src/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSw_VdwBhamSw_GeomW3P1_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_VdwBhamSw_GeomW3P1_VF_c
  35  * Electrostatics interaction: Ewald
  36  * VdW interaction:            Buckingham
  37  * Geometry:                   Water3-Particle
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecEwSw_VdwBhamSw_GeomW3P1_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              vdwjidx0;
  63     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  64     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  65     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  66     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  67     real             velec,felec,velecsum,facel,crf,krf,krf2;
  68     real             *charge;
  69     int              nvdwtype;
  70     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  71     int              *vdwtype;
  72     real             *vdwparam;
  73     int              ewitab;
  74     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
  75     real             *ewtab;
  76     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  77
  78     x                = xx[0];
  79     f                = ff[0];
  80
  81     nri              = nlist->nri;
  82     iinr             = nlist->iinr;
  83     jindex           = nlist->jindex;
  84     jjnr             = nlist->jjnr;
  85     shiftidx         = nlist->shift;
  86     gid              = nlist->gid;
  87     shiftvec         = fr->shift_vec[0];
  88     fshift           = fr->fshift[0];
  89     facel            = fr->epsfac;
  90     charge           = mdatoms->chargeA;
  91     nvdwtype         = fr->ntype;
  92     vdwparam         = fr->nbfp;
  93     vdwtype          = mdatoms->typeA;
  94
  95     sh_ewald         = fr->ic->sh_ewald;
  96     ewtab            = fr->ic->tabq_coul_FDV0;
  97     ewtabscale       = fr->ic->tabq_scale;
  98     ewtabhalfspace   = 0.5/ewtabscale;
  99
 100     /* Setup water-specific parameters */
 101     inr              = nlist->iinr[0];
 102     iq0              = facel*charge[inr+0];
 103     iq1              = facel*charge[inr+1];
 104     iq2              = facel*charge[inr+2];
 105     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 106
 107     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 108     rcutoff          = fr->rcoulomb;
 109     rcutoff2         = rcutoff*rcutoff;
 110
 111     rswitch          = fr->rcoulomb_switch;
 112     /* Setup switch parameters */
 113     d                = rcutoff-rswitch;
 114     swV3             = -10.0/(d*d*d);
 115     swV4             =  15.0/(d*d*d*d);
 116     swV5             =  -6.0/(d*d*d*d*d);
 117     swF2             = -30.0/(d*d*d);
 118     swF3             =  60.0/(d*d*d*d);
 119     swF4             = -30.0/(d*d*d*d*d);
 120
 121     outeriter        = 0;
 122     inneriter        = 0;
 123
 124     /* Start outer loop over neighborlists */
 125     for(iidx=0; iidx<nri; iidx++)
 126     {
 127         /* Load shift vector for this list */
 128         i_shift_offset   = DIM*shiftidx[iidx];
 129         shX              = shiftvec[i_shift_offset+XX];
 130         shY              = shiftvec[i_shift_offset+YY];
 131         shZ              = shiftvec[i_shift_offset+ZZ];
 132
 133         /* Load limits for loop over neighbors */
 134         j_index_start    = jindex[iidx];
 135         j_index_end      = jindex[iidx+1];
 136
 137         /* Get outer coordinate index */
 138         inr              = iinr[iidx];
 139         i_coord_offset   = DIM*inr;
 140
 141         /* Load i particle coords and add shift vector */
 142         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 143         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 144         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 145         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 146         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 147         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 148         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 149         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 150         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 151
 152         fix0             = 0.0;
 153         fiy0             = 0.0;
 154         fiz0             = 0.0;
 155         fix1             = 0.0;
 156         fiy1             = 0.0;
 157         fiz1             = 0.0;
 158         fix2             = 0.0;
 159         fiy2             = 0.0;
 160         fiz2             = 0.0;
 161
 162         /* Reset potential sums */
 163         velecsum         = 0.0;
 164         vvdwsum          = 0.0;
 165
 166         /* Start inner kernel loop */
 167         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 168         {
 169             /* Get j neighbor index, and coordinate index */
 170             jnr              = jjnr[jidx];
 171             j_coord_offset   = DIM*jnr;
 172
 173             /* load j atom coordinates */
 174             jx0              = x[j_coord_offset+DIM*0+XX];
 175             jy0              = x[j_coord_offset+DIM*0+YY];
 176             jz0              = x[j_coord_offset+DIM*0+ZZ];
 177
 178             /* Calculate displacement vector */
 179             dx00             = ix0 - jx0;
 180             dy00             = iy0 - jy0;
 181             dz00             = iz0 - jz0;
 182             dx10             = ix1 - jx0;
 183             dy10             = iy1 - jy0;
 184             dz10             = iz1 - jz0;
 185             dx20             = ix2 - jx0;
 186             dy20             = iy2 - jy0;
 187             dz20             = iz2 - jz0;
 188
 189             /* Calculate squared distance and things based on it */
 190             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 191             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 192             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 193
 194             rinv00           = gmx_invsqrt(rsq00);
 195             rinv10           = gmx_invsqrt(rsq10);
 196             rinv20           = gmx_invsqrt(rsq20);
 197
 198             rinvsq00         = rinv00*rinv00;
 199             rinvsq10         = rinv10*rinv10;
 200             rinvsq20         = rinv20*rinv20;
 201
 202             /* Load parameters for j particles */
 203             jq0              = charge[jnr+0];
 204             vdwjidx0         = 3*vdwtype[jnr+0];
 205
 206             /**************************
 207              * CALCULATE INTERACTIONS *
 208              **************************/
 209
 210             if (rsq00<rcutoff2)
 211             {
 212
 213             r00              = rsq00*rinv00;
 214
 215             qq00             = iq0*jq0;
 216             c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 217             cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 218             cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 219
 220             /* EWALD ELECTROSTATICS */
 221
 222             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 223             ewrt             = r00*ewtabscale;
 224             ewitab           = ewrt;
 225             eweps            = ewrt-ewitab;
 226             ewitab           = 4*ewitab;
 227             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 228             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 229             felec            = qq00*rinv00*(rinvsq00-felec);
 230
 231             /* BUCKINGHAM DISPERSION/REPULSION */
 232             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 233             vvdw6            = c6_00*rinvsix;
 234             br               = cexp2_00*r00;
 235             vvdwexp          = cexp1_00*exp(-br);
 236             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
 237             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 238
 239             d                = r00-rswitch;
 240             d                = (d>0.0) ? d : 0.0;
 241             d2               = d*d;
 242             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 243
 244             dsw              = d2*(swF2+d*(swF3+d*swF4));
 245
 246             /* Evaluate switch function */
 247             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 248             felec            = felec*sw - rinv00*velec*dsw;
 249             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 250             velec           *= sw;
 251             vvdw            *= sw;
 252
 253             /* Update potential sums from outer loop */
 254             velecsum        += velec;
 255             vvdwsum         += vvdw;
 256
 257             fscal            = felec+fvdw;
 258
 259             /* Calculate temporary vectorial force */
 260             tx               = fscal*dx00;
 261             ty               = fscal*dy00;
 262             tz               = fscal*dz00;
 263
 264             /* Update vectorial force */
 265             fix0            += tx;
 266             fiy0            += ty;
 267             fiz0            += tz;
 268             f[j_coord_offset+DIM*0+XX] -= tx;
 269             f[j_coord_offset+DIM*0+YY] -= ty;
 270             f[j_coord_offset+DIM*0+ZZ] -= tz;
 271
 272             }
 273
 274             /**************************
 275              * CALCULATE INTERACTIONS *
 276              **************************/
 277
 278             if (rsq10<rcutoff2)
 279             {
 280
 281             r10              = rsq10*rinv10;
 282
 283             qq10             = iq1*jq0;
 284
 285             /* EWALD ELECTROSTATICS */
 286
 287             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 288             ewrt             = r10*ewtabscale;
 289             ewitab           = ewrt;
 290             eweps            = ewrt-ewitab;
 291             ewitab           = 4*ewitab;
 292             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 293             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 294             felec            = qq10*rinv10*(rinvsq10-felec);
 295
 296             d                = r10-rswitch;
 297             d                = (d>0.0) ? d : 0.0;
 298             d2               = d*d;
 299             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 300
 301             dsw              = d2*(swF2+d*(swF3+d*swF4));
 302
 303             /* Evaluate switch function */
 304             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 305             felec            = felec*sw - rinv10*velec*dsw;
 306             velec           *= sw;
 307
 308             /* Update potential sums from outer loop */
 309             velecsum        += velec;
 310
 311             fscal            = felec;
 312
 313             /* Calculate temporary vectorial force */
 314             tx               = fscal*dx10;
 315             ty               = fscal*dy10;
 316             tz               = fscal*dz10;
 317
 318             /* Update vectorial force */
 319             fix1            += tx;
 320             fiy1            += ty;
 321             fiz1            += tz;
 322             f[j_coord_offset+DIM*0+XX] -= tx;
 323             f[j_coord_offset+DIM*0+YY] -= ty;
 324             f[j_coord_offset+DIM*0+ZZ] -= tz;
 325
 326             }
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             if (rsq20<rcutoff2)
 333             {
 334
 335             r20              = rsq20*rinv20;
 336
 337             qq20             = iq2*jq0;
 338
 339             /* EWALD ELECTROSTATICS */
 340
 341             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 342             ewrt             = r20*ewtabscale;
 343             ewitab           = ewrt;
 344             eweps            = ewrt-ewitab;
 345             ewitab           = 4*ewitab;
 346             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 347             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 348             felec            = qq20*rinv20*(rinvsq20-felec);
 349
 350             d                = r20-rswitch;
 351             d                = (d>0.0) ? d : 0.0;
 352             d2               = d*d;
 353             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 354
 355             dsw              = d2*(swF2+d*(swF3+d*swF4));
 356
 357             /* Evaluate switch function */
 358             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 359             felec            = felec*sw - rinv20*velec*dsw;
 360             velec           *= sw;
 361
 362             /* Update potential sums from outer loop */
 363             velecsum        += velec;
 364
 365             fscal            = felec;
 366
 367             /* Calculate temporary vectorial force */
 368             tx               = fscal*dx20;
 369             ty               = fscal*dy20;
 370             tz               = fscal*dz20;
 371
 372             /* Update vectorial force */
 373             fix2            += tx;
 374             fiy2            += ty;
 375             fiz2            += tz;
 376             f[j_coord_offset+DIM*0+XX] -= tx;
 377             f[j_coord_offset+DIM*0+YY] -= ty;
 378             f[j_coord_offset+DIM*0+ZZ] -= tz;
 379
 380             }
 381
 382             /* Inner loop uses 219 flops */
 383         }
 384         /* End of innermost loop */
 385
 386         tx = ty = tz = 0;
 387         f[i_coord_offset+DIM*0+XX] += fix0;
 388         f[i_coord_offset+DIM*0+YY] += fiy0;
 389         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 390         tx                         += fix0;
 391         ty                         += fiy0;
 392         tz                         += fiz0;
 393         f[i_coord_offset+DIM*1+XX] += fix1;
 394         f[i_coord_offset+DIM*1+YY] += fiy1;
 395         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 396         tx                         += fix1;
 397         ty                         += fiy1;
 398         tz                         += fiz1;
 399         f[i_coord_offset+DIM*2+XX] += fix2;
 400         f[i_coord_offset+DIM*2+YY] += fiy2;
 401         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 402         tx                         += fix2;
 403         ty                         += fiy2;
 404         tz                         += fiz2;
 405         fshift[i_shift_offset+XX]  += tx;
 406         fshift[i_shift_offset+YY]  += ty;
 407         fshift[i_shift_offset+ZZ]  += tz;
 408
 409         ggid                        = gid[iidx];
 410         /* Update potential energies */
 411         kernel_data->energygrp_elec[ggid] += velecsum;
 412         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 413
 414         /* Increment number of inner iterations */
 415         inneriter                  += j_index_end - j_index_start;
 416
 417         /* Outer loop uses 32 flops */
 418     }
 419
 420     /* Increment number of outer iterations */
 421     outeriter        += nri;
 422
 423     /* Update outer/inner flops */
 424
 425     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*32 + inneriter*219);
 426 }
 427 /*
 428  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwBhamSw_GeomW3P1_F_c
 429  * Electrostatics interaction: Ewald
 430  * VdW interaction:            Buckingham
 431  * Geometry:                   Water3-Particle
 432  * Calculate force/pot:        Force
 433  */
 434 void
 435 nb_kernel_ElecEwSw_VdwBhamSw_GeomW3P1_F_c
 436                     (t_nblist * gmx_restrict                nlist,
 437                      rvec * gmx_restrict                    xx,
 438                      rvec * gmx_restrict                    ff,
 439                      t_forcerec * gmx_restrict              fr,
 440                      t_mdatoms * gmx_restrict               mdatoms,
 441                      nb_kernel_data_t * gmx_restrict        kernel_data,
 442                      t_nrnb * gmx_restrict                  nrnb)
 443 {
 444     int              i_shift_offset,i_coord_offset,j_coord_offset;
 445     int              j_index_start,j_index_end;
 446     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 447     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 448     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 449     real             *shiftvec,*fshift,*x,*f;
 450     int              vdwioffset0;
 451     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 452     int              vdwioffset1;
 453     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 454     int              vdwioffset2;
 455     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 456     int              vdwjidx0;
 457     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 458     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 459     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 460     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 461     real             velec,felec,velecsum,facel,crf,krf,krf2;
 462     real             *charge;
 463     int              nvdwtype;
 464     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 465     int              *vdwtype;
 466     real             *vdwparam;
 467     int              ewitab;
 468     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 469     real             *ewtab;
 470     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 471
 472     x                = xx[0];
 473     f                = ff[0];
 474
 475     nri              = nlist->nri;
 476     iinr             = nlist->iinr;
 477     jindex           = nlist->jindex;
 478     jjnr             = nlist->jjnr;
 479     shiftidx         = nlist->shift;
 480     gid              = nlist->gid;
 481     shiftvec         = fr->shift_vec[0];
 482     fshift           = fr->fshift[0];
 483     facel            = fr->epsfac;
 484     charge           = mdatoms->chargeA;
 485     nvdwtype         = fr->ntype;
 486     vdwparam         = fr->nbfp;
 487     vdwtype          = mdatoms->typeA;
 488
 489     sh_ewald         = fr->ic->sh_ewald;
 490     ewtab            = fr->ic->tabq_coul_FDV0;
 491     ewtabscale       = fr->ic->tabq_scale;
 492     ewtabhalfspace   = 0.5/ewtabscale;
 493
 494     /* Setup water-specific parameters */
 495     inr              = nlist->iinr[0];
 496     iq0              = facel*charge[inr+0];
 497     iq1              = facel*charge[inr+1];
 498     iq2              = facel*charge[inr+2];
 499     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 500
 501     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 502     rcutoff          = fr->rcoulomb;
 503     rcutoff2         = rcutoff*rcutoff;
 504
 505     rswitch          = fr->rcoulomb_switch;
 506     /* Setup switch parameters */
 507     d                = rcutoff-rswitch;
 508     swV3             = -10.0/(d*d*d);
 509     swV4             =  15.0/(d*d*d*d);
 510     swV5             =  -6.0/(d*d*d*d*d);
 511     swF2             = -30.0/(d*d*d);
 512     swF3             =  60.0/(d*d*d*d);
 513     swF4             = -30.0/(d*d*d*d*d);
 514
 515     outeriter        = 0;
 516     inneriter        = 0;
 517
 518     /* Start outer loop over neighborlists */
 519     for(iidx=0; iidx<nri; iidx++)
 520     {
 521         /* Load shift vector for this list */
 522         i_shift_offset   = DIM*shiftidx[iidx];
 523         shX              = shiftvec[i_shift_offset+XX];
 524         shY              = shiftvec[i_shift_offset+YY];
 525         shZ              = shiftvec[i_shift_offset+ZZ];
 526
 527         /* Load limits for loop over neighbors */
 528         j_index_start    = jindex[iidx];
 529         j_index_end      = jindex[iidx+1];
 530
 531         /* Get outer coordinate index */
 532         inr              = iinr[iidx];
 533         i_coord_offset   = DIM*inr;
 534
 535         /* Load i particle coords and add shift vector */
 536         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 537         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 538         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 539         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 540         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 541         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 542         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 543         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 544         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 545
 546         fix0             = 0.0;
 547         fiy0             = 0.0;
 548         fiz0             = 0.0;
 549         fix1             = 0.0;
 550         fiy1             = 0.0;
 551         fiz1             = 0.0;
 552         fix2             = 0.0;
 553         fiy2             = 0.0;
 554         fiz2             = 0.0;
 555
 556         /* Start inner kernel loop */
 557         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 558         {
 559             /* Get j neighbor index, and coordinate index */
 560             jnr              = jjnr[jidx];
 561             j_coord_offset   = DIM*jnr;
 562
 563             /* load j atom coordinates */
 564             jx0              = x[j_coord_offset+DIM*0+XX];
 565             jy0              = x[j_coord_offset+DIM*0+YY];
 566             jz0              = x[j_coord_offset+DIM*0+ZZ];
 567
 568             /* Calculate displacement vector */
 569             dx00             = ix0 - jx0;
 570             dy00             = iy0 - jy0;
 571             dz00             = iz0 - jz0;
 572             dx10             = ix1 - jx0;
 573             dy10             = iy1 - jy0;
 574             dz10             = iz1 - jz0;
 575             dx20             = ix2 - jx0;
 576             dy20             = iy2 - jy0;
 577             dz20             = iz2 - jz0;
 578
 579             /* Calculate squared distance and things based on it */
 580             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 581             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 582             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 583
 584             rinv00           = gmx_invsqrt(rsq00);
 585             rinv10           = gmx_invsqrt(rsq10);
 586             rinv20           = gmx_invsqrt(rsq20);
 587
 588             rinvsq00         = rinv00*rinv00;
 589             rinvsq10         = rinv10*rinv10;
 590             rinvsq20         = rinv20*rinv20;
 591
 592             /* Load parameters for j particles */
 593             jq0              = charge[jnr+0];
 594             vdwjidx0         = 3*vdwtype[jnr+0];
 595
 596             /**************************
 597              * CALCULATE INTERACTIONS *
 598              **************************/
 599
 600             if (rsq00<rcutoff2)
 601             {
 602
 603             r00              = rsq00*rinv00;
 604
 605             qq00             = iq0*jq0;
 606             c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 607             cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 608             cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 609
 610             /* EWALD ELECTROSTATICS */
 611
 612             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 613             ewrt             = r00*ewtabscale;
 614             ewitab           = ewrt;
 615             eweps            = ewrt-ewitab;
 616             ewitab           = 4*ewitab;
 617             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 618             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 619             felec            = qq00*rinv00*(rinvsq00-felec);
 620
 621             /* BUCKINGHAM DISPERSION/REPULSION */
 622             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 623             vvdw6            = c6_00*rinvsix;
 624             br               = cexp2_00*r00;
 625             vvdwexp          = cexp1_00*exp(-br);
 626             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
 627             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 628
 629             d                = r00-rswitch;
 630             d                = (d>0.0) ? d : 0.0;
 631             d2               = d*d;
 632             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 633
 634             dsw              = d2*(swF2+d*(swF3+d*swF4));
 635
 636             /* Evaluate switch function */
 637             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 638             felec            = felec*sw - rinv00*velec*dsw;
 639             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 640
 641             fscal            = felec+fvdw;
 642
 643             /* Calculate temporary vectorial force */
 644             tx               = fscal*dx00;
 645             ty               = fscal*dy00;
 646             tz               = fscal*dz00;
 647
 648             /* Update vectorial force */
 649             fix0            += tx;
 650             fiy0            += ty;
 651             fiz0            += tz;
 652             f[j_coord_offset+DIM*0+XX] -= tx;
 653             f[j_coord_offset+DIM*0+YY] -= ty;
 654             f[j_coord_offset+DIM*0+ZZ] -= tz;
 655
 656             }
 657
 658             /**************************
 659              * CALCULATE INTERACTIONS *
 660              **************************/
 661
 662             if (rsq10<rcutoff2)
 663             {
 664
 665             r10              = rsq10*rinv10;
 666
 667             qq10             = iq1*jq0;
 668
 669             /* EWALD ELECTROSTATICS */
 670
 671             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 672             ewrt             = r10*ewtabscale;
 673             ewitab           = ewrt;
 674             eweps            = ewrt-ewitab;
 675             ewitab           = 4*ewitab;
 676             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 677             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 678             felec            = qq10*rinv10*(rinvsq10-felec);
 679
 680             d                = r10-rswitch;
 681             d                = (d>0.0) ? d : 0.0;
 682             d2               = d*d;
 683             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 684
 685             dsw              = d2*(swF2+d*(swF3+d*swF4));
 686
 687             /* Evaluate switch function */
 688             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 689             felec            = felec*sw - rinv10*velec*dsw;
 690
 691             fscal            = felec;
 692
 693             /* Calculate temporary vectorial force */
 694             tx               = fscal*dx10;
 695             ty               = fscal*dy10;
 696             tz               = fscal*dz10;
 697
 698             /* Update vectorial force */
 699             fix1            += tx;
 700             fiy1            += ty;
 701             fiz1            += tz;
 702             f[j_coord_offset+DIM*0+XX] -= tx;
 703             f[j_coord_offset+DIM*0+YY] -= ty;
 704             f[j_coord_offset+DIM*0+ZZ] -= tz;
 705
 706             }
 707
 708             /**************************
 709              * CALCULATE INTERACTIONS *
 710              **************************/
 711
 712             if (rsq20<rcutoff2)
 713             {
 714
 715             r20              = rsq20*rinv20;
 716
 717             qq20             = iq2*jq0;
 718
 719             /* EWALD ELECTROSTATICS */
 720
 721             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 722             ewrt             = r20*ewtabscale;
 723             ewitab           = ewrt;
 724             eweps            = ewrt-ewitab;
 725             ewitab           = 4*ewitab;
 726             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 727             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 728             felec            = qq20*rinv20*(rinvsq20-felec);
 729
 730             d                = r20-rswitch;
 731             d                = (d>0.0) ? d : 0.0;
 732             d2               = d*d;
 733             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 734
 735             dsw              = d2*(swF2+d*(swF3+d*swF4));
 736
 737             /* Evaluate switch function */
 738             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 739             felec            = felec*sw - rinv20*velec*dsw;
 740
 741             fscal            = felec;
 742
 743             /* Calculate temporary vectorial force */
 744             tx               = fscal*dx20;
 745             ty               = fscal*dy20;
 746             tz               = fscal*dz20;
 747
 748             /* Update vectorial force */
 749             fix2            += tx;
 750             fiy2            += ty;
 751             fiz2            += tz;
 752             f[j_coord_offset+DIM*0+XX] -= tx;
 753             f[j_coord_offset+DIM*0+YY] -= ty;
 754             f[j_coord_offset+DIM*0+ZZ] -= tz;
 755
 756             }
 757
 758             /* Inner loop uses 211 flops */
 759         }
 760         /* End of innermost loop */
 761
 762         tx = ty = tz = 0;
 763         f[i_coord_offset+DIM*0+XX] += fix0;
 764         f[i_coord_offset+DIM*0+YY] += fiy0;
 765         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 766         tx                         += fix0;
 767         ty                         += fiy0;
 768         tz                         += fiz0;
 769         f[i_coord_offset+DIM*1+XX] += fix1;
 770         f[i_coord_offset+DIM*1+YY] += fiy1;
 771         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 772         tx                         += fix1;
 773         ty                         += fiy1;
 774         tz                         += fiz1;
 775         f[i_coord_offset+DIM*2+XX] += fix2;
 776         f[i_coord_offset+DIM*2+YY] += fiy2;
 777         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 778         tx                         += fix2;
 779         ty                         += fiy2;
 780         tz                         += fiz2;
 781         fshift[i_shift_offset+XX]  += tx;
 782         fshift[i_shift_offset+YY]  += ty;
 783         fshift[i_shift_offset+ZZ]  += tz;
 784
 785         /* Increment number of inner iterations */
 786         inneriter                  += j_index_end - j_index_start;
 787
 788         /* Outer loop uses 30 flops */
 789     }
 790
 791     /* Increment number of outer iterations */
 792     outeriter        += nri;
 793
 794     /* Update outer/inner flops */
 795
 796     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*30 + inneriter*211);
 797 }