src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecRF_VdwCSTab_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_ElecRF_VdwCSTab_GeomW3W3_VF_c
  35  * Electrostatics interaction: ReactionField
  36  * VdW interaction:            CubicSplineTable
  37  * Geometry:                   Water3-Water3
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecRF_VdwCSTab_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     int              vfitab;
  84     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
  85     real             *vftab;
  86
  87     x                = xx[0];
  88     f                = ff[0];
  89
  90     nri              = nlist->nri;
  91     iinr             = nlist->iinr;
  92     jindex           = nlist->jindex;
  93     jjnr             = nlist->jjnr;
  94     shiftidx         = nlist->shift;
  95     gid              = nlist->gid;
  96     shiftvec         = fr->shift_vec[0];
  97     fshift           = fr->fshift[0];
  98     facel            = fr->epsfac;
  99     charge           = mdatoms->chargeA;
 100     krf              = fr->ic->k_rf;
 101     krf2             = krf*2.0;
 102     crf              = fr->ic->c_rf;
 103     nvdwtype         = fr->ntype;
 104     vdwparam         = fr->nbfp;
 105     vdwtype          = mdatoms->typeA;
 106
 107     vftab            = kernel_data->table_vdw->data;
 108     vftabscale       = kernel_data->table_vdw->scale;
 109
 110     /* Setup water-specific parameters */
 111     inr              = nlist->iinr[0];
 112     iq0              = facel*charge[inr+0];
 113     iq1              = facel*charge[inr+1];
 114     iq2              = facel*charge[inr+2];
 115     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 116
 117     jq0              = charge[inr+0];
 118     jq1              = charge[inr+1];
 119     jq2              = charge[inr+2];
 120     vdwjidx0         = 2*vdwtype[inr+0];
 121     qq00             = iq0*jq0;
 122     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 123     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 124     qq01             = iq0*jq1;
 125     qq02             = iq0*jq2;
 126     qq10             = iq1*jq0;
 127     qq11             = iq1*jq1;
 128     qq12             = iq1*jq2;
 129     qq20             = iq2*jq0;
 130     qq21             = iq2*jq1;
 131     qq22             = iq2*jq2;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     /* Start outer loop over neighborlists */
 137     for(iidx=0; iidx<nri; iidx++)
 138     {
 139         /* Load shift vector for this list */
 140         i_shift_offset   = DIM*shiftidx[iidx];
 141         shX              = shiftvec[i_shift_offset+XX];
 142         shY              = shiftvec[i_shift_offset+YY];
 143         shZ              = shiftvec[i_shift_offset+ZZ];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 155         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 156         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 157         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 158         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 159         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 160         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 161         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 162         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 163
 164         fix0             = 0.0;
 165         fiy0             = 0.0;
 166         fiz0             = 0.0;
 167         fix1             = 0.0;
 168         fiy1             = 0.0;
 169         fiz1             = 0.0;
 170         fix2             = 0.0;
 171         fiy2             = 0.0;
 172         fiz2             = 0.0;
 173
 174         /* Reset potential sums */
 175         velecsum         = 0.0;
 176         vvdwsum          = 0.0;
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 180         {
 181             /* Get j neighbor index, and coordinate index */
 182             jnr              = jjnr[jidx];
 183             j_coord_offset   = DIM*jnr;
 184
 185             /* load j atom coordinates */
 186             jx0              = x[j_coord_offset+DIM*0+XX];
 187             jy0              = x[j_coord_offset+DIM*0+YY];
 188             jz0              = x[j_coord_offset+DIM*0+ZZ];
 189             jx1              = x[j_coord_offset+DIM*1+XX];
 190             jy1              = x[j_coord_offset+DIM*1+YY];
 191             jz1              = x[j_coord_offset+DIM*1+ZZ];
 192             jx2              = x[j_coord_offset+DIM*2+XX];
 193             jy2              = x[j_coord_offset+DIM*2+YY];
 194             jz2              = x[j_coord_offset+DIM*2+ZZ];
 195
 196             /* Calculate displacement vector */
 197             dx00             = ix0 - jx0;
 198             dy00             = iy0 - jy0;
 199             dz00             = iz0 - jz0;
 200             dx01             = ix0 - jx1;
 201             dy01             = iy0 - jy1;
 202             dz01             = iz0 - jz1;
 203             dx02             = ix0 - jx2;
 204             dy02             = iy0 - jy2;
 205             dz02             = iz0 - jz2;
 206             dx10             = ix1 - jx0;
 207             dy10             = iy1 - jy0;
 208             dz10             = iz1 - jz0;
 209             dx11             = ix1 - jx1;
 210             dy11             = iy1 - jy1;
 211             dz11             = iz1 - jz1;
 212             dx12             = ix1 - jx2;
 213             dy12             = iy1 - jy2;
 214             dz12             = iz1 - jz2;
 215             dx20             = ix2 - jx0;
 216             dy20             = iy2 - jy0;
 217             dz20             = iz2 - jz0;
 218             dx21             = ix2 - jx1;
 219             dy21             = iy2 - jy1;
 220             dz21             = iz2 - jz1;
 221             dx22             = ix2 - jx2;
 222             dy22             = iy2 - jy2;
 223             dz22             = iz2 - jz2;
 224
 225             /* Calculate squared distance and things based on it */
 226             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 227             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 228             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 229             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 230             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 231             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 232             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 233             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 234             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 235
 236             rinv00           = gmx_invsqrt(rsq00);
 237             rinv01           = gmx_invsqrt(rsq01);
 238             rinv02           = gmx_invsqrt(rsq02);
 239             rinv10           = gmx_invsqrt(rsq10);
 240             rinv11           = gmx_invsqrt(rsq11);
 241             rinv12           = gmx_invsqrt(rsq12);
 242             rinv20           = gmx_invsqrt(rsq20);
 243             rinv21           = gmx_invsqrt(rsq21);
 244             rinv22           = gmx_invsqrt(rsq22);
 245
 246             rinvsq00         = rinv00*rinv00;
 247             rinvsq01         = rinv01*rinv01;
 248             rinvsq02         = rinv02*rinv02;
 249             rinvsq10         = rinv10*rinv10;
 250             rinvsq11         = rinv11*rinv11;
 251             rinvsq12         = rinv12*rinv12;
 252             rinvsq20         = rinv20*rinv20;
 253             rinvsq21         = rinv21*rinv21;
 254             rinvsq22         = rinv22*rinv22;
 255
 256             /**************************
 257              * CALCULATE INTERACTIONS *
 258              **************************/
 259
 260             r00              = rsq00*rinv00;
 261
 262             /* Calculate table index by multiplying r with table scale and truncate to integer */
 263             rt               = r00*vftabscale;
 264             vfitab           = rt;
 265             vfeps            = rt-vfitab;
 266             vfitab           = 2*4*vfitab;
 267
 268             /* REACTION-FIELD ELECTROSTATICS */
 269             velec            = qq00*(rinv00+krf*rsq00-crf);
 270             felec            = qq00*(rinv00*rinvsq00-krf2);
 271
 272             /* CUBIC SPLINE TABLE DISPERSION */
 273             vfitab          += 0;
 274             Y                = vftab[vfitab];
 275             F                = vftab[vfitab+1];
 276             Geps             = vfeps*vftab[vfitab+2];
 277             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 278             Fp               = F+Geps+Heps2;
 279             VV               = Y+vfeps*Fp;
 280             vvdw6            = c6_00*VV;
 281             FF               = Fp+Geps+2.0*Heps2;
 282             fvdw6            = c6_00*FF;
 283
 284             /* CUBIC SPLINE TABLE REPULSION */
 285             Y                = vftab[vfitab+4];
 286             F                = vftab[vfitab+5];
 287             Geps             = vfeps*vftab[vfitab+6];
 288             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 289             Fp               = F+Geps+Heps2;
 290             VV               = Y+vfeps*Fp;
 291             vvdw12           = c12_00*VV;
 292             FF               = Fp+Geps+2.0*Heps2;
 293             fvdw12           = c12_00*FF;
 294             vvdw             = vvdw12+vvdw6;
 295             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 296
 297             /* Update potential sums from outer loop */
 298             velecsum        += velec;
 299             vvdwsum         += vvdw;
 300
 301             fscal            = felec+fvdw;
 302
 303             /* Calculate temporary vectorial force */
 304             tx               = fscal*dx00;
 305             ty               = fscal*dy00;
 306             tz               = fscal*dz00;
 307
 308             /* Update vectorial force */
 309             fix0            += tx;
 310             fiy0            += ty;
 311             fiz0            += tz;
 312             f[j_coord_offset+DIM*0+XX] -= tx;
 313             f[j_coord_offset+DIM*0+YY] -= ty;
 314             f[j_coord_offset+DIM*0+ZZ] -= tz;
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             /* REACTION-FIELD ELECTROSTATICS */
 321             velec            = qq01*(rinv01+krf*rsq01-crf);
 322             felec            = qq01*(rinv01*rinvsq01-krf2);
 323
 324             /* Update potential sums from outer loop */
 325             velecsum        += velec;
 326
 327             fscal            = felec;
 328
 329             /* Calculate temporary vectorial force */
 330             tx               = fscal*dx01;
 331             ty               = fscal*dy01;
 332             tz               = fscal*dz01;
 333
 334             /* Update vectorial force */
 335             fix0            += tx;
 336             fiy0            += ty;
 337             fiz0            += tz;
 338             f[j_coord_offset+DIM*1+XX] -= tx;
 339             f[j_coord_offset+DIM*1+YY] -= ty;
 340             f[j_coord_offset+DIM*1+ZZ] -= tz;
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             /* REACTION-FIELD ELECTROSTATICS */
 347             velec            = qq02*(rinv02+krf*rsq02-crf);
 348             felec            = qq02*(rinv02*rinvsq02-krf2);
 349
 350             /* Update potential sums from outer loop */
 351             velecsum        += velec;
 352
 353             fscal            = felec;
 354
 355             /* Calculate temporary vectorial force */
 356             tx               = fscal*dx02;
 357             ty               = fscal*dy02;
 358             tz               = fscal*dz02;
 359
 360             /* Update vectorial force */
 361             fix0            += tx;
 362             fiy0            += ty;
 363             fiz0            += tz;
 364             f[j_coord_offset+DIM*2+XX] -= tx;
 365             f[j_coord_offset+DIM*2+YY] -= ty;
 366             f[j_coord_offset+DIM*2+ZZ] -= tz;
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             /* REACTION-FIELD ELECTROSTATICS */
 373             velec            = qq10*(rinv10+krf*rsq10-crf);
 374             felec            = qq10*(rinv10*rinvsq10-krf2);
 375
 376             /* Update potential sums from outer loop */
 377             velecsum        += velec;
 378
 379             fscal            = felec;
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = fscal*dx10;
 383             ty               = fscal*dy10;
 384             tz               = fscal*dz10;
 385
 386             /* Update vectorial force */
 387             fix1            += tx;
 388             fiy1            += ty;
 389             fiz1            += tz;
 390             f[j_coord_offset+DIM*0+XX] -= tx;
 391             f[j_coord_offset+DIM*0+YY] -= ty;
 392             f[j_coord_offset+DIM*0+ZZ] -= tz;
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             /* REACTION-FIELD ELECTROSTATICS */
 399             velec            = qq11*(rinv11+krf*rsq11-crf);
 400             felec            = qq11*(rinv11*rinvsq11-krf2);
 401
 402             /* Update potential sums from outer loop */
 403             velecsum        += velec;
 404
 405             fscal            = felec;
 406
 407             /* Calculate temporary vectorial force */
 408             tx               = fscal*dx11;
 409             ty               = fscal*dy11;
 410             tz               = fscal*dz11;
 411
 412             /* Update vectorial force */
 413             fix1            += tx;
 414             fiy1            += ty;
 415             fiz1            += tz;
 416             f[j_coord_offset+DIM*1+XX] -= tx;
 417             f[j_coord_offset+DIM*1+YY] -= ty;
 418             f[j_coord_offset+DIM*1+ZZ] -= tz;
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             /* REACTION-FIELD ELECTROSTATICS */
 425             velec            = qq12*(rinv12+krf*rsq12-crf);
 426             felec            = qq12*(rinv12*rinvsq12-krf2);
 427
 428             /* Update potential sums from outer loop */
 429             velecsum        += velec;
 430
 431             fscal            = felec;
 432
 433             /* Calculate temporary vectorial force */
 434             tx               = fscal*dx12;
 435             ty               = fscal*dy12;
 436             tz               = fscal*dz12;
 437
 438             /* Update vectorial force */
 439             fix1            += tx;
 440             fiy1            += ty;
 441             fiz1            += tz;
 442             f[j_coord_offset+DIM*2+XX] -= tx;
 443             f[j_coord_offset+DIM*2+YY] -= ty;
 444             f[j_coord_offset+DIM*2+ZZ] -= tz;
 445
 446             /**************************
 447              * CALCULATE INTERACTIONS *
 448              **************************/
 449
 450             /* REACTION-FIELD ELECTROSTATICS */
 451             velec            = qq20*(rinv20+krf*rsq20-crf);
 452             felec            = qq20*(rinv20*rinvsq20-krf2);
 453
 454             /* Update potential sums from outer loop */
 455             velecsum        += velec;
 456
 457             fscal            = felec;
 458
 459             /* Calculate temporary vectorial force */
 460             tx               = fscal*dx20;
 461             ty               = fscal*dy20;
 462             tz               = fscal*dz20;
 463
 464             /* Update vectorial force */
 465             fix2            += tx;
 466             fiy2            += ty;
 467             fiz2            += tz;
 468             f[j_coord_offset+DIM*0+XX] -= tx;
 469             f[j_coord_offset+DIM*0+YY] -= ty;
 470             f[j_coord_offset+DIM*0+ZZ] -= tz;
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             /* REACTION-FIELD ELECTROSTATICS */
 477             velec            = qq21*(rinv21+krf*rsq21-crf);
 478             felec            = qq21*(rinv21*rinvsq21-krf2);
 479
 480             /* Update potential sums from outer loop */
 481             velecsum        += velec;
 482
 483             fscal            = felec;
 484
 485             /* Calculate temporary vectorial force */
 486             tx               = fscal*dx21;
 487             ty               = fscal*dy21;
 488             tz               = fscal*dz21;
 489
 490             /* Update vectorial force */
 491             fix2            += tx;
 492             fiy2            += ty;
 493             fiz2            += tz;
 494             f[j_coord_offset+DIM*1+XX] -= tx;
 495             f[j_coord_offset+DIM*1+YY] -= ty;
 496             f[j_coord_offset+DIM*1+ZZ] -= tz;
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             /* REACTION-FIELD ELECTROSTATICS */
 503             velec            = qq22*(rinv22+krf*rsq22-crf);
 504             felec            = qq22*(rinv22*rinvsq22-krf2);
 505
 506             /* Update potential sums from outer loop */
 507             velecsum        += velec;
 508
 509             fscal            = felec;
 510
 511             /* Calculate temporary vectorial force */
 512             tx               = fscal*dx22;
 513             ty               = fscal*dy22;
 514             tz               = fscal*dz22;
 515
 516             /* Update vectorial force */
 517             fix2            += tx;
 518             fiy2            += ty;
 519             fiz2            += tz;
 520             f[j_coord_offset+DIM*2+XX] -= tx;
 521             f[j_coord_offset+DIM*2+YY] -= ty;
 522             f[j_coord_offset+DIM*2+ZZ] -= tz;
 523
 524             /* Inner loop uses 313 flops */
 525         }
 526         /* End of innermost loop */
 527
 528         tx = ty = tz = 0;
 529         f[i_coord_offset+DIM*0+XX] += fix0;
 530         f[i_coord_offset+DIM*0+YY] += fiy0;
 531         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 532         tx                         += fix0;
 533         ty                         += fiy0;
 534         tz                         += fiz0;
 535         f[i_coord_offset+DIM*1+XX] += fix1;
 536         f[i_coord_offset+DIM*1+YY] += fiy1;
 537         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 538         tx                         += fix1;
 539         ty                         += fiy1;
 540         tz                         += fiz1;
 541         f[i_coord_offset+DIM*2+XX] += fix2;
 542         f[i_coord_offset+DIM*2+YY] += fiy2;
 543         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 544         tx                         += fix2;
 545         ty                         += fiy2;
 546         tz                         += fiz2;
 547         fshift[i_shift_offset+XX]  += tx;
 548         fshift[i_shift_offset+YY]  += ty;
 549         fshift[i_shift_offset+ZZ]  += tz;
 550
 551         ggid                        = gid[iidx];
 552         /* Update potential energies */
 553         kernel_data->energygrp_elec[ggid] += velecsum;
 554         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 555
 556         /* Increment number of inner iterations */
 557         inneriter                  += j_index_end - j_index_start;
 558
 559         /* Outer loop uses 32 flops */
 560     }
 561
 562     /* Increment number of outer iterations */
 563     outeriter        += nri;
 564
 565     /* Update outer/inner flops */
 566
 567     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*313);
 568 }
 569 /*
 570  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_c
 571  * Electrostatics interaction: ReactionField
 572  * VdW interaction:            CubicSplineTable
 573  * Geometry:                   Water3-Water3
 574  * Calculate force/pot:        Force
 575  */
 576 void
 577 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_c
 578                     (t_nblist * gmx_restrict                nlist,
 579                      rvec * gmx_restrict                    xx,
 580                      rvec * gmx_restrict                    ff,
 581                      t_forcerec * gmx_restrict              fr,
 582                      t_mdatoms * gmx_restrict               mdatoms,
 583                      nb_kernel_data_t * gmx_restrict        kernel_data,
 584                      t_nrnb * gmx_restrict                  nrnb)
 585 {
 586     int              i_shift_offset,i_coord_offset,j_coord_offset;
 587     int              j_index_start,j_index_end;
 588     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 589     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 590     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 591     real             *shiftvec,*fshift,*x,*f;
 592     int              vdwioffset0;
 593     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 594     int              vdwioffset1;
 595     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 596     int              vdwioffset2;
 597     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 598     int              vdwjidx0;
 599     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 600     int              vdwjidx1;
 601     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 602     int              vdwjidx2;
 603     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 604     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 605     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 606     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 607     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 608     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 609     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 610     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 611     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 612     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 613     real             velec,felec,velecsum,facel,crf,krf,krf2;
 614     real             *charge;
 615     int              nvdwtype;
 616     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 617     int              *vdwtype;
 618     real             *vdwparam;
 619     int              vfitab;
 620     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 621     real             *vftab;
 622
 623     x                = xx[0];
 624     f                = ff[0];
 625
 626     nri              = nlist->nri;
 627     iinr             = nlist->iinr;
 628     jindex           = nlist->jindex;
 629     jjnr             = nlist->jjnr;
 630     shiftidx         = nlist->shift;
 631     gid              = nlist->gid;
 632     shiftvec         = fr->shift_vec[0];
 633     fshift           = fr->fshift[0];
 634     facel            = fr->epsfac;
 635     charge           = mdatoms->chargeA;
 636     krf              = fr->ic->k_rf;
 637     krf2             = krf*2.0;
 638     crf              = fr->ic->c_rf;
 639     nvdwtype         = fr->ntype;
 640     vdwparam         = fr->nbfp;
 641     vdwtype          = mdatoms->typeA;
 642
 643     vftab            = kernel_data->table_vdw->data;
 644     vftabscale       = kernel_data->table_vdw->scale;
 645
 646     /* Setup water-specific parameters */
 647     inr              = nlist->iinr[0];
 648     iq0              = facel*charge[inr+0];
 649     iq1              = facel*charge[inr+1];
 650     iq2              = facel*charge[inr+2];
 651     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 652
 653     jq0              = charge[inr+0];
 654     jq1              = charge[inr+1];
 655     jq2              = charge[inr+2];
 656     vdwjidx0         = 2*vdwtype[inr+0];
 657     qq00             = iq0*jq0;
 658     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 659     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 660     qq01             = iq0*jq1;
 661     qq02             = iq0*jq2;
 662     qq10             = iq1*jq0;
 663     qq11             = iq1*jq1;
 664     qq12             = iq1*jq2;
 665     qq20             = iq2*jq0;
 666     qq21             = iq2*jq1;
 667     qq22             = iq2*jq2;
 668
 669     outeriter        = 0;
 670     inneriter        = 0;
 671
 672     /* Start outer loop over neighborlists */
 673     for(iidx=0; iidx<nri; iidx++)
 674     {
 675         /* Load shift vector for this list */
 676         i_shift_offset   = DIM*shiftidx[iidx];
 677         shX              = shiftvec[i_shift_offset+XX];
 678         shY              = shiftvec[i_shift_offset+YY];
 679         shZ              = shiftvec[i_shift_offset+ZZ];
 680
 681         /* Load limits for loop over neighbors */
 682         j_index_start    = jindex[iidx];
 683         j_index_end      = jindex[iidx+1];
 684
 685         /* Get outer coordinate index */
 686         inr              = iinr[iidx];
 687         i_coord_offset   = DIM*inr;
 688
 689         /* Load i particle coords and add shift vector */
 690         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 691         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 692         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 693         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 694         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 695         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 696         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 697         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 698         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 699
 700         fix0             = 0.0;
 701         fiy0             = 0.0;
 702         fiz0             = 0.0;
 703         fix1             = 0.0;
 704         fiy1             = 0.0;
 705         fiz1             = 0.0;
 706         fix2             = 0.0;
 707         fiy2             = 0.0;
 708         fiz2             = 0.0;
 709
 710         /* Start inner kernel loop */
 711         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 712         {
 713             /* Get j neighbor index, and coordinate index */
 714             jnr              = jjnr[jidx];
 715             j_coord_offset   = DIM*jnr;
 716
 717             /* load j atom coordinates */
 718             jx0              = x[j_coord_offset+DIM*0+XX];
 719             jy0              = x[j_coord_offset+DIM*0+YY];
 720             jz0              = x[j_coord_offset+DIM*0+ZZ];
 721             jx1              = x[j_coord_offset+DIM*1+XX];
 722             jy1              = x[j_coord_offset+DIM*1+YY];
 723             jz1              = x[j_coord_offset+DIM*1+ZZ];
 724             jx2              = x[j_coord_offset+DIM*2+XX];
 725             jy2              = x[j_coord_offset+DIM*2+YY];
 726             jz2              = x[j_coord_offset+DIM*2+ZZ];
 727
 728             /* Calculate displacement vector */
 729             dx00             = ix0 - jx0;
 730             dy00             = iy0 - jy0;
 731             dz00             = iz0 - jz0;
 732             dx01             = ix0 - jx1;
 733             dy01             = iy0 - jy1;
 734             dz01             = iz0 - jz1;
 735             dx02             = ix0 - jx2;
 736             dy02             = iy0 - jy2;
 737             dz02             = iz0 - jz2;
 738             dx10             = ix1 - jx0;
 739             dy10             = iy1 - jy0;
 740             dz10             = iz1 - jz0;
 741             dx11             = ix1 - jx1;
 742             dy11             = iy1 - jy1;
 743             dz11             = iz1 - jz1;
 744             dx12             = ix1 - jx2;
 745             dy12             = iy1 - jy2;
 746             dz12             = iz1 - jz2;
 747             dx20             = ix2 - jx0;
 748             dy20             = iy2 - jy0;
 749             dz20             = iz2 - jz0;
 750             dx21             = ix2 - jx1;
 751             dy21             = iy2 - jy1;
 752             dz21             = iz2 - jz1;
 753             dx22             = ix2 - jx2;
 754             dy22             = iy2 - jy2;
 755             dz22             = iz2 - jz2;
 756
 757             /* Calculate squared distance and things based on it */
 758             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 759             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 760             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 761             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 762             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 763             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 764             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 765             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 766             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 767
 768             rinv00           = gmx_invsqrt(rsq00);
 769             rinv01           = gmx_invsqrt(rsq01);
 770             rinv02           = gmx_invsqrt(rsq02);
 771             rinv10           = gmx_invsqrt(rsq10);
 772             rinv11           = gmx_invsqrt(rsq11);
 773             rinv12           = gmx_invsqrt(rsq12);
 774             rinv20           = gmx_invsqrt(rsq20);
 775             rinv21           = gmx_invsqrt(rsq21);
 776             rinv22           = gmx_invsqrt(rsq22);
 777
 778             rinvsq00         = rinv00*rinv00;
 779             rinvsq01         = rinv01*rinv01;
 780             rinvsq02         = rinv02*rinv02;
 781             rinvsq10         = rinv10*rinv10;
 782             rinvsq11         = rinv11*rinv11;
 783             rinvsq12         = rinv12*rinv12;
 784             rinvsq20         = rinv20*rinv20;
 785             rinvsq21         = rinv21*rinv21;
 786             rinvsq22         = rinv22*rinv22;
 787
 788             /**************************
 789              * CALCULATE INTERACTIONS *
 790              **************************/
 791
 792             r00              = rsq00*rinv00;
 793
 794             /* Calculate table index by multiplying r with table scale and truncate to integer */
 795             rt               = r00*vftabscale;
 796             vfitab           = rt;
 797             vfeps            = rt-vfitab;
 798             vfitab           = 2*4*vfitab;
 799
 800             /* REACTION-FIELD ELECTROSTATICS */
 801             felec            = qq00*(rinv00*rinvsq00-krf2);
 802
 803             /* CUBIC SPLINE TABLE DISPERSION */
 804             vfitab          += 0;
 805             F                = vftab[vfitab+1];
 806             Geps             = vfeps*vftab[vfitab+2];
 807             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 808             Fp               = F+Geps+Heps2;
 809             FF               = Fp+Geps+2.0*Heps2;
 810             fvdw6            = c6_00*FF;
 811
 812             /* CUBIC SPLINE TABLE REPULSION */
 813             F                = vftab[vfitab+5];
 814             Geps             = vfeps*vftab[vfitab+6];
 815             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 816             Fp               = F+Geps+Heps2;
 817             FF               = Fp+Geps+2.0*Heps2;
 818             fvdw12           = c12_00*FF;
 819             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 820
 821             fscal            = felec+fvdw;
 822
 823             /* Calculate temporary vectorial force */
 824             tx               = fscal*dx00;
 825             ty               = fscal*dy00;
 826             tz               = fscal*dz00;
 827
 828             /* Update vectorial force */
 829             fix0            += tx;
 830             fiy0            += ty;
 831             fiz0            += tz;
 832             f[j_coord_offset+DIM*0+XX] -= tx;
 833             f[j_coord_offset+DIM*0+YY] -= ty;
 834             f[j_coord_offset+DIM*0+ZZ] -= tz;
 835
 836             /**************************
 837              * CALCULATE INTERACTIONS *
 838              **************************/
 839
 840             /* REACTION-FIELD ELECTROSTATICS */
 841             felec            = qq01*(rinv01*rinvsq01-krf2);
 842
 843             fscal            = felec;
 844
 845             /* Calculate temporary vectorial force */
 846             tx               = fscal*dx01;
 847             ty               = fscal*dy01;
 848             tz               = fscal*dz01;
 849
 850             /* Update vectorial force */
 851             fix0            += tx;
 852             fiy0            += ty;
 853             fiz0            += tz;
 854             f[j_coord_offset+DIM*1+XX] -= tx;
 855             f[j_coord_offset+DIM*1+YY] -= ty;
 856             f[j_coord_offset+DIM*1+ZZ] -= tz;
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             /* REACTION-FIELD ELECTROSTATICS */
 863             felec            = qq02*(rinv02*rinvsq02-krf2);
 864
 865             fscal            = felec;
 866
 867             /* Calculate temporary vectorial force */
 868             tx               = fscal*dx02;
 869             ty               = fscal*dy02;
 870             tz               = fscal*dz02;
 871
 872             /* Update vectorial force */
 873             fix0            += tx;
 874             fiy0            += ty;
 875             fiz0            += tz;
 876             f[j_coord_offset+DIM*2+XX] -= tx;
 877             f[j_coord_offset+DIM*2+YY] -= ty;
 878             f[j_coord_offset+DIM*2+ZZ] -= tz;
 879
 880             /**************************
 881              * CALCULATE INTERACTIONS *
 882              **************************/
 883
 884             /* REACTION-FIELD ELECTROSTATICS */
 885             felec            = qq10*(rinv10*rinvsq10-krf2);
 886
 887             fscal            = felec;
 888
 889             /* Calculate temporary vectorial force */
 890             tx               = fscal*dx10;
 891             ty               = fscal*dy10;
 892             tz               = fscal*dz10;
 893
 894             /* Update vectorial force */
 895             fix1            += tx;
 896             fiy1            += ty;
 897             fiz1            += tz;
 898             f[j_coord_offset+DIM*0+XX] -= tx;
 899             f[j_coord_offset+DIM*0+YY] -= ty;
 900             f[j_coord_offset+DIM*0+ZZ] -= tz;
 901
 902             /**************************
 903              * CALCULATE INTERACTIONS *
 904              **************************/
 905
 906             /* REACTION-FIELD ELECTROSTATICS */
 907             felec            = qq11*(rinv11*rinvsq11-krf2);
 908
 909             fscal            = felec;
 910
 911             /* Calculate temporary vectorial force */
 912             tx               = fscal*dx11;
 913             ty               = fscal*dy11;
 914             tz               = fscal*dz11;
 915
 916             /* Update vectorial force */
 917             fix1            += tx;
 918             fiy1            += ty;
 919             fiz1            += tz;
 920             f[j_coord_offset+DIM*1+XX] -= tx;
 921             f[j_coord_offset+DIM*1+YY] -= ty;
 922             f[j_coord_offset+DIM*1+ZZ] -= tz;
 923
 924             /**************************
 925              * CALCULATE INTERACTIONS *
 926              **************************/
 927
 928             /* REACTION-FIELD ELECTROSTATICS */
 929             felec            = qq12*(rinv12*rinvsq12-krf2);
 930
 931             fscal            = felec;
 932
 933             /* Calculate temporary vectorial force */
 934             tx               = fscal*dx12;
 935             ty               = fscal*dy12;
 936             tz               = fscal*dz12;
 937
 938             /* Update vectorial force */
 939             fix1            += tx;
 940             fiy1            += ty;
 941             fiz1            += tz;
 942             f[j_coord_offset+DIM*2+XX] -= tx;
 943             f[j_coord_offset+DIM*2+YY] -= ty;
 944             f[j_coord_offset+DIM*2+ZZ] -= tz;
 945
 946             /**************************
 947              * CALCULATE INTERACTIONS *
 948              **************************/
 949
 950             /* REACTION-FIELD ELECTROSTATICS */
 951             felec            = qq20*(rinv20*rinvsq20-krf2);
 952
 953             fscal            = felec;
 954
 955             /* Calculate temporary vectorial force */
 956             tx               = fscal*dx20;
 957             ty               = fscal*dy20;
 958             tz               = fscal*dz20;
 959
 960             /* Update vectorial force */
 961             fix2            += tx;
 962             fiy2            += ty;
 963             fiz2            += tz;
 964             f[j_coord_offset+DIM*0+XX] -= tx;
 965             f[j_coord_offset+DIM*0+YY] -= ty;
 966             f[j_coord_offset+DIM*0+ZZ] -= tz;
 967
 968             /**************************
 969              * CALCULATE INTERACTIONS *
 970              **************************/
 971
 972             /* REACTION-FIELD ELECTROSTATICS */
 973             felec            = qq21*(rinv21*rinvsq21-krf2);
 974
 975             fscal            = felec;
 976
 977             /* Calculate temporary vectorial force */
 978             tx               = fscal*dx21;
 979             ty               = fscal*dy21;
 980             tz               = fscal*dz21;
 981
 982             /* Update vectorial force */
 983             fix2            += tx;
 984             fiy2            += ty;
 985             fiz2            += tz;
 986             f[j_coord_offset+DIM*1+XX] -= tx;
 987             f[j_coord_offset+DIM*1+YY] -= ty;
 988             f[j_coord_offset+DIM*1+ZZ] -= tz;
 989
 990             /**************************
 991              * CALCULATE INTERACTIONS *
 992              **************************/
 993
 994             /* REACTION-FIELD ELECTROSTATICS */
 995             felec            = qq22*(rinv22*rinvsq22-krf2);
 996
 997             fscal            = felec;
 998
 999             /* Calculate temporary vectorial force */
1000             tx               = fscal*dx22;
1001             ty               = fscal*dy22;
1002             tz               = fscal*dz22;
1003
1004             /* Update vectorial force */
1005             fix2            += tx;
1006             fiy2            += ty;
1007             fiz2            += tz;
1008             f[j_coord_offset+DIM*2+XX] -= tx;
1009             f[j_coord_offset+DIM*2+YY] -= ty;
1010             f[j_coord_offset+DIM*2+ZZ] -= tz;
1011
1012             /* Inner loop uses 260 flops */
1013         }
1014         /* End of innermost loop */
1015
1016         tx = ty = tz = 0;
1017         f[i_coord_offset+DIM*0+XX] += fix0;
1018         f[i_coord_offset+DIM*0+YY] += fiy0;
1019         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1020         tx                         += fix0;
1021         ty                         += fiy0;
1022         tz                         += fiz0;
1023         f[i_coord_offset+DIM*1+XX] += fix1;
1024         f[i_coord_offset+DIM*1+YY] += fiy1;
1025         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1026         tx                         += fix1;
1027         ty                         += fiy1;
1028         tz                         += fiz1;
1029         f[i_coord_offset+DIM*2+XX] += fix2;
1030         f[i_coord_offset+DIM*2+YY] += fiy2;
1031         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1032         tx                         += fix2;
1033         ty                         += fiy2;
1034         tz                         += fiz2;
1035         fshift[i_shift_offset+XX]  += tx;
1036         fshift[i_shift_offset+YY]  += ty;
1037         fshift[i_shift_offset+ZZ]  += tz;
1038
1039         /* Increment number of inner iterations */
1040         inneriter                  += j_index_end - j_index_start;
1041
1042         /* Outer loop uses 30 flops */
1043     }
1044
1045     /* Increment number of outer iterations */
1046     outeriter        += nri;
1047
1048     /* Update outer/inner flops */
1049
1050     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*260);
1051 }