src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_c.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS c kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_VF_c
  49  * Electrostatics interaction: Coulomb
  50  * VdW interaction:            CubicSplineTable
  51  * Geometry:                   Water3-Water3
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_VF_c
  56                     (t_nblist                    * gmx_restrict       nlist,
  57                      rvec                        * gmx_restrict          xx,
  58                      rvec                        * gmx_restrict          ff,
  59                      t_forcerec                  * gmx_restrict          fr,
  60                      t_mdatoms                   * gmx_restrict     mdatoms,
  61                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  62                      t_nrnb                      * gmx_restrict        nrnb)
  63 {
  64     int              i_shift_offset,i_coord_offset,j_coord_offset;
  65     int              j_index_start,j_index_end;
  66     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  67     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  68     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  69     real             *shiftvec,*fshift,*x,*f;
  70     int              vdwioffset0;
  71     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwioffset1;
  73     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  74     int              vdwioffset2;
  75     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  76     int              vdwjidx0;
  77     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  78     int              vdwjidx1;
  79     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  80     int              vdwjidx2;
  81     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  82     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  83     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
  84     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
  85     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  86     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  87     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  88     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  89     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  90     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  91     real             velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     int              vfitab;
  98     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
  99     real             *vftab;
 100
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = fr->epsfac;
 113     charge           = mdatoms->chargeA;
 114     nvdwtype         = fr->ntype;
 115     vdwparam         = fr->nbfp;
 116     vdwtype          = mdatoms->typeA;
 117
 118     vftab            = kernel_data->table_vdw->data;
 119     vftabscale       = kernel_data->table_vdw->scale;
 120
 121     /* Setup water-specific parameters */
 122     inr              = nlist->iinr[0];
 123     iq0              = facel*charge[inr+0];
 124     iq1              = facel*charge[inr+1];
 125     iq2              = facel*charge[inr+2];
 126     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 127
 128     jq0              = charge[inr+0];
 129     jq1              = charge[inr+1];
 130     jq2              = charge[inr+2];
 131     vdwjidx0         = 2*vdwtype[inr+0];
 132     qq00             = iq0*jq0;
 133     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 134     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 135     qq01             = iq0*jq1;
 136     qq02             = iq0*jq2;
 137     qq10             = iq1*jq0;
 138     qq11             = iq1*jq1;
 139     qq12             = iq1*jq2;
 140     qq20             = iq2*jq0;
 141     qq21             = iq2*jq1;
 142     qq22             = iq2*jq2;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152         shX              = shiftvec[i_shift_offset+XX];
 153         shY              = shiftvec[i_shift_offset+YY];
 154         shZ              = shiftvec[i_shift_offset+ZZ];
 155
 156         /* Load limits for loop over neighbors */
 157         j_index_start    = jindex[iidx];
 158         j_index_end      = jindex[iidx+1];
 159
 160         /* Get outer coordinate index */
 161         inr              = iinr[iidx];
 162         i_coord_offset   = DIM*inr;
 163
 164         /* Load i particle coords and add shift vector */
 165         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 166         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 167         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 168         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 169         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 170         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 171         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 172         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 173         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 174
 175         fix0             = 0.0;
 176         fiy0             = 0.0;
 177         fiz0             = 0.0;
 178         fix1             = 0.0;
 179         fiy1             = 0.0;
 180         fiz1             = 0.0;
 181         fix2             = 0.0;
 182         fiy2             = 0.0;
 183         fiz2             = 0.0;
 184
 185         /* Reset potential sums */
 186         velecsum         = 0.0;
 187         vvdwsum          = 0.0;
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 191         {
 192             /* Get j neighbor index, and coordinate index */
 193             jnr              = jjnr[jidx];
 194             j_coord_offset   = DIM*jnr;
 195
 196             /* load j atom coordinates */
 197             jx0              = x[j_coord_offset+DIM*0+XX];
 198             jy0              = x[j_coord_offset+DIM*0+YY];
 199             jz0              = x[j_coord_offset+DIM*0+ZZ];
 200             jx1              = x[j_coord_offset+DIM*1+XX];
 201             jy1              = x[j_coord_offset+DIM*1+YY];
 202             jz1              = x[j_coord_offset+DIM*1+ZZ];
 203             jx2              = x[j_coord_offset+DIM*2+XX];
 204             jy2              = x[j_coord_offset+DIM*2+YY];
 205             jz2              = x[j_coord_offset+DIM*2+ZZ];
 206
 207             /* Calculate displacement vector */
 208             dx00             = ix0 - jx0;
 209             dy00             = iy0 - jy0;
 210             dz00             = iz0 - jz0;
 211             dx01             = ix0 - jx1;
 212             dy01             = iy0 - jy1;
 213             dz01             = iz0 - jz1;
 214             dx02             = ix0 - jx2;
 215             dy02             = iy0 - jy2;
 216             dz02             = iz0 - jz2;
 217             dx10             = ix1 - jx0;
 218             dy10             = iy1 - jy0;
 219             dz10             = iz1 - jz0;
 220             dx11             = ix1 - jx1;
 221             dy11             = iy1 - jy1;
 222             dz11             = iz1 - jz1;
 223             dx12             = ix1 - jx2;
 224             dy12             = iy1 - jy2;
 225             dz12             = iz1 - jz2;
 226             dx20             = ix2 - jx0;
 227             dy20             = iy2 - jy0;
 228             dz20             = iz2 - jz0;
 229             dx21             = ix2 - jx1;
 230             dy21             = iy2 - jy1;
 231             dz21             = iz2 - jz1;
 232             dx22             = ix2 - jx2;
 233             dy22             = iy2 - jy2;
 234             dz22             = iz2 - jz2;
 235
 236             /* Calculate squared distance and things based on it */
 237             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 238             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 239             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 240             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 241             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 242             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 243             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 244             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 245             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 246
 247             rinv00           = gmx_invsqrt(rsq00);
 248             rinv01           = gmx_invsqrt(rsq01);
 249             rinv02           = gmx_invsqrt(rsq02);
 250             rinv10           = gmx_invsqrt(rsq10);
 251             rinv11           = gmx_invsqrt(rsq11);
 252             rinv12           = gmx_invsqrt(rsq12);
 253             rinv20           = gmx_invsqrt(rsq20);
 254             rinv21           = gmx_invsqrt(rsq21);
 255             rinv22           = gmx_invsqrt(rsq22);
 256
 257             rinvsq00         = rinv00*rinv00;
 258             rinvsq01         = rinv01*rinv01;
 259             rinvsq02         = rinv02*rinv02;
 260             rinvsq10         = rinv10*rinv10;
 261             rinvsq11         = rinv11*rinv11;
 262             rinvsq12         = rinv12*rinv12;
 263             rinvsq20         = rinv20*rinv20;
 264             rinvsq21         = rinv21*rinv21;
 265             rinvsq22         = rinv22*rinv22;
 266
 267             /**************************
 268              * CALCULATE INTERACTIONS *
 269              **************************/
 270
 271             r00              = rsq00*rinv00;
 272
 273             /* Calculate table index by multiplying r with table scale and truncate to integer */
 274             rt               = r00*vftabscale;
 275             vfitab           = rt;
 276             vfeps            = rt-vfitab;
 277             vfitab           = 2*4*vfitab;
 278
 279             /* COULOMB ELECTROSTATICS */
 280             velec            = qq00*rinv00;
 281             felec            = velec*rinvsq00;
 282
 283             /* CUBIC SPLINE TABLE DISPERSION */
 284             vfitab          += 0;
 285             Y                = vftab[vfitab];
 286             F                = vftab[vfitab+1];
 287             Geps             = vfeps*vftab[vfitab+2];
 288             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 289             Fp               = F+Geps+Heps2;
 290             VV               = Y+vfeps*Fp;
 291             vvdw6            = c6_00*VV;
 292             FF               = Fp+Geps+2.0*Heps2;
 293             fvdw6            = c6_00*FF;
 294
 295             /* CUBIC SPLINE TABLE REPULSION */
 296             Y                = vftab[vfitab+4];
 297             F                = vftab[vfitab+5];
 298             Geps             = vfeps*vftab[vfitab+6];
 299             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 300             Fp               = F+Geps+Heps2;
 301             VV               = Y+vfeps*Fp;
 302             vvdw12           = c12_00*VV;
 303             FF               = Fp+Geps+2.0*Heps2;
 304             fvdw12           = c12_00*FF;
 305             vvdw             = vvdw12+vvdw6;
 306             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 307
 308             /* Update potential sums from outer loop */
 309             velecsum        += velec;
 310             vvdwsum         += vvdw;
 311
 312             fscal            = felec+fvdw;
 313
 314             /* Calculate temporary vectorial force */
 315             tx               = fscal*dx00;
 316             ty               = fscal*dy00;
 317             tz               = fscal*dz00;
 318
 319             /* Update vectorial force */
 320             fix0            += tx;
 321             fiy0            += ty;
 322             fiz0            += tz;
 323             f[j_coord_offset+DIM*0+XX] -= tx;
 324             f[j_coord_offset+DIM*0+YY] -= ty;
 325             f[j_coord_offset+DIM*0+ZZ] -= tz;
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             /* COULOMB ELECTROSTATICS */
 332             velec            = qq01*rinv01;
 333             felec            = velec*rinvsq01;
 334
 335             /* Update potential sums from outer loop */
 336             velecsum        += velec;
 337
 338             fscal            = felec;
 339
 340             /* Calculate temporary vectorial force */
 341             tx               = fscal*dx01;
 342             ty               = fscal*dy01;
 343             tz               = fscal*dz01;
 344
 345             /* Update vectorial force */
 346             fix0            += tx;
 347             fiy0            += ty;
 348             fiz0            += tz;
 349             f[j_coord_offset+DIM*1+XX] -= tx;
 350             f[j_coord_offset+DIM*1+YY] -= ty;
 351             f[j_coord_offset+DIM*1+ZZ] -= tz;
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             /* COULOMB ELECTROSTATICS */
 358             velec            = qq02*rinv02;
 359             felec            = velec*rinvsq02;
 360
 361             /* Update potential sums from outer loop */
 362             velecsum        += velec;
 363
 364             fscal            = felec;
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = fscal*dx02;
 368             ty               = fscal*dy02;
 369             tz               = fscal*dz02;
 370
 371             /* Update vectorial force */
 372             fix0            += tx;
 373             fiy0            += ty;
 374             fiz0            += tz;
 375             f[j_coord_offset+DIM*2+XX] -= tx;
 376             f[j_coord_offset+DIM*2+YY] -= ty;
 377             f[j_coord_offset+DIM*2+ZZ] -= tz;
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* COULOMB ELECTROSTATICS */
 384             velec            = qq10*rinv10;
 385             felec            = velec*rinvsq10;
 386
 387             /* Update potential sums from outer loop */
 388             velecsum        += velec;
 389
 390             fscal            = felec;
 391
 392             /* Calculate temporary vectorial force */
 393             tx               = fscal*dx10;
 394             ty               = fscal*dy10;
 395             tz               = fscal*dz10;
 396
 397             /* Update vectorial force */
 398             fix1            += tx;
 399             fiy1            += ty;
 400             fiz1            += tz;
 401             f[j_coord_offset+DIM*0+XX] -= tx;
 402             f[j_coord_offset+DIM*0+YY] -= ty;
 403             f[j_coord_offset+DIM*0+ZZ] -= tz;
 404
 405             /**************************
 406              * CALCULATE INTERACTIONS *
 407              **************************/
 408
 409             /* COULOMB ELECTROSTATICS */
 410             velec            = qq11*rinv11;
 411             felec            = velec*rinvsq11;
 412
 413             /* Update potential sums from outer loop */
 414             velecsum        += velec;
 415
 416             fscal            = felec;
 417
 418             /* Calculate temporary vectorial force */
 419             tx               = fscal*dx11;
 420             ty               = fscal*dy11;
 421             tz               = fscal*dz11;
 422
 423             /* Update vectorial force */
 424             fix1            += tx;
 425             fiy1            += ty;
 426             fiz1            += tz;
 427             f[j_coord_offset+DIM*1+XX] -= tx;
 428             f[j_coord_offset+DIM*1+YY] -= ty;
 429             f[j_coord_offset+DIM*1+ZZ] -= tz;
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             /* COULOMB ELECTROSTATICS */
 436             velec            = qq12*rinv12;
 437             felec            = velec*rinvsq12;
 438
 439             /* Update potential sums from outer loop */
 440             velecsum        += velec;
 441
 442             fscal            = felec;
 443
 444             /* Calculate temporary vectorial force */
 445             tx               = fscal*dx12;
 446             ty               = fscal*dy12;
 447             tz               = fscal*dz12;
 448
 449             /* Update vectorial force */
 450             fix1            += tx;
 451             fiy1            += ty;
 452             fiz1            += tz;
 453             f[j_coord_offset+DIM*2+XX] -= tx;
 454             f[j_coord_offset+DIM*2+YY] -= ty;
 455             f[j_coord_offset+DIM*2+ZZ] -= tz;
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             /* COULOMB ELECTROSTATICS */
 462             velec            = qq20*rinv20;
 463             felec            = velec*rinvsq20;
 464
 465             /* Update potential sums from outer loop */
 466             velecsum        += velec;
 467
 468             fscal            = felec;
 469
 470             /* Calculate temporary vectorial force */
 471             tx               = fscal*dx20;
 472             ty               = fscal*dy20;
 473             tz               = fscal*dz20;
 474
 475             /* Update vectorial force */
 476             fix2            += tx;
 477             fiy2            += ty;
 478             fiz2            += tz;
 479             f[j_coord_offset+DIM*0+XX] -= tx;
 480             f[j_coord_offset+DIM*0+YY] -= ty;
 481             f[j_coord_offset+DIM*0+ZZ] -= tz;
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             /* COULOMB ELECTROSTATICS */
 488             velec            = qq21*rinv21;
 489             felec            = velec*rinvsq21;
 490
 491             /* Update potential sums from outer loop */
 492             velecsum        += velec;
 493
 494             fscal            = felec;
 495
 496             /* Calculate temporary vectorial force */
 497             tx               = fscal*dx21;
 498             ty               = fscal*dy21;
 499             tz               = fscal*dz21;
 500
 501             /* Update vectorial force */
 502             fix2            += tx;
 503             fiy2            += ty;
 504             fiz2            += tz;
 505             f[j_coord_offset+DIM*1+XX] -= tx;
 506             f[j_coord_offset+DIM*1+YY] -= ty;
 507             f[j_coord_offset+DIM*1+ZZ] -= tz;
 508
 509             /**************************
 510              * CALCULATE INTERACTIONS *
 511              **************************/
 512
 513             /* COULOMB ELECTROSTATICS */
 514             velec            = qq22*rinv22;
 515             felec            = velec*rinvsq22;
 516
 517             /* Update potential sums from outer loop */
 518             velecsum        += velec;
 519
 520             fscal            = felec;
 521
 522             /* Calculate temporary vectorial force */
 523             tx               = fscal*dx22;
 524             ty               = fscal*dy22;
 525             tz               = fscal*dz22;
 526
 527             /* Update vectorial force */
 528             fix2            += tx;
 529             fiy2            += ty;
 530             fiz2            += tz;
 531             f[j_coord_offset+DIM*2+XX] -= tx;
 532             f[j_coord_offset+DIM*2+YY] -= ty;
 533             f[j_coord_offset+DIM*2+ZZ] -= tz;
 534
 535             /* Inner loop uses 277 flops */
 536         }
 537         /* End of innermost loop */
 538
 539         tx = ty = tz = 0;
 540         f[i_coord_offset+DIM*0+XX] += fix0;
 541         f[i_coord_offset+DIM*0+YY] += fiy0;
 542         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 543         tx                         += fix0;
 544         ty                         += fiy0;
 545         tz                         += fiz0;
 546         f[i_coord_offset+DIM*1+XX] += fix1;
 547         f[i_coord_offset+DIM*1+YY] += fiy1;
 548         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 549         tx                         += fix1;
 550         ty                         += fiy1;
 551         tz                         += fiz1;
 552         f[i_coord_offset+DIM*2+XX] += fix2;
 553         f[i_coord_offset+DIM*2+YY] += fiy2;
 554         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 555         tx                         += fix2;
 556         ty                         += fiy2;
 557         tz                         += fiz2;
 558         fshift[i_shift_offset+XX]  += tx;
 559         fshift[i_shift_offset+YY]  += ty;
 560         fshift[i_shift_offset+ZZ]  += tz;
 561
 562         ggid                        = gid[iidx];
 563         /* Update potential energies */
 564         kernel_data->energygrp_elec[ggid] += velecsum;
 565         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 566
 567         /* Increment number of inner iterations */
 568         inneriter                  += j_index_end - j_index_start;
 569
 570         /* Outer loop uses 32 flops */
 571     }
 572
 573     /* Increment number of outer iterations */
 574     outeriter        += nri;
 575
 576     /* Update outer/inner flops */
 577
 578     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*277);
 579 }
 580 /*
 581  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_F_c
 582  * Electrostatics interaction: Coulomb
 583  * VdW interaction:            CubicSplineTable
 584  * Geometry:                   Water3-Water3
 585  * Calculate force/pot:        Force
 586  */
 587 void
 588 nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_F_c
 589                     (t_nblist                    * gmx_restrict       nlist,
 590                      rvec                        * gmx_restrict          xx,
 591                      rvec                        * gmx_restrict          ff,
 592                      t_forcerec                  * gmx_restrict          fr,
 593                      t_mdatoms                   * gmx_restrict     mdatoms,
 594                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 595                      t_nrnb                      * gmx_restrict        nrnb)
 596 {
 597     int              i_shift_offset,i_coord_offset,j_coord_offset;
 598     int              j_index_start,j_index_end;
 599     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 600     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 601     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 602     real             *shiftvec,*fshift,*x,*f;
 603     int              vdwioffset0;
 604     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 605     int              vdwioffset1;
 606     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 607     int              vdwioffset2;
 608     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 609     int              vdwjidx0;
 610     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 611     int              vdwjidx1;
 612     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 613     int              vdwjidx2;
 614     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 615     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 616     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 617     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 618     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 619     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 620     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 621     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 622     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 623     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 624     real             velec,felec,velecsum,facel,crf,krf,krf2;
 625     real             *charge;
 626     int              nvdwtype;
 627     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 628     int              *vdwtype;
 629     real             *vdwparam;
 630     int              vfitab;
 631     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 632     real             *vftab;
 633
 634     x                = xx[0];
 635     f                = ff[0];
 636
 637     nri              = nlist->nri;
 638     iinr             = nlist->iinr;
 639     jindex           = nlist->jindex;
 640     jjnr             = nlist->jjnr;
 641     shiftidx         = nlist->shift;
 642     gid              = nlist->gid;
 643     shiftvec         = fr->shift_vec[0];
 644     fshift           = fr->fshift[0];
 645     facel            = fr->epsfac;
 646     charge           = mdatoms->chargeA;
 647     nvdwtype         = fr->ntype;
 648     vdwparam         = fr->nbfp;
 649     vdwtype          = mdatoms->typeA;
 650
 651     vftab            = kernel_data->table_vdw->data;
 652     vftabscale       = kernel_data->table_vdw->scale;
 653
 654     /* Setup water-specific parameters */
 655     inr              = nlist->iinr[0];
 656     iq0              = facel*charge[inr+0];
 657     iq1              = facel*charge[inr+1];
 658     iq2              = facel*charge[inr+2];
 659     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 660
 661     jq0              = charge[inr+0];
 662     jq1              = charge[inr+1];
 663     jq2              = charge[inr+2];
 664     vdwjidx0         = 2*vdwtype[inr+0];
 665     qq00             = iq0*jq0;
 666     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 667     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 668     qq01             = iq0*jq1;
 669     qq02             = iq0*jq2;
 670     qq10             = iq1*jq0;
 671     qq11             = iq1*jq1;
 672     qq12             = iq1*jq2;
 673     qq20             = iq2*jq0;
 674     qq21             = iq2*jq1;
 675     qq22             = iq2*jq2;
 676
 677     outeriter        = 0;
 678     inneriter        = 0;
 679
 680     /* Start outer loop over neighborlists */
 681     for(iidx=0; iidx<nri; iidx++)
 682     {
 683         /* Load shift vector for this list */
 684         i_shift_offset   = DIM*shiftidx[iidx];
 685         shX              = shiftvec[i_shift_offset+XX];
 686         shY              = shiftvec[i_shift_offset+YY];
 687         shZ              = shiftvec[i_shift_offset+ZZ];
 688
 689         /* Load limits for loop over neighbors */
 690         j_index_start    = jindex[iidx];
 691         j_index_end      = jindex[iidx+1];
 692
 693         /* Get outer coordinate index */
 694         inr              = iinr[iidx];
 695         i_coord_offset   = DIM*inr;
 696
 697         /* Load i particle coords and add shift vector */
 698         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 699         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 700         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 701         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 702         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 703         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 704         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 705         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 706         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 707
 708         fix0             = 0.0;
 709         fiy0             = 0.0;
 710         fiz0             = 0.0;
 711         fix1             = 0.0;
 712         fiy1             = 0.0;
 713         fiz1             = 0.0;
 714         fix2             = 0.0;
 715         fiy2             = 0.0;
 716         fiz2             = 0.0;
 717
 718         /* Start inner kernel loop */
 719         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 720         {
 721             /* Get j neighbor index, and coordinate index */
 722             jnr              = jjnr[jidx];
 723             j_coord_offset   = DIM*jnr;
 724
 725             /* load j atom coordinates */
 726             jx0              = x[j_coord_offset+DIM*0+XX];
 727             jy0              = x[j_coord_offset+DIM*0+YY];
 728             jz0              = x[j_coord_offset+DIM*0+ZZ];
 729             jx1              = x[j_coord_offset+DIM*1+XX];
 730             jy1              = x[j_coord_offset+DIM*1+YY];
 731             jz1              = x[j_coord_offset+DIM*1+ZZ];
 732             jx2              = x[j_coord_offset+DIM*2+XX];
 733             jy2              = x[j_coord_offset+DIM*2+YY];
 734             jz2              = x[j_coord_offset+DIM*2+ZZ];
 735
 736             /* Calculate displacement vector */
 737             dx00             = ix0 - jx0;
 738             dy00             = iy0 - jy0;
 739             dz00             = iz0 - jz0;
 740             dx01             = ix0 - jx1;
 741             dy01             = iy0 - jy1;
 742             dz01             = iz0 - jz1;
 743             dx02             = ix0 - jx2;
 744             dy02             = iy0 - jy2;
 745             dz02             = iz0 - jz2;
 746             dx10             = ix1 - jx0;
 747             dy10             = iy1 - jy0;
 748             dz10             = iz1 - jz0;
 749             dx11             = ix1 - jx1;
 750             dy11             = iy1 - jy1;
 751             dz11             = iz1 - jz1;
 752             dx12             = ix1 - jx2;
 753             dy12             = iy1 - jy2;
 754             dz12             = iz1 - jz2;
 755             dx20             = ix2 - jx0;
 756             dy20             = iy2 - jy0;
 757             dz20             = iz2 - jz0;
 758             dx21             = ix2 - jx1;
 759             dy21             = iy2 - jy1;
 760             dz21             = iz2 - jz1;
 761             dx22             = ix2 - jx2;
 762             dy22             = iy2 - jy2;
 763             dz22             = iz2 - jz2;
 764
 765             /* Calculate squared distance and things based on it */
 766             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 767             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 768             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 769             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 770             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 771             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 772             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 773             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 774             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 775
 776             rinv00           = gmx_invsqrt(rsq00);
 777             rinv01           = gmx_invsqrt(rsq01);
 778             rinv02           = gmx_invsqrt(rsq02);
 779             rinv10           = gmx_invsqrt(rsq10);
 780             rinv11           = gmx_invsqrt(rsq11);
 781             rinv12           = gmx_invsqrt(rsq12);
 782             rinv20           = gmx_invsqrt(rsq20);
 783             rinv21           = gmx_invsqrt(rsq21);
 784             rinv22           = gmx_invsqrt(rsq22);
 785
 786             rinvsq00         = rinv00*rinv00;
 787             rinvsq01         = rinv01*rinv01;
 788             rinvsq02         = rinv02*rinv02;
 789             rinvsq10         = rinv10*rinv10;
 790             rinvsq11         = rinv11*rinv11;
 791             rinvsq12         = rinv12*rinv12;
 792             rinvsq20         = rinv20*rinv20;
 793             rinvsq21         = rinv21*rinv21;
 794             rinvsq22         = rinv22*rinv22;
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             r00              = rsq00*rinv00;
 801
 802             /* Calculate table index by multiplying r with table scale and truncate to integer */
 803             rt               = r00*vftabscale;
 804             vfitab           = rt;
 805             vfeps            = rt-vfitab;
 806             vfitab           = 2*4*vfitab;
 807
 808             /* COULOMB ELECTROSTATICS */
 809             velec            = qq00*rinv00;
 810             felec            = velec*rinvsq00;
 811
 812             /* CUBIC SPLINE TABLE DISPERSION */
 813             vfitab          += 0;
 814             F                = vftab[vfitab+1];
 815             Geps             = vfeps*vftab[vfitab+2];
 816             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 817             Fp               = F+Geps+Heps2;
 818             FF               = Fp+Geps+2.0*Heps2;
 819             fvdw6            = c6_00*FF;
 820
 821             /* CUBIC SPLINE TABLE REPULSION */
 822             F                = vftab[vfitab+5];
 823             Geps             = vfeps*vftab[vfitab+6];
 824             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 825             Fp               = F+Geps+Heps2;
 826             FF               = Fp+Geps+2.0*Heps2;
 827             fvdw12           = c12_00*FF;
 828             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 829
 830             fscal            = felec+fvdw;
 831
 832             /* Calculate temporary vectorial force */
 833             tx               = fscal*dx00;
 834             ty               = fscal*dy00;
 835             tz               = fscal*dz00;
 836
 837             /* Update vectorial force */
 838             fix0            += tx;
 839             fiy0            += ty;
 840             fiz0            += tz;
 841             f[j_coord_offset+DIM*0+XX] -= tx;
 842             f[j_coord_offset+DIM*0+YY] -= ty;
 843             f[j_coord_offset+DIM*0+ZZ] -= tz;
 844
 845             /**************************
 846              * CALCULATE INTERACTIONS *
 847              **************************/
 848
 849             /* COULOMB ELECTROSTATICS */
 850             velec            = qq01*rinv01;
 851             felec            = velec*rinvsq01;
 852
 853             fscal            = felec;
 854
 855             /* Calculate temporary vectorial force */
 856             tx               = fscal*dx01;
 857             ty               = fscal*dy01;
 858             tz               = fscal*dz01;
 859
 860             /* Update vectorial force */
 861             fix0            += tx;
 862             fiy0            += ty;
 863             fiz0            += tz;
 864             f[j_coord_offset+DIM*1+XX] -= tx;
 865             f[j_coord_offset+DIM*1+YY] -= ty;
 866             f[j_coord_offset+DIM*1+ZZ] -= tz;
 867
 868             /**************************
 869              * CALCULATE INTERACTIONS *
 870              **************************/
 871
 872             /* COULOMB ELECTROSTATICS */
 873             velec            = qq02*rinv02;
 874             felec            = velec*rinvsq02;
 875
 876             fscal            = felec;
 877
 878             /* Calculate temporary vectorial force */
 879             tx               = fscal*dx02;
 880             ty               = fscal*dy02;
 881             tz               = fscal*dz02;
 882
 883             /* Update vectorial force */
 884             fix0            += tx;
 885             fiy0            += ty;
 886             fiz0            += tz;
 887             f[j_coord_offset+DIM*2+XX] -= tx;
 888             f[j_coord_offset+DIM*2+YY] -= ty;
 889             f[j_coord_offset+DIM*2+ZZ] -= tz;
 890
 891             /**************************
 892              * CALCULATE INTERACTIONS *
 893              **************************/
 894
 895             /* COULOMB ELECTROSTATICS */
 896             velec            = qq10*rinv10;
 897             felec            = velec*rinvsq10;
 898
 899             fscal            = felec;
 900
 901             /* Calculate temporary vectorial force */
 902             tx               = fscal*dx10;
 903             ty               = fscal*dy10;
 904             tz               = fscal*dz10;
 905
 906             /* Update vectorial force */
 907             fix1            += tx;
 908             fiy1            += ty;
 909             fiz1            += tz;
 910             f[j_coord_offset+DIM*0+XX] -= tx;
 911             f[j_coord_offset+DIM*0+YY] -= ty;
 912             f[j_coord_offset+DIM*0+ZZ] -= tz;
 913
 914             /**************************
 915              * CALCULATE INTERACTIONS *
 916              **************************/
 917
 918             /* COULOMB ELECTROSTATICS */
 919             velec            = qq11*rinv11;
 920             felec            = velec*rinvsq11;
 921
 922             fscal            = felec;
 923
 924             /* Calculate temporary vectorial force */
 925             tx               = fscal*dx11;
 926             ty               = fscal*dy11;
 927             tz               = fscal*dz11;
 928
 929             /* Update vectorial force */
 930             fix1            += tx;
 931             fiy1            += ty;
 932             fiz1            += tz;
 933             f[j_coord_offset+DIM*1+XX] -= tx;
 934             f[j_coord_offset+DIM*1+YY] -= ty;
 935             f[j_coord_offset+DIM*1+ZZ] -= tz;
 936
 937             /**************************
 938              * CALCULATE INTERACTIONS *
 939              **************************/
 940
 941             /* COULOMB ELECTROSTATICS */
 942             velec            = qq12*rinv12;
 943             felec            = velec*rinvsq12;
 944
 945             fscal            = felec;
 946
 947             /* Calculate temporary vectorial force */
 948             tx               = fscal*dx12;
 949             ty               = fscal*dy12;
 950             tz               = fscal*dz12;
 951
 952             /* Update vectorial force */
 953             fix1            += tx;
 954             fiy1            += ty;
 955             fiz1            += tz;
 956             f[j_coord_offset+DIM*2+XX] -= tx;
 957             f[j_coord_offset+DIM*2+YY] -= ty;
 958             f[j_coord_offset+DIM*2+ZZ] -= tz;
 959
 960             /**************************
 961              * CALCULATE INTERACTIONS *
 962              **************************/
 963
 964             /* COULOMB ELECTROSTATICS */
 965             velec            = qq20*rinv20;
 966             felec            = velec*rinvsq20;
 967
 968             fscal            = felec;
 969
 970             /* Calculate temporary vectorial force */
 971             tx               = fscal*dx20;
 972             ty               = fscal*dy20;
 973             tz               = fscal*dz20;
 974
 975             /* Update vectorial force */
 976             fix2            += tx;
 977             fiy2            += ty;
 978             fiz2            += tz;
 979             f[j_coord_offset+DIM*0+XX] -= tx;
 980             f[j_coord_offset+DIM*0+YY] -= ty;
 981             f[j_coord_offset+DIM*0+ZZ] -= tz;
 982
 983             /**************************
 984              * CALCULATE INTERACTIONS *
 985              **************************/
 986
 987             /* COULOMB ELECTROSTATICS */
 988             velec            = qq21*rinv21;
 989             felec            = velec*rinvsq21;
 990
 991             fscal            = felec;
 992
 993             /* Calculate temporary vectorial force */
 994             tx               = fscal*dx21;
 995             ty               = fscal*dy21;
 996             tz               = fscal*dz21;
 997
 998             /* Update vectorial force */
 999             fix2            += tx;
1000             fiy2            += ty;
1001             fiz2            += tz;
1002             f[j_coord_offset+DIM*1+XX] -= tx;
1003             f[j_coord_offset+DIM*1+YY] -= ty;
1004             f[j_coord_offset+DIM*1+ZZ] -= tz;
1005
1006             /**************************
1007              * CALCULATE INTERACTIONS *
1008              **************************/
1009
1010             /* COULOMB ELECTROSTATICS */
1011             velec            = qq22*rinv22;
1012             felec            = velec*rinvsq22;
1013
1014             fscal            = felec;
1015
1016             /* Calculate temporary vectorial force */
1017             tx               = fscal*dx22;
1018             ty               = fscal*dy22;
1019             tz               = fscal*dz22;
1020
1021             /* Update vectorial force */
1022             fix2            += tx;
1023             fiy2            += ty;
1024             fiz2            += tz;
1025             f[j_coord_offset+DIM*2+XX] -= tx;
1026             f[j_coord_offset+DIM*2+YY] -= ty;
1027             f[j_coord_offset+DIM*2+ZZ] -= tz;
1028
1029             /* Inner loop uses 260 flops */
1030         }
1031         /* End of innermost loop */
1032
1033         tx = ty = tz = 0;
1034         f[i_coord_offset+DIM*0+XX] += fix0;
1035         f[i_coord_offset+DIM*0+YY] += fiy0;
1036         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1037         tx                         += fix0;
1038         ty                         += fiy0;
1039         tz                         += fiz0;
1040         f[i_coord_offset+DIM*1+XX] += fix1;
1041         f[i_coord_offset+DIM*1+YY] += fiy1;
1042         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1043         tx                         += fix1;
1044         ty                         += fiy1;
1045         tz                         += fiz1;
1046         f[i_coord_offset+DIM*2+XX] += fix2;
1047         f[i_coord_offset+DIM*2+YY] += fiy2;
1048         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1049         tx                         += fix2;
1050         ty                         += fiy2;
1051         tz                         += fiz2;
1052         fshift[i_shift_offset+XX]  += tx;
1053         fshift[i_shift_offset+YY]  += ty;
1054         fshift[i_shift_offset+ZZ]  += tz;
1055
1056         /* Increment number of inner iterations */
1057         inneriter                  += j_index_end - j_index_start;
1058
1059         /* Outer loop uses 30 flops */
1060     }
1061
1062     /* Increment number of outer iterations */
1063     outeriter        += nri;
1064
1065     /* Update outer/inner flops */
1066
1067     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*260);
1068 }