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) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS c kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_c
  49  * Electrostatics interaction: Coulomb
  50  * VdW interaction:            LennardJones
  51  * Geometry:                   Water4-Water4
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_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              vdwioffset3;
  77     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  78     int              vdwjidx0;
  79     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  80     int              vdwjidx1;
  81     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  82     int              vdwjidx2;
  83     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  84     int              vdwjidx3;
  85     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  86     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  87     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  88     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  89     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
  90     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  91     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  92     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
  93     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
  94     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
  95     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
  96     real             velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     int              nvdwtype;
  99     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = fr->epsfac;
 115     charge           = mdatoms->chargeA;
 116     nvdwtype         = fr->ntype;
 117     vdwparam         = fr->nbfp;
 118     vdwtype          = mdatoms->typeA;
 119
 120     /* Setup water-specific parameters */
 121     inr              = nlist->iinr[0];
 122     iq1              = facel*charge[inr+1];
 123     iq2              = facel*charge[inr+2];
 124     iq3              = facel*charge[inr+3];
 125     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 126
 127     jq1              = charge[inr+1];
 128     jq2              = charge[inr+2];
 129     jq3              = charge[inr+3];
 130     vdwjidx0         = 2*vdwtype[inr+0];
 131     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 132     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 133     qq11             = iq1*jq1;
 134     qq12             = iq1*jq2;
 135     qq13             = iq1*jq3;
 136     qq21             = iq2*jq1;
 137     qq22             = iq2*jq2;
 138     qq23             = iq2*jq3;
 139     qq31             = iq3*jq1;
 140     qq32             = iq3*jq2;
 141     qq33             = iq3*jq3;
 142
 143     outeriter        = 0;
 144     inneriter        = 0;
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151         shX              = shiftvec[i_shift_offset+XX];
 152         shY              = shiftvec[i_shift_offset+YY];
 153         shZ              = shiftvec[i_shift_offset+ZZ];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 165         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 166         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 167         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 168         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 169         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 170         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 171         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 172         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 173         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 174         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 175         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 176
 177         fix0             = 0.0;
 178         fiy0             = 0.0;
 179         fiz0             = 0.0;
 180         fix1             = 0.0;
 181         fiy1             = 0.0;
 182         fiz1             = 0.0;
 183         fix2             = 0.0;
 184         fiy2             = 0.0;
 185         fiz2             = 0.0;
 186         fix3             = 0.0;
 187         fiy3             = 0.0;
 188         fiz3             = 0.0;
 189
 190         /* Reset potential sums */
 191         velecsum         = 0.0;
 192         vvdwsum          = 0.0;
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 196         {
 197             /* Get j neighbor index, and coordinate index */
 198             jnr              = jjnr[jidx];
 199             j_coord_offset   = DIM*jnr;
 200
 201             /* load j atom coordinates */
 202             jx0              = x[j_coord_offset+DIM*0+XX];
 203             jy0              = x[j_coord_offset+DIM*0+YY];
 204             jz0              = x[j_coord_offset+DIM*0+ZZ];
 205             jx1              = x[j_coord_offset+DIM*1+XX];
 206             jy1              = x[j_coord_offset+DIM*1+YY];
 207             jz1              = x[j_coord_offset+DIM*1+ZZ];
 208             jx2              = x[j_coord_offset+DIM*2+XX];
 209             jy2              = x[j_coord_offset+DIM*2+YY];
 210             jz2              = x[j_coord_offset+DIM*2+ZZ];
 211             jx3              = x[j_coord_offset+DIM*3+XX];
 212             jy3              = x[j_coord_offset+DIM*3+YY];
 213             jz3              = x[j_coord_offset+DIM*3+ZZ];
 214
 215             /* Calculate displacement vector */
 216             dx00             = ix0 - jx0;
 217             dy00             = iy0 - jy0;
 218             dz00             = iz0 - jz0;
 219             dx11             = ix1 - jx1;
 220             dy11             = iy1 - jy1;
 221             dz11             = iz1 - jz1;
 222             dx12             = ix1 - jx2;
 223             dy12             = iy1 - jy2;
 224             dz12             = iz1 - jz2;
 225             dx13             = ix1 - jx3;
 226             dy13             = iy1 - jy3;
 227             dz13             = iz1 - jz3;
 228             dx21             = ix2 - jx1;
 229             dy21             = iy2 - jy1;
 230             dz21             = iz2 - jz1;
 231             dx22             = ix2 - jx2;
 232             dy22             = iy2 - jy2;
 233             dz22             = iz2 - jz2;
 234             dx23             = ix2 - jx3;
 235             dy23             = iy2 - jy3;
 236             dz23             = iz2 - jz3;
 237             dx31             = ix3 - jx1;
 238             dy31             = iy3 - jy1;
 239             dz31             = iz3 - jz1;
 240             dx32             = ix3 - jx2;
 241             dy32             = iy3 - jy2;
 242             dz32             = iz3 - jz2;
 243             dx33             = ix3 - jx3;
 244             dy33             = iy3 - jy3;
 245             dz33             = iz3 - jz3;
 246
 247             /* Calculate squared distance and things based on it */
 248             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 249             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 250             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 251             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 252             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 253             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 254             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 255             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 256             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 257             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 258
 259             rinv11           = gmx_invsqrt(rsq11);
 260             rinv12           = gmx_invsqrt(rsq12);
 261             rinv13           = gmx_invsqrt(rsq13);
 262             rinv21           = gmx_invsqrt(rsq21);
 263             rinv22           = gmx_invsqrt(rsq22);
 264             rinv23           = gmx_invsqrt(rsq23);
 265             rinv31           = gmx_invsqrt(rsq31);
 266             rinv32           = gmx_invsqrt(rsq32);
 267             rinv33           = gmx_invsqrt(rsq33);
 268
 269             rinvsq00         = 1.0/rsq00;
 270             rinvsq11         = rinv11*rinv11;
 271             rinvsq12         = rinv12*rinv12;
 272             rinvsq13         = rinv13*rinv13;
 273             rinvsq21         = rinv21*rinv21;
 274             rinvsq22         = rinv22*rinv22;
 275             rinvsq23         = rinv23*rinv23;
 276             rinvsq31         = rinv31*rinv31;
 277             rinvsq32         = rinv32*rinv32;
 278             rinvsq33         = rinv33*rinv33;
 279
 280             /**************************
 281              * CALCULATE INTERACTIONS *
 282              **************************/
 283
 284             /* LENNARD-JONES DISPERSION/REPULSION */
 285
 286             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 287             vvdw6            = c6_00*rinvsix;
 288             vvdw12           = c12_00*rinvsix*rinvsix;
 289             vvdw             = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0);
 290             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 291
 292             /* Update potential sums from outer loop */
 293             vvdwsum         += vvdw;
 294
 295             fscal            = fvdw;
 296
 297             /* Calculate temporary vectorial force */
 298             tx               = fscal*dx00;
 299             ty               = fscal*dy00;
 300             tz               = fscal*dz00;
 301
 302             /* Update vectorial force */
 303             fix0            += tx;
 304             fiy0            += ty;
 305             fiz0            += tz;
 306             f[j_coord_offset+DIM*0+XX] -= tx;
 307             f[j_coord_offset+DIM*0+YY] -= ty;
 308             f[j_coord_offset+DIM*0+ZZ] -= tz;
 309
 310             /**************************
 311              * CALCULATE INTERACTIONS *
 312              **************************/
 313
 314             /* COULOMB ELECTROSTATICS */
 315             velec            = qq11*rinv11;
 316             felec            = velec*rinvsq11;
 317
 318             /* Update potential sums from outer loop */
 319             velecsum        += velec;
 320
 321             fscal            = felec;
 322
 323             /* Calculate temporary vectorial force */
 324             tx               = fscal*dx11;
 325             ty               = fscal*dy11;
 326             tz               = fscal*dz11;
 327
 328             /* Update vectorial force */
 329             fix1            += tx;
 330             fiy1            += ty;
 331             fiz1            += tz;
 332             f[j_coord_offset+DIM*1+XX] -= tx;
 333             f[j_coord_offset+DIM*1+YY] -= ty;
 334             f[j_coord_offset+DIM*1+ZZ] -= tz;
 335
 336             /**************************
 337              * CALCULATE INTERACTIONS *
 338              **************************/
 339
 340             /* COULOMB ELECTROSTATICS */
 341             velec            = qq12*rinv12;
 342             felec            = velec*rinvsq12;
 343
 344             /* Update potential sums from outer loop */
 345             velecsum        += velec;
 346
 347             fscal            = felec;
 348
 349             /* Calculate temporary vectorial force */
 350             tx               = fscal*dx12;
 351             ty               = fscal*dy12;
 352             tz               = fscal*dz12;
 353
 354             /* Update vectorial force */
 355             fix1            += tx;
 356             fiy1            += ty;
 357             fiz1            += tz;
 358             f[j_coord_offset+DIM*2+XX] -= tx;
 359             f[j_coord_offset+DIM*2+YY] -= ty;
 360             f[j_coord_offset+DIM*2+ZZ] -= tz;
 361
 362             /**************************
 363              * CALCULATE INTERACTIONS *
 364              **************************/
 365
 366             /* COULOMB ELECTROSTATICS */
 367             velec            = qq13*rinv13;
 368             felec            = velec*rinvsq13;
 369
 370             /* Update potential sums from outer loop */
 371             velecsum        += velec;
 372
 373             fscal            = felec;
 374
 375             /* Calculate temporary vectorial force */
 376             tx               = fscal*dx13;
 377             ty               = fscal*dy13;
 378             tz               = fscal*dz13;
 379
 380             /* Update vectorial force */
 381             fix1            += tx;
 382             fiy1            += ty;
 383             fiz1            += tz;
 384             f[j_coord_offset+DIM*3+XX] -= tx;
 385             f[j_coord_offset+DIM*3+YY] -= ty;
 386             f[j_coord_offset+DIM*3+ZZ] -= tz;
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             /* COULOMB ELECTROSTATICS */
 393             velec            = qq21*rinv21;
 394             felec            = velec*rinvsq21;
 395
 396             /* Update potential sums from outer loop */
 397             velecsum        += velec;
 398
 399             fscal            = felec;
 400
 401             /* Calculate temporary vectorial force */
 402             tx               = fscal*dx21;
 403             ty               = fscal*dy21;
 404             tz               = fscal*dz21;
 405
 406             /* Update vectorial force */
 407             fix2            += tx;
 408             fiy2            += ty;
 409             fiz2            += tz;
 410             f[j_coord_offset+DIM*1+XX] -= tx;
 411             f[j_coord_offset+DIM*1+YY] -= ty;
 412             f[j_coord_offset+DIM*1+ZZ] -= tz;
 413
 414             /**************************
 415              * CALCULATE INTERACTIONS *
 416              **************************/
 417
 418             /* COULOMB ELECTROSTATICS */
 419             velec            = qq22*rinv22;
 420             felec            = velec*rinvsq22;
 421
 422             /* Update potential sums from outer loop */
 423             velecsum        += velec;
 424
 425             fscal            = felec;
 426
 427             /* Calculate temporary vectorial force */
 428             tx               = fscal*dx22;
 429             ty               = fscal*dy22;
 430             tz               = fscal*dz22;
 431
 432             /* Update vectorial force */
 433             fix2            += tx;
 434             fiy2            += ty;
 435             fiz2            += tz;
 436             f[j_coord_offset+DIM*2+XX] -= tx;
 437             f[j_coord_offset+DIM*2+YY] -= ty;
 438             f[j_coord_offset+DIM*2+ZZ] -= tz;
 439
 440             /**************************
 441              * CALCULATE INTERACTIONS *
 442              **************************/
 443
 444             /* COULOMB ELECTROSTATICS */
 445             velec            = qq23*rinv23;
 446             felec            = velec*rinvsq23;
 447
 448             /* Update potential sums from outer loop */
 449             velecsum        += velec;
 450
 451             fscal            = felec;
 452
 453             /* Calculate temporary vectorial force */
 454             tx               = fscal*dx23;
 455             ty               = fscal*dy23;
 456             tz               = fscal*dz23;
 457
 458             /* Update vectorial force */
 459             fix2            += tx;
 460             fiy2            += ty;
 461             fiz2            += tz;
 462             f[j_coord_offset+DIM*3+XX] -= tx;
 463             f[j_coord_offset+DIM*3+YY] -= ty;
 464             f[j_coord_offset+DIM*3+ZZ] -= tz;
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             /* COULOMB ELECTROSTATICS */
 471             velec            = qq31*rinv31;
 472             felec            = velec*rinvsq31;
 473
 474             /* Update potential sums from outer loop */
 475             velecsum        += velec;
 476
 477             fscal            = felec;
 478
 479             /* Calculate temporary vectorial force */
 480             tx               = fscal*dx31;
 481             ty               = fscal*dy31;
 482             tz               = fscal*dz31;
 483
 484             /* Update vectorial force */
 485             fix3            += tx;
 486             fiy3            += ty;
 487             fiz3            += tz;
 488             f[j_coord_offset+DIM*1+XX] -= tx;
 489             f[j_coord_offset+DIM*1+YY] -= ty;
 490             f[j_coord_offset+DIM*1+ZZ] -= tz;
 491
 492             /**************************
 493              * CALCULATE INTERACTIONS *
 494              **************************/
 495
 496             /* COULOMB ELECTROSTATICS */
 497             velec            = qq32*rinv32;
 498             felec            = velec*rinvsq32;
 499
 500             /* Update potential sums from outer loop */
 501             velecsum        += velec;
 502
 503             fscal            = felec;
 504
 505             /* Calculate temporary vectorial force */
 506             tx               = fscal*dx32;
 507             ty               = fscal*dy32;
 508             tz               = fscal*dz32;
 509
 510             /* Update vectorial force */
 511             fix3            += tx;
 512             fiy3            += ty;
 513             fiz3            += tz;
 514             f[j_coord_offset+DIM*2+XX] -= tx;
 515             f[j_coord_offset+DIM*2+YY] -= ty;
 516             f[j_coord_offset+DIM*2+ZZ] -= tz;
 517
 518             /**************************
 519              * CALCULATE INTERACTIONS *
 520              **************************/
 521
 522             /* COULOMB ELECTROSTATICS */
 523             velec            = qq33*rinv33;
 524             felec            = velec*rinvsq33;
 525
 526             /* Update potential sums from outer loop */
 527             velecsum        += velec;
 528
 529             fscal            = felec;
 530
 531             /* Calculate temporary vectorial force */
 532             tx               = fscal*dx33;
 533             ty               = fscal*dy33;
 534             tz               = fscal*dz33;
 535
 536             /* Update vectorial force */
 537             fix3            += tx;
 538             fiy3            += ty;
 539             fiz3            += tz;
 540             f[j_coord_offset+DIM*3+XX] -= tx;
 541             f[j_coord_offset+DIM*3+YY] -= ty;
 542             f[j_coord_offset+DIM*3+ZZ] -= tz;
 543
 544             /* Inner loop uses 275 flops */
 545         }
 546         /* End of innermost loop */
 547
 548         tx = ty = tz = 0;
 549         f[i_coord_offset+DIM*0+XX] += fix0;
 550         f[i_coord_offset+DIM*0+YY] += fiy0;
 551         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 552         tx                         += fix0;
 553         ty                         += fiy0;
 554         tz                         += fiz0;
 555         f[i_coord_offset+DIM*1+XX] += fix1;
 556         f[i_coord_offset+DIM*1+YY] += fiy1;
 557         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 558         tx                         += fix1;
 559         ty                         += fiy1;
 560         tz                         += fiz1;
 561         f[i_coord_offset+DIM*2+XX] += fix2;
 562         f[i_coord_offset+DIM*2+YY] += fiy2;
 563         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 564         tx                         += fix2;
 565         ty                         += fiy2;
 566         tz                         += fiz2;
 567         f[i_coord_offset+DIM*3+XX] += fix3;
 568         f[i_coord_offset+DIM*3+YY] += fiy3;
 569         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 570         tx                         += fix3;
 571         ty                         += fiy3;
 572         tz                         += fiz3;
 573         fshift[i_shift_offset+XX]  += tx;
 574         fshift[i_shift_offset+YY]  += ty;
 575         fshift[i_shift_offset+ZZ]  += tz;
 576
 577         ggid                        = gid[iidx];
 578         /* Update potential energies */
 579         kernel_data->energygrp_elec[ggid] += velecsum;
 580         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 581
 582         /* Increment number of inner iterations */
 583         inneriter                  += j_index_end - j_index_start;
 584
 585         /* Outer loop uses 41 flops */
 586     }
 587
 588     /* Increment number of outer iterations */
 589     outeriter        += nri;
 590
 591     /* Update outer/inner flops */
 592
 593     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*275);
 594 }
 595 /*
 596  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_c
 597  * Electrostatics interaction: Coulomb
 598  * VdW interaction:            LennardJones
 599  * Geometry:                   Water4-Water4
 600  * Calculate force/pot:        Force
 601  */
 602 void
 603 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_c
 604                     (t_nblist                    * gmx_restrict       nlist,
 605                      rvec                        * gmx_restrict          xx,
 606                      rvec                        * gmx_restrict          ff,
 607                      t_forcerec                  * gmx_restrict          fr,
 608                      t_mdatoms                   * gmx_restrict     mdatoms,
 609                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 610                      t_nrnb                      * gmx_restrict        nrnb)
 611 {
 612     int              i_shift_offset,i_coord_offset,j_coord_offset;
 613     int              j_index_start,j_index_end;
 614     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 615     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 616     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 617     real             *shiftvec,*fshift,*x,*f;
 618     int              vdwioffset0;
 619     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 620     int              vdwioffset1;
 621     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 622     int              vdwioffset2;
 623     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 624     int              vdwioffset3;
 625     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 626     int              vdwjidx0;
 627     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 628     int              vdwjidx1;
 629     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 630     int              vdwjidx2;
 631     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 632     int              vdwjidx3;
 633     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 634     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 635     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 636     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 637     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 638     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 639     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 640     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 641     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 642     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 643     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 644     real             velec,felec,velecsum,facel,crf,krf,krf2;
 645     real             *charge;
 646     int              nvdwtype;
 647     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 648     int              *vdwtype;
 649     real             *vdwparam;
 650
 651     x                = xx[0];
 652     f                = ff[0];
 653
 654     nri              = nlist->nri;
 655     iinr             = nlist->iinr;
 656     jindex           = nlist->jindex;
 657     jjnr             = nlist->jjnr;
 658     shiftidx         = nlist->shift;
 659     gid              = nlist->gid;
 660     shiftvec         = fr->shift_vec[0];
 661     fshift           = fr->fshift[0];
 662     facel            = fr->epsfac;
 663     charge           = mdatoms->chargeA;
 664     nvdwtype         = fr->ntype;
 665     vdwparam         = fr->nbfp;
 666     vdwtype          = mdatoms->typeA;
 667
 668     /* Setup water-specific parameters */
 669     inr              = nlist->iinr[0];
 670     iq1              = facel*charge[inr+1];
 671     iq2              = facel*charge[inr+2];
 672     iq3              = facel*charge[inr+3];
 673     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 674
 675     jq1              = charge[inr+1];
 676     jq2              = charge[inr+2];
 677     jq3              = charge[inr+3];
 678     vdwjidx0         = 2*vdwtype[inr+0];
 679     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 680     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 681     qq11             = iq1*jq1;
 682     qq12             = iq1*jq2;
 683     qq13             = iq1*jq3;
 684     qq21             = iq2*jq1;
 685     qq22             = iq2*jq2;
 686     qq23             = iq2*jq3;
 687     qq31             = iq3*jq1;
 688     qq32             = iq3*jq2;
 689     qq33             = iq3*jq3;
 690
 691     outeriter        = 0;
 692     inneriter        = 0;
 693
 694     /* Start outer loop over neighborlists */
 695     for(iidx=0; iidx<nri; iidx++)
 696     {
 697         /* Load shift vector for this list */
 698         i_shift_offset   = DIM*shiftidx[iidx];
 699         shX              = shiftvec[i_shift_offset+XX];
 700         shY              = shiftvec[i_shift_offset+YY];
 701         shZ              = shiftvec[i_shift_offset+ZZ];
 702
 703         /* Load limits for loop over neighbors */
 704         j_index_start    = jindex[iidx];
 705         j_index_end      = jindex[iidx+1];
 706
 707         /* Get outer coordinate index */
 708         inr              = iinr[iidx];
 709         i_coord_offset   = DIM*inr;
 710
 711         /* Load i particle coords and add shift vector */
 712         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 713         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 714         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 715         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 716         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 717         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 718         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 719         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 720         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 721         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 722         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 723         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 724
 725         fix0             = 0.0;
 726         fiy0             = 0.0;
 727         fiz0             = 0.0;
 728         fix1             = 0.0;
 729         fiy1             = 0.0;
 730         fiz1             = 0.0;
 731         fix2             = 0.0;
 732         fiy2             = 0.0;
 733         fiz2             = 0.0;
 734         fix3             = 0.0;
 735         fiy3             = 0.0;
 736         fiz3             = 0.0;
 737
 738         /* Start inner kernel loop */
 739         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 740         {
 741             /* Get j neighbor index, and coordinate index */
 742             jnr              = jjnr[jidx];
 743             j_coord_offset   = DIM*jnr;
 744
 745             /* load j atom coordinates */
 746             jx0              = x[j_coord_offset+DIM*0+XX];
 747             jy0              = x[j_coord_offset+DIM*0+YY];
 748             jz0              = x[j_coord_offset+DIM*0+ZZ];
 749             jx1              = x[j_coord_offset+DIM*1+XX];
 750             jy1              = x[j_coord_offset+DIM*1+YY];
 751             jz1              = x[j_coord_offset+DIM*1+ZZ];
 752             jx2              = x[j_coord_offset+DIM*2+XX];
 753             jy2              = x[j_coord_offset+DIM*2+YY];
 754             jz2              = x[j_coord_offset+DIM*2+ZZ];
 755             jx3              = x[j_coord_offset+DIM*3+XX];
 756             jy3              = x[j_coord_offset+DIM*3+YY];
 757             jz3              = x[j_coord_offset+DIM*3+ZZ];
 758
 759             /* Calculate displacement vector */
 760             dx00             = ix0 - jx0;
 761             dy00             = iy0 - jy0;
 762             dz00             = iz0 - jz0;
 763             dx11             = ix1 - jx1;
 764             dy11             = iy1 - jy1;
 765             dz11             = iz1 - jz1;
 766             dx12             = ix1 - jx2;
 767             dy12             = iy1 - jy2;
 768             dz12             = iz1 - jz2;
 769             dx13             = ix1 - jx3;
 770             dy13             = iy1 - jy3;
 771             dz13             = iz1 - jz3;
 772             dx21             = ix2 - jx1;
 773             dy21             = iy2 - jy1;
 774             dz21             = iz2 - jz1;
 775             dx22             = ix2 - jx2;
 776             dy22             = iy2 - jy2;
 777             dz22             = iz2 - jz2;
 778             dx23             = ix2 - jx3;
 779             dy23             = iy2 - jy3;
 780             dz23             = iz2 - jz3;
 781             dx31             = ix3 - jx1;
 782             dy31             = iy3 - jy1;
 783             dz31             = iz3 - jz1;
 784             dx32             = ix3 - jx2;
 785             dy32             = iy3 - jy2;
 786             dz32             = iz3 - jz2;
 787             dx33             = ix3 - jx3;
 788             dy33             = iy3 - jy3;
 789             dz33             = iz3 - jz3;
 790
 791             /* Calculate squared distance and things based on it */
 792             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 793             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 794             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 795             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 796             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 797             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 798             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 799             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 800             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 801             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 802
 803             rinv11           = gmx_invsqrt(rsq11);
 804             rinv12           = gmx_invsqrt(rsq12);
 805             rinv13           = gmx_invsqrt(rsq13);
 806             rinv21           = gmx_invsqrt(rsq21);
 807             rinv22           = gmx_invsqrt(rsq22);
 808             rinv23           = gmx_invsqrt(rsq23);
 809             rinv31           = gmx_invsqrt(rsq31);
 810             rinv32           = gmx_invsqrt(rsq32);
 811             rinv33           = gmx_invsqrt(rsq33);
 812
 813             rinvsq00         = 1.0/rsq00;
 814             rinvsq11         = rinv11*rinv11;
 815             rinvsq12         = rinv12*rinv12;
 816             rinvsq13         = rinv13*rinv13;
 817             rinvsq21         = rinv21*rinv21;
 818             rinvsq22         = rinv22*rinv22;
 819             rinvsq23         = rinv23*rinv23;
 820             rinvsq31         = rinv31*rinv31;
 821             rinvsq32         = rinv32*rinv32;
 822             rinvsq33         = rinv33*rinv33;
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             /* LENNARD-JONES DISPERSION/REPULSION */
 829
 830             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 831             fvdw             = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00;
 832
 833             fscal            = fvdw;
 834
 835             /* Calculate temporary vectorial force */
 836             tx               = fscal*dx00;
 837             ty               = fscal*dy00;
 838             tz               = fscal*dz00;
 839
 840             /* Update vectorial force */
 841             fix0            += tx;
 842             fiy0            += ty;
 843             fiz0            += tz;
 844             f[j_coord_offset+DIM*0+XX] -= tx;
 845             f[j_coord_offset+DIM*0+YY] -= ty;
 846             f[j_coord_offset+DIM*0+ZZ] -= tz;
 847
 848             /**************************
 849              * CALCULATE INTERACTIONS *
 850              **************************/
 851
 852             /* COULOMB ELECTROSTATICS */
 853             velec            = qq11*rinv11;
 854             felec            = velec*rinvsq11;
 855
 856             fscal            = felec;
 857
 858             /* Calculate temporary vectorial force */
 859             tx               = fscal*dx11;
 860             ty               = fscal*dy11;
 861             tz               = fscal*dz11;
 862
 863             /* Update vectorial force */
 864             fix1            += tx;
 865             fiy1            += ty;
 866             fiz1            += tz;
 867             f[j_coord_offset+DIM*1+XX] -= tx;
 868             f[j_coord_offset+DIM*1+YY] -= ty;
 869             f[j_coord_offset+DIM*1+ZZ] -= tz;
 870
 871             /**************************
 872              * CALCULATE INTERACTIONS *
 873              **************************/
 874
 875             /* COULOMB ELECTROSTATICS */
 876             velec            = qq12*rinv12;
 877             felec            = velec*rinvsq12;
 878
 879             fscal            = felec;
 880
 881             /* Calculate temporary vectorial force */
 882             tx               = fscal*dx12;
 883             ty               = fscal*dy12;
 884             tz               = fscal*dz12;
 885
 886             /* Update vectorial force */
 887             fix1            += tx;
 888             fiy1            += ty;
 889             fiz1            += tz;
 890             f[j_coord_offset+DIM*2+XX] -= tx;
 891             f[j_coord_offset+DIM*2+YY] -= ty;
 892             f[j_coord_offset+DIM*2+ZZ] -= tz;
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             /* COULOMB ELECTROSTATICS */
 899             velec            = qq13*rinv13;
 900             felec            = velec*rinvsq13;
 901
 902             fscal            = felec;
 903
 904             /* Calculate temporary vectorial force */
 905             tx               = fscal*dx13;
 906             ty               = fscal*dy13;
 907             tz               = fscal*dz13;
 908
 909             /* Update vectorial force */
 910             fix1            += tx;
 911             fiy1            += ty;
 912             fiz1            += tz;
 913             f[j_coord_offset+DIM*3+XX] -= tx;
 914             f[j_coord_offset+DIM*3+YY] -= ty;
 915             f[j_coord_offset+DIM*3+ZZ] -= tz;
 916
 917             /**************************
 918              * CALCULATE INTERACTIONS *
 919              **************************/
 920
 921             /* COULOMB ELECTROSTATICS */
 922             velec            = qq21*rinv21;
 923             felec            = velec*rinvsq21;
 924
 925             fscal            = felec;
 926
 927             /* Calculate temporary vectorial force */
 928             tx               = fscal*dx21;
 929             ty               = fscal*dy21;
 930             tz               = fscal*dz21;
 931
 932             /* Update vectorial force */
 933             fix2            += tx;
 934             fiy2            += ty;
 935             fiz2            += tz;
 936             f[j_coord_offset+DIM*1+XX] -= tx;
 937             f[j_coord_offset+DIM*1+YY] -= ty;
 938             f[j_coord_offset+DIM*1+ZZ] -= tz;
 939
 940             /**************************
 941              * CALCULATE INTERACTIONS *
 942              **************************/
 943
 944             /* COULOMB ELECTROSTATICS */
 945             velec            = qq22*rinv22;
 946             felec            = velec*rinvsq22;
 947
 948             fscal            = felec;
 949
 950             /* Calculate temporary vectorial force */
 951             tx               = fscal*dx22;
 952             ty               = fscal*dy22;
 953             tz               = fscal*dz22;
 954
 955             /* Update vectorial force */
 956             fix2            += tx;
 957             fiy2            += ty;
 958             fiz2            += tz;
 959             f[j_coord_offset+DIM*2+XX] -= tx;
 960             f[j_coord_offset+DIM*2+YY] -= ty;
 961             f[j_coord_offset+DIM*2+ZZ] -= tz;
 962
 963             /**************************
 964              * CALCULATE INTERACTIONS *
 965              **************************/
 966
 967             /* COULOMB ELECTROSTATICS */
 968             velec            = qq23*rinv23;
 969             felec            = velec*rinvsq23;
 970
 971             fscal            = felec;
 972
 973             /* Calculate temporary vectorial force */
 974             tx               = fscal*dx23;
 975             ty               = fscal*dy23;
 976             tz               = fscal*dz23;
 977
 978             /* Update vectorial force */
 979             fix2            += tx;
 980             fiy2            += ty;
 981             fiz2            += tz;
 982             f[j_coord_offset+DIM*3+XX] -= tx;
 983             f[j_coord_offset+DIM*3+YY] -= ty;
 984             f[j_coord_offset+DIM*3+ZZ] -= tz;
 985
 986             /**************************
 987              * CALCULATE INTERACTIONS *
 988              **************************/
 989
 990             /* COULOMB ELECTROSTATICS */
 991             velec            = qq31*rinv31;
 992             felec            = velec*rinvsq31;
 993
 994             fscal            = felec;
 995
 996             /* Calculate temporary vectorial force */
 997             tx               = fscal*dx31;
 998             ty               = fscal*dy31;
 999             tz               = fscal*dz31;
1000
1001             /* Update vectorial force */
1002             fix3            += tx;
1003             fiy3            += ty;
1004             fiz3            += tz;
1005             f[j_coord_offset+DIM*1+XX] -= tx;
1006             f[j_coord_offset+DIM*1+YY] -= ty;
1007             f[j_coord_offset+DIM*1+ZZ] -= tz;
1008
1009             /**************************
1010              * CALCULATE INTERACTIONS *
1011              **************************/
1012
1013             /* COULOMB ELECTROSTATICS */
1014             velec            = qq32*rinv32;
1015             felec            = velec*rinvsq32;
1016
1017             fscal            = felec;
1018
1019             /* Calculate temporary vectorial force */
1020             tx               = fscal*dx32;
1021             ty               = fscal*dy32;
1022             tz               = fscal*dz32;
1023
1024             /* Update vectorial force */
1025             fix3            += tx;
1026             fiy3            += ty;
1027             fiz3            += tz;
1028             f[j_coord_offset+DIM*2+XX] -= tx;
1029             f[j_coord_offset+DIM*2+YY] -= ty;
1030             f[j_coord_offset+DIM*2+ZZ] -= tz;
1031
1032             /**************************
1033              * CALCULATE INTERACTIONS *
1034              **************************/
1035
1036             /* COULOMB ELECTROSTATICS */
1037             velec            = qq33*rinv33;
1038             felec            = velec*rinvsq33;
1039
1040             fscal            = felec;
1041
1042             /* Calculate temporary vectorial force */
1043             tx               = fscal*dx33;
1044             ty               = fscal*dy33;
1045             tz               = fscal*dz33;
1046
1047             /* Update vectorial force */
1048             fix3            += tx;
1049             fiy3            += ty;
1050             fiz3            += tz;
1051             f[j_coord_offset+DIM*3+XX] -= tx;
1052             f[j_coord_offset+DIM*3+YY] -= ty;
1053             f[j_coord_offset+DIM*3+ZZ] -= tz;
1054
1055             /* Inner loop uses 261 flops */
1056         }
1057         /* End of innermost loop */
1058
1059         tx = ty = tz = 0;
1060         f[i_coord_offset+DIM*0+XX] += fix0;
1061         f[i_coord_offset+DIM*0+YY] += fiy0;
1062         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1063         tx                         += fix0;
1064         ty                         += fiy0;
1065         tz                         += fiz0;
1066         f[i_coord_offset+DIM*1+XX] += fix1;
1067         f[i_coord_offset+DIM*1+YY] += fiy1;
1068         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1069         tx                         += fix1;
1070         ty                         += fiy1;
1071         tz                         += fiz1;
1072         f[i_coord_offset+DIM*2+XX] += fix2;
1073         f[i_coord_offset+DIM*2+YY] += fiy2;
1074         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1075         tx                         += fix2;
1076         ty                         += fiy2;
1077         tz                         += fiz2;
1078         f[i_coord_offset+DIM*3+XX] += fix3;
1079         f[i_coord_offset+DIM*3+YY] += fiy3;
1080         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1081         tx                         += fix3;
1082         ty                         += fiy3;
1083         tz                         += fiz3;
1084         fshift[i_shift_offset+XX]  += tx;
1085         fshift[i_shift_offset+YY]  += ty;
1086         fshift[i_shift_offset+ZZ]  += tz;
1087
1088         /* Increment number of inner iterations */
1089         inneriter                  += j_index_end - j_index_start;
1090
1091         /* Outer loop uses 39 flops */
1092     }
1093
1094     /* Increment number of outer iterations */
1095     outeriter        += nri;
1096
1097     /* Update outer/inner flops */
1098
1099     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*261);
1100 }