src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecCoul_VdwLJ_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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_VF_c
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_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      = 2*nvdwtype*vdwtype[inr+0];
 123
 124     jq0              = charge[inr+0];
 125     jq1              = charge[inr+1];
 126     jq2              = charge[inr+2];
 127     vdwjidx0         = 2*vdwtype[inr+0];
 128     qq00             = iq0*jq0;
 129     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 130     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 131     qq01             = iq0*jq1;
 132     qq02             = iq0*jq2;
 133     qq10             = iq1*jq0;
 134     qq11             = iq1*jq1;
 135     qq12             = iq1*jq2;
 136     qq20             = iq2*jq0;
 137     qq21             = iq2*jq1;
 138     qq22             = iq2*jq2;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     /* Start outer loop over neighborlists */
 144     for(iidx=0; iidx<nri; iidx++)
 145     {
 146         /* Load shift vector for this list */
 147         i_shift_offset   = DIM*shiftidx[iidx];
 148         shX              = shiftvec[i_shift_offset+XX];
 149         shY              = shiftvec[i_shift_offset+YY];
 150         shZ              = shiftvec[i_shift_offset+ZZ];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 162         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 163         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 164         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 165         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 166         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 167         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 168         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 169         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 170
 171         fix0             = 0.0;
 172         fiy0             = 0.0;
 173         fiz0             = 0.0;
 174         fix1             = 0.0;
 175         fiy1             = 0.0;
 176         fiz1             = 0.0;
 177         fix2             = 0.0;
 178         fiy2             = 0.0;
 179         fiz2             = 0.0;
 180
 181         /* Reset potential sums */
 182         velecsum         = 0.0;
 183         vvdwsum          = 0.0;
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 187         {
 188             /* Get j neighbor index, and coordinate index */
 189             jnr              = jjnr[jidx];
 190             j_coord_offset   = DIM*jnr;
 191
 192             /* load j atom coordinates */
 193             jx0              = x[j_coord_offset+DIM*0+XX];
 194             jy0              = x[j_coord_offset+DIM*0+YY];
 195             jz0              = x[j_coord_offset+DIM*0+ZZ];
 196             jx1              = x[j_coord_offset+DIM*1+XX];
 197             jy1              = x[j_coord_offset+DIM*1+YY];
 198             jz1              = x[j_coord_offset+DIM*1+ZZ];
 199             jx2              = x[j_coord_offset+DIM*2+XX];
 200             jy2              = x[j_coord_offset+DIM*2+YY];
 201             jz2              = x[j_coord_offset+DIM*2+ZZ];
 202
 203             /* Calculate displacement vector */
 204             dx00             = ix0 - jx0;
 205             dy00             = iy0 - jy0;
 206             dz00             = iz0 - jz0;
 207             dx01             = ix0 - jx1;
 208             dy01             = iy0 - jy1;
 209             dz01             = iz0 - jz1;
 210             dx02             = ix0 - jx2;
 211             dy02             = iy0 - jy2;
 212             dz02             = iz0 - jz2;
 213             dx10             = ix1 - jx0;
 214             dy10             = iy1 - jy0;
 215             dz10             = iz1 - jz0;
 216             dx11             = ix1 - jx1;
 217             dy11             = iy1 - jy1;
 218             dz11             = iz1 - jz1;
 219             dx12             = ix1 - jx2;
 220             dy12             = iy1 - jy2;
 221             dz12             = iz1 - jz2;
 222             dx20             = ix2 - jx0;
 223             dy20             = iy2 - jy0;
 224             dz20             = iz2 - jz0;
 225             dx21             = ix2 - jx1;
 226             dy21             = iy2 - jy1;
 227             dz21             = iz2 - jz1;
 228             dx22             = ix2 - jx2;
 229             dy22             = iy2 - jy2;
 230             dz22             = iz2 - jz2;
 231
 232             /* Calculate squared distance and things based on it */
 233             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 234             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 235             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 236             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 237             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 238             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 239             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 240             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 241             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 242
 243             rinv00           = gmx_invsqrt(rsq00);
 244             rinv01           = gmx_invsqrt(rsq01);
 245             rinv02           = gmx_invsqrt(rsq02);
 246             rinv10           = gmx_invsqrt(rsq10);
 247             rinv11           = gmx_invsqrt(rsq11);
 248             rinv12           = gmx_invsqrt(rsq12);
 249             rinv20           = gmx_invsqrt(rsq20);
 250             rinv21           = gmx_invsqrt(rsq21);
 251             rinv22           = gmx_invsqrt(rsq22);
 252
 253             rinvsq00         = rinv00*rinv00;
 254             rinvsq01         = rinv01*rinv01;
 255             rinvsq02         = rinv02*rinv02;
 256             rinvsq10         = rinv10*rinv10;
 257             rinvsq11         = rinv11*rinv11;
 258             rinvsq12         = rinv12*rinv12;
 259             rinvsq20         = rinv20*rinv20;
 260             rinvsq21         = rinv21*rinv21;
 261             rinvsq22         = rinv22*rinv22;
 262
 263             /**************************
 264              * CALCULATE INTERACTIONS *
 265              **************************/
 266
 267             /* COULOMB ELECTROSTATICS */
 268             velec            = qq00*rinv00;
 269             felec            = velec*rinvsq00;
 270
 271             /* LENNARD-JONES DISPERSION/REPULSION */
 272
 273             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 274             vvdw6            = c6_00*rinvsix;
 275             vvdw12           = c12_00*rinvsix*rinvsix;
 276             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 277             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 278
 279             /* Update potential sums from outer loop */
 280             velecsum        += velec;
 281             vvdwsum         += vvdw;
 282
 283             fscal            = felec+fvdw;
 284
 285             /* Calculate temporary vectorial force */
 286             tx               = fscal*dx00;
 287             ty               = fscal*dy00;
 288             tz               = fscal*dz00;
 289
 290             /* Update vectorial force */
 291             fix0            += tx;
 292             fiy0            += ty;
 293             fiz0            += tz;
 294             f[j_coord_offset+DIM*0+XX] -= tx;
 295             f[j_coord_offset+DIM*0+YY] -= ty;
 296             f[j_coord_offset+DIM*0+ZZ] -= tz;
 297
 298             /**************************
 299              * CALCULATE INTERACTIONS *
 300              **************************/
 301
 302             /* COULOMB ELECTROSTATICS */
 303             velec            = qq01*rinv01;
 304             felec            = velec*rinvsq01;
 305
 306             /* Update potential sums from outer loop */
 307             velecsum        += velec;
 308
 309             fscal            = felec;
 310
 311             /* Calculate temporary vectorial force */
 312             tx               = fscal*dx01;
 313             ty               = fscal*dy01;
 314             tz               = fscal*dz01;
 315
 316             /* Update vectorial force */
 317             fix0            += tx;
 318             fiy0            += ty;
 319             fiz0            += tz;
 320             f[j_coord_offset+DIM*1+XX] -= tx;
 321             f[j_coord_offset+DIM*1+YY] -= ty;
 322             f[j_coord_offset+DIM*1+ZZ] -= tz;
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             /* COULOMB ELECTROSTATICS */
 329             velec            = qq02*rinv02;
 330             felec            = velec*rinvsq02;
 331
 332             /* Update potential sums from outer loop */
 333             velecsum        += velec;
 334
 335             fscal            = felec;
 336
 337             /* Calculate temporary vectorial force */
 338             tx               = fscal*dx02;
 339             ty               = fscal*dy02;
 340             tz               = fscal*dz02;
 341
 342             /* Update vectorial force */
 343             fix0            += tx;
 344             fiy0            += ty;
 345             fiz0            += tz;
 346             f[j_coord_offset+DIM*2+XX] -= tx;
 347             f[j_coord_offset+DIM*2+YY] -= ty;
 348             f[j_coord_offset+DIM*2+ZZ] -= tz;
 349
 350             /**************************
 351              * CALCULATE INTERACTIONS *
 352              **************************/
 353
 354             /* COULOMB ELECTROSTATICS */
 355             velec            = qq10*rinv10;
 356             felec            = velec*rinvsq10;
 357
 358             /* Update potential sums from outer loop */
 359             velecsum        += velec;
 360
 361             fscal            = felec;
 362
 363             /* Calculate temporary vectorial force */
 364             tx               = fscal*dx10;
 365             ty               = fscal*dy10;
 366             tz               = fscal*dz10;
 367
 368             /* Update vectorial force */
 369             fix1            += tx;
 370             fiy1            += ty;
 371             fiz1            += tz;
 372             f[j_coord_offset+DIM*0+XX] -= tx;
 373             f[j_coord_offset+DIM*0+YY] -= ty;
 374             f[j_coord_offset+DIM*0+ZZ] -= tz;
 375
 376             /**************************
 377              * CALCULATE INTERACTIONS *
 378              **************************/
 379
 380             /* COULOMB ELECTROSTATICS */
 381             velec            = qq11*rinv11;
 382             felec            = velec*rinvsq11;
 383
 384             /* Update potential sums from outer loop */
 385             velecsum        += velec;
 386
 387             fscal            = felec;
 388
 389             /* Calculate temporary vectorial force */
 390             tx               = fscal*dx11;
 391             ty               = fscal*dy11;
 392             tz               = fscal*dz11;
 393
 394             /* Update vectorial force */
 395             fix1            += tx;
 396             fiy1            += ty;
 397             fiz1            += tz;
 398             f[j_coord_offset+DIM*1+XX] -= tx;
 399             f[j_coord_offset+DIM*1+YY] -= ty;
 400             f[j_coord_offset+DIM*1+ZZ] -= tz;
 401
 402             /**************************
 403              * CALCULATE INTERACTIONS *
 404              **************************/
 405
 406             /* COULOMB ELECTROSTATICS */
 407             velec            = qq12*rinv12;
 408             felec            = velec*rinvsq12;
 409
 410             /* Update potential sums from outer loop */
 411             velecsum        += velec;
 412
 413             fscal            = felec;
 414
 415             /* Calculate temporary vectorial force */
 416             tx               = fscal*dx12;
 417             ty               = fscal*dy12;
 418             tz               = fscal*dz12;
 419
 420             /* Update vectorial force */
 421             fix1            += tx;
 422             fiy1            += ty;
 423             fiz1            += tz;
 424             f[j_coord_offset+DIM*2+XX] -= tx;
 425             f[j_coord_offset+DIM*2+YY] -= ty;
 426             f[j_coord_offset+DIM*2+ZZ] -= tz;
 427
 428             /**************************
 429              * CALCULATE INTERACTIONS *
 430              **************************/
 431
 432             /* COULOMB ELECTROSTATICS */
 433             velec            = qq20*rinv20;
 434             felec            = velec*rinvsq20;
 435
 436             /* Update potential sums from outer loop */
 437             velecsum        += velec;
 438
 439             fscal            = felec;
 440
 441             /* Calculate temporary vectorial force */
 442             tx               = fscal*dx20;
 443             ty               = fscal*dy20;
 444             tz               = fscal*dz20;
 445
 446             /* Update vectorial force */
 447             fix2            += tx;
 448             fiy2            += ty;
 449             fiz2            += tz;
 450             f[j_coord_offset+DIM*0+XX] -= tx;
 451             f[j_coord_offset+DIM*0+YY] -= ty;
 452             f[j_coord_offset+DIM*0+ZZ] -= tz;
 453
 454             /**************************
 455              * CALCULATE INTERACTIONS *
 456              **************************/
 457
 458             /* COULOMB ELECTROSTATICS */
 459             velec            = qq21*rinv21;
 460             felec            = velec*rinvsq21;
 461
 462             /* Update potential sums from outer loop */
 463             velecsum        += velec;
 464
 465             fscal            = felec;
 466
 467             /* Calculate temporary vectorial force */
 468             tx               = fscal*dx21;
 469             ty               = fscal*dy21;
 470             tz               = fscal*dz21;
 471
 472             /* Update vectorial force */
 473             fix2            += tx;
 474             fiy2            += ty;
 475             fiz2            += tz;
 476             f[j_coord_offset+DIM*1+XX] -= tx;
 477             f[j_coord_offset+DIM*1+YY] -= ty;
 478             f[j_coord_offset+DIM*1+ZZ] -= tz;
 479
 480             /**************************
 481              * CALCULATE INTERACTIONS *
 482              **************************/
 483
 484             /* COULOMB ELECTROSTATICS */
 485             velec            = qq22*rinv22;
 486             felec            = velec*rinvsq22;
 487
 488             /* Update potential sums from outer loop */
 489             velecsum        += velec;
 490
 491             fscal            = felec;
 492
 493             /* Calculate temporary vectorial force */
 494             tx               = fscal*dx22;
 495             ty               = fscal*dy22;
 496             tz               = fscal*dz22;
 497
 498             /* Update vectorial force */
 499             fix2            += tx;
 500             fiy2            += ty;
 501             fiz2            += tz;
 502             f[j_coord_offset+DIM*2+XX] -= tx;
 503             f[j_coord_offset+DIM*2+YY] -= ty;
 504             f[j_coord_offset+DIM*2+ZZ] -= tz;
 505
 506             /* Inner loop uses 255 flops */
 507         }
 508         /* End of innermost loop */
 509
 510         tx = ty = tz = 0;
 511         f[i_coord_offset+DIM*0+XX] += fix0;
 512         f[i_coord_offset+DIM*0+YY] += fiy0;
 513         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 514         tx                         += fix0;
 515         ty                         += fiy0;
 516         tz                         += fiz0;
 517         f[i_coord_offset+DIM*1+XX] += fix1;
 518         f[i_coord_offset+DIM*1+YY] += fiy1;
 519         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 520         tx                         += fix1;
 521         ty                         += fiy1;
 522         tz                         += fiz1;
 523         f[i_coord_offset+DIM*2+XX] += fix2;
 524         f[i_coord_offset+DIM*2+YY] += fiy2;
 525         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 526         tx                         += fix2;
 527         ty                         += fiy2;
 528         tz                         += fiz2;
 529         fshift[i_shift_offset+XX]  += tx;
 530         fshift[i_shift_offset+YY]  += ty;
 531         fshift[i_shift_offset+ZZ]  += tz;
 532
 533         ggid                        = gid[iidx];
 534         /* Update potential energies */
 535         kernel_data->energygrp_elec[ggid] += velecsum;
 536         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 537
 538         /* Increment number of inner iterations */
 539         inneriter                  += j_index_end - j_index_start;
 540
 541         /* Outer loop uses 32 flops */
 542     }
 543
 544     /* Increment number of outer iterations */
 545     outeriter        += nri;
 546
 547     /* Update outer/inner flops */
 548
 549     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*255);
 550 }
 551 /*
 552  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_c
 553  * Electrostatics interaction: Coulomb
 554  * VdW interaction:            LennardJones
 555  * Geometry:                   Water3-Water3
 556  * Calculate force/pot:        Force
 557  */
 558 void
 559 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_c
 560                     (t_nblist                    * gmx_restrict       nlist,
 561                      rvec                        * gmx_restrict          xx,
 562                      rvec                        * gmx_restrict          ff,
 563                      t_forcerec                  * gmx_restrict          fr,
 564                      t_mdatoms                   * gmx_restrict     mdatoms,
 565                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 566                      t_nrnb                      * gmx_restrict        nrnb)
 567 {
 568     int              i_shift_offset,i_coord_offset,j_coord_offset;
 569     int              j_index_start,j_index_end;
 570     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 571     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 572     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 573     real             *shiftvec,*fshift,*x,*f;
 574     int              vdwioffset0;
 575     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 576     int              vdwioffset1;
 577     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 578     int              vdwioffset2;
 579     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 580     int              vdwjidx0;
 581     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 582     int              vdwjidx1;
 583     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 584     int              vdwjidx2;
 585     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 586     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 587     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 588     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 589     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 590     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 591     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 592     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 593     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 594     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 595     real             velec,felec,velecsum,facel,crf,krf,krf2;
 596     real             *charge;
 597     int              nvdwtype;
 598     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 599     int              *vdwtype;
 600     real             *vdwparam;
 601
 602     x                = xx[0];
 603     f                = ff[0];
 604
 605     nri              = nlist->nri;
 606     iinr             = nlist->iinr;
 607     jindex           = nlist->jindex;
 608     jjnr             = nlist->jjnr;
 609     shiftidx         = nlist->shift;
 610     gid              = nlist->gid;
 611     shiftvec         = fr->shift_vec[0];
 612     fshift           = fr->fshift[0];
 613     facel            = fr->epsfac;
 614     charge           = mdatoms->chargeA;
 615     nvdwtype         = fr->ntype;
 616     vdwparam         = fr->nbfp;
 617     vdwtype          = mdatoms->typeA;
 618
 619     /* Setup water-specific parameters */
 620     inr              = nlist->iinr[0];
 621     iq0              = facel*charge[inr+0];
 622     iq1              = facel*charge[inr+1];
 623     iq2              = facel*charge[inr+2];
 624     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 625
 626     jq0              = charge[inr+0];
 627     jq1              = charge[inr+1];
 628     jq2              = charge[inr+2];
 629     vdwjidx0         = 2*vdwtype[inr+0];
 630     qq00             = iq0*jq0;
 631     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 632     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 633     qq01             = iq0*jq1;
 634     qq02             = iq0*jq2;
 635     qq10             = iq1*jq0;
 636     qq11             = iq1*jq1;
 637     qq12             = iq1*jq2;
 638     qq20             = iq2*jq0;
 639     qq21             = iq2*jq1;
 640     qq22             = iq2*jq2;
 641
 642     outeriter        = 0;
 643     inneriter        = 0;
 644
 645     /* Start outer loop over neighborlists */
 646     for(iidx=0; iidx<nri; iidx++)
 647     {
 648         /* Load shift vector for this list */
 649         i_shift_offset   = DIM*shiftidx[iidx];
 650         shX              = shiftvec[i_shift_offset+XX];
 651         shY              = shiftvec[i_shift_offset+YY];
 652         shZ              = shiftvec[i_shift_offset+ZZ];
 653
 654         /* Load limits for loop over neighbors */
 655         j_index_start    = jindex[iidx];
 656         j_index_end      = jindex[iidx+1];
 657
 658         /* Get outer coordinate index */
 659         inr              = iinr[iidx];
 660         i_coord_offset   = DIM*inr;
 661
 662         /* Load i particle coords and add shift vector */
 663         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 664         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 665         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 666         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 667         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 668         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 669         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 670         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 671         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 672
 673         fix0             = 0.0;
 674         fiy0             = 0.0;
 675         fiz0             = 0.0;
 676         fix1             = 0.0;
 677         fiy1             = 0.0;
 678         fiz1             = 0.0;
 679         fix2             = 0.0;
 680         fiy2             = 0.0;
 681         fiz2             = 0.0;
 682
 683         /* Start inner kernel loop */
 684         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 685         {
 686             /* Get j neighbor index, and coordinate index */
 687             jnr              = jjnr[jidx];
 688             j_coord_offset   = DIM*jnr;
 689
 690             /* load j atom coordinates */
 691             jx0              = x[j_coord_offset+DIM*0+XX];
 692             jy0              = x[j_coord_offset+DIM*0+YY];
 693             jz0              = x[j_coord_offset+DIM*0+ZZ];
 694             jx1              = x[j_coord_offset+DIM*1+XX];
 695             jy1              = x[j_coord_offset+DIM*1+YY];
 696             jz1              = x[j_coord_offset+DIM*1+ZZ];
 697             jx2              = x[j_coord_offset+DIM*2+XX];
 698             jy2              = x[j_coord_offset+DIM*2+YY];
 699             jz2              = x[j_coord_offset+DIM*2+ZZ];
 700
 701             /* Calculate displacement vector */
 702             dx00             = ix0 - jx0;
 703             dy00             = iy0 - jy0;
 704             dz00             = iz0 - jz0;
 705             dx01             = ix0 - jx1;
 706             dy01             = iy0 - jy1;
 707             dz01             = iz0 - jz1;
 708             dx02             = ix0 - jx2;
 709             dy02             = iy0 - jy2;
 710             dz02             = iz0 - jz2;
 711             dx10             = ix1 - jx0;
 712             dy10             = iy1 - jy0;
 713             dz10             = iz1 - jz0;
 714             dx11             = ix1 - jx1;
 715             dy11             = iy1 - jy1;
 716             dz11             = iz1 - jz1;
 717             dx12             = ix1 - jx2;
 718             dy12             = iy1 - jy2;
 719             dz12             = iz1 - jz2;
 720             dx20             = ix2 - jx0;
 721             dy20             = iy2 - jy0;
 722             dz20             = iz2 - jz0;
 723             dx21             = ix2 - jx1;
 724             dy21             = iy2 - jy1;
 725             dz21             = iz2 - jz1;
 726             dx22             = ix2 - jx2;
 727             dy22             = iy2 - jy2;
 728             dz22             = iz2 - jz2;
 729
 730             /* Calculate squared distance and things based on it */
 731             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 732             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 733             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 734             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 735             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 736             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 737             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 738             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 739             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 740
 741             rinv00           = gmx_invsqrt(rsq00);
 742             rinv01           = gmx_invsqrt(rsq01);
 743             rinv02           = gmx_invsqrt(rsq02);
 744             rinv10           = gmx_invsqrt(rsq10);
 745             rinv11           = gmx_invsqrt(rsq11);
 746             rinv12           = gmx_invsqrt(rsq12);
 747             rinv20           = gmx_invsqrt(rsq20);
 748             rinv21           = gmx_invsqrt(rsq21);
 749             rinv22           = gmx_invsqrt(rsq22);
 750
 751             rinvsq00         = rinv00*rinv00;
 752             rinvsq01         = rinv01*rinv01;
 753             rinvsq02         = rinv02*rinv02;
 754             rinvsq10         = rinv10*rinv10;
 755             rinvsq11         = rinv11*rinv11;
 756             rinvsq12         = rinv12*rinv12;
 757             rinvsq20         = rinv20*rinv20;
 758             rinvsq21         = rinv21*rinv21;
 759             rinvsq22         = rinv22*rinv22;
 760
 761             /**************************
 762              * CALCULATE INTERACTIONS *
 763              **************************/
 764
 765             /* COULOMB ELECTROSTATICS */
 766             velec            = qq00*rinv00;
 767             felec            = velec*rinvsq00;
 768
 769             /* LENNARD-JONES DISPERSION/REPULSION */
 770
 771             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 772             fvdw             = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00;
 773
 774             fscal            = felec+fvdw;
 775
 776             /* Calculate temporary vectorial force */
 777             tx               = fscal*dx00;
 778             ty               = fscal*dy00;
 779             tz               = fscal*dz00;
 780
 781             /* Update vectorial force */
 782             fix0            += tx;
 783             fiy0            += ty;
 784             fiz0            += tz;
 785             f[j_coord_offset+DIM*0+XX] -= tx;
 786             f[j_coord_offset+DIM*0+YY] -= ty;
 787             f[j_coord_offset+DIM*0+ZZ] -= tz;
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             /* COULOMB ELECTROSTATICS */
 794             velec            = qq01*rinv01;
 795             felec            = velec*rinvsq01;
 796
 797             fscal            = felec;
 798
 799             /* Calculate temporary vectorial force */
 800             tx               = fscal*dx01;
 801             ty               = fscal*dy01;
 802             tz               = fscal*dz01;
 803
 804             /* Update vectorial force */
 805             fix0            += tx;
 806             fiy0            += ty;
 807             fiz0            += tz;
 808             f[j_coord_offset+DIM*1+XX] -= tx;
 809             f[j_coord_offset+DIM*1+YY] -= ty;
 810             f[j_coord_offset+DIM*1+ZZ] -= tz;
 811
 812             /**************************
 813              * CALCULATE INTERACTIONS *
 814              **************************/
 815
 816             /* COULOMB ELECTROSTATICS */
 817             velec            = qq02*rinv02;
 818             felec            = velec*rinvsq02;
 819
 820             fscal            = felec;
 821
 822             /* Calculate temporary vectorial force */
 823             tx               = fscal*dx02;
 824             ty               = fscal*dy02;
 825             tz               = fscal*dz02;
 826
 827             /* Update vectorial force */
 828             fix0            += tx;
 829             fiy0            += ty;
 830             fiz0            += tz;
 831             f[j_coord_offset+DIM*2+XX] -= tx;
 832             f[j_coord_offset+DIM*2+YY] -= ty;
 833             f[j_coord_offset+DIM*2+ZZ] -= tz;
 834
 835             /**************************
 836              * CALCULATE INTERACTIONS *
 837              **************************/
 838
 839             /* COULOMB ELECTROSTATICS */
 840             velec            = qq10*rinv10;
 841             felec            = velec*rinvsq10;
 842
 843             fscal            = felec;
 844
 845             /* Calculate temporary vectorial force */
 846             tx               = fscal*dx10;
 847             ty               = fscal*dy10;
 848             tz               = fscal*dz10;
 849
 850             /* Update vectorial force */
 851             fix1            += tx;
 852             fiy1            += ty;
 853             fiz1            += tz;
 854             f[j_coord_offset+DIM*0+XX] -= tx;
 855             f[j_coord_offset+DIM*0+YY] -= ty;
 856             f[j_coord_offset+DIM*0+ZZ] -= tz;
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             /* COULOMB ELECTROSTATICS */
 863             velec            = qq11*rinv11;
 864             felec            = velec*rinvsq11;
 865
 866             fscal            = felec;
 867
 868             /* Calculate temporary vectorial force */
 869             tx               = fscal*dx11;
 870             ty               = fscal*dy11;
 871             tz               = fscal*dz11;
 872
 873             /* Update vectorial force */
 874             fix1            += tx;
 875             fiy1            += ty;
 876             fiz1            += tz;
 877             f[j_coord_offset+DIM*1+XX] -= tx;
 878             f[j_coord_offset+DIM*1+YY] -= ty;
 879             f[j_coord_offset+DIM*1+ZZ] -= tz;
 880
 881             /**************************
 882              * CALCULATE INTERACTIONS *
 883              **************************/
 884
 885             /* COULOMB ELECTROSTATICS */
 886             velec            = qq12*rinv12;
 887             felec            = velec*rinvsq12;
 888
 889             fscal            = felec;
 890
 891             /* Calculate temporary vectorial force */
 892             tx               = fscal*dx12;
 893             ty               = fscal*dy12;
 894             tz               = fscal*dz12;
 895
 896             /* Update vectorial force */
 897             fix1            += tx;
 898             fiy1            += ty;
 899             fiz1            += tz;
 900             f[j_coord_offset+DIM*2+XX] -= tx;
 901             f[j_coord_offset+DIM*2+YY] -= ty;
 902             f[j_coord_offset+DIM*2+ZZ] -= tz;
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             /* COULOMB ELECTROSTATICS */
 909             velec            = qq20*rinv20;
 910             felec            = velec*rinvsq20;
 911
 912             fscal            = felec;
 913
 914             /* Calculate temporary vectorial force */
 915             tx               = fscal*dx20;
 916             ty               = fscal*dy20;
 917             tz               = fscal*dz20;
 918
 919             /* Update vectorial force */
 920             fix2            += tx;
 921             fiy2            += ty;
 922             fiz2            += tz;
 923             f[j_coord_offset+DIM*0+XX] -= tx;
 924             f[j_coord_offset+DIM*0+YY] -= ty;
 925             f[j_coord_offset+DIM*0+ZZ] -= tz;
 926
 927             /**************************
 928              * CALCULATE INTERACTIONS *
 929              **************************/
 930
 931             /* COULOMB ELECTROSTATICS */
 932             velec            = qq21*rinv21;
 933             felec            = velec*rinvsq21;
 934
 935             fscal            = felec;
 936
 937             /* Calculate temporary vectorial force */
 938             tx               = fscal*dx21;
 939             ty               = fscal*dy21;
 940             tz               = fscal*dz21;
 941
 942             /* Update vectorial force */
 943             fix2            += tx;
 944             fiy2            += ty;
 945             fiz2            += tz;
 946             f[j_coord_offset+DIM*1+XX] -= tx;
 947             f[j_coord_offset+DIM*1+YY] -= ty;
 948             f[j_coord_offset+DIM*1+ZZ] -= tz;
 949
 950             /**************************
 951              * CALCULATE INTERACTIONS *
 952              **************************/
 953
 954             /* COULOMB ELECTROSTATICS */
 955             velec            = qq22*rinv22;
 956             felec            = velec*rinvsq22;
 957
 958             fscal            = felec;
 959
 960             /* Calculate temporary vectorial force */
 961             tx               = fscal*dx22;
 962             ty               = fscal*dy22;
 963             tz               = fscal*dz22;
 964
 965             /* Update vectorial force */
 966             fix2            += tx;
 967             fiy2            += ty;
 968             fiz2            += tz;
 969             f[j_coord_offset+DIM*2+XX] -= tx;
 970             f[j_coord_offset+DIM*2+YY] -= ty;
 971             f[j_coord_offset+DIM*2+ZZ] -= tz;
 972
 973             /* Inner loop uses 241 flops */
 974         }
 975         /* End of innermost loop */
 976
 977         tx = ty = tz = 0;
 978         f[i_coord_offset+DIM*0+XX] += fix0;
 979         f[i_coord_offset+DIM*0+YY] += fiy0;
 980         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 981         tx                         += fix0;
 982         ty                         += fiy0;
 983         tz                         += fiz0;
 984         f[i_coord_offset+DIM*1+XX] += fix1;
 985         f[i_coord_offset+DIM*1+YY] += fiy1;
 986         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 987         tx                         += fix1;
 988         ty                         += fiy1;
 989         tz                         += fiz1;
 990         f[i_coord_offset+DIM*2+XX] += fix2;
 991         f[i_coord_offset+DIM*2+YY] += fiy2;
 992         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 993         tx                         += fix2;
 994         ty                         += fiy2;
 995         tz                         += fiz2;
 996         fshift[i_shift_offset+XX]  += tx;
 997         fshift[i_shift_offset+YY]  += ty;
 998         fshift[i_shift_offset+ZZ]  += tz;
 999
1000         /* Increment number of inner iterations */
1001         inneriter                  += j_index_end - j_index_start;
1002
1003         /* Outer loop uses 30 flops */
1004     }
1005
1006     /* Increment number of outer iterations */
1007     outeriter        += nri;
1008
1009     /* Update outer/inner flops */
1010
1011     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*241);
1012 }