src/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEwSw_VdwNone_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_VdwNone_GeomW3P1_VF_c
  35  * Electrostatics interaction: Ewald
  36  * VdW interaction:            None
  37  * Geometry:                   Water3-Particle
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecEwSw_VdwNone_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              ewitab;
  70     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
  71     real             *ewtab;
  72     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  73
  74     x                = xx[0];
  75     f                = ff[0];
  76
  77     nri              = nlist->nri;
  78     iinr             = nlist->iinr;
  79     jindex           = nlist->jindex;
  80     jjnr             = nlist->jjnr;
  81     shiftidx         = nlist->shift;
  82     gid              = nlist->gid;
  83     shiftvec         = fr->shift_vec[0];
  84     fshift           = fr->fshift[0];
  85     facel            = fr->epsfac;
  86     charge           = mdatoms->chargeA;
  87
  88     sh_ewald         = fr->ic->sh_ewald;
  89     ewtab            = fr->ic->tabq_coul_FDV0;
  90     ewtabscale       = fr->ic->tabq_scale;
  91     ewtabhalfspace   = 0.5/ewtabscale;
  92
  93     /* Setup water-specific parameters */
  94     inr              = nlist->iinr[0];
  95     iq0              = facel*charge[inr+0];
  96     iq1              = facel*charge[inr+1];
  97     iq2              = facel*charge[inr+2];
  98
  99     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 100     rcutoff          = fr->rcoulomb;
 101     rcutoff2         = rcutoff*rcutoff;
 102
 103     rswitch          = fr->rcoulomb_switch;
 104     /* Setup switch parameters */
 105     d                = rcutoff-rswitch;
 106     swV3             = -10.0/(d*d*d);
 107     swV4             =  15.0/(d*d*d*d);
 108     swV5             =  -6.0/(d*d*d*d*d);
 109     swF2             = -30.0/(d*d*d);
 110     swF3             =  60.0/(d*d*d*d);
 111     swF4             = -30.0/(d*d*d*d*d);
 112
 113     outeriter        = 0;
 114     inneriter        = 0;
 115
 116     /* Start outer loop over neighborlists */
 117     for(iidx=0; iidx<nri; iidx++)
 118     {
 119         /* Load shift vector for this list */
 120         i_shift_offset   = DIM*shiftidx[iidx];
 121         shX              = shiftvec[i_shift_offset+XX];
 122         shY              = shiftvec[i_shift_offset+YY];
 123         shZ              = shiftvec[i_shift_offset+ZZ];
 124
 125         /* Load limits for loop over neighbors */
 126         j_index_start    = jindex[iidx];
 127         j_index_end      = jindex[iidx+1];
 128
 129         /* Get outer coordinate index */
 130         inr              = iinr[iidx];
 131         i_coord_offset   = DIM*inr;
 132
 133         /* Load i particle coords and add shift vector */
 134         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 135         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 136         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 137         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 138         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 139         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 140         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 141         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 142         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 143
 144         fix0             = 0.0;
 145         fiy0             = 0.0;
 146         fiz0             = 0.0;
 147         fix1             = 0.0;
 148         fiy1             = 0.0;
 149         fiz1             = 0.0;
 150         fix2             = 0.0;
 151         fiy2             = 0.0;
 152         fiz2             = 0.0;
 153
 154         /* Reset potential sums */
 155         velecsum         = 0.0;
 156
 157         /* Start inner kernel loop */
 158         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 159         {
 160             /* Get j neighbor index, and coordinate index */
 161             jnr              = jjnr[jidx];
 162             j_coord_offset   = DIM*jnr;
 163
 164             /* load j atom coordinates */
 165             jx0              = x[j_coord_offset+DIM*0+XX];
 166             jy0              = x[j_coord_offset+DIM*0+YY];
 167             jz0              = x[j_coord_offset+DIM*0+ZZ];
 168
 169             /* Calculate displacement vector */
 170             dx00             = ix0 - jx0;
 171             dy00             = iy0 - jy0;
 172             dz00             = iz0 - jz0;
 173             dx10             = ix1 - jx0;
 174             dy10             = iy1 - jy0;
 175             dz10             = iz1 - jz0;
 176             dx20             = ix2 - jx0;
 177             dy20             = iy2 - jy0;
 178             dz20             = iz2 - jz0;
 179
 180             /* Calculate squared distance and things based on it */
 181             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 182             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 183             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 184
 185             rinv00           = gmx_invsqrt(rsq00);
 186             rinv10           = gmx_invsqrt(rsq10);
 187             rinv20           = gmx_invsqrt(rsq20);
 188
 189             rinvsq00         = rinv00*rinv00;
 190             rinvsq10         = rinv10*rinv10;
 191             rinvsq20         = rinv20*rinv20;
 192
 193             /* Load parameters for j particles */
 194             jq0              = charge[jnr+0];
 195
 196             /**************************
 197              * CALCULATE INTERACTIONS *
 198              **************************/
 199
 200             if (rsq00<rcutoff2)
 201             {
 202
 203             r00              = rsq00*rinv00;
 204
 205             qq00             = iq0*jq0;
 206
 207             /* EWALD ELECTROSTATICS */
 208
 209             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 210             ewrt             = r00*ewtabscale;
 211             ewitab           = ewrt;
 212             eweps            = ewrt-ewitab;
 213             ewitab           = 4*ewitab;
 214             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 215             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 216             felec            = qq00*rinv00*(rinvsq00-felec);
 217
 218             d                = r00-rswitch;
 219             d                = (d>0.0) ? d : 0.0;
 220             d2               = d*d;
 221             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 222
 223             dsw              = d2*(swF2+d*(swF3+d*swF4));
 224
 225             /* Evaluate switch function */
 226             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 227             felec            = felec*sw - rinv00*velec*dsw;
 228             velec           *= sw;
 229
 230             /* Update potential sums from outer loop */
 231             velecsum        += velec;
 232
 233             fscal            = felec;
 234
 235             /* Calculate temporary vectorial force */
 236             tx               = fscal*dx00;
 237             ty               = fscal*dy00;
 238             tz               = fscal*dz00;
 239
 240             /* Update vectorial force */
 241             fix0            += tx;
 242             fiy0            += ty;
 243             fiz0            += tz;
 244             f[j_coord_offset+DIM*0+XX] -= tx;
 245             f[j_coord_offset+DIM*0+YY] -= ty;
 246             f[j_coord_offset+DIM*0+ZZ] -= tz;
 247
 248             }
 249
 250             /**************************
 251              * CALCULATE INTERACTIONS *
 252              **************************/
 253
 254             if (rsq10<rcutoff2)
 255             {
 256
 257             r10              = rsq10*rinv10;
 258
 259             qq10             = iq1*jq0;
 260
 261             /* EWALD ELECTROSTATICS */
 262
 263             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 264             ewrt             = r10*ewtabscale;
 265             ewitab           = ewrt;
 266             eweps            = ewrt-ewitab;
 267             ewitab           = 4*ewitab;
 268             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 269             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 270             felec            = qq10*rinv10*(rinvsq10-felec);
 271
 272             d                = r10-rswitch;
 273             d                = (d>0.0) ? d : 0.0;
 274             d2               = d*d;
 275             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 276
 277             dsw              = d2*(swF2+d*(swF3+d*swF4));
 278
 279             /* Evaluate switch function */
 280             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 281             felec            = felec*sw - rinv10*velec*dsw;
 282             velec           *= sw;
 283
 284             /* Update potential sums from outer loop */
 285             velecsum        += velec;
 286
 287             fscal            = felec;
 288
 289             /* Calculate temporary vectorial force */
 290             tx               = fscal*dx10;
 291             ty               = fscal*dy10;
 292             tz               = fscal*dz10;
 293
 294             /* Update vectorial force */
 295             fix1            += tx;
 296             fiy1            += ty;
 297             fiz1            += tz;
 298             f[j_coord_offset+DIM*0+XX] -= tx;
 299             f[j_coord_offset+DIM*0+YY] -= ty;
 300             f[j_coord_offset+DIM*0+ZZ] -= tz;
 301
 302             }
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             if (rsq20<rcutoff2)
 309             {
 310
 311             r20              = rsq20*rinv20;
 312
 313             qq20             = iq2*jq0;
 314
 315             /* EWALD ELECTROSTATICS */
 316
 317             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 318             ewrt             = r20*ewtabscale;
 319             ewitab           = ewrt;
 320             eweps            = ewrt-ewitab;
 321             ewitab           = 4*ewitab;
 322             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 323             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 324             felec            = qq20*rinv20*(rinvsq20-felec);
 325
 326             d                = r20-rswitch;
 327             d                = (d>0.0) ? d : 0.0;
 328             d2               = d*d;
 329             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 330
 331             dsw              = d2*(swF2+d*(swF3+d*swF4));
 332
 333             /* Evaluate switch function */
 334             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 335             felec            = felec*sw - rinv20*velec*dsw;
 336             velec           *= sw;
 337
 338             /* Update potential sums from outer loop */
 339             velecsum        += velec;
 340
 341             fscal            = felec;
 342
 343             /* Calculate temporary vectorial force */
 344             tx               = fscal*dx20;
 345             ty               = fscal*dy20;
 346             tz               = fscal*dz20;
 347
 348             /* Update vectorial force */
 349             fix2            += tx;
 350             fiy2            += ty;
 351             fiz2            += tz;
 352             f[j_coord_offset+DIM*0+XX] -= tx;
 353             f[j_coord_offset+DIM*0+YY] -= ty;
 354             f[j_coord_offset+DIM*0+ZZ] -= tz;
 355
 356             }
 357
 358             /* Inner loop uses 177 flops */
 359         }
 360         /* End of innermost loop */
 361
 362         tx = ty = tz = 0;
 363         f[i_coord_offset+DIM*0+XX] += fix0;
 364         f[i_coord_offset+DIM*0+YY] += fiy0;
 365         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 366         tx                         += fix0;
 367         ty                         += fiy0;
 368         tz                         += fiz0;
 369         f[i_coord_offset+DIM*1+XX] += fix1;
 370         f[i_coord_offset+DIM*1+YY] += fiy1;
 371         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 372         tx                         += fix1;
 373         ty                         += fiy1;
 374         tz                         += fiz1;
 375         f[i_coord_offset+DIM*2+XX] += fix2;
 376         f[i_coord_offset+DIM*2+YY] += fiy2;
 377         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 378         tx                         += fix2;
 379         ty                         += fiy2;
 380         tz                         += fiz2;
 381         fshift[i_shift_offset+XX]  += tx;
 382         fshift[i_shift_offset+YY]  += ty;
 383         fshift[i_shift_offset+ZZ]  += tz;
 384
 385         ggid                        = gid[iidx];
 386         /* Update potential energies */
 387         kernel_data->energygrp_elec[ggid] += velecsum;
 388
 389         /* Increment number of inner iterations */
 390         inneriter                  += j_index_end - j_index_start;
 391
 392         /* Outer loop uses 31 flops */
 393     }
 394
 395     /* Increment number of outer iterations */
 396     outeriter        += nri;
 397
 398     /* Update outer/inner flops */
 399
 400     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_VF,outeriter*31 + inneriter*177);
 401 }
 402 /*
 403  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomW3P1_F_c
 404  * Electrostatics interaction: Ewald
 405  * VdW interaction:            None
 406  * Geometry:                   Water3-Particle
 407  * Calculate force/pot:        Force
 408  */
 409 void
 410 nb_kernel_ElecEwSw_VdwNone_GeomW3P1_F_c
 411                     (t_nblist * gmx_restrict                nlist,
 412                      rvec * gmx_restrict                    xx,
 413                      rvec * gmx_restrict                    ff,
 414                      t_forcerec * gmx_restrict              fr,
 415                      t_mdatoms * gmx_restrict               mdatoms,
 416                      nb_kernel_data_t * gmx_restrict        kernel_data,
 417                      t_nrnb * gmx_restrict                  nrnb)
 418 {
 419     int              i_shift_offset,i_coord_offset,j_coord_offset;
 420     int              j_index_start,j_index_end;
 421     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 422     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 423     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 424     real             *shiftvec,*fshift,*x,*f;
 425     int              vdwioffset0;
 426     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 427     int              vdwioffset1;
 428     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 429     int              vdwioffset2;
 430     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 431     int              vdwjidx0;
 432     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 433     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 434     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 435     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 436     real             velec,felec,velecsum,facel,crf,krf,krf2;
 437     real             *charge;
 438     int              ewitab;
 439     real             ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace;
 440     real             *ewtab;
 441     real             rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 442
 443     x                = xx[0];
 444     f                = ff[0];
 445
 446     nri              = nlist->nri;
 447     iinr             = nlist->iinr;
 448     jindex           = nlist->jindex;
 449     jjnr             = nlist->jjnr;
 450     shiftidx         = nlist->shift;
 451     gid              = nlist->gid;
 452     shiftvec         = fr->shift_vec[0];
 453     fshift           = fr->fshift[0];
 454     facel            = fr->epsfac;
 455     charge           = mdatoms->chargeA;
 456
 457     sh_ewald         = fr->ic->sh_ewald;
 458     ewtab            = fr->ic->tabq_coul_FDV0;
 459     ewtabscale       = fr->ic->tabq_scale;
 460     ewtabhalfspace   = 0.5/ewtabscale;
 461
 462     /* Setup water-specific parameters */
 463     inr              = nlist->iinr[0];
 464     iq0              = facel*charge[inr+0];
 465     iq1              = facel*charge[inr+1];
 466     iq2              = facel*charge[inr+2];
 467
 468     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 469     rcutoff          = fr->rcoulomb;
 470     rcutoff2         = rcutoff*rcutoff;
 471
 472     rswitch          = fr->rcoulomb_switch;
 473     /* Setup switch parameters */
 474     d                = rcutoff-rswitch;
 475     swV3             = -10.0/(d*d*d);
 476     swV4             =  15.0/(d*d*d*d);
 477     swV5             =  -6.0/(d*d*d*d*d);
 478     swF2             = -30.0/(d*d*d);
 479     swF3             =  60.0/(d*d*d*d);
 480     swF4             = -30.0/(d*d*d*d*d);
 481
 482     outeriter        = 0;
 483     inneriter        = 0;
 484
 485     /* Start outer loop over neighborlists */
 486     for(iidx=0; iidx<nri; iidx++)
 487     {
 488         /* Load shift vector for this list */
 489         i_shift_offset   = DIM*shiftidx[iidx];
 490         shX              = shiftvec[i_shift_offset+XX];
 491         shY              = shiftvec[i_shift_offset+YY];
 492         shZ              = shiftvec[i_shift_offset+ZZ];
 493
 494         /* Load limits for loop over neighbors */
 495         j_index_start    = jindex[iidx];
 496         j_index_end      = jindex[iidx+1];
 497
 498         /* Get outer coordinate index */
 499         inr              = iinr[iidx];
 500         i_coord_offset   = DIM*inr;
 501
 502         /* Load i particle coords and add shift vector */
 503         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 504         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 505         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 506         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 507         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 508         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 509         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 510         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 511         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 512
 513         fix0             = 0.0;
 514         fiy0             = 0.0;
 515         fiz0             = 0.0;
 516         fix1             = 0.0;
 517         fiy1             = 0.0;
 518         fiz1             = 0.0;
 519         fix2             = 0.0;
 520         fiy2             = 0.0;
 521         fiz2             = 0.0;
 522
 523         /* Start inner kernel loop */
 524         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 525         {
 526             /* Get j neighbor index, and coordinate index */
 527             jnr              = jjnr[jidx];
 528             j_coord_offset   = DIM*jnr;
 529
 530             /* load j atom coordinates */
 531             jx0              = x[j_coord_offset+DIM*0+XX];
 532             jy0              = x[j_coord_offset+DIM*0+YY];
 533             jz0              = x[j_coord_offset+DIM*0+ZZ];
 534
 535             /* Calculate displacement vector */
 536             dx00             = ix0 - jx0;
 537             dy00             = iy0 - jy0;
 538             dz00             = iz0 - jz0;
 539             dx10             = ix1 - jx0;
 540             dy10             = iy1 - jy0;
 541             dz10             = iz1 - jz0;
 542             dx20             = ix2 - jx0;
 543             dy20             = iy2 - jy0;
 544             dz20             = iz2 - jz0;
 545
 546             /* Calculate squared distance and things based on it */
 547             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 548             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 549             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 550
 551             rinv00           = gmx_invsqrt(rsq00);
 552             rinv10           = gmx_invsqrt(rsq10);
 553             rinv20           = gmx_invsqrt(rsq20);
 554
 555             rinvsq00         = rinv00*rinv00;
 556             rinvsq10         = rinv10*rinv10;
 557             rinvsq20         = rinv20*rinv20;
 558
 559             /* Load parameters for j particles */
 560             jq0              = charge[jnr+0];
 561
 562             /**************************
 563              * CALCULATE INTERACTIONS *
 564              **************************/
 565
 566             if (rsq00<rcutoff2)
 567             {
 568
 569             r00              = rsq00*rinv00;
 570
 571             qq00             = iq0*jq0;
 572
 573             /* EWALD ELECTROSTATICS */
 574
 575             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 576             ewrt             = r00*ewtabscale;
 577             ewitab           = ewrt;
 578             eweps            = ewrt-ewitab;
 579             ewitab           = 4*ewitab;
 580             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 581             velec            = qq00*(rinv00-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 582             felec            = qq00*rinv00*(rinvsq00-felec);
 583
 584             d                = r00-rswitch;
 585             d                = (d>0.0) ? d : 0.0;
 586             d2               = d*d;
 587             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 588
 589             dsw              = d2*(swF2+d*(swF3+d*swF4));
 590
 591             /* Evaluate switch function */
 592             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 593             felec            = felec*sw - rinv00*velec*dsw;
 594
 595             fscal            = felec;
 596
 597             /* Calculate temporary vectorial force */
 598             tx               = fscal*dx00;
 599             ty               = fscal*dy00;
 600             tz               = fscal*dz00;
 601
 602             /* Update vectorial force */
 603             fix0            += tx;
 604             fiy0            += ty;
 605             fiz0            += tz;
 606             f[j_coord_offset+DIM*0+XX] -= tx;
 607             f[j_coord_offset+DIM*0+YY] -= ty;
 608             f[j_coord_offset+DIM*0+ZZ] -= tz;
 609
 610             }
 611
 612             /**************************
 613              * CALCULATE INTERACTIONS *
 614              **************************/
 615
 616             if (rsq10<rcutoff2)
 617             {
 618
 619             r10              = rsq10*rinv10;
 620
 621             qq10             = iq1*jq0;
 622
 623             /* EWALD ELECTROSTATICS */
 624
 625             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 626             ewrt             = r10*ewtabscale;
 627             ewitab           = ewrt;
 628             eweps            = ewrt-ewitab;
 629             ewitab           = 4*ewitab;
 630             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 631             velec            = qq10*(rinv10-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 632             felec            = qq10*rinv10*(rinvsq10-felec);
 633
 634             d                = r10-rswitch;
 635             d                = (d>0.0) ? d : 0.0;
 636             d2               = d*d;
 637             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 638
 639             dsw              = d2*(swF2+d*(swF3+d*swF4));
 640
 641             /* Evaluate switch function */
 642             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 643             felec            = felec*sw - rinv10*velec*dsw;
 644
 645             fscal            = felec;
 646
 647             /* Calculate temporary vectorial force */
 648             tx               = fscal*dx10;
 649             ty               = fscal*dy10;
 650             tz               = fscal*dz10;
 651
 652             /* Update vectorial force */
 653             fix1            += tx;
 654             fiy1            += ty;
 655             fiz1            += tz;
 656             f[j_coord_offset+DIM*0+XX] -= tx;
 657             f[j_coord_offset+DIM*0+YY] -= ty;
 658             f[j_coord_offset+DIM*0+ZZ] -= tz;
 659
 660             }
 661
 662             /**************************
 663              * CALCULATE INTERACTIONS *
 664              **************************/
 665
 666             if (rsq20<rcutoff2)
 667             {
 668
 669             r20              = rsq20*rinv20;
 670
 671             qq20             = iq2*jq0;
 672
 673             /* EWALD ELECTROSTATICS */
 674
 675             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 676             ewrt             = r20*ewtabscale;
 677             ewitab           = ewrt;
 678             eweps            = ewrt-ewitab;
 679             ewitab           = 4*ewitab;
 680             felec            = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 681             velec            = qq20*(rinv20-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec)));
 682             felec            = qq20*rinv20*(rinvsq20-felec);
 683
 684             d                = r20-rswitch;
 685             d                = (d>0.0) ? d : 0.0;
 686             d2               = d*d;
 687             sw               = 1.0+d2*d*(swV3+d*(swV4+d*swV5));
 688
 689             dsw              = d2*(swF2+d*(swF3+d*swF4));
 690
 691             /* Evaluate switch function */
 692             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 693             felec            = felec*sw - rinv20*velec*dsw;
 694
 695             fscal            = felec;
 696
 697             /* Calculate temporary vectorial force */
 698             tx               = fscal*dx20;
 699             ty               = fscal*dy20;
 700             tz               = fscal*dz20;
 701
 702             /* Update vectorial force */
 703             fix2            += tx;
 704             fiy2            += ty;
 705             fiz2            += tz;
 706             f[j_coord_offset+DIM*0+XX] -= tx;
 707             f[j_coord_offset+DIM*0+YY] -= ty;
 708             f[j_coord_offset+DIM*0+ZZ] -= tz;
 709
 710             }
 711
 712             /* Inner loop uses 171 flops */
 713         }
 714         /* End of innermost loop */
 715
 716         tx = ty = tz = 0;
 717         f[i_coord_offset+DIM*0+XX] += fix0;
 718         f[i_coord_offset+DIM*0+YY] += fiy0;
 719         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 720         tx                         += fix0;
 721         ty                         += fiy0;
 722         tz                         += fiz0;
 723         f[i_coord_offset+DIM*1+XX] += fix1;
 724         f[i_coord_offset+DIM*1+YY] += fiy1;
 725         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 726         tx                         += fix1;
 727         ty                         += fiy1;
 728         tz                         += fiz1;
 729         f[i_coord_offset+DIM*2+XX] += fix2;
 730         f[i_coord_offset+DIM*2+YY] += fiy2;
 731         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 732         tx                         += fix2;
 733         ty                         += fiy2;
 734         tz                         += fiz2;
 735         fshift[i_shift_offset+XX]  += tx;
 736         fshift[i_shift_offset+YY]  += ty;
 737         fshift[i_shift_offset+ZZ]  += tz;
 738
 739         /* Increment number of inner iterations */
 740         inneriter                  += j_index_end - j_index_start;
 741
 742         /* Outer loop uses 30 flops */
 743     }
 744
 745     /* Increment number of outer iterations */
 746     outeriter        += nri;
 747
 748     /* Update outer/inner flops */
 749
 750     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_F,outeriter*30 + inneriter*171);
 751 }