ZZ] -= tz;
 477
 478             }
 479
 480             /**************************
 481              * CALCULATE INTERACTIONS *
 482              **************************/
 483
 484             if (rsq22<rcutoff2)
 485             {
 486
 487             r22              = rsq22*rinv22;
 488
 489             /* EWALD ELECTROSTATICS */
 490
 491             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 492             ewrt             = r22*ewtabscale;
 493             ewitab           = ewrt;
 494             eweps            = ewrt-ewitab;
 495             ewitab           = 4*ewitab;
 496             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 497             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 498             felec            = qq22*rinv22*(rinvsq22-felec);
 499
 500             d                = r22-rswitch;
 501             d                = (d>0.0) ? d : 0.0;
 502             d2               = d*d;
 503             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 504
 505             dsw              = d2*(swF2+d*(swF3+d*swF4));
 506
 507             /* Evaluate switch function */
 508             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 509             felec            = felec*sw - rinv22*velec*dsw;
 510             velec           *= sw;
 511
 512             /* Update potential sums from outer loop */
 513             velecsum        += velec;
 514
 515             fscal            = felec;
 516
 517             /* Calculate temporary vectorial force */
 518             tx               = fscal*dx22;
 519             ty               = fscal*dy22;
 520             tz               = fscal*dz22;
 521
 522             /* Update vectorial force */
 523             fix2            += tx;
 524             fiy2            += ty;
 525             fiz2            += tz;
 526             f[j_coord_offset+DIM*2+XX] -= tx;
 527             f[j_coord_offset+DIM*2+YY] -= ty;
 528             f[j_coord_offset+DIM*2+ZZ] -= tz;
 529
 530             }
 531
 532             /**************************
 533              * CALCULATE INTERACTIONS *
 534              **************************/
 535
 536             if (rsq23<rcutoff2)
 537             {
 538
 539             r23              = rsq23*rinv23;
 540
 541             /* EWALD ELECTROSTATICS */
 542
 543             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 544             ewrt             = r23*ewtabscale;
 545             ewitab           = ewrt;
 546             eweps            = ewrt-ewitab;
 547             ewitab           = 4*ewitab;
 548             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 549             velec            = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 550             felec            = qq23*rinv23*(rinvsq23-felec);
 551
 552             d                = r23-rswitch;
 553             d                = (d>0.0) ? d : 0.0;
 554             d2               = d*d;
 555             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 556
 557             dsw              = d2*(swF2+d*(swF3+d*swF4));
 558
 559             /* Evaluate switch function */
 560             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 561             felec            = felec*sw - rinv23*velec*dsw;
 562             velec           *= sw;
 563
 564             /* Update potential sums from outer loop */
 565             velecsum        += velec;
 566
 567             fscal            = felec;
 568
 569             /* Calculate temporary vectorial force */
 570             tx               = fscal*dx23;
 571             ty               = fscal*dy23;
 572             tz               = fscal*dz23;
 573
 574             /* Update vectorial force */
 575             fix2            += tx;
 576             fiy2            += ty;
 577             fiz2            += tz;
 578             f[j_coord_offset+DIM*3+XX] -= tx;
 579             f[j_coord_offset+DIM*3+YY] -= ty;
 580             f[j_coord_offset+DIM*3+ZZ] -= tz;
 581
 582             }
 583
 584             /**************************
 585              * CALCULATE INTERACTIONS *
 586              **************************/
 587
 588             if (rsq31<rcutoff2)
 589             {
 590
 591             r31              = rsq31*rinv31;
 592
 593             /* EWALD ELECTROSTATICS */
 594
 595             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 596             ewrt             = r31*ewtabscale;
 597             ewitab           = ewrt;
 598             eweps            = ewrt-ewitab;
 599             ewitab           = 4*ewitab;
 600             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 601             velec            = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 602             felec            = qq31*rinv31*(rinvsq31-felec);
 603
 604             d                = r31-rswitch;
 605             d                = (d>0.0) ? d : 0.0;
 606             d2               = d*d;
 607             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 608
 609             dsw              = d2*(swF2+d*(swF3+d*swF4));
 610
 611             /* Evaluate switch function */
 612             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 613             felec            = felec*sw - rinv31*velec*dsw;
 614             velec           *= sw;
 615
 616             /* Update potential sums from outer loop */
 617             velecsum        += velec;
 618
 619             fscal            = felec;
 620
 621             /* Calculate temporary vectorial force */
 622             tx               = fscal*dx31;
 623             ty               = fscal*dy31;
 624             tz               = fscal*dz31;
 625
 626             /* Update vectorial force */
 627             fix3            += tx;
 628             fiy3            += ty;
 629             fiz3            += tz;
 630             f[j_coord_offset+DIM*1+XX] -= tx;
 631             f[j_coord_offset+DIM*1+YY] -= ty;
 632             f[j_coord_offset+DIM*1+ZZ] -= tz;
 633
 634             }
 635
 636             /**************************
 637              * CALCULATE INTERACTIONS *
 638              **************************/
 639
 640             if (rsq32<rcutoff2)
 641             {
 642
 643             r32              = rsq32*rinv32;
 644
 645             /* EWALD ELECTROSTATICS */
 646
 647             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 648             ewrt             = r32*ewtabscale;
 649             ewitab           = ewrt;
 650             eweps            = ewrt-ewitab;
 651             ewitab           = 4*ewitab;
 652             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 653             velec            = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 654             felec            = qq32*rinv32*(rinvsq32-felec);
 655
 656             d                = r32-rswitch;
 657             d                = (d>0.0) ? d : 0.0;
 658             d2               = d*d;
 659             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 660
 661             dsw              = d2*(swF2+d*(swF3+d*swF4));
 662
 663             /* Evaluate switch function */
 664             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 665             felec            = felec*sw - rinv32*velec*dsw;
 666             velec           *= sw;
 667
 668             /* Update potential sums from outer loop */
 669             velecsum        += velec;
 670
 671             fscal            = felec;
 672
 673             /* Calculate temporary vectorial force */
 674             tx               = fscal*dx32;
 675             ty               = fscal*dy32;
 676             tz               = fscal*dz32;
 677
 678             /* Update vectorial force */
 679             fix3            += tx;
 680             fiy3            += ty;
 681             fiz3            += tz;
 682             f[j_coord_offset+DIM*2+XX] -= tx;
 683             f[j_coord_offset+DIM*2+YY] -= ty;
 684             f[j_coord_offset+DIM*2+ZZ] -= tz;
 685
 686             }
 687
 688             /**************************
 689              * CALCULATE INTERACTIONS *
 690              **************************/
 691
 692             if (rsq33<rcutoff2)
 693             {
 694
 695             r33              = rsq33*rinv33;
 696
 697             /* EWALD ELECTROSTATICS */
 698
 699             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 700             ewrt             = r33*ewtabscale;
 701             ewitab           = ewrt;
 702             eweps            = ewrt-ewitab;
 703             ewitab           = 4*ewitab;
 704             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 705             velec            = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 706             felec            = qq33*rinv33*(rinvsq33-felec);
 707
 708             d                = r33-rswitch;
 709             d                = (d>0.0) ? d : 0.0;
 710             d2               = d*d;
 711             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 712
 713             dsw              = d2*(swF2+d*(swF3+d*swF4));
 714
 715             /* Evaluate switch function */
 716             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 717             felec            = felec*sw - rinv33*velec*dsw;
 718             velec           *= sw;
 719
 720             /* Update potential sums from outer loop */
 721             velecsum        += velec;
 722
 723             fscal            = felec;
 724
 725             /* Calculate temporary vectorial force */
 726             tx               = fscal*dx33;
 727             ty               = fscal*dy33;
 728             tz               = fscal*dz33;
 729
 730             /* Update vectorial force */
 731             fix3            += tx;
 732             fiy3            += ty;
 733             fiz3            += tz;
 734             f[j_coord_offset+DIM*3+XX] -= tx;
 735             f[j_coord_offset+DIM*3+YY] -= ty;
 736             f[j_coord_offset+DIM*3+ZZ] -= tz;
 737
 738             }
 739
 740             /* Inner loop uses 522 flops */
 741         }
 742         /* End of innermost loop */
 743
 744         tx = ty = tz = 0;
 745         f[i_coord_offset+DIM*1+XX] += fix1;
 746         f[i_coord_offset+DIM*1+YY] += fiy1;
 747         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 748         tx                         += fix1;
 749         ty                         += fiy1;
 750         tz                         += fiz1;
 751         f[i_coord_offset+DIM*2+XX] += fix2;
 752         f[i_coord_offset+DIM*2+YY] += fiy2;
 753         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 754         tx                         += fix2;
 755         ty                         += fiy2;
 756         tz                         += fiz2;
 757         f[i_coord_offset+DIM*3+XX] += fix3;
 758         f[i_coord_offset+DIM*3+YY] += fiy3;
 759         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 760         tx                         += fix3;
 761         ty                         += fiy3;
 762         tz                         += fiz3;
 763         fshift[i_shift_offset+XX]  += tx;
 764         fshift[i_shift_offset+YY]  += ty;
 765         fshift[i_shift_offset+ZZ]  += tz;
 766
 767         ggid                        = gid[iidx];
 768         /* Update potential energies */
 769         kernel_data->energygrp_elec[ggid] += velecsum;
 770
 771         /* Increment number of inner iterations */
 772         inneriter                  += j_index_end - j_index_start;
 773
 774         /* Outer loop uses 31 flops */
 775     }
 776
 777     /* Increment number of outer iterations */
 778     outeriter        += nri;
 779
 780     /* Update outer/inner flops */
 781
 782     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*31 + inneriter*522);
 783 }
 784 /*
 785  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomW4W4_F_c
 786  * Electrostatics interaction: Ewald
 787  * VdW interaction:            None
 788  * Geometry:                   Water4-Water4
 789  * Calculate force/pot:        Force
 790  */
 791 void
 792 nb_kernel_ElecEwSw_VdwNone_GeomW4W4_F_c
 793                     (t_nblist                    * gmx_restrict       nlist,
 794                      rvec                        * gmx_restrict          xx,
 795                      rvec                        * gmx_restrict          ff,
 796                      t_forcerec                  * gmx_restrict          fr,
 797                      t_mdatoms                   * gmx_restrict     mdatoms,
 798                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 799                      t_nrnb                      * gmx_restrict        nrnb)
 800 {
 801     int              i_shift_offset,i_coord_offset,j_coord_offset;
 802     int              j_index_start,j_index_end;
 803     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 804     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 805     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 806     real             *shiftvec,*fshift,*x,*f;
 807     int              vdwioffset1;
 808     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 809     int              vdwioffset2;
 810     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 811     int              vdwioffset3;
 812     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 813     int              vdwjidx1;
 814     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 815     int              vdwjidx2;
 816     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 817     int              vdwjidx3;
 818     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 819     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 820     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 821     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 822     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 823     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 824     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 825     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 826     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 827     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 828     real             velec,felec,velecsum,facel,crf,krf,krf2;
 829     real             *charge;
 830     int              ewitab;
 831     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 832     real             *ewtab;
 833     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 834
 835     x                = xx[0];
 836     f                = ff[0];
 837
 838     nri              = nlist->nri;
 839     iinr             = nlist->iinr;
 840     jindex           = nlist->jindex;
 841     jjnr             = nlist->jjnr;
 842     shiftidx         = nlist->shift;
 843     gid              = nlist->gid;
 844     shiftvec         = fr->shift_vec[0];
 845     fshift           = fr->fshift[0];
 846     facel            = fr->epsfac;
 847     charge           = mdatoms->chargeA;
 848
 849     sh_ewald         = fr->ic->sh_ewald;
 850     ewtab            = fr->ic->tabq_coul_FDV0;
 851     ewtabscale       = fr->ic->tabq_scale;
 852     ewtabhalfspace   = 0.5/ewtabscale;
 853
 854     /* Setup water-specific parameters */
 855     inr              = nlist->iinr[0];
 856     iq1              = facel*charge[inr+1];
 857     iq2              = facel*charge[inr+2];
 858     iq3              = facel*charge[inr+3];
 859
 860     jq1              = charge[inr+1];
 861     jq2              = charge[inr+2];
 862     jq3              = charge[inr+3];
 863     qq11             = iq1*jq1;
 864     qq12             = iq1*jq2;
 865     qq13             = iq1*jq3;
 866     qq21             = iq2*jq1;
 867     qq22             = iq2*jq2;
 868     qq23             = iq2*jq3;
 869     qq31             = iq3*jq1;
 870     qq32             = iq3*jq2;
 871     qq33             = iq3*jq3;
 872
 873     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 874     rcutoff          = fr->rcoulomb;
 875     rcutoff2         = rcutoff*rcutoff;
 876
 877     rswitch          = fr->rcoulomb_switch;
 878     /* Setup switch parameters */
 879     d                = rcutoff-rswitch;
 880     swV3             = -10.0/(d*d*d);
 881     swV4             =  15.0/(d*d*d*d);
 882     swV5             =  -6.0/(d*d*d*d*d);
 883     swF2             = -30.0/(d*d*d);
 884     swF3             =  60.0/(d*d*d*d);
 885     swF4             = -30.0/(d*d*d*d*d);
 886
 887     outeriter        = 0;
 888     inneriter        = 0;
 889
 890     /* Start outer loop over neighborlists */
 891     for(iidx=0; iidx<nri; iidx++)
 892     {
 893         /* Load shift vector for this list */
 894         i_shift_offset   = DIM*shiftidx[iidx];
 895         shX              = shiftvec[i_shift_offset+XX];
 896         shY              = shiftvec[i_shift_offset+YY];
 897         shZ              = shiftvec[i_shift_offset+ZZ];
 898
 899         /* Load limits for loop over neighbors */
 900         j_index_start    = jindex[iidx];
 901         j_index_end      = jindex[iidx+1];
 902
 903         /* Get outer coordinate index */
 904         inr              = iinr[iidx];
 905         i_coord_offset   = DIM*inr;
 906
 907         /* Load i particle coords and add shift vector */
 908         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 909         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 910         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 911         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 912         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 913         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 914         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 915         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 916         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 917
 918         fix1             = 0.0;
 919         fiy1             = 0.0;
 920         fiz1             = 0.0;
 921         fix2             = 0.0;
 922         fiy2             = 0.0;
 923         fiz2             = 0.0;
 924         fix3             = 0.0;
 925         fiy3             = 0.0;
 926         fiz3             = 0.0;
 927
 928         /* Start inner kernel loop */
 929         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 930         {
 931             /* Get j neighbor index, and coordinate index */
 932             jnr              = jjnr[jidx];
 933             j_coord_offset   = DIM*jnr;
 934
 935             /* load j atom coordinates */
 936             jx1              = x[j_coord_offset+DIM*1+XX];
 937             jy1              = x[j_coord_offset+DIM*1+YY];
 938             jz1              = x[j_coord_offset+DIM*1+ZZ];
 939             jx2              = x[j_coord_offset+DIM*2+XX];
 940             jy2              = x[j_coord_offset+DIM*2+YY];
 941             jz2              = x[j_coord_offset+DIM*2+ZZ];
 942             jx3              = x[j_coord_offset+DIM*3+XX];
 943             jy3              = x[j_coord_offset+DIM*3+YY];
 944             jz3              = x[j_coord_offset+DIM*3+ZZ];
 945
 946             /* Calculate displacement vector */
 947             dx11             = ix1 - jx1;
 948             dy11             = iy1 - jy1;
 949             dz11             = iz1 - jz1;
 950             dx12             = ix1 - jx2;
 951             dy12             = iy1 - jy2;
 952             dz12             = iz1 - jz2;
 953             dx13             = ix1 - jx3;
 954             dy13             = iy1 - jy3;
 955             dz13             = iz1 - jz3;
 956             dx21             = ix2 - jx1;
 957             dy21             = iy2 - jy1;
 958             dz21             = iz2 - jz1;
 959             dx22             = ix2 - jx2;
 960             dy22             = iy2 - jy2;
 961             dz22             = iz2 - jz2;
 962             dx23             = ix2 - jx3;
 963             dy23             = iy2 - jy3;
 964             dz23             = iz2 - jz3;
 965             dx31             = ix3 - jx1;
 966             dy31             = iy3 - jy1;
 967             dz31             = iz3 - jz1;
 968             dx32             = ix3 - jx2;
 969             dy32             = iy3 - jy2;
 970             dz32             = iz3 - jz2;
 971             dx33             = ix3 - jx3;
 972             dy33             = iy3 - jy3;
 973             dz33             = iz3 - jz3;
 974
 975             /* Calculate squared distance and things based on it */
 976             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 977             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 978             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 979             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 980             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 981             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 982             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 983             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 984             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 985
 986             rinv11           = gmx_invsqrt(rsq11);
 987             rinv12           = gmx_invsqrt(rsq12);
 988             rinv13           = gmx_invsqrt(rsq13);
 989             rinv21           = gmx_invsqrt(rsq21);
 990             rinv22           = gmx_invsqrt(rsq22);
 991             rinv23           = gmx_invsqrt(rsq23);
 992             rinv31           = gmx_invsqrt(rsq31);
 993             rinv32           = gmx_invsqrt(rsq32);
 994             rinv33           = gmx_invsqrt(rsq33);
 995
 996             rinvsq11         = rinv11*rinv11;
 997             rinvsq12         = rinv12*rinv12;
 998             rinvsq13         = rinv13*rinv13;
 999             rinvsq21         = rinv21*rinv21;
1000             rinvsq22         = rinv22*rinv22;
1001             rinvsq23         = rinv23*rinv23;
1002             rinvsq31         = rinv31*rinv31;
1003             rinvsq32         = rinv32*rinv32;
1004             rinvsq33         = rinv33*rinv33;
1005
1006             /**************************
1007              * CALCULATE INTERACTIONS *
1008              **************************/
1009
1010             if (rsq11<rcutoff2)
1011             {
1012
1013             r11              = rsq11*rinv11;
1014
1015             /* EWALD ELECTROSTATICS */
1016
1017             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1018             ewrt             = r11*ewtabscale;
1019             ewitab           = ewrt;
1020             eweps            = ewrt-ewitab;
1021             ewitab           = 4*ewitab;
1022             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1023             velec            = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1024             felec            = qq11*rinv11*(rinvsq11-felec);
1025
1026             d                = r11-rswitch;
1027             d                = (d>0.0) ? d : 0.0;
1028             d2               = d*d;
1029             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1030
1031             dsw              = d2*(swF2+d*(swF3+d*swF4));
1032
1033             /* Evaluate switch function */
1034             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1035             felec            = felec*sw - rinv11*velec*dsw;
1036
1037             fscal            = felec;
1038
1039             /* Calculate temporary vectorial force */
1040             tx               = fscal*dx11;
1041             ty               = fscal*dy11;
1042             tz               = fscal*dz11;
1043
1044             /* Update vectorial force */
1045             fix1            += tx;
1046             fiy1            += ty;
1047             fiz1            += tz;
1048             f[j_coord_offset+DIM*1+XX] -= tx;
1049             f[j_coord_offset+DIM*1+YY] -= ty;
1050             f[j_coord_offset+DIM*1+ZZ] -= tz;
1051
1052             }
1053
1054             /**************************
1055              * CALCULATE INTERACTIONS *
1056              **************************/
1057
1058             if (rsq12<rcutoff2)
1059             {
1060
1061             r12              = rsq12*rinv12;
1062
1063             /* EWALD ELECTROSTATICS */
1064
1065             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1066             ewrt             = r12*ewtabscale;
1067             ewitab           = ewrt;
1068             eweps            = ewrt-ewitab;
1069             ewitab           = 4*ewitab;
1070             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1071             velec            = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1072             felec            = qq12*rinv12*(rinvsq12-felec);
1073
1074             d                = r12-rswitch;
1075             d                = (d>0.0) ? d : 0.0;
1076             d2               = d*d;
1077             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1078
1079             dsw              = d2*(swF2+d*(swF3+d*swF4));
1080
1081             /* Evaluate switch function */
1082             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1083             felec            = felec*sw - rinv12*velec*dsw;
1084
1085             fscal            = felec;
1086
1087             /* Calculate temporary vectorial force */
1088             tx               = fscal*dx12;
1089             ty               = fscal*dy12;
1090             tz               = fscal*dz12;
1091
1092             /* Update vectorial force */
1093             fix1            += tx;
1094             fiy1            += ty;
1095             fiz1            += tz;
1096             f[j_coord_offset+DIM*2+XX] -= tx;
1097             f[j_coord_offset+DIM*2+YY] -= ty;
1098             f[j_coord_offset+DIM*2+ZZ] -= tz;
1099
1100             }
1101
1102             /**************************
1103              * CALCULATE INTERACTIONS *
1104              **************************/
1105
1106             if (rsq13<rcutoff2)
1107             {
1108
1109             r13              = rsq13*rinv13;
1110
1111             /* EWALD ELECTROSTATICS */
1112
1113             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1114             ewrt             = r13*ewtabscale;
1115             ewitab           = ewrt;
1116             eweps            = ewrt-ewitab;
1117             ewitab           = 4*ewitab;
1118             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1119             velec            = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1120             felec            = qq13*rinv13*(rinvsq13-felec);
1121
1122             d                = r13-rswitch;
1123             d                = (d>0.0) ? d : 0.0;
1124             d2               = d*d;
1125             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1126
1127             dsw              = d2*(swF2+d*(swF3+d*swF4));
1128
1129             /* Evaluate switch function */
1130             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1131             felec            = felec*sw - rinv13*velec*dsw;
1132
1133             fscal            = felec;
1134
1135             /* Calculate temporary vectorial force */
1136             tx               = fscal*dx13;
1137             ty               = fscal*dy13;
1138             tz               = fscal*dz13;
1139
1140             /* Update vectorial force */
1141             fix1            += tx;
1142             fiy1            += ty;
1143             fiz1            += tz;
1144             f[j_coord_offset+DIM*3+XX] -= tx;
1145             f[j_coord_offset+DIM*3+YY] -= ty;
1146             f[j_coord_offset+DIM*3+ZZ] -= tz;
1147
1148             }
1149
1150             /**************************
1151              * CALCULATE INTERACTIONS *
1152              **************************/
1153
1154             if (rsq21<rcutoff2)
1155             {
1156
1157             r21              = rsq21*rinv21;
1158
1159             /* EWALD ELECTROSTATICS */
1160
1161             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1162             ewrt             = r21*ewtabscale;
1163             ewitab           = ewrt;
1164             eweps            = ewrt-ewitab;
1165             ewitab           = 4*ewitab;
1166             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1167             velec            = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1168             felec            = qq21*rinv21*(rinvsq21-felec);
1169
1170             d                = r21-rswitch;
1171             d                = (d>0.0) ? d : 0.0;
1172             d2               = d*d;
1173             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1174
1175             dsw              = d2*(swF2+d*(swF3+d*swF4));
1176
1177             /* Evaluate switch function */
1178             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1179             felec            = felec*sw - rinv21*velec*dsw;
1180
1181             fscal            = felec;
1182
1183             /* Calculate temporary vectorial force */
1184             tx               = fscal*dx21;
1185             ty               = fscal*dy21;
1186             tz               = fscal*dz21;
1187
1188             /* Update vectorial force */
1189             fix2            += tx;
1190             fiy2            += ty;
1191             fiz2            += tz;
1192             f[j_coord_offset+DIM*1+XX] -= tx;
1193             f[j_coord_offset+DIM*1+YY] -= ty;
1194             f[j_coord_offset+DIM*1+ZZ] -= tz;
1195
1196             }
1197
1198             /**************************
1199              * CALCULATE INTERACTIONS *
1200              **************************/
1201
1202             if (rsq22<rcutoff2)
1203             {
1204
1205             r22              = rsq22*rinv22;
1206
1207             /* EWALD ELECTROSTATICS */
1208
1209             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1210             ewrt             = r22*ewtabscale;
1211             ewitab           = ewrt;
1212             eweps            = ewrt-ewitab;
1213             ewitab           = 4*ewitab;
1214             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1215             velec            = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1216             felec            = qq22*rinv22*(rinvsq22-felec);
1217
1218             d                = r22-rswitch;
1219             d                = (d>0.0) ? d : 0.0;
1220             d2               = d*d;
1221             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1222
1223             dsw              = d2*(swF2+d*(swF3+d*swF4));
1224
1225             /* Evaluate switch function */
1226             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1227             felec            = felec*sw - rinv22*velec*dsw;
1228
1229             fscal            = felec;
1230
1231             /* Calculate temporary vectorial force */
1232             tx               = fscal*dx22;
1233             ty               = fscal*dy22;
1234             tz               = fscal*dz22;
1235
1236             /* Update vectorial force */
1237             fix2            += tx;
1238             fiy2            += ty;
1239             fiz2            += tz;
1240             f[j_coord_offset+DIM*2+XX] -= tx;
1241             f[j_coord_offset+DIM*2+YY] -= ty;
1242             f[j_coord_offset+DIM*2+ZZ] -= tz;
1243
1244             }
1245
1246             /**************************
1247              * CALCULATE INTERACTIONS *
1248              **************************/
1249
1250             if (rsq23<rcutoff2)
1251             {
1252
1253             r23              = rsq23*rinv23;
1254
1255             /* EWALD ELECTROSTATICS */
1256
1257             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1258             ewrt             = r23*ewtabscale;
1259             ewitab           = ewrt;
1260             eweps            = ewrt-ewitab;
1261             ewitab           = 4*ewitab;
1262             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1263             velec            = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1264             felec            = qq23*rinv23*(rinvsq23-felec);
1265
1266             d                = r23-rswitch;
1267             d                = (d>0.0) ? d : 0.0;
1268             d2               = d*d;
1269             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1270
1271             dsw              = d2*(swF2+d*(swF3+d*swF4));
1272
1273             /* Evaluate switch function */
1274             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1275             felec            = felec*sw - rinv23*velec*dsw;
1276
1277             fscal            = felec;
1278
1279             /* Calculate temporary vectorial force */
1280             tx               = fscal*dx23;
1281             ty               = fscal*dy23;
1282             tz               = fscal*dz23;
1283
1284             /* Update vectorial force */
1285             fix2            += tx;
1286             fiy2            += ty;
1287             fiz2            += tz;
1288             f[j_coord_offset+DIM*3+XX] -= tx;
1289             f[j_coord_offset+DIM*3+YY] -= ty;
1290             f[j_coord_offset+DIM*3+ZZ] -= tz;
1291
1292             }
1293
1294             /**************************
1295              * CALCULATE INTERACTIONS *
1296              **************************/
1297
1298             if (rsq31<rcutoff2)
1299             {
1300
1301             r31              = rsq31*rinv31;
1302
1303             /* EWALD ELECTROSTATICS */
1304
1305             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1306             ewrt             = r31*ewtabscale;
1307             ewitab           = ewrt;
1308             eweps            = ewrt-ewitab;
1309             ewitab           = 4*ewitab;
1310             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1311             velec            = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1312             felec            = qq31*rinv31*(rinvsq31-felec);
1313
1314             d                = r31-rswitch;
1315             d                = (d>0.0) ? d : 0.0;
1316             d2               = d*d;
1317             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1318
1319             dsw              = d2*(swF2+d*(swF3+d*swF4));
1320
1321             /* Evaluate switch function */
1322             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1323             felec            = felec*sw - rinv31*velec*dsw;
1324
1325             fscal            = felec;
1326
1327             /* Calculate temporary vectorial force */
1328             tx               = fscal*dx31;
1329             ty               = fscal*dy31;
1330             tz               = fscal*dz31;
1331
1332             /* Update vectorial force */
1333             fix3            += tx;
1334             fiy3            += ty;
1335             fiz3            += tz;
1336             f[j_coord_offset+DIM*1+XX] -= tx;
1337             f[j_coord_offset+DIM*1+YY] -= ty;
1338             f[j_coord_offset+DIM*1+ZZ] -= tz;
1339
1340             }
1341
1342             /**************************
1343              * CALCULATE INTERACTIONS *
1344              **************************/
1345
1346             if (rsq32<rcutoff2)
1347             {
1348
1349             r32              = rsq32*rinv32;
1350
1351             /* EWALD ELECTROSTATICS */
1352
1353             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1354             ewrt             = r32*ewtabscale;
1355             ewitab           = ewrt;
1356             eweps            = ewrt-ewitab;
1357             ewitab           = 4*ewitab;
1358             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1359             velec            = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1360             felec            = qq32*rinv32*(rinvsq32-felec);
1361
1362             d                = r32-rswitch;
1363             d                = (d>0.0) ? d : 0.0;
1364             d2               = d*d;
1365             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1366
1367             dsw              = d2*(swF2+d*(swF3+d*swF4));
1368
1369             /* Evaluate switch function */
1370             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1371             felec            = felec*sw - rinv32*velec*dsw;
1372
1373             fscal            = felec;
1374
1375             /* Calculate temporary vectorial force */
1376             tx               = fscal*dx32;
1377             ty               = fscal*dy32;
1378             tz               = fscal*dz32;
1379
1380             /* Update vectorial force */
1381             fix3            += tx;
1382             fiy3            += ty;
1383             fiz3            += tz;
1384             f[j_coord_offset+DIM*2+XX] -= tx;
1385             f[j_coord_offset+DIM*2+YY] -= ty;
1386             f[j_coord_offset+DIM*2+ZZ] -= tz;
1387
1388             }
1389
1390             /**************************
1391              * CALCULATE INTERACTIONS *
1392              **************************/
1393
1394             if (rsq33<rcutoff2)
1395             {
1396
1397             r33              = rsq33*rinv33;
1398
1399             /* EWALD ELECTROSTATICS */
1400
1401             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1402             ewrt             = r33*ewtabscale;
1403             ewitab           = ewrt;
1404             eweps            = ewrt-ewitab;
1405             ewitab           = 4*ewitab;
1406             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
1407             velec            = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
1408             felec            = qq33*rinv33*(rinvsq33-felec);
1409
1410             d                = r33-rswitch;
1411             d                = (d>0.0) ? d : 0.0;
1412             d2               = d*d;
1413             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
1414
1415             dsw              = d2*(swF2+d*(swF3+d*swF4));
1416
1417             /* Evaluate switch function */
1418             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1419             felec            = felec*sw - rinv33*velec*dsw;
1420
1421             fscal            = felec;
1422
1423             /* Calculate temporary vectorial force */
1424             tx               = fscal*dx33;
1425             ty               = fscal*dy33;
1426             tz               = fscal*dz33;
1427
1428             /* Update vectorial force */
1429             fix3            += tx;
1430             fiy3            += ty;
1431             fiz3            += tz;
1432             f[j_coord_offset+DIM*3+XX] -= tx;
1433             f[j_coord_offset+DIM*3+YY] -= ty;
1434             f[j_coord_offset+DIM*3+ZZ] -= tz;
1435
1436             }
1437
1438             /* Inner loop uses 504 flops */
1439         }
1440         /* End of innermost loop */
1441
1442         tx = ty = tz = 0;
1443         f[i_coord_offset+DIM*1+XX] += fix1;
1444         f[i_coord_offset+DIM*1+YY] += fiy1;
1445         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1446         tx                         += fix1;
1447         ty                         += fiy1;
1448         tz                         += fiz1;
1449         f[i_coord_offset+DIM*2+XX] += fix2;
1450         f[i_coord_offset+DIM*2+YY] += fiy2;
1451         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1452         tx                         += fix2;
1453         ty                         += fiy2;
1454         tz                         += fiz2;
1455         f[i_coord_offset+DIM*3+XX] += fix3;
1456         f[i_coord_offset+DIM*3+YY] += fiy3;
1457         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1458         tx                         += fix3;
1459         ty                         += fiy3;
1460         tz                         += fiz3;
1461         fshift[i_shift_offset+XX]  += tx;
1462         fshift[i_shift_offset+YY]  += ty;
1463         fshift[i_shift_offset+ZZ]  += tz;
1464
1465         /* Increment number of inner iterations */
1466         inneriter                  += j_index_end - j_index_start;
1467
1468         /* Outer loop uses 30 flops */
1469     }
1470
1471     /* Increment number of outer iterations */
1472     outeriter        += nri;
1473
1474     /* Update outer/inner flops */
1475
1476     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*30 + inneriter*504);
1477 }