src/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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_ElecCoul_VdwLJ_GeomW3W3_VF_c
  35  * Electrostatics interaction: Coulomb
  36  * VdW interaction:            LennardJones
  37  * Geometry:                   Water3-Water3
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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     int              vdwjidx1;
  65     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  66     int              vdwjidx2;
  67     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  68     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  69     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
  70     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
  71     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  72     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  73     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  74     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  75     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  76     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  77     real             velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     int              nvdwtype;
  80     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  81     int              *vdwtype;
  82     real             *vdwparam;
  83
  84     x                = xx[0];
  85     f                = ff[0];
  86
  87     nri              = nlist->nri;
  88     iinr             = nlist->iinr;
  89     jindex           = nlist->jindex;
  90     jjnr             = nlist->jjnr;
  91     shiftidx         = nlist->shift;
  92     gid              = nlist->gid;
  93     shiftvec         = fr->shift_vec[0];
  94     fshift           = fr->fshift[0];
  95     facel            = fr->epsfac;
  96     charge           = mdatoms->chargeA;
  97     nvdwtype         = fr->ntype;
  98     vdwparam         = fr->nbfp;
  99     vdwtype          = mdatoms->typeA;
 100
 101     /* Setup water-specific parameters */
 102     inr              = nlist->iinr[0];
 103     iq0              = facel*charge[inr+0];
 104     iq1              = facel*charge[inr+1];
 105     iq2              = facel*charge[inr+2];
 106     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 107
 108     jq0              = charge[inr+0];
 109     jq1              = charge[inr+1];
 110     jq2              = charge[inr+2];
 111     vdwjidx0         = 2*vdwtype[inr+0];
 112     qq00             = iq0*jq0;
 113     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 114     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 115     qq01             = iq0*jq1;
 116     qq02             = iq0*jq2;
 117     qq10             = iq1*jq0;
 118     qq11             = iq1*jq1;
 119     qq12             = iq1*jq2;
 120     qq20             = iq2*jq0;
 121     qq21             = iq2*jq1;
 122     qq22             = iq2*jq2;
 123
 124     outeriter        = 0;
 125     inneriter        = 0;
 126
 127     /* Start outer loop over neighborlists */
 128     for(iidx=0; iidx<nri; iidx++)
 129     {
 130         /* Load shift vector for this list */
 131         i_shift_offset   = DIM*shiftidx[iidx];
 132         shX              = shiftvec[i_shift_offset+XX];
 133         shY              = shiftvec[i_shift_offset+YY];
 134         shZ              = shiftvec[i_shift_offset+ZZ];
 135
 136         /* Load limits for loop over neighbors */
 137         j_index_start    = jindex[iidx];
 138         j_index_end      = jindex[iidx+1];
 139
 140         /* Get outer coordinate index */
 141         inr              = iinr[iidx];
 142         i_coord_offset   = DIM*inr;
 143
 144         /* Load i particle coords and add shift vector */
 145         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 146         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 147         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 148         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 149         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 150         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 151         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 152         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 153         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 154
 155         fix0             = 0.0;
 156         fiy0             = 0.0;
 157         fiz0             = 0.0;
 158         fix1             = 0.0;
 159         fiy1             = 0.0;
 160         fiz1             = 0.0;
 161         fix2             = 0.0;
 162         fiy2             = 0.0;
 163         fiz2             = 0.0;
 164
 165         /* Reset potential sums */
 166         velecsum         = 0.0;
 167         vvdwsum          = 0.0;
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 171         {
 172             /* Get j neighbor index, and coordinate index */
 173             jnr              = jjnr[jidx];
 174             j_coord_offset   = DIM*jnr;
 175
 176             /* load j atom coordinates */
 177             jx0              = x[j_coord_offset+DIM*0+XX];
 178             jy0              = x[j_coord_offset+DIM*0+YY];
 179             jz0              = x[j_coord_offset+DIM*0+ZZ];
 180             jx1              = x[j_coord_offset+DIM*1+XX];
 181             jy1              = x[j_coord_offset+DIM*1+YY];
 182             jz1              = x[j_coord_offset+DIM*1+ZZ];
 183             jx2              = x[j_coord_offset+DIM*2+XX];
 184             jy2              = x[j_coord_offset+DIM*2+YY];
 185             jz2              = x[j_coord_offset+DIM*2+ZZ];
 186
 187             /* Calculate displacement vector */
 188             dx00             = ix0 - jx0;
 189             dy00             = iy0 - jy0;
 190             dz00             = iz0 - jz0;
 191             dx01             = ix0 - jx1;
 192             dy01             = iy0 - jy1;
 193             dz01             = iz0 - jz1;
 194             dx02             = ix0 - jx2;
 195             dy02             = iy0 - jy2;
 196             dz02             = iz0 - jz2;
 197             dx10             = ix1 - jx0;
 198             dy10             = iy1 - jy0;
 199             dz10             = iz1 - jz0;
 200             dx11             = ix1 - jx1;
 201             dy11             = iy1 - jy1;
 202             dz11             = iz1 - jz1;
 203             dx12             = ix1 - jx2;
 204             dy12             = iy1 - jy2;
 205             dz12             = iz1 - jz2;
 206             dx20             = ix2 - jx0;
 207             dy20             = iy2 - jy0;
 208             dz20             = iz2 - jz0;
 209             dx21             = ix2 - jx1;
 210             dy21             = iy2 - jy1;
 211             dz21             = iz2 - jz1;
 212             dx22             = ix2 - jx2;
 213             dy22             = iy2 - jy2;
 214             dz22             = iz2 - jz2;
 215
 216             /* Calculate squared distance and things based on it */
 217             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 218             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 219             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 220             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 221             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 222             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 223             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 224             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 225             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 226
 227             rinv00           = gmx_invsqrt(rsq00);
 228             rinv01           = gmx_invsqrt(rsq01);
 229             rinv02           = gmx_invsqrt(rsq02);
 230             rinv10           = gmx_invsqrt(rsq10);
 231             rinv11           = gmx_invsqrt(rsq11);
 232             rinv12           = gmx_invsqrt(rsq12);
 233             rinv20           = gmx_invsqrt(rsq20);
 234             rinv21           = gmx_invsqrt(rsq21);
 235             rinv22           = gmx_invsqrt(rsq22);
 236
 237             rinvsq00         = rinv00*rinv00;
 238             rinvsq01         = rinv01*rinv01;
 239             rinvsq02         = rinv02*rinv02;
 240             rinvsq10         = rinv10*rinv10;
 241             rinvsq11         = rinv11*rinv11;
 242             rinvsq12         = rinv12*rinv12;
 243             rinvsq20         = rinv20*rinv20;
 244             rinvsq21         = rinv21*rinv21;
 245             rinvsq22         = rinv22*rinv22;
 246
 247             /**************************
 248              * CALCULATE INTERACTIONS *
 249              **************************/
 250
 251             /* COULOMB ELECTROSTATICS */
 252             velec            = qq00*rinv00;
 253             felec            = velec*rinvsq00;
 254
 255             /* LENNARD-JONES DISPERSION/REPULSION */
 256
 257             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 258             vvdw6            = c6_00*rinvsix;
 259             vvdw12           = c12_00*rinvsix*rinvsix;
 260             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 261             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 262
 263             /* Update potential sums from outer loop */
 264             velecsum        += velec;
 265             vvdwsum         += vvdw;
 266
 267             fscal            = felec+fvdw;
 268
 269             /* Calculate temporary vectorial force */
 270             tx               = fscal*dx00;
 271             ty               = fscal*dy00;
 272             tz               = fscal*dz00;
 273
 274             /* Update vectorial force */
 275             fix0            += tx;
 276             fiy0            += ty;
 277             fiz0            += tz;
 278             f[j_coord_offset+DIM*0+XX] -= tx;
 279             f[j_coord_offset+DIM*0+YY] -= ty;
 280             f[j_coord_offset+DIM*0+ZZ] -= tz;
 281
 282             /**************************
 283              * CALCULATE INTERACTIONS *
 284              **************************/
 285
 286             /* COULOMB ELECTROSTATICS */
 287             velec            = qq01*rinv01;
 288             felec            = velec*rinvsq01;
 289
 290             /* Update potential sums from outer loop */
 291             velecsum        += velec;
 292
 293             fscal            = felec;
 294
 295             /* Calculate temporary vectorial force */
 296             tx               = fscal*dx01;
 297             ty               = fscal*dy01;
 298             tz               = fscal*dz01;
 299
 300             /* Update vectorial force */
 301             fix0            += tx;
 302             fiy0            += ty;
 303             fiz0            += tz;
 304             f[j_coord_offset+DIM*1+XX] -= tx;
 305             f[j_coord_offset+DIM*1+YY] -= ty;
 306             f[j_coord_offset+DIM*1+ZZ] -= tz;
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             /* COULOMB ELECTROSTATICS */
 313             velec            = qq02*rinv02;
 314             felec            = velec*rinvsq02;
 315
 316             /* Update potential sums from outer loop */
 317             velecsum        += velec;
 318
 319             fscal            = felec;
 320
 321             /* Calculate temporary vectorial force */
 322             tx               = fscal*dx02;
 323             ty               = fscal*dy02;
 324             tz               = fscal*dz02;
 325
 326             /* Update vectorial force */
 327             fix0            += tx;
 328             fiy0            += ty;
 329             fiz0            += tz;
 330             f[j_coord_offset+DIM*2+XX] -= tx;
 331             f[j_coord_offset+DIM*2+YY] -= ty;
 332             f[j_coord_offset+DIM*2+ZZ] -= tz;
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             /* COULOMB ELECTROSTATICS */
 339             velec            = qq10*rinv10;
 340             felec            = velec*rinvsq10;
 341
 342             /* Update potential sums from outer loop */
 343             velecsum        += velec;
 344
 345             fscal            = felec;
 346
 347             /* Calculate temporary vectorial force */
 348             tx               = fscal*dx10;
 349             ty               = fscal*dy10;
 350             tz               = fscal*dz10;
 351
 352             /* Update vectorial force */
 353             fix1            += tx;
 354             fiy1            += ty;
 355             fiz1            += tz;
 356             f[j_coord_offset+DIM*0+XX] -= tx;
 357             f[j_coord_offset+DIM*0+YY] -= ty;
 358             f[j_coord_offset+DIM*0+ZZ] -= tz;
 359
 360             /**************************
 361              * CALCULATE INTERACTIONS *
 362              **************************/
 363
 364             /* COULOMB ELECTROSTATICS */
 365             velec            = qq11*rinv11;
 366             felec            = velec*rinvsq11;
 367
 368             /* Update potential sums from outer loop */
 369             velecsum        += velec;
 370
 371             fscal            = felec;
 372
 373             /* Calculate temporary vectorial force */
 374             tx               = fscal*dx11;
 375             ty               = fscal*dy11;
 376             tz               = fscal*dz11;
 377
 378             /* Update vectorial force */
 379             fix1            += tx;
 380             fiy1            += ty;
 381             fiz1            += tz;
 382             f[j_coord_offset+DIM*1+XX] -= tx;
 383             f[j_coord_offset+DIM*1+YY] -= ty;
 384             f[j_coord_offset+DIM*1+ZZ] -= tz;
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             /* COULOMB ELECTROSTATICS */
 391             velec            = qq12*rinv12;
 392             felec            = velec*rinvsq12;
 393
 394             /* Update potential sums from outer loop */
 395             velecsum        += velec;
 396
 397             fscal            = felec;
 398
 399             /* Calculate temporary vectorial force */
 400             tx               = fscal*dx12;
 401             ty               = fscal*dy12;
 402             tz               = fscal*dz12;
 403
 404             /* Update vectorial force */
 405             fix1            += tx;
 406             fiy1            += ty;
 407             fiz1            += tz;
 408             f[j_coord_offset+DIM*2+XX] -= tx;
 409             f[j_coord_offset+DIM*2+YY] -= ty;
 410             f[j_coord_offset+DIM*2+ZZ] -= tz;
 411
 412             /**************************
 413              * CALCULATE INTERACTIONS *
 414              **************************/
 415
 416             /* COULOMB ELECTROSTATICS */
 417             velec            = qq20*rinv20;
 418             felec            = velec*rinvsq20;
 419
 420             /* Update potential sums from outer loop */
 421             velecsum        += velec;
 422
 423             fscal            = felec;
 424
 425             /* Calculate temporary vectorial force */
 426             tx               = fscal*dx20;
 427             ty               = fscal*dy20;
 428             tz               = fscal*dz20;
 429
 430             /* Update vectorial force */
 431             fix2            += tx;
 432             fiy2            += ty;
 433             fiz2            += tz;
 434             f[j_coord_offset+DIM*0+XX] -= tx;
 435             f[j_coord_offset+DIM*0+YY] -= ty;
 436             f[j_coord_offset+DIM*0+ZZ] -= tz;
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             /* COULOMB ELECTROSTATICS */
 443             velec            = qq21*rinv21;
 444             felec            = velec*rinvsq21;
 445
 446             /* Update potential sums from outer loop */
 447             velecsum        += velec;
 448
 449             fscal            = felec;
 450
 451             /* Calculate temporary vectorial force */
 452             tx               = fscal*dx21;
 453             ty               = fscal*dy21;
 454             tz               = fscal*dz21;
 455
 456             /* Update vectorial force */
 457             fix2            += tx;
 458             fiy2            += ty;
 459             fiz2            += tz;
 460             f[j_coord_offset+DIM*1+XX] -= tx;
 461             f[j_coord_offset+DIM*1+YY] -= ty;
 462             f[j_coord_offset+DIM*1+ZZ] -= tz;
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             /* COULOMB ELECTROSTATICS */
 469             velec            = qq22*rinv22;
 470             felec            = velec*rinvsq22;
 471
 472             /* Update potential sums from outer loop */
 473             velecsum        += velec;
 474
 475             fscal            = felec;
 476
 477             /* Calculate temporary vectorial force */
 478             tx               = fscal*dx22;
 479             ty               = fscal*dy22;
 480             tz               = fscal*dz22;
 481
 482             /* Update vectorial force */
 483             fix2            += tx;
 484             fiy2            += ty;
 485             fiz2            += tz;
 486             f[j_coord_offset+DIM*2+XX] -= tx;
 487             f[j_coord_offset+DIM*2+YY] -= ty;
 488             f[j_coord_offset+DIM*2+ZZ] -= tz;
 489
 490             /* Inner loop uses 255 flops */
 491         }
 492         /* End of innermost loop */
 493
 494         tx = ty = tz = 0;
 495         f[i_coord_offset+DIM*0+XX] += fix0;
 496         f[i_coord_offset+DIM*0+YY] += fiy0;
 497         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 498         tx                         += fix0;
 499         ty                         += fiy0;
 500         tz                         += fiz0;
 501         f[i_coord_offset+DIM*1+XX] += fix1;
 502         f[i_coord_offset+DIM*1+YY] += fiy1;
 503         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 504         tx                         += fix1;
 505         ty                         += fiy1;
 506         tz                         += fiz1;
 507         f[i_coord_offset+DIM*2+XX] += fix2;
 508         f[i_coord_offset+DIM*2+YY] += fiy2;
 509         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 510         tx                         += fix2;
 511         ty                         += fiy2;
 512         tz                         += fiz2;
 513         fshift[i_shift_offset+XX]  += tx;
 514         fshift[i_shift_offset+YY]  += ty;
 515         fshift[i_shift_offset+ZZ]  += tz;
 516
 517         ggid                        = gid[iidx];
 518         /* Update potential energies */
 519         kernel_data->energygrp_elec[ggid] += velecsum;
 520         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 521
 522         /* Increment number of inner iterations */
 523         inneriter                  += j_index_end - j_index_start;
 524
 525         /* Outer loop uses 32 flops */
 526     }
 527
 528     /* Increment number of outer iterations */
 529     outeriter        += nri;
 530
 531     /* Update outer/inner flops */
 532
 533     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*255);
 534 }
 535 /*
 536  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_c
 537  * Electrostatics interaction: Coulomb
 538  * VdW interaction:            LennardJones
 539  * Geometry:                   Water3-Water3
 540  * Calculate force/pot:        Force
 541  */
 542 void
 543 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_c
 544                     (t_nblist * gmx_restrict                nlist,
 545                      rvec * gmx_restrict                    xx,
 546                      rvec * gmx_restrict                    ff,
 547                      t_forcerec * gmx_restrict              fr,
 548                      t_mdatoms * gmx_restrict               mdatoms,
 549                      nb_kernel_data_t * gmx_restrict        kernel_data,
 550                      t_nrnb * gmx_restrict                  nrnb)
 551 {
 552     int              i_shift_offset,i_coord_offset,j_coord_offset;
 553     int              j_index_start,j_index_end;
 554     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 555     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 556     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 557     real             *shiftvec,*fshift,*x,*f;
 558     int              vdwioffset0;
 559     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 560     int              vdwioffset1;
 561     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 562     int              vdwioffset2;
 563     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 564     int              vdwjidx0;
 565     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 566     int              vdwjidx1;
 567     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 568     int              vdwjidx2;
 569     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 570     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 571     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 572     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 573     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 574     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 575     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 576     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 577     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 578     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 579     real             velec,felec,velecsum,facel,crf,krf,krf2;
 580     real             *charge;
 581     int              nvdwtype;
 582     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 583     int              *vdwtype;
 584     real             *vdwparam;
 585
 586     x                = xx[0];
 587     f                = ff[0];
 588
 589     nri              = nlist->nri;
 590     iinr             = nlist->iinr;
 591     jindex           = nlist->jindex;
 592     jjnr             = nlist->jjnr;
 593     shiftidx         = nlist->shift;
 594     gid              = nlist->gid;
 595     shiftvec         = fr->shift_vec[0];
 596     fshift           = fr->fshift[0];
 597     facel            = fr->epsfac;
 598     charge           = mdatoms->chargeA;
 599     nvdwtype         = fr->ntype;
 600     vdwparam         = fr->nbfp;
 601     vdwtype          = mdatoms->typeA;
 602
 603     /* Setup water-specific parameters */
 604     inr              = nlist->iinr[0];
 605     iq0              = facel*charge[inr+0];
 606     iq1              = facel*charge[inr+1];
 607     iq2              = facel*charge[inr+2];
 608     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 609
 610     jq0              = charge[inr+0];
 611     jq1              = charge[inr+1];
 612     jq2              = charge[inr+2];
 613     vdwjidx0         = 2*vdwtype[inr+0];
 614     qq00             = iq0*jq0;
 615     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 616     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 617     qq01             = iq0*jq1;
 618     qq02             = iq0*jq2;
 619     qq10             = iq1*jq0;
 620     qq11             = iq1*jq1;
 621     qq12             = iq1*jq2;
 622     qq20             = iq2*jq0;
 623     qq21             = iq2*jq1;
 624     qq22             = iq2*jq2;
 625
 626     outeriter        = 0;
 627     inneriter        = 0;
 628
 629     /* Start outer loop over neighborlists */
 630     for(iidx=0; iidx<nri; iidx++)
 631     {
 632         /* Load shift vector for this list */
 633         i_shift_offset   = DIM*shiftidx[iidx];
 634         shX              = shiftvec[i_shift_offset+XX];
 635         shY              = shiftvec[i_shift_offset+YY];
 636         shZ              = shiftvec[i_shift_offset+ZZ];
 637
 638         /* Load limits for loop over neighbors */
 639         j_index_start    = jindex[iidx];
 640         j_index_end      = jindex[iidx+1];
 641
 642         /* Get outer coordinate index */
 643         inr              = iinr[iidx];
 644         i_coord_offset   = DIM*inr;
 645
 646         /* Load i particle coords and add shift vector */
 647         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 648         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 649         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 650         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 651         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 652         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 653         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 654         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 655         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 656
 657         fix0             = 0.0;
 658         fiy0             = 0.0;
 659         fiz0             = 0.0;
 660         fix1             = 0.0;
 661         fiy1             = 0.0;
 662         fiz1             = 0.0;
 663         fix2             = 0.0;
 664         fiy2             = 0.0;
 665         fiz2             = 0.0;
 666
 667         /* Start inner kernel loop */
 668         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 669         {
 670             /* Get j neighbor index, and coordinate index */
 671             jnr              = jjnr[jidx];
 672             j_coord_offset   = DIM*jnr;
 673
 674             /* load j atom coordinates */
 675             jx0              = x[j_coord_offset+DIM*0+XX];
 676             jy0              = x[j_coord_offset+DIM*0+YY];
 677             jz0              = x[j_coord_offset+DIM*0+ZZ];
 678             jx1              = x[j_coord_offset+DIM*1+XX];
 679             jy1              = x[j_coord_offset+DIM*1+YY];
 680             jz1              = x[j_coord_offset+DIM*1+ZZ];
 681             jx2              = x[j_coord_offset+DIM*2+XX];
 682             jy2              = x[j_coord_offset+DIM*2+YY];
 683             jz2              = x[j_coord_offset+DIM*2+ZZ];
 684
 685             /* Calculate displacement vector */
 686             dx00             = ix0 - jx0;
 687             dy00             = iy0 - jy0;
 688             dz00             = iz0 - jz0;
 689             dx01             = ix0 - jx1;
 690             dy01             = iy0 - jy1;
 691             dz01             = iz0 - jz1;
 692             dx02             = ix0 - jx2;
 693             dy02             = iy0 - jy2;
 694             dz02             = iz0 - jz2;
 695             dx10             = ix1 - jx0;
 696             dy10             = iy1 - jy0;
 697             dz10             = iz1 - jz0;
 698             dx11             = ix1 - jx1;
 699             dy11             = iy1 - jy1;
 700             dz11             = iz1 - jz1;
 701             dx12             = ix1 - jx2;
 702             dy12             = iy1 - jy2;
 703             dz12             = iz1 - jz2;
 704             dx20             = ix2 - jx0;
 705             dy20             = iy2 - jy0;
 706             dz20             = iz2 - jz0;
 707             dx21             = ix2 - jx1;
 708             dy21             = iy2 - jy1;
 709             dz21             = iz2 - jz1;
 710             dx22             = ix2 - jx2;
 711             dy22             = iy2 - jy2;
 712             dz22             = iz2 - jz2;
 713
 714             /* Calculate squared distance and things based on it */
 715             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 716             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 717             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 718             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 719             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 720             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 721             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 722             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 723             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 724
 725             rinv00           = gmx_invsqrt(rsq00);
 726             rinv01           = gmx_invsqrt(rsq01);
 727             rinv02           = gmx_invsqrt(rsq02);
 728             rinv10           = gmx_invsqrt(rsq10);
 729             rinv11           = gmx_invsqrt(rsq11);
 730             rinv12           = gmx_invsqrt(rsq12);
 731             rinv20           = gmx_invsqrt(rsq20);
 732             rinv21           = gmx_invsqrt(rsq21);
 733             rinv22           = gmx_invsqrt(rsq22);
 734
 735             rinvsq00         = rinv00*rinv00;
 736             rinvsq01         = rinv01*rinv01;
 737             rinvsq02         = rinv02*rinv02;
 738             rinvsq10         = rinv10*rinv10;
 739             rinvsq11         = rinv11*rinv11;
 740             rinvsq12         = rinv12*rinv12;
 741             rinvsq20         = rinv20*rinv20;
 742             rinvsq21         = rinv21*rinv21;
 743             rinvsq22         = rinv22*rinv22;
 744
 745             /**************************
 746              * CALCULATE INTERACTIONS *
 747              **************************/
 748
 749             /* COULOMB ELECTROSTATICS */
 750             velec            = qq00*rinv00;
 751             felec            = velec*rinvsq00;
 752
 753             /* LENNARD-JONES DISPERSION/REPULSION */
 754
 755             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 756             fvdw             = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00;
 757
 758             fscal            = felec+fvdw;
 759
 760             /* Calculate temporary vectorial force */
 761             tx               = fscal*dx00;
 762             ty               = fscal*dy00;
 763             tz               = fscal*dz00;
 764
 765             /* Update vectorial force */
 766             fix0            += tx;
 767             fiy0            += ty;
 768             fiz0            += tz;
 769             f[j_coord_offset+DIM*0+XX] -= tx;
 770             f[j_coord_offset+DIM*0+YY] -= ty;
 771             f[j_coord_offset+DIM*0+ZZ] -= tz;
 772
 773             /**************************
 774              * CALCULATE INTERACTIONS *
 775              **************************/
 776
 777             /* COULOMB ELECTROSTATICS */
 778             velec            = qq01*rinv01;
 779             felec            = velec*rinvsq01;
 780
 781             fscal            = felec;
 782
 783             /* Calculate temporary vectorial force */
 784             tx               = fscal*dx01;
 785             ty               = fscal*dy01;
 786             tz               = fscal*dz01;
 787
 788             /* Update vectorial force */
 789             fix0            += tx;
 790             fiy0            += ty;
 791             fiz0            += tz;
 792             f[j_coord_offset+DIM*1+XX] -= tx;
 793             f[j_coord_offset+DIM*1+YY] -= ty;
 794             f[j_coord_offset+DIM*1+ZZ] -= tz;
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             /* COULOMB ELECTROSTATICS */
 801             velec            = qq02*rinv02;
 802             felec            = velec*rinvsq02;
 803
 804             fscal            = felec;
 805
 806             /* Calculate temporary vectorial force */
 807             tx               = fscal*dx02;
 808             ty               = fscal*dy02;
 809             tz               = fscal*dz02;
 810
 811             /* Update vectorial force */
 812             fix0            += tx;
 813             fiy0            += ty;
 814             fiz0            += tz;
 815             f[j_coord_offset+DIM*2+XX] -= tx;
 816             f[j_coord_offset+DIM*2+YY] -= ty;
 817             f[j_coord_offset+DIM*2+ZZ] -= tz;
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             /* COULOMB ELECTROSTATICS */
 824             velec            = qq10*rinv10;
 825             felec            = velec*rinvsq10;
 826
 827             fscal            = felec;
 828
 829             /* Calculate temporary vectorial force */
 830             tx               = fscal*dx10;
 831             ty               = fscal*dy10;
 832             tz               = fscal*dz10;
 833
 834             /* Update vectorial force */
 835             fix1            += tx;
 836             fiy1            += ty;
 837             fiz1            += tz;
 838             f[j_coord_offset+DIM*0+XX] -= tx;
 839             f[j_coord_offset+DIM*0+YY] -= ty;
 840             f[j_coord_offset+DIM*0+ZZ] -= tz;
 841
 842             /**************************
 843              * CALCULATE INTERACTIONS *
 844              **************************/
 845
 846             /* COULOMB ELECTROSTATICS */
 847             velec            = qq11*rinv11;
 848             felec            = velec*rinvsq11;
 849
 850             fscal            = felec;
 851
 852             /* Calculate temporary vectorial force */
 853             tx               = fscal*dx11;
 854             ty               = fscal*dy11;
 855             tz               = fscal*dz11;
 856
 857             /* Update vectorial force */
 858             fix1            += tx;
 859             fiy1            += ty;
 860             fiz1            += tz;
 861             f[j_coord_offset+DIM*1+XX] -= tx;
 862             f[j_coord_offset+DIM*1+YY] -= ty;
 863             f[j_coord_offset+DIM*1+ZZ] -= tz;
 864
 865             /**************************
 866              * CALCULATE INTERACTIONS *
 867              **************************/
 868
 869             /* COULOMB ELECTROSTATICS */
 870             velec            = qq12*rinv12;
 871             felec            = velec*rinvsq12;
 872
 873             fscal            = felec;
 874
 875             /* Calculate temporary vectorial force */
 876             tx               = fscal*dx12;
 877             ty               = fscal*dy12;
 878             tz               = fscal*dz12;
 879
 880             /* Update vectorial force */
 881             fix1            += tx;
 882             fiy1            += ty;
 883             fiz1            += tz;
 884             f[j_coord_offset+DIM*2+XX] -= tx;
 885             f[j_coord_offset+DIM*2+YY] -= ty;
 886             f[j_coord_offset+DIM*2+ZZ] -= tz;
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             /* COULOMB ELECTROSTATICS */
 893             velec            = qq20*rinv20;
 894             felec            = velec*rinvsq20;
 895
 896             fscal            = felec;
 897
 898             /* Calculate temporary vectorial force */
 899             tx               = fscal*dx20;
 900             ty               = fscal*dy20;
 901             tz               = fscal*dz20;
 902
 903             /* Update vectorial force */
 904             fix2            += tx;
 905             fiy2            += ty;
 906             fiz2            += tz;
 907             f[j_coord_offset+DIM*0+XX] -= tx;
 908             f[j_coord_offset+DIM*0+YY] -= ty;
 909             f[j_coord_offset+DIM*0+ZZ] -= tz;
 910
 911             /**************************
 912              * CALCULATE INTERACTIONS *
 913              **************************/
 914
 915             /* COULOMB ELECTROSTATICS */
 916             velec            = qq21*rinv21;
 917             felec            = velec*rinvsq21;
 918
 919             fscal            = felec;
 920
 921             /* Calculate temporary vectorial force */
 922             tx               = fscal*dx21;
 923             ty               = fscal*dy21;
 924             tz               = fscal*dz21;
 925
 926             /* Update vectorial force */
 927             fix2            += tx;
 928             fiy2            += ty;
 929             fiz2            += tz;
 930             f[j_coord_offset+DIM*1+XX] -= tx;
 931             f[j_coord_offset+DIM*1+YY] -= ty;
 932             f[j_coord_offset+DIM*1+ZZ] -= tz;
 933
 934             /**************************
 935              * CALCULATE INTERACTIONS *
 936              **************************/
 937
 938             /* COULOMB ELECTROSTATICS */
 939             velec            = qq22*rinv22;
 940             felec            = velec*rinvsq22;
 941
 942             fscal            = felec;
 943
 944             /* Calculate temporary vectorial force */
 945             tx               = fscal*dx22;
 946             ty               = fscal*dy22;
 947             tz               = fscal*dz22;
 948
 949             /* Update vectorial force */
 950             fix2            += tx;
 951             fiy2            += ty;
 952             fiz2            += tz;
 953             f[j_coord_offset+DIM*2+XX] -= tx;
 954             f[j_coord_offset+DIM*2+YY] -= ty;
 955             f[j_coord_offset+DIM*2+ZZ] -= tz;
 956
 957             /* Inner loop uses 241 flops */
 958         }
 959         /* End of innermost loop */
 960
 961         tx = ty = tz = 0;
 962         f[i_coord_offset+DIM*0+XX] += fix0;
 963         f[i_coord_offset+DIM*0+YY] += fiy0;
 964         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 965         tx                         += fix0;
 966         ty                         += fiy0;
 967         tz                         += fiz0;
 968         f[i_coord_offset+DIM*1+XX] += fix1;
 969         f[i_coord_offset+DIM*1+YY] += fiy1;
 970         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 971         tx                         += fix1;
 972         ty                         += fiy1;
 973         tz                         += fiz1;
 974         f[i_coord_offset+DIM*2+XX] += fix2;
 975         f[i_coord_offset+DIM*2+YY] += fiy2;
 976         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 977         tx                         += fix2;
 978         ty                         += fiy2;
 979         tz                         += fiz2;
 980         fshift[i_shift_offset+XX]  += tx;
 981         fshift[i_shift_offset+YY]  += ty;
 982         fshift[i_shift_offset+ZZ]  += tz;
 983
 984         /* Increment number of inner iterations */
 985         inneriter                  += j_index_end - j_index_start;
 986
 987         /* Outer loop uses 30 flops */
 988     }
 989
 990     /* Increment number of outer iterations */
 991     outeriter        += nri;
 992
 993     /* Update outer/inner flops */
 994
 995     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*241);
 996 }