src/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecCoul_VdwBham_GeomW4W4_c.c

   1 /*
   2  * Note: this file was generated by the Gromacs c kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 /*
  34  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwBham_GeomW4W4_VF_c
  35  * Electrostatics interaction: Coulomb
  36  * VdW interaction:            Buckingham
  37  * Geometry:                   Water4-Water4
  38  * Calculate force/pot:        PotentialAndForce
  39  */
  40 void
  41 nb_kernel_ElecCoul_VdwBham_GeomW4W4_VF_c
  42                     (t_nblist * gmx_restrict                nlist,
  43                      rvec * gmx_restrict                    xx,
  44                      rvec * gmx_restrict                    ff,
  45                      t_forcerec * gmx_restrict              fr,
  46                      t_mdatoms * gmx_restrict               mdatoms,
  47                      nb_kernel_data_t * gmx_restrict        kernel_data,
  48                      t_nrnb * gmx_restrict                  nrnb)
  49 {
  50     int              i_shift_offset,i_coord_offset,j_coord_offset;
  51     int              j_index_start,j_index_end;
  52     int              nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter;
  53     real             shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2;
  54     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  55     real             *shiftvec,*fshift,*x,*f;
  56     int              vdwioffset0;
  57     real             ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  58     int              vdwioffset1;
  59     real             ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  60     int              vdwioffset2;
  61     real             ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  62     int              vdwioffset3;
  63     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  64     int              vdwjidx0;
  65     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  66     int              vdwjidx1;
  67     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  68     int              vdwjidx2;
  69     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  70     int              vdwjidx3;
  71     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  72     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
  73     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
  74     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
  75     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
  76     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
  77     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
  78     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
  79     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
  80     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
  81     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
  82     real             velec,felec,velecsum,facel,crf,krf,krf2;
  83     real             *charge;
  84     int              nvdwtype;
  85     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
  86     int              *vdwtype;
  87     real             *vdwparam;
  88
  89     x                = xx[0];
  90     f                = ff[0];
  91
  92     nri              = nlist->nri;
  93     iinr             = nlist->iinr;
  94     jindex           = nlist->jindex;
  95     jjnr             = nlist->jjnr;
  96     shiftidx         = nlist->shift;
  97     gid              = nlist->gid;
  98     shiftvec         = fr->shift_vec[0];
  99     fshift           = fr->fshift[0];
 100     facel            = fr->epsfac;
 101     charge           = mdatoms->chargeA;
 102     nvdwtype         = fr->ntype;
 103     vdwparam         = fr->nbfp;
 104     vdwtype          = mdatoms->typeA;
 105
 106     /* Setup water-specific parameters */
 107     inr              = nlist->iinr[0];
 108     iq1              = facel*charge[inr+1];
 109     iq2              = facel*charge[inr+2];
 110     iq3              = facel*charge[inr+3];
 111     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 112
 113     jq1              = charge[inr+1];
 114     jq2              = charge[inr+2];
 115     jq3              = charge[inr+3];
 116     vdwjidx0         = 3*vdwtype[inr+0];
 117     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 118     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 119     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 120     qq11             = iq1*jq1;
 121     qq12             = iq1*jq2;
 122     qq13             = iq1*jq3;
 123     qq21             = iq2*jq1;
 124     qq22             = iq2*jq2;
 125     qq23             = iq2*jq3;
 126     qq31             = iq3*jq1;
 127     qq32             = iq3*jq2;
 128     qq33             = iq3*jq3;
 129
 130     outeriter        = 0;
 131     inneriter        = 0;
 132
 133     /* Start outer loop over neighborlists */
 134     for(iidx=0; iidx<nri; iidx++)
 135     {
 136         /* Load shift vector for this list */
 137         i_shift_offset   = DIM*shiftidx[iidx];
 138         shX              = shiftvec[i_shift_offset+XX];
 139         shY              = shiftvec[i_shift_offset+YY];
 140         shZ              = shiftvec[i_shift_offset+ZZ];
 141
 142         /* Load limits for loop over neighbors */
 143         j_index_start    = jindex[iidx];
 144         j_index_end      = jindex[iidx+1];
 145
 146         /* Get outer coordinate index */
 147         inr              = iinr[iidx];
 148         i_coord_offset   = DIM*inr;
 149
 150         /* Load i particle coords and add shift vector */
 151         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 152         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 153         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 154         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 155         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 156         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 157         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 158         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 159         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 160         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 161         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 162         iz3              = shZ + x[i_coord_offset+DIM*3+ZZ];
 163
 164         fix0             = 0.0;
 165         fiy0             = 0.0;
 166         fiz0             = 0.0;
 167         fix1             = 0.0;
 168         fiy1             = 0.0;
 169         fiz1             = 0.0;
 170         fix2             = 0.0;
 171         fiy2             = 0.0;
 172         fiz2             = 0.0;
 173         fix3             = 0.0;
 174         fiy3             = 0.0;
 175         fiz3             = 0.0;
 176
 177         /* Reset potential sums */
 178         velecsum         = 0.0;
 179         vvdwsum          = 0.0;
 180
 181         /* Start inner kernel loop */
 182         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 183         {
 184             /* Get j neighbor index, and coordinate index */
 185             jnr              = jjnr[jidx];
 186             j_coord_offset   = DIM*jnr;
 187
 188             /* load j atom coordinates */
 189             jx0              = x[j_coord_offset+DIM*0+XX];
 190             jy0              = x[j_coord_offset+DIM*0+YY];
 191             jz0              = x[j_coord_offset+DIM*0+ZZ];
 192             jx1              = x[j_coord_offset+DIM*1+XX];
 193             jy1              = x[j_coord_offset+DIM*1+YY];
 194             jz1              = x[j_coord_offset+DIM*1+ZZ];
 195             jx2              = x[j_coord_offset+DIM*2+XX];
 196             jy2              = x[j_coord_offset+DIM*2+YY];
 197             jz2              = x[j_coord_offset+DIM*2+ZZ];
 198             jx3              = x[j_coord_offset+DIM*3+XX];
 199             jy3              = x[j_coord_offset+DIM*3+YY];
 200             jz3              = x[j_coord_offset+DIM*3+ZZ];
 201
 202             /* Calculate displacement vector */
 203             dx00             = ix0 - jx0;
 204             dy00             = iy0 - jy0;
 205             dz00             = iz0 - jz0;
 206             dx11             = ix1 - jx1;
 207             dy11             = iy1 - jy1;
 208             dz11             = iz1 - jz1;
 209             dx12             = ix1 - jx2;
 210             dy12             = iy1 - jy2;
 211             dz12             = iz1 - jz2;
 212             dx13             = ix1 - jx3;
 213             dy13             = iy1 - jy3;
 214             dz13             = iz1 - jz3;
 215             dx21             = ix2 - jx1;
 216             dy21             = iy2 - jy1;
 217             dz21             = iz2 - jz1;
 218             dx22             = ix2 - jx2;
 219             dy22             = iy2 - jy2;
 220             dz22             = iz2 - jz2;
 221             dx23             = ix2 - jx3;
 222             dy23             = iy2 - jy3;
 223             dz23             = iz2 - jz3;
 224             dx31             = ix3 - jx1;
 225             dy31             = iy3 - jy1;
 226             dz31             = iz3 - jz1;
 227             dx32             = ix3 - jx2;
 228             dy32             = iy3 - jy2;
 229             dz32             = iz3 - jz2;
 230             dx33             = ix3 - jx3;
 231             dy33             = iy3 - jy3;
 232             dz33             = iz3 - jz3;
 233
 234             /* Calculate squared distance and things based on it */
 235             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 236             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 237             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 238             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 239             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 240             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 241             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 242             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 243             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 244             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 245
 246             rinv00           = gmx_invsqrt(rsq00);
 247             rinv11           = gmx_invsqrt(rsq11);
 248             rinv12           = gmx_invsqrt(rsq12);
 249             rinv13           = gmx_invsqrt(rsq13);
 250             rinv21           = gmx_invsqrt(rsq21);
 251             rinv22           = gmx_invsqrt(rsq22);
 252             rinv23           = gmx_invsqrt(rsq23);
 253             rinv31           = gmx_invsqrt(rsq31);
 254             rinv32           = gmx_invsqrt(rsq32);
 255             rinv33           = gmx_invsqrt(rsq33);
 256
 257             rinvsq00         = rinv00*rinv00;
 258             rinvsq11         = rinv11*rinv11;
 259             rinvsq12         = rinv12*rinv12;
 260             rinvsq13         = rinv13*rinv13;
 261             rinvsq21         = rinv21*rinv21;
 262             rinvsq22         = rinv22*rinv22;
 263             rinvsq23         = rinv23*rinv23;
 264             rinvsq31         = rinv31*rinv31;
 265             rinvsq32         = rinv32*rinv32;
 266             rinvsq33         = rinv33*rinv33;
 267
 268             /**************************
 269              * CALCULATE INTERACTIONS *
 270              **************************/
 271
 272             r00              = rsq00*rinv00;
 273
 274             /* BUCKINGHAM DISPERSION/REPULSION */
 275             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 276             vvdw6            = c6_00*rinvsix;
 277             br               = cexp2_00*r00;
 278             vvdwexp          = cexp1_00*exp(-br);
 279             vvdw             = vvdwexp - vvdw6*(1.0/6.0);
 280             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 281
 282             /* Update potential sums from outer loop */
 283             vvdwsum         += vvdw;
 284
 285             fscal            = fvdw;
 286
 287             /* Calculate temporary vectorial force */
 288             tx               = fscal*dx00;
 289             ty               = fscal*dy00;
 290             tz               = fscal*dz00;
 291
 292             /* Update vectorial force */
 293             fix0            += tx;
 294             fiy0            += ty;
 295             fiz0            += tz;
 296             f[j_coord_offset+DIM*0+XX] -= tx;
 297             f[j_coord_offset+DIM*0+YY] -= ty;
 298             f[j_coord_offset+DIM*0+ZZ] -= tz;
 299
 300             /**************************
 301              * CALCULATE INTERACTIONS *
 302              **************************/
 303
 304             /* COULOMB ELECTROSTATICS */
 305             velec            = qq11*rinv11;
 306             felec            = velec*rinvsq11;
 307
 308             /* Update potential sums from outer loop */
 309             velecsum        += velec;
 310
 311             fscal            = felec;
 312
 313             /* Calculate temporary vectorial force */
 314             tx               = fscal*dx11;
 315             ty               = fscal*dy11;
 316             tz               = fscal*dz11;
 317
 318             /* Update vectorial force */
 319             fix1            += tx;
 320             fiy1            += ty;
 321             fiz1            += tz;
 322             f[j_coord_offset+DIM*1+XX] -= tx;
 323             f[j_coord_offset+DIM*1+YY] -= ty;
 324             f[j_coord_offset+DIM*1+ZZ] -= tz;
 325
 326             /**************************
 327              * CALCULATE INTERACTIONS *
 328              **************************/
 329
 330             /* COULOMB ELECTROSTATICS */
 331             velec            = qq12*rinv12;
 332             felec            = velec*rinvsq12;
 333
 334             /* Update potential sums from outer loop */
 335             velecsum        += velec;
 336
 337             fscal            = felec;
 338
 339             /* Calculate temporary vectorial force */
 340             tx               = fscal*dx12;
 341             ty               = fscal*dy12;
 342             tz               = fscal*dz12;
 343
 344             /* Update vectorial force */
 345             fix1            += tx;
 346             fiy1            += ty;
 347             fiz1            += tz;
 348             f[j_coord_offset+DIM*2+XX] -= tx;
 349             f[j_coord_offset+DIM*2+YY] -= ty;
 350             f[j_coord_offset+DIM*2+ZZ] -= tz;
 351
 352             /**************************
 353              * CALCULATE INTERACTIONS *
 354              **************************/
 355
 356             /* COULOMB ELECTROSTATICS */
 357             velec            = qq13*rinv13;
 358             felec            = velec*rinvsq13;
 359
 360             /* Update potential sums from outer loop */
 361             velecsum        += velec;
 362
 363             fscal            = felec;
 364
 365             /* Calculate temporary vectorial force */
 366             tx               = fscal*dx13;
 367             ty               = fscal*dy13;
 368             tz               = fscal*dz13;
 369
 370             /* Update vectorial force */
 371             fix1            += tx;
 372             fiy1            += ty;
 373             fiz1            += tz;
 374             f[j_coord_offset+DIM*3+XX] -= tx;
 375             f[j_coord_offset+DIM*3+YY] -= ty;
 376             f[j_coord_offset+DIM*3+ZZ] -= tz;
 377
 378             /**************************
 379              * CALCULATE INTERACTIONS *
 380              **************************/
 381
 382             /* COULOMB ELECTROSTATICS */
 383             velec            = qq21*rinv21;
 384             felec            = velec*rinvsq21;
 385
 386             /* Update potential sums from outer loop */
 387             velecsum        += velec;
 388
 389             fscal            = felec;
 390
 391             /* Calculate temporary vectorial force */
 392             tx               = fscal*dx21;
 393             ty               = fscal*dy21;
 394             tz               = fscal*dz21;
 395
 396             /* Update vectorial force */
 397             fix2            += tx;
 398             fiy2            += ty;
 399             fiz2            += tz;
 400             f[j_coord_offset+DIM*1+XX] -= tx;
 401             f[j_coord_offset+DIM*1+YY] -= ty;
 402             f[j_coord_offset+DIM*1+ZZ] -= tz;
 403
 404             /**************************
 405              * CALCULATE INTERACTIONS *
 406              **************************/
 407
 408             /* COULOMB ELECTROSTATICS */
 409             velec            = qq22*rinv22;
 410             felec            = velec*rinvsq22;
 411
 412             /* Update potential sums from outer loop */
 413             velecsum        += velec;
 414
 415             fscal            = felec;
 416
 417             /* Calculate temporary vectorial force */
 418             tx               = fscal*dx22;
 419             ty               = fscal*dy22;
 420             tz               = fscal*dz22;
 421
 422             /* Update vectorial force */
 423             fix2            += tx;
 424             fiy2            += ty;
 425             fiz2            += tz;
 426             f[j_coord_offset+DIM*2+XX] -= tx;
 427             f[j_coord_offset+DIM*2+YY] -= ty;
 428             f[j_coord_offset+DIM*2+ZZ] -= tz;
 429
 430             /**************************
 431              * CALCULATE INTERACTIONS *
 432              **************************/
 433
 434             /* COULOMB ELECTROSTATICS */
 435             velec            = qq23*rinv23;
 436             felec            = velec*rinvsq23;
 437
 438             /* Update potential sums from outer loop */
 439             velecsum        += velec;
 440
 441             fscal            = felec;
 442
 443             /* Calculate temporary vectorial force */
 444             tx               = fscal*dx23;
 445             ty               = fscal*dy23;
 446             tz               = fscal*dz23;
 447
 448             /* Update vectorial force */
 449             fix2            += tx;
 450             fiy2            += ty;
 451             fiz2            += tz;
 452             f[j_coord_offset+DIM*3+XX] -= tx;
 453             f[j_coord_offset+DIM*3+YY] -= ty;
 454             f[j_coord_offset+DIM*3+ZZ] -= tz;
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             /* COULOMB ELECTROSTATICS */
 461             velec            = qq31*rinv31;
 462             felec            = velec*rinvsq31;
 463
 464             /* Update potential sums from outer loop */
 465             velecsum        += velec;
 466
 467             fscal            = felec;
 468
 469             /* Calculate temporary vectorial force */
 470             tx               = fscal*dx31;
 471             ty               = fscal*dy31;
 472             tz               = fscal*dz31;
 473
 474             /* Update vectorial force */
 475             fix3            += tx;
 476             fiy3            += ty;
 477             fiz3            += tz;
 478             f[j_coord_offset+DIM*1+XX] -= tx;
 479             f[j_coord_offset+DIM*1+YY] -= ty;
 480             f[j_coord_offset+DIM*1+ZZ] -= tz;
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             /* COULOMB ELECTROSTATICS */
 487             velec            = qq32*rinv32;
 488             felec            = velec*rinvsq32;
 489
 490             /* Update potential sums from outer loop */
 491             velecsum        += velec;
 492
 493             fscal            = felec;
 494
 495             /* Calculate temporary vectorial force */
 496             tx               = fscal*dx32;
 497             ty               = fscal*dy32;
 498             tz               = fscal*dz32;
 499
 500             /* Update vectorial force */
 501             fix3            += tx;
 502             fiy3            += ty;
 503             fiz3            += tz;
 504             f[j_coord_offset+DIM*2+XX] -= tx;
 505             f[j_coord_offset+DIM*2+YY] -= ty;
 506             f[j_coord_offset+DIM*2+ZZ] -= tz;
 507
 508             /**************************
 509              * CALCULATE INTERACTIONS *
 510              **************************/
 511
 512             /* COULOMB ELECTROSTATICS */
 513             velec            = qq33*rinv33;
 514             felec            = velec*rinvsq33;
 515
 516             /* Update potential sums from outer loop */
 517             velecsum        += velec;
 518
 519             fscal            = felec;
 520
 521             /* Calculate temporary vectorial force */
 522             tx               = fscal*dx33;
 523             ty               = fscal*dy33;
 524             tz               = fscal*dz33;
 525
 526             /* Update vectorial force */
 527             fix3            += tx;
 528             fiy3            += ty;
 529             fiz3            += tz;
 530             f[j_coord_offset+DIM*3+XX] -= tx;
 531             f[j_coord_offset+DIM*3+YY] -= ty;
 532             f[j_coord_offset+DIM*3+ZZ] -= tz;
 533
 534             /* Inner loop uses 304 flops */
 535         }
 536         /* End of innermost loop */
 537
 538         tx = ty = tz = 0;
 539         f[i_coord_offset+DIM*0+XX] += fix0;
 540         f[i_coord_offset+DIM*0+YY] += fiy0;
 541         f[i_coord_offset+DIM*0+ZZ] += fiz0;
 542         tx                         += fix0;
 543         ty                         += fiy0;
 544         tz                         += fiz0;
 545         f[i_coord_offset+DIM*1+XX] += fix1;
 546         f[i_coord_offset+DIM*1+YY] += fiy1;
 547         f[i_coord_offset+DIM*1+ZZ] += fiz1;
 548         tx                         += fix1;
 549         ty                         += fiy1;
 550         tz                         += fiz1;
 551         f[i_coord_offset+DIM*2+XX] += fix2;
 552         f[i_coord_offset+DIM*2+YY] += fiy2;
 553         f[i_coord_offset+DIM*2+ZZ] += fiz2;
 554         tx                         += fix2;
 555         ty                         += fiy2;
 556         tz                         += fiz2;
 557         f[i_coord_offset+DIM*3+XX] += fix3;
 558         f[i_coord_offset+DIM*3+YY] += fiy3;
 559         f[i_coord_offset+DIM*3+ZZ] += fiz3;
 560         tx                         += fix3;
 561         ty                         += fiy3;
 562         tz                         += fiz3;
 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 41 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_W4W4_VF,outeriter*41 + inneriter*304);
 584 }
 585 /*
 586  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwBham_GeomW4W4_F_c
 587  * Electrostatics interaction: Coulomb
 588  * VdW interaction:            Buckingham
 589  * Geometry:                   Water4-Water4
 590  * Calculate force/pot:        Force
 591  */
 592 void
 593 nb_kernel_ElecCoul_VdwBham_GeomW4W4_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_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              vdwioffset3;
 615     real             ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 616     int              vdwjidx0;
 617     real             jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 618     int              vdwjidx1;
 619     real             jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 620     int              vdwjidx2;
 621     real             jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 622     int              vdwjidx3;
 623     real             jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 624     real             dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00;
 625     real             dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11;
 626     real             dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12;
 627     real             dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13;
 628     real             dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21;
 629     real             dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22;
 630     real             dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23;
 631     real             dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31;
 632     real             dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32;
 633     real             dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33;
 634     real             velec,felec,velecsum,facel,crf,krf,krf2;
 635     real             *charge;
 636     int              nvdwtype;
 637     real             rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6;
 638     int              *vdwtype;
 639     real             *vdwparam;
 640
 641     x                = xx[0];
 642     f                = ff[0];
 643
 644     nri              = nlist->nri;
 645     iinr             = nlist->iinr;
 646     jindex           = nlist->jindex;
 647     jjnr             = nlist->jjnr;
 648     shiftidx         = nlist->shift;
 649     gid              = nlist->gid;
 650     shiftvec         = fr->shift_vec[0];
 651     fshift           = fr->fshift[0];
 652     facel            = fr->epsfac;
 653     charge           = mdatoms->chargeA;
 654     nvdwtype         = fr->ntype;
 655     vdwparam         = fr->nbfp;
 656     vdwtype          = mdatoms->typeA;
 657
 658     /* Setup water-specific parameters */
 659     inr              = nlist->iinr[0];
 660     iq1              = facel*charge[inr+1];
 661     iq2              = facel*charge[inr+2];
 662     iq3              = facel*charge[inr+3];
 663     vdwioffset0      = 3*nvdwtype*vdwtype[inr+0];
 664
 665     jq1              = charge[inr+1];
 666     jq2              = charge[inr+2];
 667     jq3              = charge[inr+3];
 668     vdwjidx0         = 3*vdwtype[inr+0];
 669     c6_00            = vdwparam[vdwioffset0+vdwjidx0];
 670     cexp1_00         = vdwparam[vdwioffset0+vdwjidx0+1];
 671     cexp2_00         = vdwparam[vdwioffset0+vdwjidx0+2];
 672     qq11             = iq1*jq1;
 673     qq12             = iq1*jq2;
 674     qq13             = iq1*jq3;
 675     qq21             = iq2*jq1;
 676     qq22             = iq2*jq2;
 677     qq23             = iq2*jq3;
 678     qq31             = iq3*jq1;
 679     qq32             = iq3*jq2;
 680     qq33             = iq3*jq3;
 681
 682     outeriter        = 0;
 683     inneriter        = 0;
 684
 685     /* Start outer loop over neighborlists */
 686     for(iidx=0; iidx<nri; iidx++)
 687     {
 688         /* Load shift vector for this list */
 689         i_shift_offset   = DIM*shiftidx[iidx];
 690         shX              = shiftvec[i_shift_offset+XX];
 691         shY              = shiftvec[i_shift_offset+YY];
 692         shZ              = shiftvec[i_shift_offset+ZZ];
 693
 694         /* Load limits for loop over neighbors */
 695         j_index_start    = jindex[iidx];
 696         j_index_end      = jindex[iidx+1];
 697
 698         /* Get outer coordinate index */
 699         inr              = iinr[iidx];
 700         i_coord_offset   = DIM*inr;
 701
 702         /* Load i particle coords and add shift vector */
 703         ix0              = shX + x[i_coord_offset+DIM*0+XX];
 704         iy0              = shY + x[i_coord_offset+DIM*0+YY];
 705         iz0              = shZ + x[i_coord_offset+DIM*0+ZZ];
 706         ix1              = shX + x[i_coord_offset+DIM*1+XX];
 707         iy1              = shY + x[i_coord_offset+DIM*1+YY];
 708         iz1              = shZ + x[i_coord_offset+DIM*1+ZZ];
 709         ix2              = shX + x[i_coord_offset+DIM*2+XX];
 710         iy2              = shY + x[i_coord_offset+DIM*2+YY];
 711         iz2              = shZ + x[i_coord_offset+DIM*2+ZZ];
 712         ix3              = shX + x[i_coord_offset+DIM*3+XX];
 713         iy3              = shY + x[i_coord_offset+DIM*3+YY];
 714         iz3              = shZ + x[i_coord_offset+DIM*3+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         fix3             = 0.0;
 726         fiy3             = 0.0;
 727         fiz3             = 0.0;
 728
 729         /* Start inner kernel loop */
 730         for(jidx=j_index_start; jidx<j_index_end; jidx++)
 731         {
 732             /* Get j neighbor index, and coordinate index */
 733             jnr              = jjnr[jidx];
 734             j_coord_offset   = DIM*jnr;
 735
 736             /* load j atom coordinates */
 737             jx0              = x[j_coord_offset+DIM*0+XX];
 738             jy0              = x[j_coord_offset+DIM*0+YY];
 739             jz0              = x[j_coord_offset+DIM*0+ZZ];
 740             jx1              = x[j_coord_offset+DIM*1+XX];
 741             jy1              = x[j_coord_offset+DIM*1+YY];
 742             jz1              = x[j_coord_offset+DIM*1+ZZ];
 743             jx2              = x[j_coord_offset+DIM*2+XX];
 744             jy2              = x[j_coord_offset+DIM*2+YY];
 745             jz2              = x[j_coord_offset+DIM*2+ZZ];
 746             jx3              = x[j_coord_offset+DIM*3+XX];
 747             jy3              = x[j_coord_offset+DIM*3+YY];
 748             jz3              = x[j_coord_offset+DIM*3+ZZ];
 749
 750             /* Calculate displacement vector */
 751             dx00             = ix0 - jx0;
 752             dy00             = iy0 - jy0;
 753             dz00             = iz0 - jz0;
 754             dx11             = ix1 - jx1;
 755             dy11             = iy1 - jy1;
 756             dz11             = iz1 - jz1;
 757             dx12             = ix1 - jx2;
 758             dy12             = iy1 - jy2;
 759             dz12             = iz1 - jz2;
 760             dx13             = ix1 - jx3;
 761             dy13             = iy1 - jy3;
 762             dz13             = iz1 - jz3;
 763             dx21             = ix2 - jx1;
 764             dy21             = iy2 - jy1;
 765             dz21             = iz2 - jz1;
 766             dx22             = ix2 - jx2;
 767             dy22             = iy2 - jy2;
 768             dz22             = iz2 - jz2;
 769             dx23             = ix2 - jx3;
 770             dy23             = iy2 - jy3;
 771             dz23             = iz2 - jz3;
 772             dx31             = ix3 - jx1;
 773             dy31             = iy3 - jy1;
 774             dz31             = iz3 - jz1;
 775             dx32             = ix3 - jx2;
 776             dy32             = iy3 - jy2;
 777             dz32             = iz3 - jz2;
 778             dx33             = ix3 - jx3;
 779             dy33             = iy3 - jy3;
 780             dz33             = iz3 - jz3;
 781
 782             /* Calculate squared distance and things based on it */
 783             rsq00            = dx00*dx00+dy00*dy00+dz00*dz00;
 784             rsq11            = dx11*dx11+dy11*dy11+dz11*dz11;
 785             rsq12            = dx12*dx12+dy12*dy12+dz12*dz12;
 786             rsq13            = dx13*dx13+dy13*dy13+dz13*dz13;
 787             rsq21            = dx21*dx21+dy21*dy21+dz21*dz21;
 788             rsq22            = dx22*dx22+dy22*dy22+dz22*dz22;
 789             rsq23            = dx23*dx23+dy23*dy23+dz23*dz23;
 790             rsq31            = dx31*dx31+dy31*dy31+dz31*dz31;
 791             rsq32            = dx32*dx32+dy32*dy32+dz32*dz32;
 792             rsq33            = dx33*dx33+dy33*dy33+dz33*dz33;
 793
 794             rinv00           = gmx_invsqrt(rsq00);
 795             rinv11           = gmx_invsqrt(rsq11);
 796             rinv12           = gmx_invsqrt(rsq12);
 797             rinv13           = gmx_invsqrt(rsq13);
 798             rinv21           = gmx_invsqrt(rsq21);
 799             rinv22           = gmx_invsqrt(rsq22);
 800             rinv23           = gmx_invsqrt(rsq23);
 801             rinv31           = gmx_invsqrt(rsq31);
 802             rinv32           = gmx_invsqrt(rsq32);
 803             rinv33           = gmx_invsqrt(rsq33);
 804
 805             rinvsq00         = rinv00*rinv00;
 806             rinvsq11         = rinv11*rinv11;
 807             rinvsq12         = rinv12*rinv12;
 808             rinvsq13         = rinv13*rinv13;
 809             rinvsq21         = rinv21*rinv21;
 810             rinvsq22         = rinv22*rinv22;
 811             rinvsq23         = rinv23*rinv23;
 812             rinvsq31         = rinv31*rinv31;
 813             rinvsq32         = rinv32*rinv32;
 814             rinvsq33         = rinv33*rinv33;
 815
 816             /**************************
 817              * CALCULATE INTERACTIONS *
 818              **************************/
 819
 820             r00              = rsq00*rinv00;
 821
 822             /* BUCKINGHAM DISPERSION/REPULSION */
 823             rinvsix          = rinvsq00*rinvsq00*rinvsq00;
 824             vvdw6            = c6_00*rinvsix;
 825             br               = cexp2_00*r00;
 826             vvdwexp          = cexp1_00*exp(-br);
 827             fvdw             = (br*vvdwexp-vvdw6)*rinvsq00;
 828
 829             fscal            = fvdw;
 830
 831             /* Calculate temporary vectorial force */
 832             tx               = fscal*dx00;
 833             ty               = fscal*dy00;
 834             tz               = fscal*dz00;
 835
 836             /* Update vectorial force */
 837             fix0            += tx;
 838             fiy0            += ty;
 839             fiz0            += tz;
 840             f[j_coord_offset+DIM*0+XX] -= tx;
 841             f[j_coord_offset+DIM*0+YY] -= ty;
 842             f[j_coord_offset+DIM*0+ZZ] -= tz;
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* COULOMB ELECTROSTATICS */
 849             velec            = qq11*rinv11;
 850             felec            = velec*rinvsq11;
 851
 852             fscal            = felec;
 853
 854             /* Calculate temporary vectorial force */
 855             tx               = fscal*dx11;
 856             ty               = fscal*dy11;
 857             tz               = fscal*dz11;
 858
 859             /* Update vectorial force */
 860             fix1            += tx;
 861             fiy1            += ty;
 862             fiz1            += tz;
 863             f[j_coord_offset+DIM*1+XX] -= tx;
 864             f[j_coord_offset+DIM*1+YY] -= ty;
 865             f[j_coord_offset+DIM*1+ZZ] -= tz;
 866
 867             /**************************
 868              * CALCULATE INTERACTIONS *
 869              **************************/
 870
 871             /* COULOMB ELECTROSTATICS */
 872             velec            = qq12*rinv12;
 873             felec            = velec*rinvsq12;
 874
 875             fscal            = felec;
 876
 877             /* Calculate temporary vectorial force */
 878             tx               = fscal*dx12;
 879             ty               = fscal*dy12;
 880             tz               = fscal*dz12;
 881
 882             /* Update vectorial force */
 883             fix1            += tx;
 884             fiy1            += ty;
 885             fiz1            += tz;
 886             f[j_coord_offset+DIM*2+XX] -= tx;
 887             f[j_coord_offset+DIM*2+YY] -= ty;
 888             f[j_coord_offset+DIM*2+ZZ] -= tz;
 889
 890             /**************************
 891              * CALCULATE INTERACTIONS *
 892              **************************/
 893
 894             /* COULOMB ELECTROSTATICS */
 895             velec            = qq13*rinv13;
 896             felec            = velec*rinvsq13;
 897
 898             fscal            = felec;
 899
 900             /* Calculate temporary vectorial force */
 901             tx               = fscal*dx13;
 902             ty               = fscal*dy13;
 903             tz               = fscal*dz13;
 904
 905             /* Update vectorial force */
 906             fix1            += tx;
 907             fiy1            += ty;
 908             fiz1            += tz;
 909             f[j_coord_offset+DIM*3+XX] -= tx;
 910             f[j_coord_offset+DIM*3+YY] -= ty;
 911             f[j_coord_offset+DIM*3+ZZ] -= tz;
 912
 913             /**************************
 914              * CALCULATE INTERACTIONS *
 915              **************************/
 916
 917             /* COULOMB ELECTROSTATICS */
 918             velec            = qq21*rinv21;
 919             felec            = velec*rinvsq21;
 920
 921             fscal            = felec;
 922
 923             /* Calculate temporary vectorial force */
 924             tx               = fscal*dx21;
 925             ty               = fscal*dy21;
 926             tz               = fscal*dz21;
 927
 928             /* Update vectorial force */
 929             fix2            += tx;
 930             fiy2            += ty;
 931             fiz2            += tz;
 932             f[j_coord_offset+DIM*1+XX] -= tx;
 933             f[j_coord_offset+DIM*1+YY] -= ty;
 934             f[j_coord_offset+DIM*1+ZZ] -= tz;
 935
 936             /**************************
 937              * CALCULATE INTERACTIONS *
 938              **************************/
 939
 940             /* COULOMB ELECTROSTATICS */
 941             velec            = qq22*rinv22;
 942             felec            = velec*rinvsq22;
 943
 944             fscal            = felec;
 945
 946             /* Calculate temporary vectorial force */
 947             tx               = fscal*dx22;
 948             ty               = fscal*dy22;
 949             tz               = fscal*dz22;
 950
 951             /* Update vectorial force */
 952             fix2            += tx;
 953             fiy2            += ty;
 954             fiz2            += tz;
 955             f[j_coord_offset+DIM*2+XX] -= tx;
 956             f[j_coord_offset+DIM*2+YY] -= ty;
 957             f[j_coord_offset+DIM*2+ZZ] -= tz;
 958
 959             /**************************
 960              * CALCULATE INTERACTIONS *
 961              **************************/
 962
 963             /* COULOMB ELECTROSTATICS */
 964             velec            = qq23*rinv23;
 965             felec            = velec*rinvsq23;
 966
 967             fscal            = felec;
 968
 969             /* Calculate temporary vectorial force */
 970             tx               = fscal*dx23;
 971             ty               = fscal*dy23;
 972             tz               = fscal*dz23;
 973
 974             /* Update vectorial force */
 975             fix2            += tx;
 976             fiy2            += ty;
 977             fiz2            += tz;
 978             f[j_coord_offset+DIM*3+XX] -= tx;
 979             f[j_coord_offset+DIM*3+YY] -= ty;
 980             f[j_coord_offset+DIM*3+ZZ] -= tz;
 981
 982             /**************************
 983              * CALCULATE INTERACTIONS *
 984              **************************/
 985
 986             /* COULOMB ELECTROSTATICS */
 987             velec            = qq31*rinv31;
 988             felec            = velec*rinvsq31;
 989
 990             fscal            = felec;
 991
 992             /* Calculate temporary vectorial force */
 993             tx               = fscal*dx31;
 994             ty               = fscal*dy31;
 995             tz               = fscal*dz31;
 996
 997             /* Update vectorial force */
 998             fix3            += tx;
 999             fiy3            += ty;
1000             fiz3            += tz;
1001             f[j_coord_offset+DIM*1+XX] -= tx;
1002             f[j_coord_offset+DIM*1+YY] -= ty;
1003             f[j_coord_offset+DIM*1+ZZ] -= tz;
1004
1005             /**************************
1006              * CALCULATE INTERACTIONS *
1007              **************************/
1008
1009             /* COULOMB ELECTROSTATICS */
1010             velec            = qq32*rinv32;
1011             felec            = velec*rinvsq32;
1012
1013             fscal            = felec;
1014
1015             /* Calculate temporary vectorial force */
1016             tx               = fscal*dx32;
1017             ty               = fscal*dy32;
1018             tz               = fscal*dz32;
1019
1020             /* Update vectorial force */
1021             fix3            += tx;
1022             fiy3            += ty;
1023             fiz3            += 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             /**************************
1029              * CALCULATE INTERACTIONS *
1030              **************************/
1031
1032             /* COULOMB ELECTROSTATICS */
1033             velec            = qq33*rinv33;
1034             felec            = velec*rinvsq33;
1035
1036             fscal            = felec;
1037
1038             /* Calculate temporary vectorial force */
1039             tx               = fscal*dx33;
1040             ty               = fscal*dy33;
1041             tz               = fscal*dz33;
1042
1043             /* Update vectorial force */
1044             fix3            += tx;
1045             fiy3            += ty;
1046             fiz3            += tz;
1047             f[j_coord_offset+DIM*3+XX] -= tx;
1048             f[j_coord_offset+DIM*3+YY] -= ty;
1049             f[j_coord_offset+DIM*3+ZZ] -= tz;
1050
1051             /* Inner loop uses 292 flops */
1052         }
1053         /* End of innermost loop */
1054
1055         tx = ty = tz = 0;
1056         f[i_coord_offset+DIM*0+XX] += fix0;
1057         f[i_coord_offset+DIM*0+YY] += fiy0;
1058         f[i_coord_offset+DIM*0+ZZ] += fiz0;
1059         tx                         += fix0;
1060         ty                         += fiy0;
1061         tz                         += fiz0;
1062         f[i_coord_offset+DIM*1+XX] += fix1;
1063         f[i_coord_offset+DIM*1+YY] += fiy1;
1064         f[i_coord_offset+DIM*1+ZZ] += fiz1;
1065         tx                         += fix1;
1066         ty                         += fiy1;
1067         tz                         += fiz1;
1068         f[i_coord_offset+DIM*2+XX] += fix2;
1069         f[i_coord_offset+DIM*2+YY] += fiy2;
1070         f[i_coord_offset+DIM*2+ZZ] += fiz2;
1071         tx                         += fix2;
1072         ty                         += fiy2;
1073         tz                         += fiz2;
1074         f[i_coord_offset+DIM*3+XX] += fix3;
1075         f[i_coord_offset+DIM*3+YY] += fiy3;
1076         f[i_coord_offset+DIM*3+ZZ] += fiz3;
1077         tx                         += fix3;
1078         ty                         += fiy3;
1079         tz                         += fiz3;
1080         fshift[i_shift_offset+XX]  += tx;
1081         fshift[i_shift_offset+YY]  += ty;
1082         fshift[i_shift_offset+ZZ]  += tz;
1083
1084         /* Increment number of inner iterations */
1085         inneriter                  += j_index_end - j_index_start;
1086
1087         /* Outer loop uses 39 flops */
1088     }
1089
1090     /* Increment number of outer iterations */
1091     outeriter        += nri;
1092
1093     /* Update outer/inner flops */
1094
1095     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*292);
1096 }