src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSw_VdwLJSw_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_VdwLJSw_GeomW3P1_VF_c
  35  * Electrostatics interaction: Ewald
  36  * VdW interaction:            LennardJones
  37  * Geometry:                   Water3-Particle
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecEwSw_VdwLJSw_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      = 2*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         = 2*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             c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 218
 219             /* EWALD ELECTROSTATICS */
 220
 221             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 222             ewrt             = r00*ewtabscale;
 223             ewitab           = ewrt;
 224             eweps            = ewrt-ewitab;
 225             ewitab           = 4*ewitab;
 226             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 227             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 228             felec            = qq00*rinv00*(rinvsq00-felec);
 229
 230             /* LENNARD-JONES DISPERSION/REPULSION */
 231
 232             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 233             vvdw6            = c6_00*rinvsix;
 234             vvdw12           = c12_00*rinvsix*rinvsix;
 235             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 236             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 237
 238             d                = r00-rswitch;
 239             d                = (d>0.0) ? d : 0.0;
 240             d2               = d*d;
 241             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 242
 243             dsw              = d2*(swF2+d*(swF3+d*swF4));
 244
 245             /* Evaluate switch function */
 246             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 247             felec            = felec*sw - rinv00*velec*dsw;
 248             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 249             velec           *= sw;
 250             vvdw            *= sw;
 251
 252             /* Update potential sums from outer loop */
 253             velecsum        += velec;
 254             vvdwsum         += vvdw;
 255
 256             fscal            = felec+fvdw;
 257
 258             /* Calculate temporary vectorial force */
 259             tx               = fscal*dx00;
 260             ty               = fscal*dy00;
 261             tz               = fscal*dz00;
 262
 263             /* Update vectorial force */
 264             fix0            += tx;
 265             fiy0            += ty;
 266             fiz0            += tz;
 267             f[j_coord_offset+DIM*0+XX] -= tx;
 268             f[j_coord_offset+DIM*0+YY] -= ty;
 269             f[j_coord_offset+DIM*0+ZZ] -= tz;
 270
 271             }
 272
 273             /**************************
 274              * CALCULATE INTERACTIONS *
 275              **************************/
 276
 277             if (rsq10<rcutoff2)
 278             {
 279
 280             r10              = rsq10*rinv10;
 281
 282             qq10             = iq1*jq0;
 283
 284             /* EWALD ELECTROSTATICS */
 285
 286             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 287             ewrt             = r10*ewtabscale;
 288             ewitab           = ewrt;
 289             eweps            = ewrt-ewitab;
 290             ewitab           = 4*ewitab;
 291             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 292             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 293             felec            = qq10*rinv10*(rinvsq10-felec);
 294
 295             d                = r10-rswitch;
 296             d                = (d>0.0) ? d : 0.0;
 297             d2               = d*d;
 298             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 299
 300             dsw              = d2*(swF2+d*(swF3+d*swF4));
 301
 302             /* Evaluate switch function */
 303             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 304             felec            = felec*sw - rinv10*velec*dsw;
 305             velec           *= sw;
 306
 307             /* Update potential sums from outer loop */
 308             velecsum        += velec;
 309
 310             fscal            = felec;
 311
 312             /* Calculate temporary vectorial force */
 313             tx               = fscal*dx10;
 314             ty               = fscal*dy10;
 315             tz               = fscal*dz10;
 316
 317             /* Update vectorial force */
 318             fix1            += tx;
 319             fiy1            += ty;
 320             fiz1            += tz;
 321             f[j_coord_offset+DIM*0+XX] -= tx;
 322             f[j_coord_offset+DIM*0+YY] -= ty;
 323             f[j_coord_offset+DIM*0+ZZ] -= tz;
 324
 325             }
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             if (rsq20<rcutoff2)
 332             {
 333
 334             r20              = rsq20*rinv20;
 335
 336             qq20             = iq2*jq0;
 337
 338             /* EWALD ELECTROSTATICS */
 339
 340             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 341             ewrt             = r20*ewtabscale;
 342             ewitab           = ewrt;
 343             eweps            = ewrt-ewitab;
 344             ewitab           = 4*ewitab;
 345             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 346             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 347             felec            = qq20*rinv20*(rinvsq20-felec);
 348
 349             d                = r20-rswitch;
 350             d                = (d>0.0) ? d : 0.0;
 351             d2               = d*d;
 352             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 353
 354             dsw              = d2*(swF2+d*(swF3+d*swF4));
 355
 356             /* Evaluate switch function */
 357             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 358             felec            = felec*sw - rinv20*velec*dsw;
 359             velec           *= sw;
 360
 361             /* Update potential sums from outer loop */
 362             velecsum        += velec;
 363
 364             fscal            = felec;
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = fscal*dx20;
 368             ty               = fscal*dy20;
 369             tz               = fscal*dz20;
 370
 371             /* Update vectorial force */
 372             fix2            += tx;
 373             fiy2            += ty;
 374             fiz2            += tz;
 375             f[j_coord_offset+DIM*0+XX] -= tx;
 376             f[j_coord_offset+DIM*0+YY] -= ty;
 377             f[j_coord_offset+DIM*0+ZZ] -= tz;
 378
 379             }
 380
 381             /* Inner loop uses 193 flops */
 382         }
 383         /* End of innermost loop */
 384
 385         tx = ty = tz = 0;
 386         f[i_coord_offset+DIM*0+XX] += fix0;
 387         f[i_coord_offset+DIM*0+YY] += fiy0;
 388         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 389         tx                         += fix0;
 390         ty                         += fiy0;
 391         tz                         += fiz0;
 392         f[i_coord_offset+DIM*1+XX] += fix1;
 393         f[i_coord_offset+DIM*1+YY] += fiy1;
 394         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 395         tx                         += fix1;
 396         ty                         += fiy1;
 397         tz                         += fiz1;
 398         f[i_coord_offset+DIM*2+XX] += fix2;
 399         f[i_coord_offset+DIM*2+YY] += fiy2;
 400         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 401         tx                         += fix2;
 402         ty                         += fiy2;
 403         tz                         += fiz2;
 404         fshift[i_shift_offset+XX]  += tx;
 405         fshift[i_shift_offset+YY]  += ty;
 406         fshift[i_shift_offset+ZZ]  += tz;
 407
 408         ggid                        = gid[iidx];
 409         /* Update potential energies */
 410         kernel_data->energygrp_elec[ggid] += velecsum;
 411         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 412
 413         /* Increment number of inner iterations */
 414         inneriter                  += j_index_end - j_index_start;
 415
 416         /* Outer loop uses 32 flops */
 417     }
 418
 419     /* Increment number of outer iterations */
 420     outeriter        += nri;
 421
 422     /* Update outer/inner flops */
 423
 424     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*32 + inneriter*193);
 425 }
 426 /*
 427  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW3P1_F_c
 428  * Electrostatics interaction: Ewald
 429  * VdW interaction:            LennardJones
 430  * Geometry:                   Water3-Particle
 431  * Calculate force/pot:        Force
 432  */
 433 void
 434 nb_kernel_ElecEwSw_VdwLJSw_GeomW3P1_F_c
 435                     (t_nblist * gmx_restrict                nlist,
 436                      rvec * gmx_restrict                    xx,
 437                      rvec * gmx_restrict                    ff,
 438                      t_forcerec * gmx_restrict              fr,
 439                      t_mdatoms * gmx_restrict               mdatoms,
 440                      nb_kernel_data_t * gmx_restrict        kernel_data,
 441                      t_nrnb * gmx_restrict                  nrnb)
 442 {
 443     int              i_shift_offset,i_coord_offset,j_coord_offset;
 444     int              j_index_start,j_index_end;
 445     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 446     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 447     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 448     real             *shiftvec,*fshift,*x,*f;
 449     int              vdwioffset0;
 450     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 451     int              vdwioffset1;
 452     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 453     int              vdwioffset2;
 454     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 455     int              vdwjidx0;
 456     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 457     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 458     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 459     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 460     real             velec,felec,velecsum,facel,crf,krf,krf2;
 461     real             *charge;
 462     int              nvdwtype;
 463     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 464     int              *vdwtype;
 465     real             *vdwparam;
 466     int              ewitab;
 467     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 468     real             *ewtab;
 469     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 470
 471     x                = xx[0];
 472     f                = ff[0];
 473
 474     nri              = nlist->nri;
 475     iinr             = nlist->iinr;
 476     jindex           = nlist->jindex;
 477     jjnr             = nlist->jjnr;
 478     shiftidx         = nlist->shift;
 479     gid              = nlist->gid;
 480     shiftvec         = fr->shift_vec[0];
 481     fshift           = fr->fshift[0];
 482     facel            = fr->epsfac;
 483     charge           = mdatoms->chargeA;
 484     nvdwtype         = fr->ntype;
 485     vdwparam         = fr->nbfp;
 486     vdwtype          = mdatoms->typeA;
 487
 488     sh_ewald         = fr->ic->sh_ewald;
 489     ewtab            = fr->ic->tabq_coul_FDV0;
 490     ewtabscale       = fr->ic->tabq_scale;
 491     ewtabhalfspace   = 0.5/ewtabscale;
 492
 493     /* Setup water-specific parameters */
 494     inr              = nlist->iinr[0];
 495     iq0              = facel*charge[inr+0];
 496     iq1              = facel*charge[inr+1];
 497     iq2              = facel*charge[inr+2];
 498     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 499
 500     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 501     rcutoff          = fr->rcoulomb;
 502     rcutoff2         = rcutoff*rcutoff;
 503
 504     rswitch          = fr->rcoulomb_switch;
 505     /* Setup switch parameters */
 506     d                = rcutoff-rswitch;
 507     swV3             = -10.0/(d*d*d);
 508     swV4             =  15.0/(d*d*d*d);
 509     swV5             =  -6.0/(d*d*d*d*d);
 510     swF2             = -30.0/(d*d*d);
 511     swF3             =  60.0/(d*d*d*d);
 512     swF4             = -30.0/(d*d*d*d*d);
 513
 514     outeriter        = 0;
 515     inneriter        = 0;
 516
 517     /* Start outer loop over neighborlists */
 518     for(iidx=0; iidx<nri; iidx++)
 519     {
 520         /* Load shift vector for this list */
 521         i_shift_offset   = DIM*shiftidx[iidx];
 522         shX              = shiftvec[i_shift_offset+XX];
 523         shY              = shiftvec[i_shift_offset+YY];
 524         shZ              = shiftvec[i_shift_offset+ZZ];
 525
 526         /* Load limits for loop over neighbors */
 527         j_index_start    = jindex[iidx];
 528         j_index_end      = jindex[iidx+1];
 529
 530         /* Get outer coordinate index */
 531         inr              = iinr[iidx];
 532         i_coord_offset   = DIM*inr;
 533
 534         /* Load i particle coords and add shift vector */
 535         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 536         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 537         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 538         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 539         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 540         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 541         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 542         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 543         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 544
 545         fix0             = 0.0;
 546         fiy0             = 0.0;
 547         fiz0             = 0.0;
 548         fix1             = 0.0;
 549         fiy1             = 0.0;
 550         fiz1             = 0.0;
 551         fix2             = 0.0;
 552         fiy2             = 0.0;
 553         fiz2             = 0.0;
 554
 555         /* Start inner kernel loop */
 556         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 557         {
 558             /* Get j neighbor index, and coordinate index */
 559             jnr              = jjnr[jidx];
 560             j_coord_offset   = DIM*jnr;
 561
 562             /* load j atom coordinates */
 563             jx0              = x[j_coord_offset+DIM*0+XX];
 564             jy0              = x[j_coord_offset+DIM*0+YY];
 565             jz0              = x[j_coord_offset+DIM*0+ZZ];
 566
 567             /* Calculate displacement vector */
 568             dx00             = ix0 - jx0;
 569             dy00             = iy0 - jy0;
 570             dz00             = iz0 - jz0;
 571             dx10             = ix1 - jx0;
 572             dy10             = iy1 - jy0;
 573             dz10             = iz1 - jz0;
 574             dx20             = ix2 - jx0;
 575             dy20             = iy2 - jy0;
 576             dz20             = iz2 - jz0;
 577
 578             /* Calculate squared distance and things based on it */
 579             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 580             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 581             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 582
 583             rinv00           = gmx_invsqrt(rsq00);
 584             rinv10           = gmx_invsqrt(rsq10);
 585             rinv20           = gmx_invsqrt(rsq20);
 586
 587             rinvsq00         = rinv00*rinv00;
 588             rinvsq10         = rinv10*rinv10;
 589             rinvsq20         = rinv20*rinv20;
 590
 591             /* Load parameters for j particles */
 592             jq0              = charge[jnr+0];
 593             vdwjidx0         = 2*vdwtype[jnr+0];
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             if (rsq00<rcutoff2)
 600             {
 601
 602             r00              = rsq00*rinv00;
 603
 604             qq00             = iq0*jq0;
 605             c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 606             c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 607
 608             /* EWALD ELECTROSTATICS */
 609
 610             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 611             ewrt             = r00*ewtabscale;
 612             ewitab           = ewrt;
 613             eweps            = ewrt-ewitab;
 614             ewitab           = 4*ewitab;
 615             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 616             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 617             felec            = qq00*rinv00*(rinvsq00-felec);
 618
 619             /* LENNARD-JONES DISPERSION/REPULSION */
 620
 621             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 622             vvdw6            = c6_00*rinvsix;
 623             vvdw12           = c12_00*rinvsix*rinvsix;
 624             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 625             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 626
 627             d                = r00-rswitch;
 628             d                = (d>0.0) ? d : 0.0;
 629             d2               = d*d;
 630             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 631
 632             dsw              = d2*(swF2+d*(swF3+d*swF4));
 633
 634             /* Evaluate switch function */
 635             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 636             felec            = felec*sw - rinv00*velec*dsw;
 637             fvdw             = fvdw*sw - rinv00*vvdw*dsw;
 638
 639             fscal            = felec+fvdw;
 640
 641             /* Calculate temporary vectorial force */
 642             tx               = fscal*dx00;
 643             ty               = fscal*dy00;
 644             tz               = fscal*dz00;
 645
 646             /* Update vectorial force */
 647             fix0            += tx;
 648             fiy0            += ty;
 649             fiz0            += tz;
 650             f[j_coord_offset+DIM*0+XX] -= tx;
 651             f[j_coord_offset+DIM*0+YY] -= ty;
 652             f[j_coord_offset+DIM*0+ZZ] -= tz;
 653
 654             }
 655
 656             /**************************
 657              * CALCULATE INTERACTIONS *
 658              **************************/
 659
 660             if (rsq10<rcutoff2)
 661             {
 662
 663             r10              = rsq10*rinv10;
 664
 665             qq10             = iq1*jq0;
 666
 667             /* EWALD ELECTROSTATICS */
 668
 669             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 670             ewrt             = r10*ewtabscale;
 671             ewitab           = ewrt;
 672             eweps            = ewrt-ewitab;
 673             ewitab           = 4*ewitab;
 674             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 675             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 676             felec            = qq10*rinv10*(rinvsq10-felec);
 677
 678             d                = r10-rswitch;
 679             d                = (d>0.0) ? d : 0.0;
 680             d2               = d*d;
 681             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 682
 683             dsw              = d2*(swF2+d*(swF3+d*swF4));
 684
 685             /* Evaluate switch function */
 686             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 687             felec            = felec*sw - rinv10*velec*dsw;
 688
 689             fscal            = felec;
 690
 691             /* Calculate temporary vectorial force */
 692             tx               = fscal*dx10;
 693             ty               = fscal*dy10;
 694             tz               = fscal*dz10;
 695
 696             /* Update vectorial force */
 697             fix1            += tx;
 698             fiy1            += ty;
 699             fiz1            += tz;
 700             f[j_coord_offset+DIM*0+XX] -= tx;
 701             f[j_coord_offset+DIM*0+YY] -= ty;
 702             f[j_coord_offset+DIM*0+ZZ] -= tz;
 703
 704             }
 705
 706             /**************************
 707              * CALCULATE INTERACTIONS *
 708              **************************/
 709
 710             if (rsq20<rcutoff2)
 711             {
 712
 713             r20              = rsq20*rinv20;
 714
 715             qq20             = iq2*jq0;
 716
 717             /* EWALD ELECTROSTATICS */
 718
 719             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 720             ewrt             = r20*ewtabscale;
 721             ewitab           = ewrt;
 722             eweps            = ewrt-ewitab;
 723             ewitab           = 4*ewitab;
 724             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 725             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 726             felec            = qq20*rinv20*(rinvsq20-felec);
 727
 728             d                = r20-rswitch;
 729             d                = (d>0.0) ? d : 0.0;
 730             d2               = d*d;
 731             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 732
 733             dsw              = d2*(swF2+d*(swF3+d*swF4));
 734
 735             /* Evaluate switch function */
 736             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 737             felec            = felec*sw - rinv20*velec*dsw;
 738
 739             fscal            = felec;
 740
 741             /* Calculate temporary vectorial force */
 742             tx               = fscal*dx20;
 743             ty               = fscal*dy20;
 744             tz               = fscal*dz20;
 745
 746             /* Update vectorial force */
 747             fix2            += tx;
 748             fiy2            += ty;
 749             fiz2            += tz;
 750             f[j_coord_offset+DIM*0+XX] -= tx;
 751             f[j_coord_offset+DIM*0+YY] -= ty;
 752             f[j_coord_offset+DIM*0+ZZ] -= tz;
 753
 754             }
 755
 756             /* Inner loop uses 185 flops */
 757         }
 758         /* End of innermost loop */
 759
 760         tx = ty = tz = 0;
 761         f[i_coord_offset+DIM*0+XX] += fix0;
 762         f[i_coord_offset+DIM*0+YY] += fiy0;
 763         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 764         tx                         += fix0;
 765         ty                         += fiy0;
 766         tz                         += fiz0;
 767         f[i_coord_offset+DIM*1+XX] += fix1;
 768         f[i_coord_offset+DIM*1+YY] += fiy1;
 769         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 770         tx                         += fix1;
 771         ty                         += fiy1;
 772         tz                         += fiz1;
 773         f[i_coord_offset+DIM*2+XX] += fix2;
 774         f[i_coord_offset+DIM*2+YY] += fiy2;
 775         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 776         tx                         += fix2;
 777         ty                         += fiy2;
 778         tz                         += fiz2;
 779         fshift[i_shift_offset+XX]  += tx;
 780         fshift[i_shift_offset+YY]  += ty;
 781         fshift[i_shift_offset+ZZ]  += tz;
 782
 783         /* Increment number of inner iterations */
 784         inneriter                  += j_index_end - j_index_start;
 785
 786         /* Outer loop uses 30 flops */
 787     }
 788
 789     /* Increment number of outer iterations */
 790     outeriter        += nri;
 791
 792     /* Update outer/inner flops */
 793
 794     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*30 + inneriter*185);
 795 }