src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_c.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 "vec.h"
  47 #include "nrnb.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_c
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_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              vdwioffset3;
  79     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  80     int              vdwjidx0;
  81     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  82     int              vdwjidx1;
  83     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  84     int              vdwjidx2;
  85     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  86     int              vdwjidx3;
  87     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  88     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  89     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  90     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  91     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
  92     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  93     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  94     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
  95     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
  96     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
  97     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
  98     real             velec,felec,velecsum,facel,crf,krf,krf2;
  99     real             *charge;
 100     int              nvdwtype;
 101     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 102     int              *vdwtype;
 103     real             *vdwparam;
 104
 105     x                = xx[0];
 106     f                = ff[0];
 107
 108     nri              = nlist->nri;
 109     iinr             = nlist->iinr;
 110     jindex           = nlist->jindex;
 111     jjnr             = nlist->jjnr;
 112     shiftidx         = nlist->shift;
 113     gid              = nlist->gid;
 114     shiftvec         = fr->shift_vec[0];
 115     fshift           = fr->fshift[0];
 116     facel            = fr->epsfac;
 117     charge           = mdatoms->chargeA;
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     /* Setup water-specific parameters */
 123     inr              = nlist->iinr[0];
 124     iq1              = facel*charge[inr+1];
 125     iq2              = facel*charge[inr+2];
 126     iq3              = facel*charge[inr+3];
 127     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 128
 129     jq1              = charge[inr+1];
 130     jq2              = charge[inr+2];
 131     jq3              = charge[inr+3];
 132     vdwjidx0         = 2*vdwtype[inr+0];
 133     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 134     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 135     qq11             = iq1*jq1;
 136     qq12             = iq1*jq2;
 137     qq13             = iq1*jq3;
 138     qq21             = iq2*jq1;
 139     qq22             = iq2*jq2;
 140     qq23             = iq2*jq3;
 141     qq31             = iq3*jq1;
 142     qq32             = iq3*jq2;
 143     qq33             = iq3*jq3;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153         shX              = shiftvec[i_shift_offset+XX];
 154         shY              = shiftvec[i_shift_offset+YY];
 155         shZ              = shiftvec[i_shift_offset+ZZ];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 167         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 168         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 169         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 170         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 171         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 172         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 173         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 174         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 175         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 176         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 177         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 178
 179         fix0             = 0.0;
 180         fiy0             = 0.0;
 181         fiz0             = 0.0;
 182         fix1             = 0.0;
 183         fiy1             = 0.0;
 184         fiz1             = 0.0;
 185         fix2             = 0.0;
 186         fiy2             = 0.0;
 187         fiz2             = 0.0;
 188         fix3             = 0.0;
 189         fiy3             = 0.0;
 190         fiz3             = 0.0;
 191
 192         /* Reset potential sums */
 193         velecsum         = 0.0;
 194         vvdwsum          = 0.0;
 195
 196         /* Start inner kernel loop */
 197         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 198         {
 199             /* Get j neighbor index, and coordinate index */
 200             jnr              = jjnr[jidx];
 201             j_coord_offset   = DIM*jnr;
 202
 203             /* load j atom coordinates */
 204             jx0              = x[j_coord_offset+DIM*0+XX];
 205             jy0              = x[j_coord_offset+DIM*0+YY];
 206             jz0              = x[j_coord_offset+DIM*0+ZZ];
 207             jx1              = x[j_coord_offset+DIM*1+XX];
 208             jy1              = x[j_coord_offset+DIM*1+YY];
 209             jz1              = x[j_coord_offset+DIM*1+ZZ];
 210             jx2              = x[j_coord_offset+DIM*2+XX];
 211             jy2              = x[j_coord_offset+DIM*2+YY];
 212             jz2              = x[j_coord_offset+DIM*2+ZZ];
 213             jx3              = x[j_coord_offset+DIM*3+XX];
 214             jy3              = x[j_coord_offset+DIM*3+YY];
 215             jz3              = x[j_coord_offset+DIM*3+ZZ];
 216
 217             /* Calculate displacement vector */
 218             dx00             = ix0 - jx0;
 219             dy00             = iy0 - jy0;
 220             dz00             = iz0 - jz0;
 221             dx11             = ix1 - jx1;
 222             dy11             = iy1 - jy1;
 223             dz11             = iz1 - jz1;
 224             dx12             = ix1 - jx2;
 225             dy12             = iy1 - jy2;
 226             dz12             = iz1 - jz2;
 227             dx13             = ix1 - jx3;
 228             dy13             = iy1 - jy3;
 229             dz13             = iz1 - jz3;
 230             dx21             = ix2 - jx1;
 231             dy21             = iy2 - jy1;
 232             dz21             = iz2 - jz1;
 233             dx22             = ix2 - jx2;
 234             dy22             = iy2 - jy2;
 235             dz22             = iz2 - jz2;
 236             dx23             = ix2 - jx3;
 237             dy23             = iy2 - jy3;
 238             dz23             = iz2 - jz3;
 239             dx31             = ix3 - jx1;
 240             dy31             = iy3 - jy1;
 241             dz31             = iz3 - jz1;
 242             dx32             = ix3 - jx2;
 243             dy32             = iy3 - jy2;
 244             dz32             = iz3 - jz2;
 245             dx33             = ix3 - jx3;
 246             dy33             = iy3 - jy3;
 247             dz33             = iz3 - jz3;
 248
 249             /* Calculate squared distance and things based on it */
 250             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 251             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 252             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 253             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 254             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 255             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 256             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 257             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 258             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 259             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 260
 261             rinv11           = gmx_invsqrt(rsq11);
 262             rinv12           = gmx_invsqrt(rsq12);
 263             rinv13           = gmx_invsqrt(rsq13);
 264             rinv21           = gmx_invsqrt(rsq21);
 265             rinv22           = gmx_invsqrt(rsq22);
 266             rinv23           = gmx_invsqrt(rsq23);
 267             rinv31           = gmx_invsqrt(rsq31);
 268             rinv32           = gmx_invsqrt(rsq32);
 269             rinv33           = gmx_invsqrt(rsq33);
 270
 271             rinvsq00         = 1.0/rsq00;
 272             rinvsq11         = rinv11*rinv11;
 273             rinvsq12         = rinv12*rinv12;
 274             rinvsq13         = rinv13*rinv13;
 275             rinvsq21         = rinv21*rinv21;
 276             rinvsq22         = rinv22*rinv22;
 277             rinvsq23         = rinv23*rinv23;
 278             rinvsq31         = rinv31*rinv31;
 279             rinvsq32         = rinv32*rinv32;
 280             rinvsq33         = rinv33*rinv33;
 281
 282             /**************************
 283              * CALCULATE INTERACTIONS *
 284              **************************/
 285
 286             /* LENNARD-JONES DISPERSION/REPULSION */
 287
 288             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 289             vvdw6            = c6_00*rinvsix;
 290             vvdw12           = c12_00*rinvsix*rinvsix;
 291             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 292             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 293
 294             /* Update potential sums from outer loop */
 295             vvdwsum         += vvdw;
 296
 297             fscal            = fvdw;
 298
 299             /* Calculate temporary vectorial force */
 300             tx               = fscal*dx00;
 301             ty               = fscal*dy00;
 302             tz               = fscal*dz00;
 303
 304             /* Update vectorial force */
 305             fix0            += tx;
 306             fiy0            += ty;
 307             fiz0            += tz;
 308             f[j_coord_offset+DIM*0+XX] -= tx;
 309             f[j_coord_offset+DIM*0+YY] -= ty;
 310             f[j_coord_offset+DIM*0+ZZ] -= tz;
 311
 312             /**************************
 313              * CALCULATE INTERACTIONS *
 314              **************************/
 315
 316             /* COULOMB ELECTROSTATICS */
 317             velec            = qq11*rinv11;
 318             felec            = velec*rinvsq11;
 319
 320             /* Update potential sums from outer loop */
 321             velecsum        += velec;
 322
 323             fscal            = felec;
 324
 325             /* Calculate temporary vectorial force */
 326             tx               = fscal*dx11;
 327             ty               = fscal*dy11;
 328             tz               = fscal*dz11;
 329
 330             /* Update vectorial force */
 331             fix1            += tx;
 332             fiy1            += ty;
 333             fiz1            += tz;
 334             f[j_coord_offset+DIM*1+XX] -= tx;
 335             f[j_coord_offset+DIM*1+YY] -= ty;
 336             f[j_coord_offset+DIM*1+ZZ] -= tz;
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             /* COULOMB ELECTROSTATICS */
 343             velec            = qq12*rinv12;
 344             felec            = velec*rinvsq12;
 345
 346             /* Update potential sums from outer loop */
 347             velecsum        += velec;
 348
 349             fscal            = felec;
 350
 351             /* Calculate temporary vectorial force */
 352             tx               = fscal*dx12;
 353             ty               = fscal*dy12;
 354             tz               = fscal*dz12;
 355
 356             /* Update vectorial force */
 357             fix1            += tx;
 358             fiy1            += ty;
 359             fiz1            += tz;
 360             f[j_coord_offset+DIM*2+XX] -= tx;
 361             f[j_coord_offset+DIM*2+YY] -= ty;
 362             f[j_coord_offset+DIM*2+ZZ] -= tz;
 363
 364             /**************************
 365              * CALCULATE INTERACTIONS *
 366              **************************/
 367
 368             /* COULOMB ELECTROSTATICS */
 369             velec            = qq13*rinv13;
 370             felec            = velec*rinvsq13;
 371
 372             /* Update potential sums from outer loop */
 373             velecsum        += velec;
 374
 375             fscal            = felec;
 376
 377             /* Calculate temporary vectorial force */
 378             tx               = fscal*dx13;
 379             ty               = fscal*dy13;
 380             tz               = fscal*dz13;
 381
 382             /* Update vectorial force */
 383             fix1            += tx;
 384             fiy1            += ty;
 385             fiz1            += tz;
 386             f[j_coord_offset+DIM*3+XX] -= tx;
 387             f[j_coord_offset+DIM*3+YY] -= ty;
 388             f[j_coord_offset+DIM*3+ZZ] -= tz;
 389
 390             /**************************
 391              * CALCULATE INTERACTIONS *
 392              **************************/
 393
 394             /* COULOMB ELECTROSTATICS */
 395             velec            = qq21*rinv21;
 396             felec            = velec*rinvsq21;
 397
 398             /* Update potential sums from outer loop */
 399             velecsum        += velec;
 400
 401             fscal            = felec;
 402
 403             /* Calculate temporary vectorial force */
 404             tx               = fscal*dx21;
 405             ty               = fscal*dy21;
 406             tz               = fscal*dz21;
 407
 408             /* Update vectorial force */
 409             fix2            += tx;
 410             fiy2            += ty;
 411             fiz2            += tz;
 412             f[j_coord_offset+DIM*1+XX] -= tx;
 413             f[j_coord_offset+DIM*1+YY] -= ty;
 414             f[j_coord_offset+DIM*1+ZZ] -= tz;
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             /* COULOMB ELECTROSTATICS */
 421             velec            = qq22*rinv22;
 422             felec            = velec*rinvsq22;
 423
 424             /* Update potential sums from outer loop */
 425             velecsum        += velec;
 426
 427             fscal            = felec;
 428
 429             /* Calculate temporary vectorial force */
 430             tx               = fscal*dx22;
 431             ty               = fscal*dy22;
 432             tz               = fscal*dz22;
 433
 434             /* Update vectorial force */
 435             fix2            += tx;
 436             fiy2            += ty;
 437             fiz2            += tz;
 438             f[j_coord_offset+DIM*2+XX] -= tx;
 439             f[j_coord_offset+DIM*2+YY] -= ty;
 440             f[j_coord_offset+DIM*2+ZZ] -= tz;
 441
 442             /**************************
 443              * CALCULATE INTERACTIONS *
 444              **************************/
 445
 446             /* COULOMB ELECTROSTATICS */
 447             velec            = qq23*rinv23;
 448             felec            = velec*rinvsq23;
 449
 450             /* Update potential sums from outer loop */
 451             velecsum        += velec;
 452
 453             fscal            = felec;
 454
 455             /* Calculate temporary vectorial force */
 456             tx               = fscal*dx23;
 457             ty               = fscal*dy23;
 458             tz               = fscal*dz23;
 459
 460             /* Update vectorial force */
 461             fix2            += tx;
 462             fiy2            += ty;
 463             fiz2            += tz;
 464             f[j_coord_offset+DIM*3+XX] -= tx;
 465             f[j_coord_offset+DIM*3+YY] -= ty;
 466             f[j_coord_offset+DIM*3+ZZ] -= tz;
 467
 468             /**************************
 469              * CALCULATE INTERACTIONS *
 470              **************************/
 471
 472             /* COULOMB ELECTROSTATICS */
 473             velec            = qq31*rinv31;
 474             felec            = velec*rinvsq31;
 475
 476             /* Update potential sums from outer loop */
 477             velecsum        += velec;
 478
 479             fscal            = felec;
 480
 481             /* Calculate temporary vectorial force */
 482             tx               = fscal*dx31;
 483             ty               = fscal*dy31;
 484             tz               = fscal*dz31;
 485
 486             /* Update vectorial force */
 487             fix3            += tx;
 488             fiy3            += ty;
 489             fiz3            += tz;
 490             f[j_coord_offset+DIM*1+XX] -= tx;
 491             f[j_coord_offset+DIM*1+YY] -= ty;
 492             f[j_coord_offset+DIM*1+ZZ] -= tz;
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             /* COULOMB ELECTROSTATICS */
 499             velec            = qq32*rinv32;
 500             felec            = velec*rinvsq32;
 501
 502             /* Update potential sums from outer loop */
 503             velecsum        += velec;
 504
 505             fscal            = felec;
 506
 507             /* Calculate temporary vectorial force */
 508             tx               = fscal*dx32;
 509             ty               = fscal*dy32;
 510             tz               = fscal*dz32;
 511
 512             /* Update vectorial force */
 513             fix3            += tx;
 514             fiy3            += ty;
 515             fiz3            += tz;
 516             f[j_coord_offset+DIM*2+XX] -= tx;
 517             f[j_coord_offset+DIM*2+YY] -= ty;
 518             f[j_coord_offset+DIM*2+ZZ] -= tz;
 519
 520             /**************************
 521              * CALCULATE INTERACTIONS *
 522              **************************/
 523
 524             /* COULOMB ELECTROSTATICS */
 525             velec            = qq33*rinv33;
 526             felec            = velec*rinvsq33;
 527
 528             /* Update potential sums from outer loop */
 529             velecsum        += velec;
 530
 531             fscal            = felec;
 532
 533             /* Calculate temporary vectorial force */
 534             tx               = fscal*dx33;
 535             ty               = fscal*dy33;
 536             tz               = fscal*dz33;
 537
 538             /* Update vectorial force */
 539             fix3            += tx;
 540             fiy3            += ty;
 541             fiz3            += tz;
 542             f[j_coord_offset+DIM*3+XX] -= tx;
 543             f[j_coord_offset+DIM*3+YY] -= ty;
 544             f[j_coord_offset+DIM*3+ZZ] -= tz;
 545
 546             /* Inner loop uses 275 flops */
 547         }
 548         /* End of innermost loop */
 549
 550         tx = ty = tz = 0;
 551         f[i_coord_offset+DIM*0+XX] += fix0;
 552         f[i_coord_offset+DIM*0+YY] += fiy0;
 553         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 554         tx                         += fix0;
 555         ty                         += fiy0;
 556         tz                         += fiz0;
 557         f[i_coord_offset+DIM*1+XX] += fix1;
 558         f[i_coord_offset+DIM*1+YY] += fiy1;
 559         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 560         tx                         += fix1;
 561         ty                         += fiy1;
 562         tz                         += fiz1;
 563         f[i_coord_offset+DIM*2+XX] += fix2;
 564         f[i_coord_offset+DIM*2+YY] += fiy2;
 565         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 566         tx                         += fix2;
 567         ty                         += fiy2;
 568         tz                         += fiz2;
 569         f[i_coord_offset+DIM*3+XX] += fix3;
 570         f[i_coord_offset+DIM*3+YY] += fiy3;
 571         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 572         tx                         += fix3;
 573         ty                         += fiy3;
 574         tz                         += fiz3;
 575         fshift[i_shift_offset+XX]  += tx;
 576         fshift[i_shift_offset+YY]  += ty;
 577         fshift[i_shift_offset+ZZ]  += tz;
 578
 579         ggid                        = gid[iidx];
 580         /* Update potential energies */
 581         kernel_data->energygrp_elec[ggid] += velecsum;
 582         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 583
 584         /* Increment number of inner iterations */
 585         inneriter                  += j_index_end - j_index_start;
 586
 587         /* Outer loop uses 41 flops */
 588     }
 589
 590     /* Increment number of outer iterations */
 591     outeriter        += nri;
 592
 593     /* Update outer/inner flops */
 594
 595     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*275);
 596 }
 597 /*
 598  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_c
 599  * Electrostatics interaction: Coulomb
 600  * VdW interaction:            LennardJones
 601  * Geometry:                   Water4-Water4
 602  * Calculate force/pot:        Force
 603  */
 604 void
 605 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_c
 606                     (t_nblist                    * gmx_restrict       nlist,
 607                      rvec                        * gmx_restrict          xx,
 608                      rvec                        * gmx_restrict          ff,
 609                      t_forcerec                  * gmx_restrict          fr,
 610                      t_mdatoms                   * gmx_restrict     mdatoms,
 611                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 612                      t_nrnb                      * gmx_restrict        nrnb)
 613 {
 614     int              i_shift_offset,i_coord_offset,j_coord_offset;
 615     int              j_index_start,j_index_end;
 616     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 617     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 618     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 619     real             *shiftvec,*fshift,*x,*f;
 620     int              vdwioffset0;
 621     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 622     int              vdwioffset1;
 623     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 624     int              vdwioffset2;
 625     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 626     int              vdwioffset3;
 627     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 628     int              vdwjidx0;
 629     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 630     int              vdwjidx1;
 631     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 632     int              vdwjidx2;
 633     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 634     int              vdwjidx3;
 635     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 636     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 637     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 638     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 639     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 640     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 641     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 642     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 643     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 644     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 645     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 646     real             velec,felec,velecsum,facel,crf,krf,krf2;
 647     real             *charge;
 648     int              nvdwtype;
 649     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 650     int              *vdwtype;
 651     real             *vdwparam;
 652
 653     x                = xx[0];
 654     f                = ff[0];
 655
 656     nri              = nlist->nri;
 657     iinr             = nlist->iinr;
 658     jindex           = nlist->jindex;
 659     jjnr             = nlist->jjnr;
 660     shiftidx         = nlist->shift;
 661     gid              = nlist->gid;
 662     shiftvec         = fr->shift_vec[0];
 663     fshift           = fr->fshift[0];
 664     facel            = fr->epsfac;
 665     charge           = mdatoms->chargeA;
 666     nvdwtype         = fr->ntype;
 667     vdwparam         = fr->nbfp;
 668     vdwtype          = mdatoms->typeA;
 669
 670     /* Setup water-specific parameters */
 671     inr              = nlist->iinr[0];
 672     iq1              = facel*charge[inr+1];
 673     iq2              = facel*charge[inr+2];
 674     iq3              = facel*charge[inr+3];
 675     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 676
 677     jq1              = charge[inr+1];
 678     jq2              = charge[inr+2];
 679     jq3              = charge[inr+3];
 680     vdwjidx0         = 2*vdwtype[inr+0];
 681     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 682     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 683     qq11             = iq1*jq1;
 684     qq12             = iq1*jq2;
 685     qq13             = iq1*jq3;
 686     qq21             = iq2*jq1;
 687     qq22             = iq2*jq2;
 688     qq23             = iq2*jq3;
 689     qq31             = iq3*jq1;
 690     qq32             = iq3*jq2;
 691     qq33             = iq3*jq3;
 692
 693     outeriter        = 0;
 694     inneriter        = 0;
 695
 696     /* Start outer loop over neighborlists */
 697     for(iidx=0; iidx<nri; iidx++)
 698     {
 699         /* Load shift vector for this list */
 700         i_shift_offset   = DIM*shiftidx[iidx];
 701         shX              = shiftvec[i_shift_offset+XX];
 702         shY              = shiftvec[i_shift_offset+YY];
 703         shZ              = shiftvec[i_shift_offset+ZZ];
 704
 705         /* Load limits for loop over neighbors */
 706         j_index_start    = jindex[iidx];
 707         j_index_end      = jindex[iidx+1];
 708
 709         /* Get outer coordinate index */
 710         inr              = iinr[iidx];
 711         i_coord_offset   = DIM*inr;
 712
 713         /* Load i particle coords and add shift vector */
 714         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 715         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 716         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 717         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 718         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 719         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 720         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 721         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 722         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 723         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 724         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 725         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 726
 727         fix0             = 0.0;
 728         fiy0             = 0.0;
 729         fiz0             = 0.0;
 730         fix1             = 0.0;
 731         fiy1             = 0.0;
 732         fiz1             = 0.0;
 733         fix2             = 0.0;
 734         fiy2             = 0.0;
 735         fiz2             = 0.0;
 736         fix3             = 0.0;
 737         fiy3             = 0.0;
 738         fiz3             = 0.0;
 739
 740         /* Start inner kernel loop */
 741         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 742         {
 743             /* Get j neighbor index, and coordinate index */
 744             jnr              = jjnr[jidx];
 745             j_coord_offset   = DIM*jnr;
 746
 747             /* load j atom coordinates */
 748             jx0              = x[j_coord_offset+DIM*0+XX];
 749             jy0              = x[j_coord_offset+DIM*0+YY];
 750             jz0              = x[j_coord_offset+DIM*0+ZZ];
 751             jx1              = x[j_coord_offset+DIM*1+XX];
 752             jy1              = x[j_coord_offset+DIM*1+YY];
 753             jz1              = x[j_coord_offset+DIM*1+ZZ];
 754             jx2              = x[j_coord_offset+DIM*2+XX];
 755             jy2              = x[j_coord_offset+DIM*2+YY];
 756             jz2              = x[j_coord_offset+DIM*2+ZZ];
 757             jx3              = x[j_coord_offset+DIM*3+XX];
 758             jy3              = x[j_coord_offset+DIM*3+YY];
 759             jz3              = x[j_coord_offset+DIM*3+ZZ];
 760
 761             /* Calculate displacement vector */
 762             dx00             = ix0 - jx0;
 763             dy00             = iy0 - jy0;
 764             dz00             = iz0 - jz0;
 765             dx11             = ix1 - jx1;
 766             dy11             = iy1 - jy1;
 767             dz11             = iz1 - jz1;
 768             dx12             = ix1 - jx2;
 769             dy12             = iy1 - jy2;
 770             dz12             = iz1 - jz2;
 771             dx13             = ix1 - jx3;
 772             dy13             = iy1 - jy3;
 773             dz13             = iz1 - jz3;
 774             dx21             = ix2 - jx1;
 775             dy21             = iy2 - jy1;
 776             dz21             = iz2 - jz1;
 777             dx22             = ix2 - jx2;
 778             dy22             = iy2 - jy2;
 779             dz22             = iz2 - jz2;
 780             dx23             = ix2 - jx3;
 781             dy23             = iy2 - jy3;
 782             dz23             = iz2 - jz3;
 783             dx31             = ix3 - jx1;
 784             dy31             = iy3 - jy1;
 785             dz31             = iz3 - jz1;
 786             dx32             = ix3 - jx2;
 787             dy32             = iy3 - jy2;
 788             dz32             = iz3 - jz2;
 789             dx33             = ix3 - jx3;
 790             dy33             = iy3 - jy3;
 791             dz33             = iz3 - jz3;
 792
 793             /* Calculate squared distance and things based on it */
 794             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 795             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 796             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 797             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 798             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 799             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 800             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 801             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 802             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 803             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 804
 805             rinv11           = gmx_invsqrt(rsq11);
 806             rinv12           = gmx_invsqrt(rsq12);
 807             rinv13           = gmx_invsqrt(rsq13);
 808             rinv21           = gmx_invsqrt(rsq21);
 809             rinv22           = gmx_invsqrt(rsq22);
 810             rinv23           = gmx_invsqrt(rsq23);
 811             rinv31           = gmx_invsqrt(rsq31);
 812             rinv32           = gmx_invsqrt(rsq32);
 813             rinv33           = gmx_invsqrt(rsq33);
 814
 815             rinvsq00         = 1.0/rsq00;
 816             rinvsq11         = rinv11*rinv11;
 817             rinvsq12         = rinv12*rinv12;
 818             rinvsq13         = rinv13*rinv13;
 819             rinvsq21         = rinv21*rinv21;
 820             rinvsq22         = rinv22*rinv22;
 821             rinvsq23         = rinv23*rinv23;
 822             rinvsq31         = rinv31*rinv31;
 823             rinvsq32         = rinv32*rinv32;
 824             rinvsq33         = rinv33*rinv33;
 825
 826             /**************************
 827              * CALCULATE INTERACTIONS *
 828              **************************/
 829
 830             /* LENNARD-JONES DISPERSION/REPULSION */
 831
 832             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 833             fvdw             = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00;
 834
 835             fscal            = fvdw;
 836
 837             /* Calculate temporary vectorial force */
 838             tx               = fscal*dx00;
 839             ty               = fscal*dy00;
 840             tz               = fscal*dz00;
 841
 842             /* Update vectorial force */
 843             fix0            += tx;
 844             fiy0            += ty;
 845             fiz0            += tz;
 846             f[j_coord_offset+DIM*0+XX] -= tx;
 847             f[j_coord_offset+DIM*0+YY] -= ty;
 848             f[j_coord_offset+DIM*0+ZZ] -= tz;
 849
 850             /**************************
 851              * CALCULATE INTERACTIONS *
 852              **************************/
 853
 854             /* COULOMB ELECTROSTATICS */
 855             velec            = qq11*rinv11;
 856             felec            = velec*rinvsq11;
 857
 858             fscal            = felec;
 859
 860             /* Calculate temporary vectorial force */
 861             tx               = fscal*dx11;
 862             ty               = fscal*dy11;
 863             tz               = fscal*dz11;
 864
 865             /* Update vectorial force */
 866             fix1            += tx;
 867             fiy1            += ty;
 868             fiz1            += tz;
 869             f[j_coord_offset+DIM*1+XX] -= tx;
 870             f[j_coord_offset+DIM*1+YY] -= ty;
 871             f[j_coord_offset+DIM*1+ZZ] -= tz;
 872
 873             /**************************
 874              * CALCULATE INTERACTIONS *
 875              **************************/
 876
 877             /* COULOMB ELECTROSTATICS */
 878             velec            = qq12*rinv12;
 879             felec            = velec*rinvsq12;
 880
 881             fscal            = felec;
 882
 883             /* Calculate temporary vectorial force */
 884             tx               = fscal*dx12;
 885             ty               = fscal*dy12;
 886             tz               = fscal*dz12;
 887
 888             /* Update vectorial force */
 889             fix1            += tx;
 890             fiy1            += ty;
 891             fiz1            += tz;
 892             f[j_coord_offset+DIM*2+XX] -= tx;
 893             f[j_coord_offset+DIM*2+YY] -= ty;
 894             f[j_coord_offset+DIM*2+ZZ] -= tz;
 895
 896             /**************************
 897              * CALCULATE INTERACTIONS *
 898              **************************/
 899
 900             /* COULOMB ELECTROSTATICS */
 901             velec            = qq13*rinv13;
 902             felec            = velec*rinvsq13;
 903
 904             fscal            = felec;
 905
 906             /* Calculate temporary vectorial force */
 907             tx               = fscal*dx13;
 908             ty               = fscal*dy13;
 909             tz               = fscal*dz13;
 910
 911             /* Update vectorial force */
 912             fix1            += tx;
 913             fiy1            += ty;
 914             fiz1            += tz;
 915             f[j_coord_offset+DIM*3+XX] -= tx;
 916             f[j_coord_offset+DIM*3+YY] -= ty;
 917             f[j_coord_offset+DIM*3+ZZ] -= tz;
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             /* COULOMB ELECTROSTATICS */
 924             velec            = qq21*rinv21;
 925             felec            = velec*rinvsq21;
 926
 927             fscal            = felec;
 928
 929             /* Calculate temporary vectorial force */
 930             tx               = fscal*dx21;
 931             ty               = fscal*dy21;
 932             tz               = fscal*dz21;
 933
 934             /* Update vectorial force */
 935             fix2            += tx;
 936             fiy2            += ty;
 937             fiz2            += tz;
 938             f[j_coord_offset+DIM*1+XX] -= tx;
 939             f[j_coord_offset+DIM*1+YY] -= ty;
 940             f[j_coord_offset+DIM*1+ZZ] -= tz;
 941
 942             /**************************
 943              * CALCULATE INTERACTIONS *
 944              **************************/
 945
 946             /* COULOMB ELECTROSTATICS */
 947             velec            = qq22*rinv22;
 948             felec            = velec*rinvsq22;
 949
 950             fscal            = felec;
 951
 952             /* Calculate temporary vectorial force */
 953             tx               = fscal*dx22;
 954             ty               = fscal*dy22;
 955             tz               = fscal*dz22;
 956
 957             /* Update vectorial force */
 958             fix2            += tx;
 959             fiy2            += ty;
 960             fiz2            += tz;
 961             f[j_coord_offset+DIM*2+XX] -= tx;
 962             f[j_coord_offset+DIM*2+YY] -= ty;
 963             f[j_coord_offset+DIM*2+ZZ] -= tz;
 964
 965             /**************************
 966              * CALCULATE INTERACTIONS *
 967              **************************/
 968
 969             /* COULOMB ELECTROSTATICS */
 970             velec            = qq23*rinv23;
 971             felec            = velec*rinvsq23;
 972
 973             fscal            = felec;
 974
 975             /* Calculate temporary vectorial force */
 976             tx               = fscal*dx23;
 977             ty               = fscal*dy23;
 978             tz               = fscal*dz23;
 979
 980             /* Update vectorial force */
 981             fix2            += tx;
 982             fiy2            += ty;
 983             fiz2            += tz;
 984             f[j_coord_offset+DIM*3+XX] -= tx;
 985             f[j_coord_offset+DIM*3+YY] -= ty;
 986             f[j_coord_offset+DIM*3+ZZ] -= tz;
 987
 988             /**************************
 989              * CALCULATE INTERACTIONS *
 990              **************************/
 991
 992             /* COULOMB ELECTROSTATICS */
 993             velec            = qq31*rinv31;
 994             felec            = velec*rinvsq31;
 995
 996             fscal            = felec;
 997
 998             /* Calculate temporary vectorial force */
 999             tx               = fscal*dx31;
1000             ty               = fscal*dy31;
1001             tz               = fscal*dz31;
1002
1003             /* Update vectorial force */
1004             fix3            += tx;
1005             fiy3            += ty;
1006             fiz3            += tz;
1007             f[j_coord_offset+DIM*1+XX] -= tx;
1008             f[j_coord_offset+DIM*1+YY] -= ty;
1009             f[j_coord_offset+DIM*1+ZZ] -= tz;
1010
1011             /**************************
1012              * CALCULATE INTERACTIONS *
1013              **************************/
1014
1015             /* COULOMB ELECTROSTATICS */
1016             velec            = qq32*rinv32;
1017             felec            = velec*rinvsq32;
1018
1019             fscal            = felec;
1020
1021             /* Calculate temporary vectorial force */
1022             tx               = fscal*dx32;
1023             ty               = fscal*dy32;
1024             tz               = fscal*dz32;
1025
1026             /* Update vectorial force */
1027             fix3            += tx;
1028             fiy3            += ty;
1029             fiz3            += tz;
1030             f[j_coord_offset+DIM*2+XX] -= tx;
1031             f[j_coord_offset+DIM*2+YY] -= ty;
1032             f[j_coord_offset+DIM*2+ZZ] -= tz;
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             /* COULOMB ELECTROSTATICS */
1039             velec            = qq33*rinv33;
1040             felec            = velec*rinvsq33;
1041
1042             fscal            = felec;
1043
1044             /* Calculate temporary vectorial force */
1045             tx               = fscal*dx33;
1046             ty               = fscal*dy33;
1047             tz               = fscal*dz33;
1048
1049             /* Update vectorial force */
1050             fix3            += tx;
1051             fiy3            += ty;
1052             fiz3            += tz;
1053             f[j_coord_offset+DIM*3+XX] -= tx;
1054             f[j_coord_offset+DIM*3+YY] -= ty;
1055             f[j_coord_offset+DIM*3+ZZ] -= tz;
1056
1057             /* Inner loop uses 261 flops */
1058         }
1059         /* End of innermost loop */
1060
1061         tx = ty = tz = 0;
1062         f[i_coord_offset+DIM*0+XX] += fix0;
1063         f[i_coord_offset+DIM*0+YY] += fiy0;
1064         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1065         tx                         += fix0;
1066         ty                         += fiy0;
1067         tz                         += fiz0;
1068         f[i_coord_offset+DIM*1+XX] += fix1;
1069         f[i_coord_offset+DIM*1+YY] += fiy1;
1070         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1071         tx                         += fix1;
1072         ty                         += fiy1;
1073         tz                         += fiz1;
1074         f[i_coord_offset+DIM*2+XX] += fix2;
1075         f[i_coord_offset+DIM*2+YY] += fiy2;
1076         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1077         tx                         += fix2;
1078         ty                         += fiy2;
1079         tz                         += fiz2;
1080         f[i_coord_offset+DIM*3+XX] += fix3;
1081         f[i_coord_offset+DIM*3+YY] += fiy3;
1082         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1083         tx                         += fix3;
1084         ty                         += fiy3;
1085         tz                         += fiz3;
1086         fshift[i_shift_offset+XX]  += tx;
1087         fshift[i_shift_offset+YY]  += ty;
1088         fshift[i_shift_offset+ZZ]  += tz;
1089
1090         /* Increment number of inner iterations */
1091         inneriter                  += j_index_end - j_index_start;
1092
1093         /* Outer loop uses 39 flops */
1094     }
1095
1096     /* Increment number of outer iterations */
1097     outeriter        += nri;
1098
1099     /* Update outer/inner flops */
1100
1101     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*261);
1102 }