src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_c.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS c kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 /*
  48  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_c
  49  * Electrostatics interaction: ReactionField
  50  * VdW interaction:            LennardJones
  51  * Geometry:                   Water3-Water3
  52  * Calculate force/pot:        PotentialAndForce
  53  */
  54 void
  55 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_c
  56                     (t_nblist                    * gmx_restrict       nlist,
  57                      rvec                        * gmx_restrict          xx,
  58                      rvec                        * gmx_restrict          ff,
  59                      t_forcerec                  * gmx_restrict          fr,
  60                      t_mdatoms                   * gmx_restrict     mdatoms,
  61                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  62                      t_nrnb                      * gmx_restrict        nrnb)
  63 {
  64     int              i_shift_offset,i_coord_offset,j_coord_offset;
  65     int              j_index_start,j_index_end;
  66     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  67     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  68     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  69     real             *shiftvec,*fshift,*x,*f;
  70     int              vdwioffset0;
  71     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwioffset1;
  73     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  74     int              vdwioffset2;
  75     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  76     int              vdwjidx0;
  77     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  78     int              vdwjidx1;
  79     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  80     int              vdwjidx2;
  81     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  82     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  83     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
  84     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
  85     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  86     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  87     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  88     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  89     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  90     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  91     real             velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97
  98     x                = xx[0];
  99     f                = ff[0];
 100
 101     nri              = nlist->nri;
 102     iinr             = nlist->iinr;
 103     jindex           = nlist->jindex;
 104     jjnr             = nlist->jjnr;
 105     shiftidx         = nlist->shift;
 106     gid              = nlist->gid;
 107     shiftvec         = fr->shift_vec[0];
 108     fshift           = fr->fshift[0];
 109     facel            = fr->epsfac;
 110     charge           = mdatoms->chargeA;
 111     krf              = fr->ic->k_rf;
 112     krf2             = krf*2.0;
 113     crf              = fr->ic->c_rf;
 114     nvdwtype         = fr->ntype;
 115     vdwparam         = fr->nbfp;
 116     vdwtype          = mdatoms->typeA;
 117
 118     /* Setup water-specific parameters */
 119     inr              = nlist->iinr[0];
 120     iq0              = facel*charge[inr+0];
 121     iq1              = facel*charge[inr+1];
 122     iq2              = facel*charge[inr+2];
 123     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 124
 125     jq0              = charge[inr+0];
 126     jq1              = charge[inr+1];
 127     jq2              = charge[inr+2];
 128     vdwjidx0         = 2*vdwtype[inr+0];
 129     qq00             = iq0*jq0;
 130     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 131     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 132     qq01             = iq0*jq1;
 133     qq02             = iq0*jq2;
 134     qq10             = iq1*jq0;
 135     qq11             = iq1*jq1;
 136     qq12             = iq1*jq2;
 137     qq20             = iq2*jq0;
 138     qq21             = iq2*jq1;
 139     qq22             = iq2*jq2;
 140
 141     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 142     rcutoff          = fr->rcoulomb;
 143     rcutoff2         = rcutoff*rcutoff;
 144
 145     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 146     rvdw             = fr->rvdw;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156         shX              = shiftvec[i_shift_offset+XX];
 157         shY              = shiftvec[i_shift_offset+YY];
 158         shZ              = shiftvec[i_shift_offset+ZZ];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 170         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 171         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 172         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 173         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 174         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 175         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 176         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 177         iz2              = shZ + x[i_coord_offset+DIM*2+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
 189         /* Reset potential sums */
 190         velecsum         = 0.0;
 191         vvdwsum          = 0.0;
 192
 193         /* Start inner kernel loop */
 194         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 195         {
 196             /* Get j neighbor index, and coordinate index */
 197             jnr              = jjnr[jidx];
 198             j_coord_offset   = DIM*jnr;
 199
 200             /* load j atom coordinates */
 201             jx0              = x[j_coord_offset+DIM*0+XX];
 202             jy0              = x[j_coord_offset+DIM*0+YY];
 203             jz0              = x[j_coord_offset+DIM*0+ZZ];
 204             jx1              = x[j_coord_offset+DIM*1+XX];
 205             jy1              = x[j_coord_offset+DIM*1+YY];
 206             jz1              = x[j_coord_offset+DIM*1+ZZ];
 207             jx2              = x[j_coord_offset+DIM*2+XX];
 208             jy2              = x[j_coord_offset+DIM*2+YY];
 209             jz2              = x[j_coord_offset+DIM*2+ZZ];
 210
 211             /* Calculate displacement vector */
 212             dx00             = ix0 - jx0;
 213             dy00             = iy0 - jy0;
 214             dz00             = iz0 - jz0;
 215             dx01             = ix0 - jx1;
 216             dy01             = iy0 - jy1;
 217             dz01             = iz0 - jz1;
 218             dx02             = ix0 - jx2;
 219             dy02             = iy0 - jy2;
 220             dz02             = iz0 - jz2;
 221             dx10             = ix1 - jx0;
 222             dy10             = iy1 - jy0;
 223             dz10             = iz1 - jz0;
 224             dx11             = ix1 - jx1;
 225             dy11             = iy1 - jy1;
 226             dz11             = iz1 - jz1;
 227             dx12             = ix1 - jx2;
 228             dy12             = iy1 - jy2;
 229             dz12             = iz1 - jz2;
 230             dx20             = ix2 - jx0;
 231             dy20             = iy2 - jy0;
 232             dz20             = iz2 - jz0;
 233             dx21             = ix2 - jx1;
 234             dy21             = iy2 - jy1;
 235             dz21             = iz2 - jz1;
 236             dx22             = ix2 - jx2;
 237             dy22             = iy2 - jy2;
 238             dz22             = iz2 - jz2;
 239
 240             /* Calculate squared distance and things based on it */
 241             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 242             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 243             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 244             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 245             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 246             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 247             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 248             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 249             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 250
 251             rinv00           = gmx_invsqrt(rsq00);
 252             rinv01           = gmx_invsqrt(rsq01);
 253             rinv02           = gmx_invsqrt(rsq02);
 254             rinv10           = gmx_invsqrt(rsq10);
 255             rinv11           = gmx_invsqrt(rsq11);
 256             rinv12           = gmx_invsqrt(rsq12);
 257             rinv20           = gmx_invsqrt(rsq20);
 258             rinv21           = gmx_invsqrt(rsq21);
 259             rinv22           = gmx_invsqrt(rsq22);
 260
 261             rinvsq00         = rinv00*rinv00;
 262             rinvsq01         = rinv01*rinv01;
 263             rinvsq02         = rinv02*rinv02;
 264             rinvsq10         = rinv10*rinv10;
 265             rinvsq11         = rinv11*rinv11;
 266             rinvsq12         = rinv12*rinv12;
 267             rinvsq20         = rinv20*rinv20;
 268             rinvsq21         = rinv21*rinv21;
 269             rinvsq22         = rinv22*rinv22;
 270
 271             /**************************
 272              * CALCULATE INTERACTIONS *
 273              **************************/
 274
 275             if (rsq00<rcutoff2)
 276             {
 277
 278             /* REACTION-FIELD ELECTROSTATICS */
 279             velec            = qq00*(rinv00+krf*rsq00-crf);
 280             felec            = qq00*(rinv00*rinvsq00-krf2);
 281
 282             /* LENNARD-JONES DISPERSION/REPULSION */
 283
 284             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 285             vvdw6            = c6_00*rinvsix;
 286             vvdw12           = c12_00*rinvsix*rinvsix;
 287             vvdw             = (vvdw12 - c12_00*sh_vdw_invrcut6*sh_vdw_invrcut6)*(1.0/12.0) - (vvdw6 - c6_00*sh_vdw_invrcut6)*(1.0/6.0);
 288             fvdw             = (vvdw12-vvdw6)*rinvsq00;
 289
 290             /* Update potential sums from outer loop */
 291             velecsum        += velec;
 292             vvdwsum         += vvdw;
 293
 294             fscal            = felec+fvdw;
 295
 296             /* Calculate temporary vectorial force */
 297             tx               = fscal*dx00;
 298             ty               = fscal*dy00;
 299             tz               = fscal*dz00;
 300
 301             /* Update vectorial force */
 302             fix0            += tx;
 303             fiy0            += ty;
 304             fiz0            += tz;
 305             f[j_coord_offset+DIM*0+XX] -= tx;
 306             f[j_coord_offset+DIM*0+YY] -= ty;
 307             f[j_coord_offset+DIM*0+ZZ] -= tz;
 308
 309             }
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             if (rsq01<rcutoff2)
 316             {
 317
 318             /* REACTION-FIELD ELECTROSTATICS */
 319             velec            = qq01*(rinv01+krf*rsq01-crf);
 320             felec            = qq01*(rinv01*rinvsq01-krf2);
 321
 322             /* Update potential sums from outer loop */
 323             velecsum        += velec;
 324
 325             fscal            = felec;
 326
 327             /* Calculate temporary vectorial force */
 328             tx               = fscal*dx01;
 329             ty               = fscal*dy01;
 330             tz               = fscal*dz01;
 331
 332             /* Update vectorial force */
 333             fix0            += tx;
 334             fiy0            += ty;
 335             fiz0            += tz;
 336             f[j_coord_offset+DIM*1+XX] -= tx;
 337             f[j_coord_offset+DIM*1+YY] -= ty;
 338             f[j_coord_offset+DIM*1+ZZ] -= tz;
 339
 340             }
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             if (rsq02<rcutoff2)
 347             {
 348
 349             /* REACTION-FIELD ELECTROSTATICS */
 350             velec            = qq02*(rinv02+krf*rsq02-crf);
 351             felec            = qq02*(rinv02*rinvsq02-krf2);
 352
 353             /* Update potential sums from outer loop */
 354             velecsum        += velec;
 355
 356             fscal            = felec;
 357
 358             /* Calculate temporary vectorial force */
 359             tx               = fscal*dx02;
 360             ty               = fscal*dy02;
 361             tz               = fscal*dz02;
 362
 363             /* Update vectorial force */
 364             fix0            += tx;
 365             fiy0            += ty;
 366             fiz0            += tz;
 367             f[j_coord_offset+DIM*2+XX] -= tx;
 368             f[j_coord_offset+DIM*2+YY] -= ty;
 369             f[j_coord_offset+DIM*2+ZZ] -= tz;
 370
 371             }
 372
 373             /**************************
 374              * CALCULATE INTERACTIONS *
 375              **************************/
 376
 377             if (rsq10<rcutoff2)
 378             {
 379
 380             /* REACTION-FIELD ELECTROSTATICS */
 381             velec            = qq10*(rinv10+krf*rsq10-crf);
 382             felec            = qq10*(rinv10*rinvsq10-krf2);
 383
 384             /* Update potential sums from outer loop */
 385             velecsum        += velec;
 386
 387             fscal            = felec;
 388
 389             /* Calculate temporary vectorial force */
 390             tx               = fscal*dx10;
 391             ty               = fscal*dy10;
 392             tz               = fscal*dz10;
 393
 394             /* Update vectorial force */
 395             fix1            += tx;
 396             fiy1            += ty;
 397             fiz1            += tz;
 398             f[j_coord_offset+DIM*0+XX] -= tx;
 399             f[j_coord_offset+DIM*0+YY] -= ty;
 400             f[j_coord_offset+DIM*0+ZZ] -= tz;
 401
 402             }
 403
 404             /**************************
 405              * CALCULATE INTERACTIONS *
 406              **************************/
 407
 408             if (rsq11<rcutoff2)
 409             {
 410
 411             /* REACTION-FIELD ELECTROSTATICS */
 412             velec            = qq11*(rinv11+krf*rsq11-crf);
 413             felec            = qq11*(rinv11*rinvsq11-krf2);
 414
 415             /* Update potential sums from outer loop */
 416             velecsum        += velec;
 417
 418             fscal            = felec;
 419
 420             /* Calculate temporary vectorial force */
 421             tx               = fscal*dx11;
 422             ty               = fscal*dy11;
 423             tz               = fscal*dz11;
 424
 425             /* Update vectorial force */
 426             fix1            += tx;
 427             fiy1            += ty;
 428             fiz1            += tz;
 429             f[j_coord_offset+DIM*1+XX] -= tx;
 430             f[j_coord_offset+DIM*1+YY] -= ty;
 431             f[j_coord_offset+DIM*1+ZZ] -= tz;
 432
 433             }
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             if (rsq12<rcutoff2)
 440             {
 441
 442             /* REACTION-FIELD ELECTROSTATICS */
 443             velec            = qq12*(rinv12+krf*rsq12-crf);
 444             felec            = qq12*(rinv12*rinvsq12-krf2);
 445
 446             /* Update potential sums from outer loop */
 447             velecsum        += velec;
 448
 449             fscal            = felec;
 450
 451             /* Calculate temporary vectorial force */
 452             tx               = fscal*dx12;
 453             ty               = fscal*dy12;
 454             tz               = fscal*dz12;
 455
 456             /* Update vectorial force */
 457             fix1            += tx;
 458             fiy1            += ty;
 459             fiz1            += tz;
 460             f[j_coord_offset+DIM*2+XX] -= tx;
 461             f[j_coord_offset+DIM*2+YY] -= ty;
 462             f[j_coord_offset+DIM*2+ZZ] -= tz;
 463
 464             }
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             if (rsq20<rcutoff2)
 471             {
 472
 473             /* REACTION-FIELD ELECTROSTATICS */
 474             velec            = qq20*(rinv20+krf*rsq20-crf);
 475             felec            = qq20*(rinv20*rinvsq20-krf2);
 476
 477             /* Update potential sums from outer loop */
 478             velecsum        += velec;
 479
 480             fscal            = felec;
 481
 482             /* Calculate temporary vectorial force */
 483             tx               = fscal*dx20;
 484             ty               = fscal*dy20;
 485             tz               = fscal*dz20;
 486
 487             /* Update vectorial force */
 488             fix2            += tx;
 489             fiy2            += ty;
 490             fiz2            += tz;
 491             f[j_coord_offset+DIM*0+XX] -= tx;
 492             f[j_coord_offset+DIM*0+YY] -= ty;
 493             f[j_coord_offset+DIM*0+ZZ] -= tz;
 494
 495             }
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             if (rsq21<rcutoff2)
 502             {
 503
 504             /* REACTION-FIELD ELECTROSTATICS */
 505             velec            = qq21*(rinv21+krf*rsq21-crf);
 506             felec            = qq21*(rinv21*rinvsq21-krf2);
 507
 508             /* Update potential sums from outer loop */
 509             velecsum        += velec;
 510
 511             fscal            = felec;
 512
 513             /* Calculate temporary vectorial force */
 514             tx               = fscal*dx21;
 515             ty               = fscal*dy21;
 516             tz               = fscal*dz21;
 517
 518             /* Update vectorial force */
 519             fix2            += tx;
 520             fiy2            += ty;
 521             fiz2            += tz;
 522             f[j_coord_offset+DIM*1+XX] -= tx;
 523             f[j_coord_offset+DIM*1+YY] -= ty;
 524             f[j_coord_offset+DIM*1+ZZ] -= tz;
 525
 526             }
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             if (rsq22<rcutoff2)
 533             {
 534
 535             /* REACTION-FIELD ELECTROSTATICS */
 536             velec            = qq22*(rinv22+krf*rsq22-crf);
 537             felec            = qq22*(rinv22*rinvsq22-krf2);
 538
 539             /* Update potential sums from outer loop */
 540             velecsum        += velec;
 541
 542             fscal            = felec;
 543
 544             /* Calculate temporary vectorial force */
 545             tx               = fscal*dx22;
 546             ty               = fscal*dy22;
 547             tz               = fscal*dz22;
 548
 549             /* Update vectorial force */
 550             fix2            += tx;
 551             fiy2            += ty;
 552             fiz2            += tz;
 553             f[j_coord_offset+DIM*2+XX] -= tx;
 554             f[j_coord_offset+DIM*2+YY] -= ty;
 555             f[j_coord_offset+DIM*2+ZZ] -= tz;
 556
 557             }
 558
 559             /* Inner loop uses 296 flops */
 560         }
 561         /* End of innermost loop */
 562
 563         tx = ty = tz = 0;
 564         f[i_coord_offset+DIM*0+XX] += fix0;
 565         f[i_coord_offset+DIM*0+YY] += fiy0;
 566         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 567         tx                         += fix0;
 568         ty                         += fiy0;
 569         tz                         += fiz0;
 570         f[i_coord_offset+DIM*1+XX] += fix1;
 571         f[i_coord_offset+DIM*1+YY] += fiy1;
 572         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 573         tx                         += fix1;
 574         ty                         += fiy1;
 575         tz                         += fiz1;
 576         f[i_coord_offset+DIM*2+XX] += fix2;
 577         f[i_coord_offset+DIM*2+YY] += fiy2;
 578         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 579         tx                         += fix2;
 580         ty                         += fiy2;
 581         tz                         += fiz2;
 582         fshift[i_shift_offset+XX]  += tx;
 583         fshift[i_shift_offset+YY]  += ty;
 584         fshift[i_shift_offset+ZZ]  += tz;
 585
 586         ggid                        = gid[iidx];
 587         /* Update potential energies */
 588         kernel_data->energygrp_elec[ggid] += velecsum;
 589         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 590
 591         /* Increment number of inner iterations */
 592         inneriter                  += j_index_end - j_index_start;
 593
 594         /* Outer loop uses 32 flops */
 595     }
 596
 597     /* Increment number of outer iterations */
 598     outeriter        += nri;
 599
 600     /* Update outer/inner flops */
 601
 602     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*296);
 603 }
 604 /*
 605  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_c
 606  * Electrostatics interaction: ReactionField
 607  * VdW interaction:            LennardJones
 608  * Geometry:                   Water3-Water3
 609  * Calculate force/pot:        Force
 610  */
 611 void
 612 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_c
 613                     (t_nblist                    * gmx_restrict       nlist,
 614                      rvec                        * gmx_restrict          xx,
 615                      rvec                        * gmx_restrict          ff,
 616                      t_forcerec                  * gmx_restrict          fr,
 617                      t_mdatoms                   * gmx_restrict     mdatoms,
 618                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 619                      t_nrnb                      * gmx_restrict        nrnb)
 620 {
 621     int              i_shift_offset,i_coord_offset,j_coord_offset;
 622     int              j_index_start,j_index_end;
 623     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 624     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 625     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 626     real             *shiftvec,*fshift,*x,*f;
 627     int              vdwioffset0;
 628     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 629     int              vdwioffset1;
 630     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 631     int              vdwioffset2;
 632     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 633     int              vdwjidx0;
 634     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 635     int              vdwjidx1;
 636     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 637     int              vdwjidx2;
 638     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 639     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 640     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 641     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 642     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 643     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 644     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 645     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 646     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 647     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 648     real             velec,felec,velecsum,facel,crf,krf,krf2;
 649     real             *charge;
 650     int              nvdwtype;
 651     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 652     int              *vdwtype;
 653     real             *vdwparam;
 654
 655     x                = xx[0];
 656     f                = ff[0];
 657
 658     nri              = nlist->nri;
 659     iinr             = nlist->iinr;
 660     jindex           = nlist->jindex;
 661     jjnr             = nlist->jjnr;
 662     shiftidx         = nlist->shift;
 663     gid              = nlist->gid;
 664     shiftvec         = fr->shift_vec[0];
 665     fshift           = fr->fshift[0];
 666     facel            = fr->epsfac;
 667     charge           = mdatoms->chargeA;
 668     krf              = fr->ic->k_rf;
 669     krf2             = krf*2.0;
 670     crf              = fr->ic->c_rf;
 671     nvdwtype         = fr->ntype;
 672     vdwparam         = fr->nbfp;
 673     vdwtype          = mdatoms->typeA;
 674
 675     /* Setup water-specific parameters */
 676     inr              = nlist->iinr[0];
 677     iq0              = facel*charge[inr+0];
 678     iq1              = facel*charge[inr+1];
 679     iq2              = facel*charge[inr+2];
 680     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 681
 682     jq0              = charge[inr+0];
 683     jq1              = charge[inr+1];
 684     jq2              = charge[inr+2];
 685     vdwjidx0         = 2*vdwtype[inr+0];
 686     qq00             = iq0*jq0;
 687     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 688     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 689     qq01             = iq0*jq1;
 690     qq02             = iq0*jq2;
 691     qq10             = iq1*jq0;
 692     qq11             = iq1*jq1;
 693     qq12             = iq1*jq2;
 694     qq20             = iq2*jq0;
 695     qq21             = iq2*jq1;
 696     qq22             = iq2*jq2;
 697
 698     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 699     rcutoff          = fr->rcoulomb;
 700     rcutoff2         = rcutoff*rcutoff;
 701
 702     sh_vdw_invrcut6  = fr->ic->sh_invrc6;
 703     rvdw             = fr->rvdw;
 704
 705     outeriter        = 0;
 706     inneriter        = 0;
 707
 708     /* Start outer loop over neighborlists */
 709     for(iidx=0; iidx<nri; iidx++)
 710     {
 711         /* Load shift vector for this list */
 712         i_shift_offset   = DIM*shiftidx[iidx];
 713         shX              = shiftvec[i_shift_offset+XX];
 714         shY              = shiftvec[i_shift_offset+YY];
 715         shZ              = shiftvec[i_shift_offset+ZZ];
 716
 717         /* Load limits for loop over neighbors */
 718         j_index_start    = jindex[iidx];
 719         j_index_end      = jindex[iidx+1];
 720
 721         /* Get outer coordinate index */
 722         inr              = iinr[iidx];
 723         i_coord_offset   = DIM*inr;
 724
 725         /* Load i particle coords and add shift vector */
 726         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 727         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 728         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 729         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 730         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 731         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 732         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 733         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 734         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 735
 736         fix0             = 0.0;
 737         fiy0             = 0.0;
 738         fiz0             = 0.0;
 739         fix1             = 0.0;
 740         fiy1             = 0.0;
 741         fiz1             = 0.0;
 742         fix2             = 0.0;
 743         fiy2             = 0.0;
 744         fiz2             = 0.0;
 745
 746         /* Start inner kernel loop */
 747         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 748         {
 749             /* Get j neighbor index, and coordinate index */
 750             jnr              = jjnr[jidx];
 751             j_coord_offset   = DIM*jnr;
 752
 753             /* load j atom coordinates */
 754             jx0              = x[j_coord_offset+DIM*0+XX];
 755             jy0              = x[j_coord_offset+DIM*0+YY];
 756             jz0              = x[j_coord_offset+DIM*0+ZZ];
 757             jx1              = x[j_coord_offset+DIM*1+XX];
 758             jy1              = x[j_coord_offset+DIM*1+YY];
 759             jz1              = x[j_coord_offset+DIM*1+ZZ];
 760             jx2              = x[j_coord_offset+DIM*2+XX];
 761             jy2              = x[j_coord_offset+DIM*2+YY];
 762             jz2              = x[j_coord_offset+DIM*2+ZZ];
 763
 764             /* Calculate displacement vector */
 765             dx00             = ix0 - jx0;
 766             dy00             = iy0 - jy0;
 767             dz00             = iz0 - jz0;
 768             dx01             = ix0 - jx1;
 769             dy01             = iy0 - jy1;
 770             dz01             = iz0 - jz1;
 771             dx02             = ix0 - jx2;
 772             dy02             = iy0 - jy2;
 773             dz02             = iz0 - jz2;
 774             dx10             = ix1 - jx0;
 775             dy10             = iy1 - jy0;
 776             dz10             = iz1 - jz0;
 777             dx11             = ix1 - jx1;
 778             dy11             = iy1 - jy1;
 779             dz11             = iz1 - jz1;
 780             dx12             = ix1 - jx2;
 781             dy12             = iy1 - jy2;
 782             dz12             = iz1 - jz2;
 783             dx20             = ix2 - jx0;
 784             dy20             = iy2 - jy0;
 785             dz20             = iz2 - jz0;
 786             dx21             = ix2 - jx1;
 787             dy21             = iy2 - jy1;
 788             dz21             = iz2 - jz1;
 789             dx22             = ix2 - jx2;
 790             dy22             = iy2 - jy2;
 791             dz22             = iz2 - jz2;
 792
 793             /* Calculate squared distance and things based on it */
 794             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 795             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 796             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 797             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 798             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 799             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 800             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 801             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 802             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 803
 804             rinv00           = gmx_invsqrt(rsq00);
 805             rinv01           = gmx_invsqrt(rsq01);
 806             rinv02           = gmx_invsqrt(rsq02);
 807             rinv10           = gmx_invsqrt(rsq10);
 808             rinv11           = gmx_invsqrt(rsq11);
 809             rinv12           = gmx_invsqrt(rsq12);
 810             rinv20           = gmx_invsqrt(rsq20);
 811             rinv21           = gmx_invsqrt(rsq21);
 812             rinv22           = gmx_invsqrt(rsq22);
 813
 814             rinvsq00         = rinv00*rinv00;
 815             rinvsq01         = rinv01*rinv01;
 816             rinvsq02         = rinv02*rinv02;
 817             rinvsq10         = rinv10*rinv10;
 818             rinvsq11         = rinv11*rinv11;
 819             rinvsq12         = rinv12*rinv12;
 820             rinvsq20         = rinv20*rinv20;
 821             rinvsq21         = rinv21*rinv21;
 822             rinvsq22         = rinv22*rinv22;
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (rsq00<rcutoff2)
 829             {
 830
 831             /* REACTION-FIELD ELECTROSTATICS */
 832             felec            = qq00*(rinv00*rinvsq00-krf2);
 833
 834             /* LENNARD-JONES DISPERSION/REPULSION */
 835
 836             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 837             fvdw             = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00;
 838
 839             fscal            = felec+fvdw;
 840
 841             /* Calculate temporary vectorial force */
 842             tx               = fscal*dx00;
 843             ty               = fscal*dy00;
 844             tz               = fscal*dz00;
 845
 846             /* Update vectorial force */
 847             fix0            += tx;
 848             fiy0            += ty;
 849             fiz0            += tz;
 850             f[j_coord_offset+DIM*0+XX] -= tx;
 851             f[j_coord_offset+DIM*0+YY] -= ty;
 852             f[j_coord_offset+DIM*0+ZZ] -= tz;
 853
 854             }
 855
 856             /**************************
 857              * CALCULATE INTERACTIONS *
 858              **************************/
 859
 860             if (rsq01<rcutoff2)
 861             {
 862
 863             /* REACTION-FIELD ELECTROSTATICS */
 864             felec            = qq01*(rinv01*rinvsq01-krf2);
 865
 866             fscal            = felec;
 867
 868             /* Calculate temporary vectorial force */
 869             tx               = fscal*dx01;
 870             ty               = fscal*dy01;
 871             tz               = fscal*dz01;
 872
 873             /* Update vectorial force */
 874             fix0            += tx;
 875             fiy0            += ty;
 876             fiz0            += tz;
 877             f[j_coord_offset+DIM*1+XX] -= tx;
 878             f[j_coord_offset+DIM*1+YY] -= ty;
 879             f[j_coord_offset+DIM*1+ZZ] -= tz;
 880
 881             }
 882
 883             /**************************
 884              * CALCULATE INTERACTIONS *
 885              **************************/
 886
 887             if (rsq02<rcutoff2)
 888             {
 889
 890             /* REACTION-FIELD ELECTROSTATICS */
 891             felec            = qq02*(rinv02*rinvsq02-krf2);
 892
 893             fscal            = felec;
 894
 895             /* Calculate temporary vectorial force */
 896             tx               = fscal*dx02;
 897             ty               = fscal*dy02;
 898             tz               = fscal*dz02;
 899
 900             /* Update vectorial force */
 901             fix0            += tx;
 902             fiy0            += ty;
 903             fiz0            += tz;
 904             f[j_coord_offset+DIM*2+XX] -= tx;
 905             f[j_coord_offset+DIM*2+YY] -= ty;
 906             f[j_coord_offset+DIM*2+ZZ] -= tz;
 907
 908             }
 909
 910             /**************************
 911              * CALCULATE INTERACTIONS *
 912              **************************/
 913
 914             if (rsq10<rcutoff2)
 915             {
 916
 917             /* REACTION-FIELD ELECTROSTATICS */
 918             felec            = qq10*(rinv10*rinvsq10-krf2);
 919
 920             fscal            = felec;
 921
 922             /* Calculate temporary vectorial force */
 923             tx               = fscal*dx10;
 924             ty               = fscal*dy10;
 925             tz               = fscal*dz10;
 926
 927             /* Update vectorial force */
 928             fix1            += tx;
 929             fiy1            += ty;
 930             fiz1            += tz;
 931             f[j_coord_offset+DIM*0+XX] -= tx;
 932             f[j_coord_offset+DIM*0+YY] -= ty;
 933             f[j_coord_offset+DIM*0+ZZ] -= tz;
 934
 935             }
 936
 937             /**************************
 938              * CALCULATE INTERACTIONS *
 939              **************************/
 940
 941             if (rsq11<rcutoff2)
 942             {
 943
 944             /* REACTION-FIELD ELECTROSTATICS */
 945             felec            = qq11*(rinv11*rinvsq11-krf2);
 946
 947             fscal            = felec;
 948
 949             /* Calculate temporary vectorial force */
 950             tx               = fscal*dx11;
 951             ty               = fscal*dy11;
 952             tz               = fscal*dz11;
 953
 954             /* Update vectorial force */
 955             fix1            += tx;
 956             fiy1            += ty;
 957             fiz1            += tz;
 958             f[j_coord_offset+DIM*1+XX] -= tx;
 959             f[j_coord_offset+DIM*1+YY] -= ty;
 960             f[j_coord_offset+DIM*1+ZZ] -= tz;
 961
 962             }
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             if (rsq12<rcutoff2)
 969             {
 970
 971             /* REACTION-FIELD ELECTROSTATICS */
 972             felec            = qq12*(rinv12*rinvsq12-krf2);
 973
 974             fscal            = felec;
 975
 976             /* Calculate temporary vectorial force */
 977             tx               = fscal*dx12;
 978             ty               = fscal*dy12;
 979             tz               = fscal*dz12;
 980
 981             /* Update vectorial force */
 982             fix1            += tx;
 983             fiy1            += ty;
 984             fiz1            += tz;
 985             f[j_coord_offset+DIM*2+XX] -= tx;
 986             f[j_coord_offset+DIM*2+YY] -= ty;
 987             f[j_coord_offset+DIM*2+ZZ] -= tz;
 988
 989             }
 990
 991             /**************************
 992              * CALCULATE INTERACTIONS *
 993              **************************/
 994
 995             if (rsq20<rcutoff2)
 996             {
 997
 998             /* REACTION-FIELD ELECTROSTATICS */
 999             felec            = qq20*(rinv20*rinvsq20-krf2);
1000
1001             fscal            = felec;
1002
1003             /* Calculate temporary vectorial force */
1004             tx               = fscal*dx20;
1005             ty               = fscal*dy20;
1006             tz               = fscal*dz20;
1007
1008             /* Update vectorial force */
1009             fix2            += tx;
1010             fiy2            += ty;
1011             fiz2            += tz;
1012             f[j_coord_offset+DIM*0+XX] -= tx;
1013             f[j_coord_offset+DIM*0+YY] -= ty;
1014             f[j_coord_offset+DIM*0+ZZ] -= tz;
1015
1016             }
1017
1018             /**************************
1019              * CALCULATE INTERACTIONS *
1020              **************************/
1021
1022             if (rsq21<rcutoff2)
1023             {
1024
1025             /* REACTION-FIELD ELECTROSTATICS */
1026             felec            = qq21*(rinv21*rinvsq21-krf2);
1027
1028             fscal            = felec;
1029
1030             /* Calculate temporary vectorial force */
1031             tx               = fscal*dx21;
1032             ty               = fscal*dy21;
1033             tz               = fscal*dz21;
1034
1035             /* Update vectorial force */
1036             fix2            += tx;
1037             fiy2            += ty;
1038             fiz2            += tz;
1039             f[j_coord_offset+DIM*1+XX] -= tx;
1040             f[j_coord_offset+DIM*1+YY] -= ty;
1041             f[j_coord_offset+DIM*1+ZZ] -= tz;
1042
1043             }
1044
1045             /**************************
1046              * CALCULATE INTERACTIONS *
1047              **************************/
1048
1049             if (rsq22<rcutoff2)
1050             {
1051
1052             /* REACTION-FIELD ELECTROSTATICS */
1053             felec            = qq22*(rinv22*rinvsq22-krf2);
1054
1055             fscal            = felec;
1056
1057             /* Calculate temporary vectorial force */
1058             tx               = fscal*dx22;
1059             ty               = fscal*dy22;
1060             tz               = fscal*dz22;
1061
1062             /* Update vectorial force */
1063             fix2            += tx;
1064             fiy2            += ty;
1065             fiz2            += tz;
1066             f[j_coord_offset+DIM*2+XX] -= tx;
1067             f[j_coord_offset+DIM*2+YY] -= ty;
1068             f[j_coord_offset+DIM*2+ZZ] -= tz;
1069
1070             }
1071
1072             /* Inner loop uses 241 flops */
1073         }
1074         /* End of innermost loop */
1075
1076         tx = ty = tz = 0;
1077         f[i_coord_offset+DIM*0+XX] += fix0;
1078         f[i_coord_offset+DIM*0+YY] += fiy0;
1079         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1080         tx                         += fix0;
1081         ty                         += fiy0;
1082         tz                         += fiz0;
1083         f[i_coord_offset+DIM*1+XX] += fix1;
1084         f[i_coord_offset+DIM*1+YY] += fiy1;
1085         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1086         tx                         += fix1;
1087         ty                         += fiy1;
1088         tz                         += fiz1;
1089         f[i_coord_offset+DIM*2+XX] += fix2;
1090         f[i_coord_offset+DIM*2+YY] += fiy2;
1091         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1092         tx                         += fix2;
1093         ty                         += fiy2;
1094         tz                         += fiz2;
1095         fshift[i_shift_offset+XX]  += tx;
1096         fshift[i_shift_offset+YY]  += ty;
1097         fshift[i_shift_offset+ZZ]  += tz;
1098
1099         /* Increment number of inner iterations */
1100         inneriter                  += j_index_end - j_index_start;
1101
1102         /* Outer loop uses 30 flops */
1103     }
1104
1105     /* Increment number of outer iterations */
1106     outeriter        += nri;
1107
1108     /* Update outer/inner flops */
1109
1110     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*241);
1111 }