src/gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecRF_VdwCSTab_GeomW3W3_c.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS c kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_c
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_c
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     int              i_shift_offset,i_coord_offset,j_coord_offset;
  67     int              j_index_start,j_index_end;
  68     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  69     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  70     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  71     real             *shiftvec,*fshift,*x,*f;
  72     int              vdwioffset0;
  73     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwioffset1;
  75     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  76     int              vdwioffset2;
  77     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  78     int              vdwjidx0;
  79     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  80     int              vdwjidx1;
  81     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  82     int              vdwjidx2;
  83     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  84     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  85     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
  86     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
  87     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
  88     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  89     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  90     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
  91     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  92     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  93     real             velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     int              vfitab;
 100     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 101     real             *vftab;
 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     krf              = fr->ic->k_rf;
 117     krf2             = krf*2.0;
 118     crf              = fr->ic->c_rf;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_vdw->data;
 124     vftabscale       = kernel_data->table_vdw->scale;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = facel*charge[inr+0];
 129     iq1              = facel*charge[inr+1];
 130     iq2              = facel*charge[inr+2];
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     jq0              = charge[inr+0];
 134     jq1              = charge[inr+1];
 135     jq2              = charge[inr+2];
 136     vdwjidx0         = 2*vdwtype[inr+0];
 137     qq00             = iq0*jq0;
 138     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 139     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 140     qq01             = iq0*jq1;
 141     qq02             = iq0*jq2;
 142     qq10             = iq1*jq0;
 143     qq11             = iq1*jq1;
 144     qq12             = iq1*jq2;
 145     qq20             = iq2*jq0;
 146     qq21             = iq2*jq1;
 147     qq22             = iq2*jq2;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157         shX              = shiftvec[i_shift_offset+XX];
 158         shY              = shiftvec[i_shift_offset+YY];
 159         shZ              = shiftvec[i_shift_offset+ZZ];
 160
 161         /* Load limits for loop over neighbors */
 162         j_index_start    = jindex[iidx];
 163         j_index_end      = jindex[iidx+1];
 164
 165         /* Get outer coordinate index */
 166         inr              = iinr[iidx];
 167         i_coord_offset   = DIM*inr;
 168
 169         /* Load i particle coords and add shift vector */
 170         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 171         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 172         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 173         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 174         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 175         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 176         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 177         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 178         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 179
 180         fix0             = 0.0;
 181         fiy0             = 0.0;
 182         fiz0             = 0.0;
 183         fix1             = 0.0;
 184         fiy1             = 0.0;
 185         fiz1             = 0.0;
 186         fix2             = 0.0;
 187         fiy2             = 0.0;
 188         fiz2             = 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
 212             /* Calculate displacement vector */
 213             dx00             = ix0 - jx0;
 214             dy00             = iy0 - jy0;
 215             dz00             = iz0 - jz0;
 216             dx01             = ix0 - jx1;
 217             dy01             = iy0 - jy1;
 218             dz01             = iz0 - jz1;
 219             dx02             = ix0 - jx2;
 220             dy02             = iy0 - jy2;
 221             dz02             = iz0 - jz2;
 222             dx10             = ix1 - jx0;
 223             dy10             = iy1 - jy0;
 224             dz10             = iz1 - jz0;
 225             dx11             = ix1 - jx1;
 226             dy11             = iy1 - jy1;
 227             dz11             = iz1 - jz1;
 228             dx12             = ix1 - jx2;
 229             dy12             = iy1 - jy2;
 230             dz12             = iz1 - jz2;
 231             dx20             = ix2 - jx0;
 232             dy20             = iy2 - jy0;
 233             dz20             = iz2 - jz0;
 234             dx21             = ix2 - jx1;
 235             dy21             = iy2 - jy1;
 236             dz21             = iz2 - jz1;
 237             dx22             = ix2 - jx2;
 238             dy22             = iy2 - jy2;
 239             dz22             = iz2 - jz2;
 240
 241             /* Calculate squared distance and things based on it */
 242             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 243             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 244             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 245             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 246             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 247             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 248             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 249             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 250             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 251
 252             rinv00           = gmx_invsqrt(rsq00);
 253             rinv01           = gmx_invsqrt(rsq01);
 254             rinv02           = gmx_invsqrt(rsq02);
 255             rinv10           = gmx_invsqrt(rsq10);
 256             rinv11           = gmx_invsqrt(rsq11);
 257             rinv12           = gmx_invsqrt(rsq12);
 258             rinv20           = gmx_invsqrt(rsq20);
 259             rinv21           = gmx_invsqrt(rsq21);
 260             rinv22           = gmx_invsqrt(rsq22);
 261
 262             rinvsq00         = rinv00*rinv00;
 263             rinvsq01         = rinv01*rinv01;
 264             rinvsq02         = rinv02*rinv02;
 265             rinvsq10         = rinv10*rinv10;
 266             rinvsq11         = rinv11*rinv11;
 267             rinvsq12         = rinv12*rinv12;
 268             rinvsq20         = rinv20*rinv20;
 269             rinvsq21         = rinv21*rinv21;
 270             rinvsq22         = rinv22*rinv22;
 271
 272             /**************************
 273              * CALCULATE INTERACTIONS *
 274              **************************/
 275
 276             r00              = rsq00*rinv00;
 277
 278             /* Calculate table index by multiplying r with table scale and truncate to integer */
 279             rt               = r00*vftabscale;
 280             vfitab           = rt;
 281             vfeps            = rt-vfitab;
 282             vfitab           = 2*4*vfitab;
 283
 284             /* REACTION-FIELD ELECTROSTATICS */
 285             velec            = qq00*(rinv00+krf*rsq00-crf);
 286             felec            = qq00*(rinv00*rinvsq00-krf2);
 287
 288             /* CUBIC SPLINE TABLE DISPERSION */
 289             vfitab          += 0;
 290             Y                = vftab[vfitab];
 291             F                = vftab[vfitab+1];
 292             Geps             = vfeps*vftab[vfitab+2];
 293             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 294             Fp               = F+Geps+Heps2;
 295             VV               = Y+vfeps*Fp;
 296             vvdw6            = c6_00*VV;
 297             FF               = Fp+Geps+2.0*Heps2;
 298             fvdw6            = c6_00*FF;
 299
 300             /* CUBIC SPLINE TABLE REPULSION */
 301             Y                = vftab[vfitab+4];
 302             F                = vftab[vfitab+5];
 303             Geps             = vfeps*vftab[vfitab+6];
 304             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 305             Fp               = F+Geps+Heps2;
 306             VV               = Y+vfeps*Fp;
 307             vvdw12           = c12_00*VV;
 308             FF               = Fp+Geps+2.0*Heps2;
 309             fvdw12           = c12_00*FF;
 310             vvdw             = vvdw12+vvdw6;
 311             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 312
 313             /* Update potential sums from outer loop */
 314             velecsum        += velec;
 315             vvdwsum         += vvdw;
 316
 317             fscal            = felec+fvdw;
 318
 319             /* Calculate temporary vectorial force */
 320             tx               = fscal*dx00;
 321             ty               = fscal*dy00;
 322             tz               = fscal*dz00;
 323
 324             /* Update vectorial force */
 325             fix0            += tx;
 326             fiy0            += ty;
 327             fiz0            += tz;
 328             f[j_coord_offset+DIM*0+XX] -= tx;
 329             f[j_coord_offset+DIM*0+YY] -= ty;
 330             f[j_coord_offset+DIM*0+ZZ] -= tz;
 331
 332             /**************************
 333              * CALCULATE INTERACTIONS *
 334              **************************/
 335
 336             /* REACTION-FIELD ELECTROSTATICS */
 337             velec            = qq01*(rinv01+krf*rsq01-crf);
 338             felec            = qq01*(rinv01*rinvsq01-krf2);
 339
 340             /* Update potential sums from outer loop */
 341             velecsum        += velec;
 342
 343             fscal            = felec;
 344
 345             /* Calculate temporary vectorial force */
 346             tx               = fscal*dx01;
 347             ty               = fscal*dy01;
 348             tz               = fscal*dz01;
 349
 350             /* Update vectorial force */
 351             fix0            += tx;
 352             fiy0            += ty;
 353             fiz0            += tz;
 354             f[j_coord_offset+DIM*1+XX] -= tx;
 355             f[j_coord_offset+DIM*1+YY] -= ty;
 356             f[j_coord_offset+DIM*1+ZZ] -= tz;
 357
 358             /**************************
 359              * CALCULATE INTERACTIONS *
 360              **************************/
 361
 362             /* REACTION-FIELD ELECTROSTATICS */
 363             velec            = qq02*(rinv02+krf*rsq02-crf);
 364             felec            = qq02*(rinv02*rinvsq02-krf2);
 365
 366             /* Update potential sums from outer loop */
 367             velecsum        += velec;
 368
 369             fscal            = felec;
 370
 371             /* Calculate temporary vectorial force */
 372             tx               = fscal*dx02;
 373             ty               = fscal*dy02;
 374             tz               = fscal*dz02;
 375
 376             /* Update vectorial force */
 377             fix0            += tx;
 378             fiy0            += ty;
 379             fiz0            += tz;
 380             f[j_coord_offset+DIM*2+XX] -= tx;
 381             f[j_coord_offset+DIM*2+YY] -= ty;
 382             f[j_coord_offset+DIM*2+ZZ] -= tz;
 383
 384             /**************************
 385              * CALCULATE INTERACTIONS *
 386              **************************/
 387
 388             /* REACTION-FIELD ELECTROSTATICS */
 389             velec            = qq10*(rinv10+krf*rsq10-crf);
 390             felec            = qq10*(rinv10*rinvsq10-krf2);
 391
 392             /* Update potential sums from outer loop */
 393             velecsum        += velec;
 394
 395             fscal            = felec;
 396
 397             /* Calculate temporary vectorial force */
 398             tx               = fscal*dx10;
 399             ty               = fscal*dy10;
 400             tz               = fscal*dz10;
 401
 402             /* Update vectorial force */
 403             fix1            += tx;
 404             fiy1            += ty;
 405             fiz1            += tz;
 406             f[j_coord_offset+DIM*0+XX] -= tx;
 407             f[j_coord_offset+DIM*0+YY] -= ty;
 408             f[j_coord_offset+DIM*0+ZZ] -= tz;
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = qq11*(rinv11+krf*rsq11-crf);
 416             felec            = qq11*(rinv11*rinvsq11-krf2);
 417
 418             /* Update potential sums from outer loop */
 419             velecsum        += velec;
 420
 421             fscal            = felec;
 422
 423             /* Calculate temporary vectorial force */
 424             tx               = fscal*dx11;
 425             ty               = fscal*dy11;
 426             tz               = fscal*dz11;
 427
 428             /* Update vectorial force */
 429             fix1            += tx;
 430             fiy1            += ty;
 431             fiz1            += tz;
 432             f[j_coord_offset+DIM*1+XX] -= tx;
 433             f[j_coord_offset+DIM*1+YY] -= ty;
 434             f[j_coord_offset+DIM*1+ZZ] -= tz;
 435
 436             /**************************
 437              * CALCULATE INTERACTIONS *
 438              **************************/
 439
 440             /* REACTION-FIELD ELECTROSTATICS */
 441             velec            = qq12*(rinv12+krf*rsq12-crf);
 442             felec            = qq12*(rinv12*rinvsq12-krf2);
 443
 444             /* Update potential sums from outer loop */
 445             velecsum        += velec;
 446
 447             fscal            = felec;
 448
 449             /* Calculate temporary vectorial force */
 450             tx               = fscal*dx12;
 451             ty               = fscal*dy12;
 452             tz               = fscal*dz12;
 453
 454             /* Update vectorial force */
 455             fix1            += tx;
 456             fiy1            += ty;
 457             fiz1            += tz;
 458             f[j_coord_offset+DIM*2+XX] -= tx;
 459             f[j_coord_offset+DIM*2+YY] -= ty;
 460             f[j_coord_offset+DIM*2+ZZ] -= tz;
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             /* REACTION-FIELD ELECTROSTATICS */
 467             velec            = qq20*(rinv20+krf*rsq20-crf);
 468             felec            = qq20*(rinv20*rinvsq20-krf2);
 469
 470             /* Update potential sums from outer loop */
 471             velecsum        += velec;
 472
 473             fscal            = felec;
 474
 475             /* Calculate temporary vectorial force */
 476             tx               = fscal*dx20;
 477             ty               = fscal*dy20;
 478             tz               = fscal*dz20;
 479
 480             /* Update vectorial force */
 481             fix2            += tx;
 482             fiy2            += ty;
 483             fiz2            += tz;
 484             f[j_coord_offset+DIM*0+XX] -= tx;
 485             f[j_coord_offset+DIM*0+YY] -= ty;
 486             f[j_coord_offset+DIM*0+ZZ] -= tz;
 487
 488             /**************************
 489              * CALCULATE INTERACTIONS *
 490              **************************/
 491
 492             /* REACTION-FIELD ELECTROSTATICS */
 493             velec            = qq21*(rinv21+krf*rsq21-crf);
 494             felec            = qq21*(rinv21*rinvsq21-krf2);
 495
 496             /* Update potential sums from outer loop */
 497             velecsum        += velec;
 498
 499             fscal            = felec;
 500
 501             /* Calculate temporary vectorial force */
 502             tx               = fscal*dx21;
 503             ty               = fscal*dy21;
 504             tz               = fscal*dz21;
 505
 506             /* Update vectorial force */
 507             fix2            += tx;
 508             fiy2            += ty;
 509             fiz2            += tz;
 510             f[j_coord_offset+DIM*1+XX] -= tx;
 511             f[j_coord_offset+DIM*1+YY] -= ty;
 512             f[j_coord_offset+DIM*1+ZZ] -= tz;
 513
 514             /**************************
 515              * CALCULATE INTERACTIONS *
 516              **************************/
 517
 518             /* REACTION-FIELD ELECTROSTATICS */
 519             velec            = qq22*(rinv22+krf*rsq22-crf);
 520             felec            = qq22*(rinv22*rinvsq22-krf2);
 521
 522             /* Update potential sums from outer loop */
 523             velecsum        += velec;
 524
 525             fscal            = felec;
 526
 527             /* Calculate temporary vectorial force */
 528             tx               = fscal*dx22;
 529             ty               = fscal*dy22;
 530             tz               = fscal*dz22;
 531
 532             /* Update vectorial force */
 533             fix2            += tx;
 534             fiy2            += ty;
 535             fiz2            += tz;
 536             f[j_coord_offset+DIM*2+XX] -= tx;
 537             f[j_coord_offset+DIM*2+YY] -= ty;
 538             f[j_coord_offset+DIM*2+ZZ] -= tz;
 539
 540             /* Inner loop uses 313 flops */
 541         }
 542         /* End of innermost loop */
 543
 544         tx = ty = tz = 0;
 545         f[i_coord_offset+DIM*0+XX] += fix0;
 546         f[i_coord_offset+DIM*0+YY] += fiy0;
 547         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 548         tx                         += fix0;
 549         ty                         += fiy0;
 550         tz                         += fiz0;
 551         f[i_coord_offset+DIM*1+XX] += fix1;
 552         f[i_coord_offset+DIM*1+YY] += fiy1;
 553         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 554         tx                         += fix1;
 555         ty                         += fiy1;
 556         tz                         += fiz1;
 557         f[i_coord_offset+DIM*2+XX] += fix2;
 558         f[i_coord_offset+DIM*2+YY] += fiy2;
 559         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 560         tx                         += fix2;
 561         ty                         += fiy2;
 562         tz                         += fiz2;
 563         fshift[i_shift_offset+XX]  += tx;
 564         fshift[i_shift_offset+YY]  += ty;
 565         fshift[i_shift_offset+ZZ]  += tz;
 566
 567         ggid                        = gid[iidx];
 568         /* Update potential energies */
 569         kernel_data->energygrp_elec[ggid] += velecsum;
 570         kernel_data->energygrp_vdw[ggid] += vvdwsum;
 571
 572         /* Increment number of inner iterations */
 573         inneriter                  += j_index_end - j_index_start;
 574
 575         /* Outer loop uses 32 flops */
 576     }
 577
 578     /* Increment number of outer iterations */
 579     outeriter        += nri;
 580
 581     /* Update outer/inner flops */
 582
 583     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*32 + inneriter*313);
 584 }
 585 /*
 586  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_c
 587  * Electrostatics interaction: ReactionField
 588  * VdW interaction:            CubicSplineTable
 589  * Geometry:                   Water3-Water3
 590  * Calculate force/pot:        Force
 591  */
 592 void
 593 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_c
 594                     (t_nblist                    * gmx_restrict       nlist,
 595                      rvec                        * gmx_restrict          xx,
 596                      rvec                        * gmx_restrict          ff,
 597                      t_forcerec                  * gmx_restrict          fr,
 598                      t_mdatoms                   * gmx_restrict     mdatoms,
 599                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 600                      t_nrnb                      * gmx_restrict        nrnb)
 601 {
 602     int              i_shift_offset,i_coord_offset,j_coord_offset;
 603     int              j_index_start,j_index_end;
 604     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
 605     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
 606     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 607     real             *shiftvec,*fshift,*x,*f;
 608     int              vdwioffset0;
 609     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 610     int              vdwioffset1;
 611     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 612     int              vdwioffset2;
 613     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 614     int              vdwjidx0;
 615     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 616     int              vdwjidx1;
 617     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 618     int              vdwjidx2;
 619     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 620     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 621     real             dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01,cexp1_01,cexp2_01;
 622     real             dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02,cexp1_02,cexp2_02;
 623     real             dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10,cexp1_10,cexp2_10;
 624     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 625     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 626     real             dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20,cexp1_20,cexp2_20;
 627     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 628     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 629     real             velec,felec,velecsum,facel,crf,krf,krf2;
 630     real             *charge;
 631     int              nvdwtype;
 632     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 633     int              *vdwtype;
 634     real             *vdwparam;
 635     int              vfitab;
 636     real             rt,vfeps,vftabscale,Y,F,Geps,Heps2,Fp,VV,FF;
 637     real             *vftab;
 638
 639     x                = xx[0];
 640     f                = ff[0];
 641
 642     nri              = nlist->nri;
 643     iinr             = nlist->iinr;
 644     jindex           = nlist->jindex;
 645     jjnr             = nlist->jjnr;
 646     shiftidx         = nlist->shift;
 647     gid              = nlist->gid;
 648     shiftvec         = fr->shift_vec[0];
 649     fshift           = fr->fshift[0];
 650     facel            = fr->epsfac;
 651     charge           = mdatoms->chargeA;
 652     krf              = fr->ic->k_rf;
 653     krf2             = krf*2.0;
 654     crf              = fr->ic->c_rf;
 655     nvdwtype         = fr->ntype;
 656     vdwparam         = fr->nbfp;
 657     vdwtype          = mdatoms->typeA;
 658
 659     vftab            = kernel_data->table_vdw->data;
 660     vftabscale       = kernel_data->table_vdw->scale;
 661
 662     /* Setup water-specific parameters */
 663     inr              = nlist->iinr[0];
 664     iq0              = facel*charge[inr+0];
 665     iq1              = facel*charge[inr+1];
 666     iq2              = facel*charge[inr+2];
 667     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 668
 669     jq0              = charge[inr+0];
 670     jq1              = charge[inr+1];
 671     jq2              = charge[inr+2];
 672     vdwjidx0         = 2*vdwtype[inr+0];
 673     qq00             = iq0*jq0;
 674     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 675     c12_00           = vdwparam[vdwioffset0+vdwjidx0+1];
 676     qq01             = iq0*jq1;
 677     qq02             = iq0*jq2;
 678     qq10             = iq1*jq0;
 679     qq11             = iq1*jq1;
 680     qq12             = iq1*jq2;
 681     qq20             = iq2*jq0;
 682     qq21             = iq2*jq1;
 683     qq22             = iq2*jq2;
 684
 685     outeriter        = 0;
 686     inneriter        = 0;
 687
 688     /* Start outer loop over neighborlists */
 689     for(iidx=0; iidx<nri; iidx++)
 690     {
 691         /* Load shift vector for this list */
 692         i_shift_offset   = DIM*shiftidx[iidx];
 693         shX              = shiftvec[i_shift_offset+XX];
 694         shY              = shiftvec[i_shift_offset+YY];
 695         shZ              = shiftvec[i_shift_offset+ZZ];
 696
 697         /* Load limits for loop over neighbors */
 698         j_index_start    = jindex[iidx];
 699         j_index_end      = jindex[iidx+1];
 700
 701         /* Get outer coordinate index */
 702         inr              = iinr[iidx];
 703         i_coord_offset   = DIM*inr;
 704
 705         /* Load i particle coords and add shift vector */
 706         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 707         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 708         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 709         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 710         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 711         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 712         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 713         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 714         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 715
 716         fix0             = 0.0;
 717         fiy0             = 0.0;
 718         fiz0             = 0.0;
 719         fix1             = 0.0;
 720         fiy1             = 0.0;
 721         fiz1             = 0.0;
 722         fix2             = 0.0;
 723         fiy2             = 0.0;
 724         fiz2             = 0.0;
 725
 726         /* Start inner kernel loop */
 727         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 728         {
 729             /* Get j neighbor index, and coordinate index */
 730             jnr              = jjnr[jidx];
 731             j_coord_offset   = DIM*jnr;
 732
 733             /* load j atom coordinates */
 734             jx0              = x[j_coord_offset+DIM*0+XX];
 735             jy0              = x[j_coord_offset+DIM*0+YY];
 736             jz0              = x[j_coord_offset+DIM*0+ZZ];
 737             jx1              = x[j_coord_offset+DIM*1+XX];
 738             jy1              = x[j_coord_offset+DIM*1+YY];
 739             jz1              = x[j_coord_offset+DIM*1+ZZ];
 740             jx2              = x[j_coord_offset+DIM*2+XX];
 741             jy2              = x[j_coord_offset+DIM*2+YY];
 742             jz2              = x[j_coord_offset+DIM*2+ZZ];
 743
 744             /* Calculate displacement vector */
 745             dx00             = ix0 - jx0;
 746             dy00             = iy0 - jy0;
 747             dz00             = iz0 - jz0;
 748             dx01             = ix0 - jx1;
 749             dy01             = iy0 - jy1;
 750             dz01             = iz0 - jz1;
 751             dx02             = ix0 - jx2;
 752             dy02             = iy0 - jy2;
 753             dz02             = iz0 - jz2;
 754             dx10             = ix1 - jx0;
 755             dy10             = iy1 - jy0;
 756             dz10             = iz1 - jz0;
 757             dx11             = ix1 - jx1;
 758             dy11             = iy1 - jy1;
 759             dz11             = iz1 - jz1;
 760             dx12             = ix1 - jx2;
 761             dy12             = iy1 - jy2;
 762             dz12             = iz1 - jz2;
 763             dx20             = ix2 - jx0;
 764             dy20             = iy2 - jy0;
 765             dz20             = iz2 - jz0;
 766             dx21             = ix2 - jx1;
 767             dy21             = iy2 - jy1;
 768             dz21             = iz2 - jz1;
 769             dx22             = ix2 - jx2;
 770             dy22             = iy2 - jy2;
 771             dz22             = iz2 - jz2;
 772
 773             /* Calculate squared distance and things based on it */
 774             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 775             rsq01            = dx01*dx01+dy01*dy01+dz01*dz01;
 776             rsq02            = dx02*dx02+dy02*dy02+dz02*dz02;
 777             rsq10            = dx10*dx10+dy10*dy10+dz10*dz10;
 778             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 779             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 780             rsq20            = dx20*dx20+dy20*dy20+dz20*dz20;
 781             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 782             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 783
 784             rinv00           = gmx_invsqrt(rsq00);
 785             rinv01           = gmx_invsqrt(rsq01);
 786             rinv02           = gmx_invsqrt(rsq02);
 787             rinv10           = gmx_invsqrt(rsq10);
 788             rinv11           = gmx_invsqrt(rsq11);
 789             rinv12           = gmx_invsqrt(rsq12);
 790             rinv20           = gmx_invsqrt(rsq20);
 791             rinv21           = gmx_invsqrt(rsq21);
 792             rinv22           = gmx_invsqrt(rsq22);
 793
 794             rinvsq00         = rinv00*rinv00;
 795             rinvsq01         = rinv01*rinv01;
 796             rinvsq02         = rinv02*rinv02;
 797             rinvsq10         = rinv10*rinv10;
 798             rinvsq11         = rinv11*rinv11;
 799             rinvsq12         = rinv12*rinv12;
 800             rinvsq20         = rinv20*rinv20;
 801             rinvsq21         = rinv21*rinv21;
 802             rinvsq22         = rinv22*rinv22;
 803
 804             /**************************
 805              * CALCULATE INTERACTIONS *
 806              **************************/
 807
 808             r00              = rsq00*rinv00;
 809
 810             /* Calculate table index by multiplying r with table scale and truncate to integer */
 811             rt               = r00*vftabscale;
 812             vfitab           = rt;
 813             vfeps            = rt-vfitab;
 814             vfitab           = 2*4*vfitab;
 815
 816             /* REACTION-FIELD ELECTROSTATICS */
 817             felec            = qq00*(rinv00*rinvsq00-krf2);
 818
 819             /* CUBIC SPLINE TABLE DISPERSION */
 820             vfitab          += 0;
 821             F                = vftab[vfitab+1];
 822             Geps             = vfeps*vftab[vfitab+2];
 823             Heps2            = vfeps*vfeps*vftab[vfitab+3];
 824             Fp               = F+Geps+Heps2;
 825             FF               = Fp+Geps+2.0*Heps2;
 826             fvdw6            = c6_00*FF;
 827
 828             /* CUBIC SPLINE TABLE REPULSION */
 829             F                = vftab[vfitab+5];
 830             Geps             = vfeps*vftab[vfitab+6];
 831             Heps2            = vfeps*vfeps*vftab[vfitab+7];
 832             Fp               = F+Geps+Heps2;
 833             FF               = Fp+Geps+2.0*Heps2;
 834             fvdw12           = c12_00*FF;
 835             fvdw             = -(fvdw6+fvdw12)*vftabscale*rinv00;
 836
 837             fscal            = felec+fvdw;
 838
 839             /* Calculate temporary vectorial force */
 840             tx               = fscal*dx00;
 841             ty               = fscal*dy00;
 842             tz               = fscal*dz00;
 843
 844             /* Update vectorial force */
 845             fix0            += tx;
 846             fiy0            += ty;
 847             fiz0            += tz;
 848             f[j_coord_offset+DIM*0+XX] -= tx;
 849             f[j_coord_offset+DIM*0+YY] -= ty;
 850             f[j_coord_offset+DIM*0+ZZ] -= tz;
 851
 852             /**************************
 853              * CALCULATE INTERACTIONS *
 854              **************************/
 855
 856             /* REACTION-FIELD ELECTROSTATICS */
 857             felec            = qq01*(rinv01*rinvsq01-krf2);
 858
 859             fscal            = felec;
 860
 861             /* Calculate temporary vectorial force */
 862             tx               = fscal*dx01;
 863             ty               = fscal*dy01;
 864             tz               = fscal*dz01;
 865
 866             /* Update vectorial force */
 867             fix0            += tx;
 868             fiy0            += ty;
 869             fiz0            += tz;
 870             f[j_coord_offset+DIM*1+XX] -= tx;
 871             f[j_coord_offset+DIM*1+YY] -= ty;
 872             f[j_coord_offset+DIM*1+ZZ] -= tz;
 873
 874             /**************************
 875              * CALCULATE INTERACTIONS *
 876              **************************/
 877
 878             /* REACTION-FIELD ELECTROSTATICS */
 879             felec            = qq02*(rinv02*rinvsq02-krf2);
 880
 881             fscal            = felec;
 882
 883             /* Calculate temporary vectorial force */
 884             tx               = fscal*dx02;
 885             ty               = fscal*dy02;
 886             tz               = fscal*dz02;
 887
 888             /* Update vectorial force */
 889             fix0            += tx;
 890             fiy0            += ty;
 891             fiz0            += tz;
 892             f[j_coord_offset+DIM*2+XX] -= tx;
 893             f[j_coord_offset+DIM*2+YY] -= ty;
 894             f[j_coord_offset+DIM*2+ZZ] -= tz;
 895
 896             /**************************
 897              * CALCULATE INTERACTIONS *
 898              **************************/
 899
 900             /* REACTION-FIELD ELECTROSTATICS */
 901             felec            = qq10*(rinv10*rinvsq10-krf2);
 902
 903             fscal            = felec;
 904
 905             /* Calculate temporary vectorial force */
 906             tx               = fscal*dx10;
 907             ty               = fscal*dy10;
 908             tz               = fscal*dz10;
 909
 910             /* Update vectorial force */
 911             fix1            += tx;
 912             fiy1            += ty;
 913             fiz1            += tz;
 914             f[j_coord_offset+DIM*0+XX] -= tx;
 915             f[j_coord_offset+DIM*0+YY] -= ty;
 916             f[j_coord_offset+DIM*0+ZZ] -= tz;
 917
 918             /**************************
 919              * CALCULATE INTERACTIONS *
 920              **************************/
 921
 922             /* REACTION-FIELD ELECTROSTATICS */
 923             felec            = qq11*(rinv11*rinvsq11-krf2);
 924
 925             fscal            = felec;
 926
 927             /* Calculate temporary vectorial force */
 928             tx               = fscal*dx11;
 929             ty               = fscal*dy11;
 930             tz               = fscal*dz11;
 931
 932             /* Update vectorial force */
 933             fix1            += tx;
 934             fiy1            += ty;
 935             fiz1            += 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             /* REACTION-FIELD ELECTROSTATICS */
 945             felec            = qq12*(rinv12*rinvsq12-krf2);
 946
 947             fscal            = felec;
 948
 949             /* Calculate temporary vectorial force */
 950             tx               = fscal*dx12;
 951             ty               = fscal*dy12;
 952             tz               = fscal*dz12;
 953
 954             /* Update vectorial force */
 955             fix1            += tx;
 956             fiy1            += ty;
 957             fiz1            += tz;
 958             f[j_coord_offset+DIM*2+XX] -= tx;
 959             f[j_coord_offset+DIM*2+YY] -= ty;
 960             f[j_coord_offset+DIM*2+ZZ] -= tz;
 961
 962             /**************************
 963              * CALCULATE INTERACTIONS *
 964              **************************/
 965
 966             /* REACTION-FIELD ELECTROSTATICS */
 967             felec            = qq20*(rinv20*rinvsq20-krf2);
 968
 969             fscal            = felec;
 970
 971             /* Calculate temporary vectorial force */
 972             tx               = fscal*dx20;
 973             ty               = fscal*dy20;
 974             tz               = fscal*dz20;
 975
 976             /* Update vectorial force */
 977             fix2            += tx;
 978             fiy2            += ty;
 979             fiz2            += tz;
 980             f[j_coord_offset+DIM*0+XX] -= tx;
 981             f[j_coord_offset+DIM*0+YY] -= ty;
 982             f[j_coord_offset+DIM*0+ZZ] -= tz;
 983
 984             /**************************
 985              * CALCULATE INTERACTIONS *
 986              **************************/
 987
 988             /* REACTION-FIELD ELECTROSTATICS */
 989             felec            = qq21*(rinv21*rinvsq21-krf2);
 990
 991             fscal            = felec;
 992
 993             /* Calculate temporary vectorial force */
 994             tx               = fscal*dx21;
 995             ty               = fscal*dy21;
 996             tz               = fscal*dz21;
 997
 998             /* Update vectorial force */
 999             fix2            += tx;
1000             fiy2            += ty;
1001             fiz2            += tz;
1002             f[j_coord_offset+DIM*1+XX] -= tx;
1003             f[j_coord_offset+DIM*1+YY] -= ty;
1004             f[j_coord_offset+DIM*1+ZZ] -= tz;
1005
1006             /**************************
1007              * CALCULATE INTERACTIONS *
1008              **************************/
1009
1010             /* REACTION-FIELD ELECTROSTATICS */
1011             felec            = qq22*(rinv22*rinvsq22-krf2);
1012
1013             fscal            = felec;
1014
1015             /* Calculate temporary vectorial force */
1016             tx               = fscal*dx22;
1017             ty               = fscal*dy22;
1018             tz               = fscal*dz22;
1019
1020             /* Update vectorial force */
1021             fix2            += tx;
1022             fiy2            += ty;
1023             fiz2            += tz;
1024             f[j_coord_offset+DIM*2+XX] -= tx;
1025             f[j_coord_offset+DIM*2+YY] -= ty;
1026             f[j_coord_offset+DIM*2+ZZ] -= tz;
1027
1028             /* Inner loop uses 260 flops */
1029         }
1030         /* End of innermost loop */
1031
1032         tx = ty = tz = 0;
1033         f[i_coord_offset+DIM*0+XX] += fix0;
1034         f[i_coord_offset+DIM*0+YY] += fiy0;
1035         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1036         tx                         += fix0;
1037         ty                         += fiy0;
1038         tz                         += fiz0;
1039         f[i_coord_offset+DIM*1+XX] += fix1;
1040         f[i_coord_offset+DIM*1+YY] += fiy1;
1041         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1042         tx                         += fix1;
1043         ty                         += fiy1;
1044         tz                         += fiz1;
1045         f[i_coord_offset+DIM*2+XX] += fix2;
1046         f[i_coord_offset+DIM*2+YY] += fiy2;
1047         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1048         tx                         += fix2;
1049         ty                         += fiy2;
1050         tz                         += fiz2;
1051         fshift[i_shift_offset+XX]  += tx;
1052         fshift[i_shift_offset+YY]  += ty;
1053         fshift[i_shift_offset+ZZ]  += tz;
1054
1055         /* Increment number of inner iterations */
1056         inneriter                  += j_index_end - j_index_start;
1057
1058         /* Outer loop uses 30 flops */
1059     }
1060
1061     /* Increment number of outer iterations */
1062     outeriter        += nri;
1063
1064     /* Update outer/inner flops */
1065
1066     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*30 + inneriter*260);
1067 }