src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecCoul_VdwBham_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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "nrnb.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwBham_GeomW3W3_VF_c
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            Buckingham
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwBham_GeomW3W3_VF_c
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     int              i_shift_offset,i_coord_offset,j_coord_offset;
  67     int              j_index_start,j_index_end;
  68     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  69     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  70     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  71     real             *shiftvec,*fshift,*x,*f;
  72     int              vdwioffset0;
  73     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwioffset1;
  75     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  76     int              vdwioffset2;
  77     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  78     int              vdwjidx0;
  79     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  80     int              vdwjidx1;
  81     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  82     int              vdwjidx2;
  83     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  84     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  85     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
  86     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
  87     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  88     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  89     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  90     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  91     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  92     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  93     real             velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = fr->epsfac;
 112     charge           = mdatoms->chargeA;
 113     nvdwtype         = fr->ntype;
 114     vdwparam         = fr->nbfp;
 115     vdwtype          = mdatoms->typeA;
 116
 117     /* Setup water-specific parameters */
 118     inr              = nlist->iinr[0];
 119     iq0              = facel*charge[inr+0];
 120     iq1              = facel*charge[inr+1];
 121     iq2              = facel*charge[inr+2];
 122     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 123
 124     jq0              = charge[inr+0];
 125     jq1              = charge[inr+1];
 126     jq2              = charge[inr+2];
 127     vdwjidx0         = 3*vdwtype[inr+0];
 128     qq00             = iq0*jq0;
 129     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 130     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 131     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 132     qq01             = iq0*jq1;
 133     qq02             = iq0*jq2;
 134     qq10             = iq1*jq0;
 135     qq11             = iq1*jq1;
 136     qq12             = iq1*jq2;
 137     qq20             = iq2*jq0;
 138     qq21             = iq2*jq1;
 139     qq22             = iq2*jq2;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149         shX              = shiftvec[i_shift_offset+XX];
 150         shY              = shiftvec[i_shift_offset+YY];
 151         shZ              = shiftvec[i_shift_offset+ZZ];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 163         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 164         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 165         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 166         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 167         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 168         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 169         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 170         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 171
 172         fix0             = 0.0;
 173         fiy0             = 0.0;
 174         fiz0             = 0.0;
 175         fix1             = 0.0;
 176         fiy1             = 0.0;
 177         fiz1             = 0.0;
 178         fix2             = 0.0;
 179         fiy2             = 0.0;
 180         fiz2             = 0.0;
 181
 182         /* Reset potential sums */
 183         velecsum         = 0.0;
 184         vvdwsum          = 0.0;
 185
 186         /* Start inner kernel loop */
 187         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 188         {
 189             /* Get j neighbor index, and coordinate index */
 190             jnr              = jjnr[jidx];
 191             j_coord_offset   = DIM*jnr;
 192
 193             /* load j atom coordinates */
 194             jx0              = x[j_coord_offset+DIM*0+XX];
 195             jy0              = x[j_coord_offset+DIM*0+YY];
 196             jz0              = x[j_coord_offset+DIM*0+ZZ];
 197             jx1              = x[j_coord_offset+DIM*1+XX];
 198             jy1              = x[j_coord_offset+DIM*1+YY];
 199             jz1              = x[j_coord_offset+DIM*1+ZZ];
 200             jx2              = x[j_coord_offset+DIM*2+XX];
 201             jy2              = x[j_coord_offset+DIM*2+YY];
 202             jz2              = x[j_coord_offset+DIM*2+ZZ];
 203
 204             /* Calculate displacement vector */
 205             dx00             = ix0 - jx0;
 206             dy00             = iy0 - jy0;
 207             dz00             = iz0 - jz0;
 208             dx01             = ix0 - jx1;
 209             dy01             = iy0 - jy1;
 210             dz01             = iz0 - jz1;
 211             dx02             = ix0 - jx2;
 212             dy02             = iy0 - jy2;
 213             dz02             = iz0 - jz2;
 214             dx10             = ix1 - jx0;
 215             dy10             = iy1 - jy0;
 216             dz10             = iz1 - jz0;
 217             dx11             = ix1 - jx1;
 218             dy11             = iy1 - jy1;
 219             dz11             = iz1 - jz1;
 220             dx12             = ix1 - jx2;
 221             dy12             = iy1 - jy2;
 222             dz12             = iz1 - jz2;
 223             dx20             = ix2 - jx0;
 224             dy20             = iy2 - jy0;
 225             dz20             = iz2 - jz0;
 226             dx21             = ix2 - jx1;
 227             dy21             = iy2 - jy1;
 228             dz21             = iz2 - jz1;
 229             dx22             = ix2 - jx2;
 230             dy22             = iy2 - jy2;
 231             dz22             = iz2 - jz2;
 232
 233             /* Calculate squared distance and things based on it */
 234             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 235             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 236             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 237             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 238             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 239             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 240             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 241             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 242             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 243
 244             rinv00           = gmx_invsqrt(rsq00);
 245             rinv01           = gmx_invsqrt(rsq01);
 246             rinv02           = gmx_invsqrt(rsq02);
 247             rinv10           = gmx_invsqrt(rsq10);
 248             rinv11           = gmx_invsqrt(rsq11);
 249             rinv12           = gmx_invsqrt(rsq12);
 250             rinv20           = gmx_invsqrt(rsq20);
 251             rinv21           = gmx_invsqrt(rsq21);
 252             rinv22           = gmx_invsqrt(rsq22);
 253
 254             rinvsq00         = rinv00*rinv00;
 255             rinvsq01         = rinv01*rinv01;
 256             rinvsq02         = rinv02*rinv02;
 257             rinvsq10         = rinv10*rinv10;
 258             rinvsq11         = rinv11*rinv11;
 259             rinvsq12         = rinv12*rinv12;
 260             rinvsq20         = rinv20*rinv20;
 261             rinvsq21         = rinv21*rinv21;
 262             rinvsq22         = rinv22*rinv22;
 263
 264             /**************************
 265              * CALCULATE INTERACTIONS *
 266              **************************/
 267
 268             r00              = rsq00*rinv00;
 269
 270             /* COULOMB ELECTROSTATICS */
 271             velec            = qq00*rinv00;
 272             felec            = velec*rinvsq00;
 273
 274             /* BUCKINGHAM DISPERSION/REPULSION */
 275             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 276             vvdw6            = c6_00*rinvsix;
 277             br               = cexp2_00*r00;
 278             vvdwexp          = cexp1_00*exp(-br);
 279             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
 280             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 281
 282             /* Update potential sums from outer loop */
 283             velecsum        += velec;
 284             vvdwsum         += vvdw;
 285
 286             fscal            = felec+fvdw;
 287
 288             /* Calculate temporary vectorial force */
 289             tx               = fscal*dx00;
 290             ty               = fscal*dy00;
 291             tz               = fscal*dz00;
 292
 293             /* Update vectorial force */
 294             fix0            += tx;
 295             fiy0            += ty;
 296             fiz0            += tz;
 297             f[j_coord_offset+DIM*0+XX] -= tx;
 298             f[j_coord_offset+DIM*0+YY] -= ty;
 299             f[j_coord_offset+DIM*0+ZZ] -= tz;
 300
 301             /**************************
 302              * CALCULATE INTERACTIONS *
 303              **************************/
 304
 305             /* COULOMB ELECTROSTATICS */
 306             velec            = qq01*rinv01;
 307             felec            = velec*rinvsq01;
 308
 309             /* Update potential sums from outer loop */
 310             velecsum        += velec;
 311
 312             fscal            = felec;
 313
 314             /* Calculate temporary vectorial force */
 315             tx               = fscal*dx01;
 316             ty               = fscal*dy01;
 317             tz               = fscal*dz01;
 318
 319             /* Update vectorial force */
 320             fix0            += tx;
 321             fiy0            += ty;
 322             fiz0            += tz;
 323             f[j_coord_offset+DIM*1+XX] -= tx;
 324             f[j_coord_offset+DIM*1+YY] -= ty;
 325             f[j_coord_offset+DIM*1+ZZ] -= tz;
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             /* COULOMB ELECTROSTATICS */
 332             velec            = qq02*rinv02;
 333             felec            = velec*rinvsq02;
 334
 335             /* Update potential sums from outer loop */
 336             velecsum        += velec;
 337
 338             fscal            = felec;
 339
 340             /* Calculate temporary vectorial force */
 341             tx               = fscal*dx02;
 342             ty               = fscal*dy02;
 343             tz               = fscal*dz02;
 344
 345             /* Update vectorial force */
 346             fix0            += tx;
 347             fiy0            += ty;
 348             fiz0            += tz;
 349             f[j_coord_offset+DIM*2+XX] -= tx;
 350             f[j_coord_offset+DIM*2+YY] -= ty;
 351             f[j_coord_offset+DIM*2+ZZ] -= tz;
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             /* COULOMB ELECTROSTATICS */
 358             velec            = qq10*rinv10;
 359             felec            = velec*rinvsq10;
 360
 361             /* Update potential sums from outer loop */
 362             velecsum        += velec;
 363
 364             fscal            = felec;
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = fscal*dx10;
 368             ty               = fscal*dy10;
 369             tz               = fscal*dz10;
 370
 371             /* Update vectorial force */
 372             fix1            += tx;
 373             fiy1            += ty;
 374             fiz1            += tz;
 375             f[j_coord_offset+DIM*0+XX] -= tx;
 376             f[j_coord_offset+DIM*0+YY] -= ty;
 377             f[j_coord_offset+DIM*0+ZZ] -= tz;
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* COULOMB ELECTROSTATICS */
 384             velec            = qq11*rinv11;
 385             felec            = velec*rinvsq11;
 386
 387             /* Update potential sums from outer loop */
 388             velecsum        += velec;
 389
 390             fscal            = felec;
 391
 392             /* Calculate temporary vectorial force */
 393             tx               = fscal*dx11;
 394             ty               = fscal*dy11;
 395             tz               = fscal*dz11;
 396
 397             /* Update vectorial force */
 398             fix1            += tx;
 399             fiy1            += ty;
 400             fiz1            += tz;
 401             f[j_coord_offset+DIM*1+XX] -= tx;
 402             f[j_coord_offset+DIM*1+YY] -= ty;
 403             f[j_coord_offset+DIM*1+ZZ] -= tz;
 404
 405             /**************************
 406              * CALCULATE INTERACTIONS *
 407              **************************/
 408
 409             /* COULOMB ELECTROSTATICS */
 410             velec            = qq12*rinv12;
 411             felec            = velec*rinvsq12;
 412
 413             /* Update potential sums from outer loop */
 414             velecsum        += velec;
 415
 416             fscal            = felec;
 417
 418             /* Calculate temporary vectorial force */
 419             tx               = fscal*dx12;
 420             ty               = fscal*dy12;
 421             tz               = fscal*dz12;
 422
 423             /* Update vectorial force */
 424             fix1            += tx;
 425             fiy1            += ty;
 426             fiz1            += tz;
 427             f[j_coord_offset+DIM*2+XX] -= tx;
 428             f[j_coord_offset+DIM*2+YY] -= ty;
 429             f[j_coord_offset+DIM*2+ZZ] -= tz;
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             /* COULOMB ELECTROSTATICS */
 436             velec            = qq20*rinv20;
 437             felec            = velec*rinvsq20;
 438
 439             /* Update potential sums from outer loop */
 440             velecsum        += velec;
 441
 442             fscal            = felec;
 443
 444             /* Calculate temporary vectorial force */
 445             tx               = fscal*dx20;
 446             ty               = fscal*dy20;
 447             tz               = fscal*dz20;
 448
 449             /* Update vectorial force */
 450             fix2            += tx;
 451             fiy2            += ty;
 452             fiz2            += tz;
 453             f[j_coord_offset+DIM*0+XX] -= tx;
 454             f[j_coord_offset+DIM*0+YY] -= ty;
 455             f[j_coord_offset+DIM*0+ZZ] -= tz;
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             /* COULOMB ELECTROSTATICS */
 462             velec            = qq21*rinv21;
 463             felec            = velec*rinvsq21;
 464
 465             /* Update potential sums from outer loop */
 466             velecsum        += velec;
 467
 468             fscal            = felec;
 469
 470             /* Calculate temporary vectorial force */
 471             tx               = fscal*dx21;
 472             ty               = fscal*dy21;
 473             tz               = fscal*dz21;
 474
 475             /* Update vectorial force */
 476             fix2            += tx;
 477             fiy2            += ty;
 478             fiz2            += tz;
 479             f[j_coord_offset+DIM*1+XX] -= tx;
 480             f[j_coord_offset+DIM*1+YY] -= ty;
 481             f[j_coord_offset+DIM*1+ZZ] -= tz;
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             /* COULOMB ELECTROSTATICS */
 488             velec            = qq22*rinv22;
 489             felec            = velec*rinvsq22;
 490
 491             /* Update potential sums from outer loop */
 492             velecsum        += velec;
 493
 494             fscal            = felec;
 495
 496             /* Calculate temporary vectorial force */
 497             tx               = fscal*dx22;
 498             ty               = fscal*dy22;
 499             tz               = fscal*dz22;
 500
 501             /* Update vectorial force */
 502             fix2            += tx;
 503             fiy2            += ty;
 504             fiz2            += tz;
 505             f[j_coord_offset+DIM*2+XX] -= tx;
 506             f[j_coord_offset+DIM*2+YY] -= ty;
 507             f[j_coord_offset+DIM*2+ZZ] -= tz;
 508
 509             /* Inner loop uses 282 flops */
 510         }
 511         /* End of innermost loop */
 512
 513         tx = ty = tz = 0;
 514         f[i_coord_offset+DIM*0+XX] += fix0;
 515         f[i_coord_offset+DIM*0+YY] += fiy0;
 516         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 517         tx                         += fix0;
 518         ty                         += fiy0;
 519         tz                         += fiz0;
 520         f[i_coord_offset+DIM*1+XX] += fix1;
 521         f[i_coord_offset+DIM*1+YY] += fiy1;
 522         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 523         tx                         += fix1;
 524         ty                         += fiy1;
 525         tz                         += fiz1;
 526         f[i_coord_offset+DIM*2+XX] += fix2;
 527         f[i_coord_offset+DIM*2+YY] += fiy2;
 528         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 529         tx                         += fix2;
 530         ty                         += fiy2;
 531         tz                         += fiz2;
 532         fshift[i_shift_offset+XX]  += tx;
 533         fshift[i_shift_offset+YY]  += ty;
 534         fshift[i_shift_offset+ZZ]  += tz;
 535
 536         ggid                        = gid[iidx];
 537         /* Update potential energies */
 538         kernel_data->energygrp_elec[ggid] += velecsum;
 539         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 540
 541         /* Increment number of inner iterations */
 542         inneriter                  += j_index_end - j_index_start;
 543
 544         /* Outer loop uses 32 flops */
 545     }
 546
 547     /* Increment number of outer iterations */
 548     outeriter        += nri;
 549
 550     /* Update outer/inner flops */
 551
 552     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*282);
 553 }
 554 /*
 555  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwBham_GeomW3W3_F_c
 556  * Electrostatics interaction: Coulomb
 557  * VdW interaction:            Buckingham
 558  * Geometry:                   Water3-Water3
 559  * Calculate force/pot:        Force
 560  */
 561 void
 562 nb_kernel_ElecCoul_VdwBham_GeomW3W3_F_c
 563                     (t_nblist                    * gmx_restrict       nlist,
 564                      rvec                        * gmx_restrict          xx,
 565                      rvec                        * gmx_restrict          ff,
 566                      t_forcerec                  * gmx_restrict          fr,
 567                      t_mdatoms                   * gmx_restrict     mdatoms,
 568                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 569                      t_nrnb                      * gmx_restrict        nrnb)
 570 {
 571     int              i_shift_offset,i_coord_offset,j_coord_offset;
 572     int              j_index_start,j_index_end;
 573     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 574     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 575     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 576     real             *shiftvec,*fshift,*x,*f;
 577     int              vdwioffset0;
 578     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 579     int              vdwioffset1;
 580     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 581     int              vdwioffset2;
 582     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 583     int              vdwjidx0;
 584     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 585     int              vdwjidx1;
 586     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 587     int              vdwjidx2;
 588     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 589     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 590     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 591     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 592     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 593     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 594     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 595     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 596     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 597     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 598     real             velec,felec,velecsum,facel,crf,krf,krf2;
 599     real             *charge;
 600     int              nvdwtype;
 601     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 602     int              *vdwtype;
 603     real             *vdwparam;
 604
 605     x                = xx[0];
 606     f                = ff[0];
 607
 608     nri              = nlist->nri;
 609     iinr             = nlist->iinr;
 610     jindex           = nlist->jindex;
 611     jjnr             = nlist->jjnr;
 612     shiftidx         = nlist->shift;
 613     gid              = nlist->gid;
 614     shiftvec         = fr->shift_vec[0];
 615     fshift           = fr->fshift[0];
 616     facel            = fr->epsfac;
 617     charge           = mdatoms->chargeA;
 618     nvdwtype         = fr->ntype;
 619     vdwparam         = fr->nbfp;
 620     vdwtype          = mdatoms->typeA;
 621
 622     /* Setup water-specific parameters */
 623     inr              = nlist->iinr[0];
 624     iq0              = facel*charge[inr+0];
 625     iq1              = facel*charge[inr+1];
 626     iq2              = facel*charge[inr+2];
 627     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 628
 629     jq0              = charge[inr+0];
 630     jq1              = charge[inr+1];
 631     jq2              = charge[inr+2];
 632     vdwjidx0         = 3*vdwtype[inr+0];
 633     qq00             = iq0*jq0;
 634     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 635     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 636     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 637     qq01             = iq0*jq1;
 638     qq02             = iq0*jq2;
 639     qq10             = iq1*jq0;
 640     qq11             = iq1*jq1;
 641     qq12             = iq1*jq2;
 642     qq20             = iq2*jq0;
 643     qq21             = iq2*jq1;
 644     qq22             = iq2*jq2;
 645
 646     outeriter        = 0;
 647     inneriter        = 0;
 648
 649     /* Start outer loop over neighborlists */
 650     for(iidx=0; iidx<nri; iidx++)
 651     {
 652         /* Load shift vector for this list */
 653         i_shift_offset   = DIM*shiftidx[iidx];
 654         shX              = shiftvec[i_shift_offset+XX];
 655         shY              = shiftvec[i_shift_offset+YY];
 656         shZ              = shiftvec[i_shift_offset+ZZ];
 657
 658         /* Load limits for loop over neighbors */
 659         j_index_start    = jindex[iidx];
 660         j_index_end      = jindex[iidx+1];
 661
 662         /* Get outer coordinate index */
 663         inr              = iinr[iidx];
 664         i_coord_offset   = DIM*inr;
 665
 666         /* Load i particle coords and add shift vector */
 667         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 668         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 669         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 670         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 671         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 672         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 673         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 674         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 675         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 676
 677         fix0             = 0.0;
 678         fiy0             = 0.0;
 679         fiz0             = 0.0;
 680         fix1             = 0.0;
 681         fiy1             = 0.0;
 682         fiz1             = 0.0;
 683         fix2             = 0.0;
 684         fiy2             = 0.0;
 685         fiz2             = 0.0;
 686
 687         /* Start inner kernel loop */
 688         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 689         {
 690             /* Get j neighbor index, and coordinate index */
 691             jnr              = jjnr[jidx];
 692             j_coord_offset   = DIM*jnr;
 693
 694             /* load j atom coordinates */
 695             jx0              = x[j_coord_offset+DIM*0+XX];
 696             jy0              = x[j_coord_offset+DIM*0+YY];
 697             jz0              = x[j_coord_offset+DIM*0+ZZ];
 698             jx1              = x[j_coord_offset+DIM*1+XX];
 699             jy1              = x[j_coord_offset+DIM*1+YY];
 700             jz1              = x[j_coord_offset+DIM*1+ZZ];
 701             jx2              = x[j_coord_offset+DIM*2+XX];
 702             jy2              = x[j_coord_offset+DIM*2+YY];
 703             jz2              = x[j_coord_offset+DIM*2+ZZ];
 704
 705             /* Calculate displacement vector */
 706             dx00             = ix0 - jx0;
 707             dy00             = iy0 - jy0;
 708             dz00             = iz0 - jz0;
 709             dx01             = ix0 - jx1;
 710             dy01             = iy0 - jy1;
 711             dz01             = iz0 - jz1;
 712             dx02             = ix0 - jx2;
 713             dy02             = iy0 - jy2;
 714             dz02             = iz0 - jz2;
 715             dx10             = ix1 - jx0;
 716             dy10             = iy1 - jy0;
 717             dz10             = iz1 - jz0;
 718             dx11             = ix1 - jx1;
 719             dy11             = iy1 - jy1;
 720             dz11             = iz1 - jz1;
 721             dx12             = ix1 - jx2;
 722             dy12             = iy1 - jy2;
 723             dz12             = iz1 - jz2;
 724             dx20             = ix2 - jx0;
 725             dy20             = iy2 - jy0;
 726             dz20             = iz2 - jz0;
 727             dx21             = ix2 - jx1;
 728             dy21             = iy2 - jy1;
 729             dz21             = iz2 - jz1;
 730             dx22             = ix2 - jx2;
 731             dy22             = iy2 - jy2;
 732             dz22             = iz2 - jz2;
 733
 734             /* Calculate squared distance and things based on it */
 735             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 736             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 737             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 738             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 739             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 740             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 741             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 742             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 743             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 744
 745             rinv00           = gmx_invsqrt(rsq00);
 746             rinv01           = gmx_invsqrt(rsq01);
 747             rinv02           = gmx_invsqrt(rsq02);
 748             rinv10           = gmx_invsqrt(rsq10);
 749             rinv11           = gmx_invsqrt(rsq11);
 750             rinv12           = gmx_invsqrt(rsq12);
 751             rinv20           = gmx_invsqrt(rsq20);
 752             rinv21           = gmx_invsqrt(rsq21);
 753             rinv22           = gmx_invsqrt(rsq22);
 754
 755             rinvsq00         = rinv00*rinv00;
 756             rinvsq01         = rinv01*rinv01;
 757             rinvsq02         = rinv02*rinv02;
 758             rinvsq10         = rinv10*rinv10;
 759             rinvsq11         = rinv11*rinv11;
 760             rinvsq12         = rinv12*rinv12;
 761             rinvsq20         = rinv20*rinv20;
 762             rinvsq21         = rinv21*rinv21;
 763             rinvsq22         = rinv22*rinv22;
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             r00              = rsq00*rinv00;
 770
 771             /* COULOMB ELECTROSTATICS */
 772             velec            = qq00*rinv00;
 773             felec            = velec*rinvsq00;
 774
 775             /* BUCKINGHAM DISPERSION/REPULSION */
 776             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 777             vvdw6            = c6_00*rinvsix;
 778             br               = cexp2_00*r00;
 779             vvdwexp          = cexp1_00*exp(-br);
 780             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 781
 782             fscal            = felec+fvdw;
 783
 784             /* Calculate temporary vectorial force */
 785             tx               = fscal*dx00;
 786             ty               = fscal*dy00;
 787             tz               = fscal*dz00;
 788
 789             /* Update vectorial force */
 790             fix0            += tx;
 791             fiy0            += ty;
 792             fiz0            += tz;
 793             f[j_coord_offset+DIM*0+XX] -= tx;
 794             f[j_coord_offset+DIM*0+YY] -= ty;
 795             f[j_coord_offset+DIM*0+ZZ] -= tz;
 796
 797             /**************************
 798              * CALCULATE INTERACTIONS *
 799              **************************/
 800
 801             /* COULOMB ELECTROSTATICS */
 802             velec            = qq01*rinv01;
 803             felec            = velec*rinvsq01;
 804
 805             fscal            = felec;
 806
 807             /* Calculate temporary vectorial force */
 808             tx               = fscal*dx01;
 809             ty               = fscal*dy01;
 810             tz               = fscal*dz01;
 811
 812             /* Update vectorial force */
 813             fix0            += tx;
 814             fiy0            += ty;
 815             fiz0            += tz;
 816             f[j_coord_offset+DIM*1+XX] -= tx;
 817             f[j_coord_offset+DIM*1+YY] -= ty;
 818             f[j_coord_offset+DIM*1+ZZ] -= tz;
 819
 820             /**************************
 821              * CALCULATE INTERACTIONS *
 822              **************************/
 823
 824             /* COULOMB ELECTROSTATICS */
 825             velec            = qq02*rinv02;
 826             felec            = velec*rinvsq02;
 827
 828             fscal            = felec;
 829
 830             /* Calculate temporary vectorial force */
 831             tx               = fscal*dx02;
 832             ty               = fscal*dy02;
 833             tz               = fscal*dz02;
 834
 835             /* Update vectorial force */
 836             fix0            += tx;
 837             fiy0            += ty;
 838             fiz0            += tz;
 839             f[j_coord_offset+DIM*2+XX] -= tx;
 840             f[j_coord_offset+DIM*2+YY] -= ty;
 841             f[j_coord_offset+DIM*2+ZZ] -= tz;
 842
 843             /**************************
 844              * CALCULATE INTERACTIONS *
 845              **************************/
 846
 847             /* COULOMB ELECTROSTATICS */
 848             velec            = qq10*rinv10;
 849             felec            = velec*rinvsq10;
 850
 851             fscal            = felec;
 852
 853             /* Calculate temporary vectorial force */
 854             tx               = fscal*dx10;
 855             ty               = fscal*dy10;
 856             tz               = fscal*dz10;
 857
 858             /* Update vectorial force */
 859             fix1            += tx;
 860             fiy1            += ty;
 861             fiz1            += tz;
 862             f[j_coord_offset+DIM*0+XX] -= tx;
 863             f[j_coord_offset+DIM*0+YY] -= ty;
 864             f[j_coord_offset+DIM*0+ZZ] -= tz;
 865
 866             /**************************
 867              * CALCULATE INTERACTIONS *
 868              **************************/
 869
 870             /* COULOMB ELECTROSTATICS */
 871             velec            = qq11*rinv11;
 872             felec            = velec*rinvsq11;
 873
 874             fscal            = felec;
 875
 876             /* Calculate temporary vectorial force */
 877             tx               = fscal*dx11;
 878             ty               = fscal*dy11;
 879             tz               = fscal*dz11;
 880
 881             /* Update vectorial force */
 882             fix1            += tx;
 883             fiy1            += ty;
 884             fiz1            += tz;
 885             f[j_coord_offset+DIM*1+XX] -= tx;
 886             f[j_coord_offset+DIM*1+YY] -= ty;
 887             f[j_coord_offset+DIM*1+ZZ] -= tz;
 888
 889             /**************************
 890              * CALCULATE INTERACTIONS *
 891              **************************/
 892
 893             /* COULOMB ELECTROSTATICS */
 894             velec            = qq12*rinv12;
 895             felec            = velec*rinvsq12;
 896
 897             fscal            = felec;
 898
 899             /* Calculate temporary vectorial force */
 900             tx               = fscal*dx12;
 901             ty               = fscal*dy12;
 902             tz               = fscal*dz12;
 903
 904             /* Update vectorial force */
 905             fix1            += tx;
 906             fiy1            += ty;
 907             fiz1            += tz;
 908             f[j_coord_offset+DIM*2+XX] -= tx;
 909             f[j_coord_offset+DIM*2+YY] -= ty;
 910             f[j_coord_offset+DIM*2+ZZ] -= tz;
 911
 912             /**************************
 913              * CALCULATE INTERACTIONS *
 914              **************************/
 915
 916             /* COULOMB ELECTROSTATICS */
 917             velec            = qq20*rinv20;
 918             felec            = velec*rinvsq20;
 919
 920             fscal            = felec;
 921
 922             /* Calculate temporary vectorial force */
 923             tx               = fscal*dx20;
 924             ty               = fscal*dy20;
 925             tz               = fscal*dz20;
 926
 927             /* Update vectorial force */
 928             fix2            += tx;
 929             fiy2            += ty;
 930             fiz2            += tz;
 931             f[j_coord_offset+DIM*0+XX] -= tx;
 932             f[j_coord_offset+DIM*0+YY] -= ty;
 933             f[j_coord_offset+DIM*0+ZZ] -= tz;
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             /* COULOMB ELECTROSTATICS */
 940             velec            = qq21*rinv21;
 941             felec            = velec*rinvsq21;
 942
 943             fscal            = felec;
 944
 945             /* Calculate temporary vectorial force */
 946             tx               = fscal*dx21;
 947             ty               = fscal*dy21;
 948             tz               = fscal*dz21;
 949
 950             /* Update vectorial force */
 951             fix2            += tx;
 952             fiy2            += ty;
 953             fiz2            += tz;
 954             f[j_coord_offset+DIM*1+XX] -= tx;
 955             f[j_coord_offset+DIM*1+YY] -= ty;
 956             f[j_coord_offset+DIM*1+ZZ] -= tz;
 957
 958             /**************************
 959              * CALCULATE INTERACTIONS *
 960              **************************/
 961
 962             /* COULOMB ELECTROSTATICS */
 963             velec            = qq22*rinv22;
 964             felec            = velec*rinvsq22;
 965
 966             fscal            = felec;
 967
 968             /* Calculate temporary vectorial force */
 969             tx               = fscal*dx22;
 970             ty               = fscal*dy22;
 971             tz               = fscal*dz22;
 972
 973             /* Update vectorial force */
 974             fix2            += tx;
 975             fiy2            += ty;
 976             fiz2            += tz;
 977             f[j_coord_offset+DIM*2+XX] -= tx;
 978             f[j_coord_offset+DIM*2+YY] -= ty;
 979             f[j_coord_offset+DIM*2+ZZ] -= tz;
 980
 981             /* Inner loop uses 270 flops */
 982         }
 983         /* End of innermost loop */
 984
 985         tx = ty = tz = 0;
 986         f[i_coord_offset+DIM*0+XX] += fix0;
 987         f[i_coord_offset+DIM*0+YY] += fiy0;
 988         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 989         tx                         += fix0;
 990         ty                         += fiy0;
 991         tz                         += fiz0;
 992         f[i_coord_offset+DIM*1+XX] += fix1;
 993         f[i_coord_offset+DIM*1+YY] += fiy1;
 994         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 995         tx                         += fix1;
 996         ty                         += fiy1;
 997         tz                         += fiz1;
 998         f[i_coord_offset+DIM*2+XX] += fix2;
 999         f[i_coord_offset+DIM*2+YY] += fiy2;
1000         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1001         tx                         += fix2;
1002         ty                         += fiy2;
1003         tz                         += fiz2;
1004         fshift[i_shift_offset+XX]  += tx;
1005         fshift[i_shift_offset+YY]  += ty;
1006         fshift[i_shift_offset+ZZ]  += tz;
1007
1008         /* Increment number of inner iterations */
1009         inneriter                  += j_index_end - j_index_start;
1010
1011         /* Outer loop uses 30 flops */
1012     }
1013
1014     /* Increment number of outer iterations */
1015     outeriter        += nri;
1016
1017     /* Update outer/inner flops */
1018
1019     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*270);
1020 }