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