src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecRF_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_ElecRF_VdwCSTab_GeomW3W3_VF_c
  49  * Electrostatics interaction: ReactionField
  50  * VdW interaction:            CubicSplineTable
  51  * Geometry:                   Water3-Water3
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecRF_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     krf              = fr->ic->k_rf;
 115     krf2             = krf*2.0;
 116     crf              = fr->ic->c_rf;
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     vftab            = kernel_data->table_vdw->data;
 122     vftabscale       = kernel_data->table_vdw->scale;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq0              = facel*charge[inr+0];
 127     iq1              = facel*charge[inr+1];
 128     iq2              = facel*charge[inr+2];
 129     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 130
 131     jq0              = charge[inr+0];
 132     jq1              = charge[inr+1];
 133     jq2              = charge[inr+2];
 134     vdwjidx0         = 2*vdwtype[inr+0];
 135     qq00             = iq0*jq0;
 136     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 137     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 138     qq01             = iq0*jq1;
 139     qq02             = iq0*jq2;
 140     qq10             = iq1*jq0;
 141     qq11             = iq1*jq1;
 142     qq12             = iq1*jq2;
 143     qq20             = iq2*jq0;
 144     qq21             = iq2*jq1;
 145     qq22             = iq2*jq2;
 146
 147     outeriter        = 0;
 148     inneriter        = 0;
 149
 150     /* Start outer loop over neighborlists */
 151     for(iidx=0; iidx<nri; iidx++)
 152     {
 153         /* Load shift vector for this list */
 154         i_shift_offset   = DIM*shiftidx[iidx];
 155         shX              = shiftvec[i_shift_offset+XX];
 156         shY              = shiftvec[i_shift_offset+YY];
 157         shZ              = shiftvec[i_shift_offset+ZZ];
 158
 159         /* Load limits for loop over neighbors */
 160         j_index_start    = jindex[iidx];
 161         j_index_end      = jindex[iidx+1];
 162
 163         /* Get outer coordinate index */
 164         inr              = iinr[iidx];
 165         i_coord_offset   = DIM*inr;
 166
 167         /* Load i particle coords and add shift vector */
 168         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 169         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 170         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 171         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 172         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 173         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 174         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 175         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 176         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 177
 178         fix0             = 0.0;
 179         fiy0             = 0.0;
 180         fiz0             = 0.0;
 181         fix1             = 0.0;
 182         fiy1             = 0.0;
 183         fiz1             = 0.0;
 184         fix2             = 0.0;
 185         fiy2             = 0.0;
 186         fiz2             = 0.0;
 187
 188         /* Reset potential sums */
 189         velecsum         = 0.0;
 190         vvdwsum          = 0.0;
 191
 192         /* Start inner kernel loop */
 193         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 194         {
 195             /* Get j neighbor index, and coordinate index */
 196             jnr              = jjnr[jidx];
 197             j_coord_offset   = DIM*jnr;
 198
 199             /* load j atom coordinates */
 200             jx0              = x[j_coord_offset+DIM*0+XX];
 201             jy0              = x[j_coord_offset+DIM*0+YY];
 202             jz0              = x[j_coord_offset+DIM*0+ZZ];
 203             jx1              = x[j_coord_offset+DIM*1+XX];
 204             jy1              = x[j_coord_offset+DIM*1+YY];
 205             jz1              = x[j_coord_offset+DIM*1+ZZ];
 206             jx2              = x[j_coord_offset+DIM*2+XX];
 207             jy2              = x[j_coord_offset+DIM*2+YY];
 208             jz2              = x[j_coord_offset+DIM*2+ZZ];
 209
 210             /* Calculate displacement vector */
 211             dx00             = ix0 - jx0;
 212             dy00             = iy0 - jy0;
 213             dz00             = iz0 - jz0;
 214             dx01             = ix0 - jx1;
 215             dy01             = iy0 - jy1;
 216             dz01             = iz0 - jz1;
 217             dx02             = ix0 - jx2;
 218             dy02             = iy0 - jy2;
 219             dz02             = iz0 - jz2;
 220             dx10             = ix1 - jx0;
 221             dy10             = iy1 - jy0;
 222             dz10             = iz1 - jz0;
 223             dx11             = ix1 - jx1;
 224             dy11             = iy1 - jy1;
 225             dz11             = iz1 - jz1;
 226             dx12             = ix1 - jx2;
 227             dy12             = iy1 - jy2;
 228             dz12             = iz1 - jz2;
 229             dx20             = ix2 - jx0;
 230             dy20             = iy2 - jy0;
 231             dz20             = iz2 - jz0;
 232             dx21             = ix2 - jx1;
 233             dy21             = iy2 - jy1;
 234             dz21             = iz2 - jz1;
 235             dx22             = ix2 - jx2;
 236             dy22             = iy2 - jy2;
 237             dz22             = iz2 - jz2;
 238
 239             /* Calculate squared distance and things based on it */
 240             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 241             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 242             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 243             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 244             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 245             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 246             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 247             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 248             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 249
 250             rinv00           = gmx_invsqrt(rsq00);
 251             rinv01           = gmx_invsqrt(rsq01);
 252             rinv02           = gmx_invsqrt(rsq02);
 253             rinv10           = gmx_invsqrt(rsq10);
 254             rinv11           = gmx_invsqrt(rsq11);
 255             rinv12           = gmx_invsqrt(rsq12);
 256             rinv20           = gmx_invsqrt(rsq20);
 257             rinv21           = gmx_invsqrt(rsq21);
 258             rinv22           = gmx_invsqrt(rsq22);
 259
 260             rinvsq00         = rinv00*rinv00;
 261             rinvsq01         = rinv01*rinv01;
 262             rinvsq02         = rinv02*rinv02;
 263             rinvsq10         = rinv10*rinv10;
 264             rinvsq11         = rinv11*rinv11;
 265             rinvsq12         = rinv12*rinv12;
 266             rinvsq20         = rinv20*rinv20;
 267             rinvsq21         = rinv21*rinv21;
 268             rinvsq22         = rinv22*rinv22;
 269
 270             /**************************
 271              * CALCULATE INTERACTIONS *
 272              **************************/
 273
 274             r00              = rsq00*rinv00;
 275
 276             /* Calculate table index by multiplying r with table scale and truncate to integer */
 277             rt               = r00*vftabscale;
 278             vfitab           = rt;
 279             vfeps            = rt-vfitab;
 280             vfitab           = 2*4*vfitab;
 281
 282             /* REACTION-FIELD ELECTROSTATICS */
 283             velec            = qq00*(rinv00+krf*rsq00-crf);
 284             felec            = qq00*(rinv00*rinvsq00-krf2);
 285
 286             /* CUBIC SPLINE TABLE DISPERSION */
 287             vfitab          += 0;
 288             Y                = vftab[vfitab];
 289             F                = vftab[vfitab+1];
 290             Geps             = vfeps*vftab[vfitab+2];
 291             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 292             Fp               = F+Geps+Heps2;
 293             VV               = Y+vfeps*Fp;
 294             vvdw6            = c6_00*VV;
 295             FF               = Fp+Geps+2.0*Heps2;
 296             fvdw6            = c6_00*FF;
 297
 298             /* CUBIC SPLINE TABLE REPULSION */
 299             Y                = vftab[vfitab+4];
 300             F                = vftab[vfitab+5];
 301             Geps             = vfeps*vftab[vfitab+6];
 302             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 303             Fp               = F+Geps+Heps2;
 304             VV               = Y+vfeps*Fp;
 305             vvdw12           = c12_00*VV;
 306             FF               = Fp+Geps+2.0*Heps2;
 307             fvdw12           = c12_00*FF;
 308             vvdw             = vvdw12+vvdw6;
 309             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 310
 311             /* Update potential sums from outer loop */
 312             velecsum        += velec;
 313             vvdwsum         += vvdw;
 314
 315             fscal            = felec+fvdw;
 316
 317             /* Calculate temporary vectorial force */
 318             tx               = fscal*dx00;
 319             ty               = fscal*dy00;
 320             tz               = fscal*dz00;
 321
 322             /* Update vectorial force */
 323             fix0            += tx;
 324             fiy0            += ty;
 325             fiz0            += tz;
 326             f[j_coord_offset+DIM*0+XX] -= tx;
 327             f[j_coord_offset+DIM*0+YY] -= ty;
 328             f[j_coord_offset+DIM*0+ZZ] -= tz;
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             /* REACTION-FIELD ELECTROSTATICS */
 335             velec            = qq01*(rinv01+krf*rsq01-crf);
 336             felec            = qq01*(rinv01*rinvsq01-krf2);
 337
 338             /* Update potential sums from outer loop */
 339             velecsum        += velec;
 340
 341             fscal            = felec;
 342
 343             /* Calculate temporary vectorial force */
 344             tx               = fscal*dx01;
 345             ty               = fscal*dy01;
 346             tz               = fscal*dz01;
 347
 348             /* Update vectorial force */
 349             fix0            += tx;
 350             fiy0            += ty;
 351             fiz0            += tz;
 352             f[j_coord_offset+DIM*1+XX] -= tx;
 353             f[j_coord_offset+DIM*1+YY] -= ty;
 354             f[j_coord_offset+DIM*1+ZZ] -= tz;
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = qq02*(rinv02+krf*rsq02-crf);
 362             felec            = qq02*(rinv02*rinvsq02-krf2);
 363
 364             /* Update potential sums from outer loop */
 365             velecsum        += velec;
 366
 367             fscal            = felec;
 368
 369             /* Calculate temporary vectorial force */
 370             tx               = fscal*dx02;
 371             ty               = fscal*dy02;
 372             tz               = fscal*dz02;
 373
 374             /* Update vectorial force */
 375             fix0            += tx;
 376             fiy0            += ty;
 377             fiz0            += tz;
 378             f[j_coord_offset+DIM*2+XX] -= tx;
 379             f[j_coord_offset+DIM*2+YY] -= ty;
 380             f[j_coord_offset+DIM*2+ZZ] -= tz;
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             /* REACTION-FIELD ELECTROSTATICS */
 387             velec            = qq10*(rinv10+krf*rsq10-crf);
 388             felec            = qq10*(rinv10*rinvsq10-krf2);
 389
 390             /* Update potential sums from outer loop */
 391             velecsum        += velec;
 392
 393             fscal            = felec;
 394
 395             /* Calculate temporary vectorial force */
 396             tx               = fscal*dx10;
 397             ty               = fscal*dy10;
 398             tz               = fscal*dz10;
 399
 400             /* Update vectorial force */
 401             fix1            += tx;
 402             fiy1            += ty;
 403             fiz1            += tz;
 404             f[j_coord_offset+DIM*0+XX] -= tx;
 405             f[j_coord_offset+DIM*0+YY] -= ty;
 406             f[j_coord_offset+DIM*0+ZZ] -= tz;
 407
 408             /**************************
 409              * CALCULATE INTERACTIONS *
 410              **************************/
 411
 412             /* REACTION-FIELD ELECTROSTATICS */
 413             velec            = qq11*(rinv11+krf*rsq11-crf);
 414             felec            = qq11*(rinv11*rinvsq11-krf2);
 415
 416             /* Update potential sums from outer loop */
 417             velecsum        += velec;
 418
 419             fscal            = felec;
 420
 421             /* Calculate temporary vectorial force */
 422             tx               = fscal*dx11;
 423             ty               = fscal*dy11;
 424             tz               = fscal*dz11;
 425
 426             /* Update vectorial force */
 427             fix1            += tx;
 428             fiy1            += ty;
 429             fiz1            += tz;
 430             f[j_coord_offset+DIM*1+XX] -= tx;
 431             f[j_coord_offset+DIM*1+YY] -= ty;
 432             f[j_coord_offset+DIM*1+ZZ] -= tz;
 433
 434             /**************************
 435              * CALCULATE INTERACTIONS *
 436              **************************/
 437
 438             /* REACTION-FIELD ELECTROSTATICS */
 439             velec            = qq12*(rinv12+krf*rsq12-crf);
 440             felec            = qq12*(rinv12*rinvsq12-krf2);
 441
 442             /* Update potential sums from outer loop */
 443             velecsum        += velec;
 444
 445             fscal            = felec;
 446
 447             /* Calculate temporary vectorial force */
 448             tx               = fscal*dx12;
 449             ty               = fscal*dy12;
 450             tz               = fscal*dz12;
 451
 452             /* Update vectorial force */
 453             fix1            += tx;
 454             fiy1            += ty;
 455             fiz1            += tz;
 456             f[j_coord_offset+DIM*2+XX] -= tx;
 457             f[j_coord_offset+DIM*2+YY] -= ty;
 458             f[j_coord_offset+DIM*2+ZZ] -= tz;
 459
 460             /**************************
 461              * CALCULATE INTERACTIONS *
 462              **************************/
 463
 464             /* REACTION-FIELD ELECTROSTATICS */
 465             velec            = qq20*(rinv20+krf*rsq20-crf);
 466             felec            = qq20*(rinv20*rinvsq20-krf2);
 467
 468             /* Update potential sums from outer loop */
 469             velecsum        += velec;
 470
 471             fscal            = felec;
 472
 473             /* Calculate temporary vectorial force */
 474             tx               = fscal*dx20;
 475             ty               = fscal*dy20;
 476             tz               = fscal*dz20;
 477
 478             /* Update vectorial force */
 479             fix2            += tx;
 480             fiy2            += ty;
 481             fiz2            += tz;
 482             f[j_coord_offset+DIM*0+XX] -= tx;
 483             f[j_coord_offset+DIM*0+YY] -= ty;
 484             f[j_coord_offset+DIM*0+ZZ] -= tz;
 485
 486             /**************************
 487              * CALCULATE INTERACTIONS *
 488              **************************/
 489
 490             /* REACTION-FIELD ELECTROSTATICS */
 491             velec            = qq21*(rinv21+krf*rsq21-crf);
 492             felec            = qq21*(rinv21*rinvsq21-krf2);
 493
 494             /* Update potential sums from outer loop */
 495             velecsum        += velec;
 496
 497             fscal            = felec;
 498
 499             /* Calculate temporary vectorial force */
 500             tx               = fscal*dx21;
 501             ty               = fscal*dy21;
 502             tz               = fscal*dz21;
 503
 504             /* Update vectorial force */
 505             fix2            += tx;
 506             fiy2            += ty;
 507             fiz2            += tz;
 508             f[j_coord_offset+DIM*1+XX] -= tx;
 509             f[j_coord_offset+DIM*1+YY] -= ty;
 510             f[j_coord_offset+DIM*1+ZZ] -= tz;
 511
 512             /**************************
 513              * CALCULATE INTERACTIONS *
 514              **************************/
 515
 516             /* REACTION-FIELD ELECTROSTATICS */
 517             velec            = qq22*(rinv22+krf*rsq22-crf);
 518             felec            = qq22*(rinv22*rinvsq22-krf2);
 519
 520             /* Update potential sums from outer loop */
 521             velecsum        += velec;
 522
 523             fscal            = felec;
 524
 525             /* Calculate temporary vectorial force */
 526             tx               = fscal*dx22;
 527             ty               = fscal*dy22;
 528             tz               = fscal*dz22;
 529
 530             /* Update vectorial force */
 531             fix2            += tx;
 532             fiy2            += ty;
 533             fiz2            += tz;
 534             f[j_coord_offset+DIM*2+XX] -= tx;
 535             f[j_coord_offset+DIM*2+YY] -= ty;
 536             f[j_coord_offset+DIM*2+ZZ] -= tz;
 537
 538             /* Inner loop uses 313 flops */
 539         }
 540         /* End of innermost loop */
 541
 542         tx = ty = tz = 0;
 543         f[i_coord_offset+DIM*0+XX] += fix0;
 544         f[i_coord_offset+DIM*0+YY] += fiy0;
 545         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 546         tx                         += fix0;
 547         ty                         += fiy0;
 548         tz                         += fiz0;
 549         f[i_coord_offset+DIM*1+XX] += fix1;
 550         f[i_coord_offset+DIM*1+YY] += fiy1;
 551         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 552         tx                         += fix1;
 553         ty                         += fiy1;
 554         tz                         += fiz1;
 555         f[i_coord_offset+DIM*2+XX] += fix2;
 556         f[i_coord_offset+DIM*2+YY] += fiy2;
 557         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 558         tx                         += fix2;
 559         ty                         += fiy2;
 560         tz                         += fiz2;
 561         fshift[i_shift_offset+XX]  += tx;
 562         fshift[i_shift_offset+YY]  += ty;
 563         fshift[i_shift_offset+ZZ]  += tz;
 564
 565         ggid                        = gid[iidx];
 566         /* Update potential energies */
 567         kernel_data->energygrp_elec[ggid] += velecsum;
 568         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 569
 570         /* Increment number of inner iterations */
 571         inneriter                  += j_index_end - j_index_start;
 572
 573         /* Outer loop uses 32 flops */
 574     }
 575
 576     /* Increment number of outer iterations */
 577     outeriter        += nri;
 578
 579     /* Update outer/inner flops */
 580
 581     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*313);
 582 }
 583 /*
 584  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_c
 585  * Electrostatics interaction: ReactionField
 586  * VdW interaction:            CubicSplineTable
 587  * Geometry:                   Water3-Water3
 588  * Calculate force/pot:        Force
 589  */
 590 void
 591 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_c
 592                     (t_nblist                    * gmx_restrict       nlist,
 593                      rvec                        * gmx_restrict          xx,
 594                      rvec                        * gmx_restrict          ff,
 595                      t_forcerec                  * gmx_restrict          fr,
 596                      t_mdatoms                   * gmx_restrict     mdatoms,
 597                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 598                      t_nrnb                      * gmx_restrict        nrnb)
 599 {
 600     int              i_shift_offset,i_coord_offset,j_coord_offset;
 601     int              j_index_start,j_index_end;
 602     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 603     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 604     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 605     real             *shiftvec,*fshift,*x,*f;
 606     int              vdwioffset0;
 607     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 608     int              vdwioffset1;
 609     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 610     int              vdwioffset2;
 611     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 612     int              vdwjidx0;
 613     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 614     int              vdwjidx1;
 615     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 616     int              vdwjidx2;
 617     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 618     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 619     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 620     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 621     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 622     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 623     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 624     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 625     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 626     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 627     real             velec,felec,velecsum,facel,crf,krf,krf2;
 628     real             *charge;
 629     int              nvdwtype;
 630     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 631     int              *vdwtype;
 632     real             *vdwparam;
 633     int              vfitab;
 634     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 635     real             *vftab;
 636
 637     x                = xx[0];
 638     f                = ff[0];
 639
 640     nri              = nlist->nri;
 641     iinr             = nlist->iinr;
 642     jindex           = nlist->jindex;
 643     jjnr             = nlist->jjnr;
 644     shiftidx         = nlist->shift;
 645     gid              = nlist->gid;
 646     shiftvec         = fr->shift_vec[0];
 647     fshift           = fr->fshift[0];
 648     facel            = fr->epsfac;
 649     charge           = mdatoms->chargeA;
 650     krf              = fr->ic->k_rf;
 651     krf2             = krf*2.0;
 652     crf              = fr->ic->c_rf;
 653     nvdwtype         = fr->ntype;
 654     vdwparam         = fr->nbfp;
 655     vdwtype          = mdatoms->typeA;
 656
 657     vftab            = kernel_data->table_vdw->data;
 658     vftabscale       = kernel_data->table_vdw->scale;
 659
 660     /* Setup water-specific parameters */
 661     inr              = nlist->iinr[0];
 662     iq0              = facel*charge[inr+0];
 663     iq1              = facel*charge[inr+1];
 664     iq2              = facel*charge[inr+2];
 665     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 666
 667     jq0              = charge[inr+0];
 668     jq1              = charge[inr+1];
 669     jq2              = charge[inr+2];
 670     vdwjidx0         = 2*vdwtype[inr+0];
 671     qq00             = iq0*jq0;
 672     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 673     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 674     qq01             = iq0*jq1;
 675     qq02             = iq0*jq2;
 676     qq10             = iq1*jq0;
 677     qq11             = iq1*jq1;
 678     qq12             = iq1*jq2;
 679     qq20             = iq2*jq0;
 680     qq21             = iq2*jq1;
 681     qq22             = iq2*jq2;
 682
 683     outeriter        = 0;
 684     inneriter        = 0;
 685
 686     /* Start outer loop over neighborlists */
 687     for(iidx=0; iidx<nri; iidx++)
 688     {
 689         /* Load shift vector for this list */
 690         i_shift_offset   = DIM*shiftidx[iidx];
 691         shX              = shiftvec[i_shift_offset+XX];
 692         shY              = shiftvec[i_shift_offset+YY];
 693         shZ              = shiftvec[i_shift_offset+ZZ];
 694
 695         /* Load limits for loop over neighbors */
 696         j_index_start    = jindex[iidx];
 697         j_index_end      = jindex[iidx+1];
 698
 699         /* Get outer coordinate index */
 700         inr              = iinr[iidx];
 701         i_coord_offset   = DIM*inr;
 702
 703         /* Load i particle coords and add shift vector */
 704         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 705         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 706         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 707         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 708         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 709         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 710         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 711         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 712         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 713
 714         fix0             = 0.0;
 715         fiy0             = 0.0;
 716         fiz0             = 0.0;
 717         fix1             = 0.0;
 718         fiy1             = 0.0;
 719         fiz1             = 0.0;
 720         fix2             = 0.0;
 721         fiy2             = 0.0;
 722         fiz2             = 0.0;
 723
 724         /* Start inner kernel loop */
 725         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 726         {
 727             /* Get j neighbor index, and coordinate index */
 728             jnr              = jjnr[jidx];
 729             j_coord_offset   = DIM*jnr;
 730
 731             /* load j atom coordinates */
 732             jx0              = x[j_coord_offset+DIM*0+XX];
 733             jy0              = x[j_coord_offset+DIM*0+YY];
 734             jz0              = x[j_coord_offset+DIM*0+ZZ];
 735             jx1              = x[j_coord_offset+DIM*1+XX];
 736             jy1              = x[j_coord_offset+DIM*1+YY];
 737             jz1              = x[j_coord_offset+DIM*1+ZZ];
 738             jx2              = x[j_coord_offset+DIM*2+XX];
 739             jy2              = x[j_coord_offset+DIM*2+YY];
 740             jz2              = x[j_coord_offset+DIM*2+ZZ];
 741
 742             /* Calculate displacement vector */
 743             dx00             = ix0 - jx0;
 744             dy00             = iy0 - jy0;
 745             dz00             = iz0 - jz0;
 746             dx01             = ix0 - jx1;
 747             dy01             = iy0 - jy1;
 748             dz01             = iz0 - jz1;
 749             dx02             = ix0 - jx2;
 750             dy02             = iy0 - jy2;
 751             dz02             = iz0 - jz2;
 752             dx10             = ix1 - jx0;
 753             dy10             = iy1 - jy0;
 754             dz10             = iz1 - jz0;
 755             dx11             = ix1 - jx1;
 756             dy11             = iy1 - jy1;
 757             dz11             = iz1 - jz1;
 758             dx12             = ix1 - jx2;
 759             dy12             = iy1 - jy2;
 760             dz12             = iz1 - jz2;
 761             dx20             = ix2 - jx0;
 762             dy20             = iy2 - jy0;
 763             dz20             = iz2 - jz0;
 764             dx21             = ix2 - jx1;
 765             dy21             = iy2 - jy1;
 766             dz21             = iz2 - jz1;
 767             dx22             = ix2 - jx2;
 768             dy22             = iy2 - jy2;
 769             dz22             = iz2 - jz2;
 770
 771             /* Calculate squared distance and things based on it */
 772             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 773             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 774             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 775             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 776             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 777             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 778             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 779             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 780             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 781
 782             rinv00           = gmx_invsqrt(rsq00);
 783             rinv01           = gmx_invsqrt(rsq01);
 784             rinv02           = gmx_invsqrt(rsq02);
 785             rinv10           = gmx_invsqrt(rsq10);
 786             rinv11           = gmx_invsqrt(rsq11);
 787             rinv12           = gmx_invsqrt(rsq12);
 788             rinv20           = gmx_invsqrt(rsq20);
 789             rinv21           = gmx_invsqrt(rsq21);
 790             rinv22           = gmx_invsqrt(rsq22);
 791
 792             rinvsq00         = rinv00*rinv00;
 793             rinvsq01         = rinv01*rinv01;
 794             rinvsq02         = rinv02*rinv02;
 795             rinvsq10         = rinv10*rinv10;
 796             rinvsq11         = rinv11*rinv11;
 797             rinvsq12         = rinv12*rinv12;
 798             rinvsq20         = rinv20*rinv20;
 799             rinvsq21         = rinv21*rinv21;
 800             rinvsq22         = rinv22*rinv22;
 801
 802             /**************************
 803              * CALCULATE INTERACTIONS *
 804              **************************/
 805
 806             r00              = rsq00*rinv00;
 807
 808             /* Calculate table index by multiplying r with table scale and truncate to integer */
 809             rt               = r00*vftabscale;
 810             vfitab           = rt;
 811             vfeps            = rt-vfitab;
 812             vfitab           = 2*4*vfitab;
 813
 814             /* REACTION-FIELD ELECTROSTATICS */
 815             felec            = qq00*(rinv00*rinvsq00-krf2);
 816
 817             /* CUBIC SPLINE TABLE DISPERSION */
 818             vfitab          += 0;
 819             F                = vftab[vfitab+1];
 820             Geps             = vfeps*vftab[vfitab+2];
 821             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 822             Fp               = F+Geps+Heps2;
 823             FF               = Fp+Geps+2.0*Heps2;
 824             fvdw6            = c6_00*FF;
 825
 826             /* CUBIC SPLINE TABLE REPULSION */
 827             F                = vftab[vfitab+5];
 828             Geps             = vfeps*vftab[vfitab+6];
 829             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 830             Fp               = F+Geps+Heps2;
 831             FF               = Fp+Geps+2.0*Heps2;
 832             fvdw12           = c12_00*FF;
 833             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 834
 835             fscal            = felec+fvdw;
 836
 837             /* Calculate temporary vectorial force */
 838             tx               = fscal*dx00;
 839             ty               = fscal*dy00;
 840             tz               = fscal*dz00;
 841
 842             /* Update vectorial force */
 843             fix0            += tx;
 844             fiy0            += ty;
 845             fiz0            += tz;
 846             f[j_coord_offset+DIM*0+XX] -= tx;
 847             f[j_coord_offset+DIM*0+YY] -= ty;
 848             f[j_coord_offset+DIM*0+ZZ] -= tz;
 849
 850             /**************************
 851              * CALCULATE INTERACTIONS *
 852              **************************/
 853
 854             /* REACTION-FIELD ELECTROSTATICS */
 855             felec            = qq01*(rinv01*rinvsq01-krf2);
 856
 857             fscal            = felec;
 858
 859             /* Calculate temporary vectorial force */
 860             tx               = fscal*dx01;
 861             ty               = fscal*dy01;
 862             tz               = fscal*dz01;
 863
 864             /* Update vectorial force */
 865             fix0            += tx;
 866             fiy0            += ty;
 867             fiz0            += tz;
 868             f[j_coord_offset+DIM*1+XX] -= tx;
 869             f[j_coord_offset+DIM*1+YY] -= ty;
 870             f[j_coord_offset+DIM*1+ZZ] -= tz;
 871
 872             /**************************
 873              * CALCULATE INTERACTIONS *
 874              **************************/
 875
 876             /* REACTION-FIELD ELECTROSTATICS */
 877             felec            = qq02*(rinv02*rinvsq02-krf2);
 878
 879             fscal            = felec;
 880
 881             /* Calculate temporary vectorial force */
 882             tx               = fscal*dx02;
 883             ty               = fscal*dy02;
 884             tz               = fscal*dz02;
 885
 886             /* Update vectorial force */
 887             fix0            += tx;
 888             fiy0            += ty;
 889             fiz0            += tz;
 890             f[j_coord_offset+DIM*2+XX] -= tx;
 891             f[j_coord_offset+DIM*2+YY] -= ty;
 892             f[j_coord_offset+DIM*2+ZZ] -= tz;
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             /* REACTION-FIELD ELECTROSTATICS */
 899             felec            = qq10*(rinv10*rinvsq10-krf2);
 900
 901             fscal            = felec;
 902
 903             /* Calculate temporary vectorial force */
 904             tx               = fscal*dx10;
 905             ty               = fscal*dy10;
 906             tz               = fscal*dz10;
 907
 908             /* Update vectorial force */
 909             fix1            += tx;
 910             fiy1            += ty;
 911             fiz1            += tz;
 912             f[j_coord_offset+DIM*0+XX] -= tx;
 913             f[j_coord_offset+DIM*0+YY] -= ty;
 914             f[j_coord_offset+DIM*0+ZZ] -= tz;
 915
 916             /**************************
 917              * CALCULATE INTERACTIONS *
 918              **************************/
 919
 920             /* REACTION-FIELD ELECTROSTATICS */
 921             felec            = qq11*(rinv11*rinvsq11-krf2);
 922
 923             fscal            = felec;
 924
 925             /* Calculate temporary vectorial force */
 926             tx               = fscal*dx11;
 927             ty               = fscal*dy11;
 928             tz               = fscal*dz11;
 929
 930             /* Update vectorial force */
 931             fix1            += tx;
 932             fiy1            += ty;
 933             fiz1            += tz;
 934             f[j_coord_offset+DIM*1+XX] -= tx;
 935             f[j_coord_offset+DIM*1+YY] -= ty;
 936             f[j_coord_offset+DIM*1+ZZ] -= tz;
 937
 938             /**************************
 939              * CALCULATE INTERACTIONS *
 940              **************************/
 941
 942             /* REACTION-FIELD ELECTROSTATICS */
 943             felec            = qq12*(rinv12*rinvsq12-krf2);
 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             /* REACTION-FIELD ELECTROSTATICS */
 965             felec            = qq20*(rinv20*rinvsq20-krf2);
 966
 967             fscal            = felec;
 968
 969             /* Calculate temporary vectorial force */
 970             tx               = fscal*dx20;
 971             ty               = fscal*dy20;
 972             tz               = fscal*dz20;
 973
 974             /* Update vectorial force */
 975             fix2            += tx;
 976             fiy2            += ty;
 977             fiz2            += tz;
 978             f[j_coord_offset+DIM*0+XX] -= tx;
 979             f[j_coord_offset+DIM*0+YY] -= ty;
 980             f[j_coord_offset+DIM*0+ZZ] -= tz;
 981
 982             /**************************
 983              * CALCULATE INTERACTIONS *
 984              **************************/
 985
 986             /* REACTION-FIELD ELECTROSTATICS */
 987             felec            = qq21*(rinv21*rinvsq21-krf2);
 988
 989             fscal            = felec;
 990
 991             /* Calculate temporary vectorial force */
 992             tx               = fscal*dx21;
 993             ty               = fscal*dy21;
 994             tz               = fscal*dz21;
 995
 996             /* Update vectorial force */
 997             fix2            += tx;
 998             fiy2            += ty;
 999             fiz2            += tz;
1000             f[j_coord_offset+DIM*1+XX] -= tx;
1001             f[j_coord_offset+DIM*1+YY] -= ty;
1002             f[j_coord_offset+DIM*1+ZZ] -= tz;
1003
1004             /**************************
1005              * CALCULATE INTERACTIONS *
1006              **************************/
1007
1008             /* REACTION-FIELD ELECTROSTATICS */
1009             felec            = qq22*(rinv22*rinvsq22-krf2);
1010
1011             fscal            = felec;
1012
1013             /* Calculate temporary vectorial force */
1014             tx               = fscal*dx22;
1015             ty               = fscal*dy22;
1016             tz               = fscal*dz22;
1017
1018             /* Update vectorial force */
1019             fix2            += tx;
1020             fiy2            += ty;
1021             fiz2            += tz;
1022             f[j_coord_offset+DIM*2+XX] -= tx;
1023             f[j_coord_offset+DIM*2+YY] -= ty;
1024             f[j_coord_offset+DIM*2+ZZ] -= tz;
1025
1026             /* Inner loop uses 260 flops */
1027         }
1028         /* End of innermost loop */
1029
1030         tx = ty = tz = 0;
1031         f[i_coord_offset+DIM*0+XX] += fix0;
1032         f[i_coord_offset+DIM*0+YY] += fiy0;
1033         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1034         tx                         += fix0;
1035         ty                         += fiy0;
1036         tz                         += fiz0;
1037         f[i_coord_offset+DIM*1+XX] += fix1;
1038         f[i_coord_offset+DIM*1+YY] += fiy1;
1039         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1040         tx                         += fix1;
1041         ty                         += fiy1;
1042         tz                         += fiz1;
1043         f[i_coord_offset+DIM*2+XX] += fix2;
1044         f[i_coord_offset+DIM*2+YY] += fiy2;
1045         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1046         tx                         += fix2;
1047         ty                         += fiy2;
1048         tz                         += fiz2;
1049         fshift[i_shift_offset+XX]  += tx;
1050         fshift[i_shift_offset+YY]  += ty;
1051         fshift[i_shift_offset+ZZ]  += tz;
1052
1053         /* Increment number of inner iterations */
1054         inneriter                  += j_index_end - j_index_start;
1055
1056         /* Outer loop uses 30 flops */
1057     }
1058
1059     /* Increment number of outer iterations */
1060     outeriter        += nri;
1061
1062     /* Update outer/inner flops */
1063
1064     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*260);
1065 }