src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecCoul_VdwLJ_GeomW4P1_sparc64_hpc_ace_double.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 sparc64_hpc_ace_double 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 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4P1_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: Coulomb
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water4-Particle
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecCoul_VdwLJ_GeomW4P1_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwioffset3;
  88     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  89     int              vdwjidx0A,vdwjidx0B;
  90     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  93     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  94     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  95     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 102     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 103     _fjsp_v2r8       itab_tmp;
 104     _fjsp_v2r8       dummy_mask,cutoff_mask;
 105     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 106     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 107     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 108
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 129     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 130     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     /* Avoid stupid compiler warnings */
 134     jnrA = jnrB = 0;
 135     j_coord_offsetA = 0;
 136     j_coord_offsetB = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 157                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 158
 159         fix0             = _fjsp_setzero_v2r8();
 160         fiy0             = _fjsp_setzero_v2r8();
 161         fiz0             = _fjsp_setzero_v2r8();
 162         fix1             = _fjsp_setzero_v2r8();
 163         fiy1             = _fjsp_setzero_v2r8();
 164         fiz1             = _fjsp_setzero_v2r8();
 165         fix2             = _fjsp_setzero_v2r8();
 166         fiy2             = _fjsp_setzero_v2r8();
 167         fiz2             = _fjsp_setzero_v2r8();
 168         fix3             = _fjsp_setzero_v2r8();
 169         fiy3             = _fjsp_setzero_v2r8();
 170         fiz3             = _fjsp_setzero_v2r8();
 171
 172         /* Reset potential sums */
 173         velecsum         = _fjsp_setzero_v2r8();
 174         vvdwsum          = _fjsp_setzero_v2r8();
 175
 176         /* Start inner kernel loop */
 177         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 178         {
 179
 180             /* Get j neighbor index, and coordinate index */
 181             jnrA             = jjnr[jidx];
 182             jnrB             = jjnr[jidx+1];
 183             j_coord_offsetA  = DIM*jnrA;
 184             j_coord_offsetB  = DIM*jnrB;
 185
 186             /* load j atom coordinates */
 187             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 188                                               &jx0,&jy0,&jz0);
 189
 190             /* Calculate displacement vector */
 191             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 192             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 193             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 194             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 195             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 196             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 197             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 198             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 199             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 200             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 201             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 202             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 203
 204             /* Calculate squared distance and things based on it */
 205             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 206             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 207             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 208             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 209
 210             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 211             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 212             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 213
 214             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 215             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 216             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 217             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 218
 219             /* Load parameters for j particles */
 220             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 221             vdwjidx0A        = 2*vdwtype[jnrA+0];
 222             vdwjidx0B        = 2*vdwtype[jnrB+0];
 223
 224             fjx0             = _fjsp_setzero_v2r8();
 225             fjy0             = _fjsp_setzero_v2r8();
 226             fjz0             = _fjsp_setzero_v2r8();
 227
 228             /**************************
 229              * CALCULATE INTERACTIONS *
 230              **************************/
 231
 232             /* Compute parameters for interactions between i and j atoms */
 233             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 234                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 235
 236             /* LENNARD-JONES DISPERSION/REPULSION */
 237
 238             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 239             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 240             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 241             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 242             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 243
 244             /* Update potential sum for this i atom from the interaction with this j atom. */
 245             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 246
 247             fscal            = fvdw;
 248
 249             /* Update vectorial force */
 250             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 251             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 252             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 253
 254             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 255             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 256             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 257
 258             /**************************
 259              * CALCULATE INTERACTIONS *
 260              **************************/
 261
 262             /* Compute parameters for interactions between i and j atoms */
 263             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 264
 265             /* COULOMB ELECTROSTATICS */
 266             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 267             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 268
 269             /* Update potential sum for this i atom from the interaction with this j atom. */
 270             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 271
 272             fscal            = felec;
 273
 274             /* Update vectorial force */
 275             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 276             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 277             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 278
 279             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 280             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 281             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 282
 283             /**************************
 284              * CALCULATE INTERACTIONS *
 285              **************************/
 286
 287             /* Compute parameters for interactions between i and j atoms */
 288             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 289
 290             /* COULOMB ELECTROSTATICS */
 291             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 292             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 293
 294             /* Update potential sum for this i atom from the interaction with this j atom. */
 295             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 296
 297             fscal            = felec;
 298
 299             /* Update vectorial force */
 300             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 301             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 302             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 303
 304             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 305             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 306             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             /* Compute parameters for interactions between i and j atoms */
 313             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 314
 315             /* COULOMB ELECTROSTATICS */
 316             velec            = _fjsp_mul_v2r8(qq30,rinv30);
 317             felec            = _fjsp_mul_v2r8(velec,rinvsq30);
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324             /* Update vectorial force */
 325             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 326             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 327             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 328
 329             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 330             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 331             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 332
 333             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 334
 335             /* Inner loop uses 131 flops */
 336         }
 337
 338         if(jidx<j_index_end)
 339         {
 340
 341             jnrA             = jjnr[jidx];
 342             j_coord_offsetA  = DIM*jnrA;
 343
 344             /* load j atom coordinates */
 345             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 346                                               &jx0,&jy0,&jz0);
 347
 348             /* Calculate displacement vector */
 349             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 350             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 351             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 352             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 353             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 354             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 355             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 356             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 357             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 358             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 359             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 360             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 361
 362             /* Calculate squared distance and things based on it */
 363             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 364             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 365             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 366             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 367
 368             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 369             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 370             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 371
 372             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 373             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 374             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 375             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 376
 377             /* Load parameters for j particles */
 378             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 379             vdwjidx0A        = 2*vdwtype[jnrA+0];
 380
 381             fjx0             = _fjsp_setzero_v2r8();
 382             fjy0             = _fjsp_setzero_v2r8();
 383             fjz0             = _fjsp_setzero_v2r8();
 384
 385             /**************************
 386              * CALCULATE INTERACTIONS *
 387              **************************/
 388
 389             /* Compute parameters for interactions between i and j atoms */
 390             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 391
 392             /* LENNARD-JONES DISPERSION/REPULSION */
 393
 394             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 395             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 396             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 397             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 398             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 402             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 403
 404             fscal            = fvdw;
 405
 406             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 407
 408             /* Update vectorial force */
 409             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 410             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 411             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 412
 413             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 414             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 415             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 416
 417             /**************************
 418              * CALCULATE INTERACTIONS *
 419              **************************/
 420
 421             /* Compute parameters for interactions between i and j atoms */
 422             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 423
 424             /* COULOMB ELECTROSTATICS */
 425             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 426             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 427
 428             /* Update potential sum for this i atom from the interaction with this j atom. */
 429             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 430             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 435
 436             /* Update vectorial force */
 437             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 438             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 439             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 440
 441             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 442             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 443             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* Compute parameters for interactions between i and j atoms */
 450             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 451
 452             /* COULOMB ELECTROSTATICS */
 453             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 454             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 458             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 459
 460             fscal            = felec;
 461
 462             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 463
 464             /* Update vectorial force */
 465             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 466             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 467             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 468
 469             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 470             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 471             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 472
 473             /**************************
 474              * CALCULATE INTERACTIONS *
 475              **************************/
 476
 477             /* Compute parameters for interactions between i and j atoms */
 478             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 479
 480             /* COULOMB ELECTROSTATICS */
 481             velec            = _fjsp_mul_v2r8(qq30,rinv30);
 482             felec            = _fjsp_mul_v2r8(velec,rinvsq30);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 486             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 487
 488             fscal            = felec;
 489
 490             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 491
 492             /* Update vectorial force */
 493             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 494             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 495             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 496
 497             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 498             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 499             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 500
 501             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 502
 503             /* Inner loop uses 131 flops */
 504         }
 505
 506         /* End of innermost loop */
 507
 508         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 509                                               f+i_coord_offset,fshift+i_shift_offset);
 510
 511         ggid                        = gid[iidx];
 512         /* Update potential energies */
 513         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 514         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 515
 516         /* Increment number of inner iterations */
 517         inneriter                  += j_index_end - j_index_start;
 518
 519         /* Outer loop uses 26 flops */
 520     }
 521
 522     /* Increment number of outer iterations */
 523     outeriter        += nri;
 524
 525     /* Update outer/inner flops */
 526
 527     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*131);
 528 }
 529 /*
 530  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 531  * Electrostatics interaction: Coulomb
 532  * VdW interaction:            LennardJones
 533  * Geometry:                   Water4-Particle
 534  * Calculate force/pot:        Force
 535  */
 536 void
 537 nb_kernel_ElecCoul_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 538                     (t_nblist                    * gmx_restrict       nlist,
 539                      rvec                        * gmx_restrict          xx,
 540                      rvec                        * gmx_restrict          ff,
 541                      t_forcerec                  * gmx_restrict          fr,
 542                      t_mdatoms                   * gmx_restrict     mdatoms,
 543                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 544                      t_nrnb                      * gmx_restrict        nrnb)
 545 {
 546     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 547      * just 0 for non-waters.
 548      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 549      * jnr indices corresponding to data put in the four positions in the SIMD register.
 550      */
 551     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 552     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 553     int              jnrA,jnrB;
 554     int              j_coord_offsetA,j_coord_offsetB;
 555     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 556     real             rcutoff_scalar;
 557     real             *shiftvec,*fshift,*x,*f;
 558     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 559     int              vdwioffset0;
 560     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 561     int              vdwioffset1;
 562     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 563     int              vdwioffset2;
 564     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 565     int              vdwioffset3;
 566     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 567     int              vdwjidx0A,vdwjidx0B;
 568     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 569     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 570     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 571     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 572     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 573     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 574     real             *charge;
 575     int              nvdwtype;
 576     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 577     int              *vdwtype;
 578     real             *vdwparam;
 579     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 580     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 581     _fjsp_v2r8       itab_tmp;
 582     _fjsp_v2r8       dummy_mask,cutoff_mask;
 583     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 584     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 585     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 586
 587     x                = xx[0];
 588     f                = ff[0];
 589
 590     nri              = nlist->nri;
 591     iinr             = nlist->iinr;
 592     jindex           = nlist->jindex;
 593     jjnr             = nlist->jjnr;
 594     shiftidx         = nlist->shift;
 595     gid              = nlist->gid;
 596     shiftvec         = fr->shift_vec[0];
 597     fshift           = fr->fshift[0];
 598     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 599     charge           = mdatoms->chargeA;
 600     nvdwtype         = fr->ntype;
 601     vdwparam         = fr->nbfp;
 602     vdwtype          = mdatoms->typeA;
 603
 604     /* Setup water-specific parameters */
 605     inr              = nlist->iinr[0];
 606     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 607     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 608     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 609     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 610
 611     /* Avoid stupid compiler warnings */
 612     jnrA = jnrB = 0;
 613     j_coord_offsetA = 0;
 614     j_coord_offsetB = 0;
 615
 616     outeriter        = 0;
 617     inneriter        = 0;
 618
 619     /* Start outer loop over neighborlists */
 620     for(iidx=0; iidx<nri; iidx++)
 621     {
 622         /* Load shift vector for this list */
 623         i_shift_offset   = DIM*shiftidx[iidx];
 624
 625         /* Load limits for loop over neighbors */
 626         j_index_start    = jindex[iidx];
 627         j_index_end      = jindex[iidx+1];
 628
 629         /* Get outer coordinate index */
 630         inr              = iinr[iidx];
 631         i_coord_offset   = DIM*inr;
 632
 633         /* Load i particle coords and add shift vector */
 634         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 635                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 636
 637         fix0             = _fjsp_setzero_v2r8();
 638         fiy0             = _fjsp_setzero_v2r8();
 639         fiz0             = _fjsp_setzero_v2r8();
 640         fix1             = _fjsp_setzero_v2r8();
 641         fiy1             = _fjsp_setzero_v2r8();
 642         fiz1             = _fjsp_setzero_v2r8();
 643         fix2             = _fjsp_setzero_v2r8();
 644         fiy2             = _fjsp_setzero_v2r8();
 645         fiz2             = _fjsp_setzero_v2r8();
 646         fix3             = _fjsp_setzero_v2r8();
 647         fiy3             = _fjsp_setzero_v2r8();
 648         fiz3             = _fjsp_setzero_v2r8();
 649
 650         /* Start inner kernel loop */
 651         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 652         {
 653
 654             /* Get j neighbor index, and coordinate index */
 655             jnrA             = jjnr[jidx];
 656             jnrB             = jjnr[jidx+1];
 657             j_coord_offsetA  = DIM*jnrA;
 658             j_coord_offsetB  = DIM*jnrB;
 659
 660             /* load j atom coordinates */
 661             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 662                                               &jx0,&jy0,&jz0);
 663
 664             /* Calculate displacement vector */
 665             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 666             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 667             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 668             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 669             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 670             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 671             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 672             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 673             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 674             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 675             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 676             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 677
 678             /* Calculate squared distance and things based on it */
 679             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 680             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 681             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 682             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 683
 684             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 685             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 686             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 687
 688             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 689             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 690             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 691             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 692
 693             /* Load parameters for j particles */
 694             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 695             vdwjidx0A        = 2*vdwtype[jnrA+0];
 696             vdwjidx0B        = 2*vdwtype[jnrB+0];
 697
 698             fjx0             = _fjsp_setzero_v2r8();
 699             fjy0             = _fjsp_setzero_v2r8();
 700             fjz0             = _fjsp_setzero_v2r8();
 701
 702             /**************************
 703              * CALCULATE INTERACTIONS *
 704              **************************/
 705
 706             /* Compute parameters for interactions between i and j atoms */
 707             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 708                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 709
 710             /* LENNARD-JONES DISPERSION/REPULSION */
 711
 712             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 713             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 714
 715             fscal            = fvdw;
 716
 717             /* Update vectorial force */
 718             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 719             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 720             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 721
 722             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 723             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 724             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 725
 726             /**************************
 727              * CALCULATE INTERACTIONS *
 728              **************************/
 729
 730             /* Compute parameters for interactions between i and j atoms */
 731             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 732
 733             /* COULOMB ELECTROSTATICS */
 734             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 735             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 736
 737             fscal            = felec;
 738
 739             /* Update vectorial force */
 740             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 741             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 742             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 743
 744             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 745             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 746             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             /* Compute parameters for interactions between i and j atoms */
 753             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 754
 755             /* COULOMB ELECTROSTATICS */
 756             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 757             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 758
 759             fscal            = felec;
 760
 761             /* Update vectorial force */
 762             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 763             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 764             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 765
 766             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 767             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 768             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 769
 770             /**************************
 771              * CALCULATE INTERACTIONS *
 772              **************************/
 773
 774             /* Compute parameters for interactions between i and j atoms */
 775             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 776
 777             /* COULOMB ELECTROSTATICS */
 778             velec            = _fjsp_mul_v2r8(qq30,rinv30);
 779             felec            = _fjsp_mul_v2r8(velec,rinvsq30);
 780
 781             fscal            = felec;
 782
 783             /* Update vectorial force */
 784             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 785             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 786             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 787
 788             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 789             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 790             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 791
 792             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 793
 794             /* Inner loop uses 123 flops */
 795         }
 796
 797         if(jidx<j_index_end)
 798         {
 799
 800             jnrA             = jjnr[jidx];
 801             j_coord_offsetA  = DIM*jnrA;
 802
 803             /* load j atom coordinates */
 804             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 805                                               &jx0,&jy0,&jz0);
 806
 807             /* Calculate displacement vector */
 808             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 809             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 810             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 811             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 812             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 813             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 814             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 815             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 816             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 817             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 818             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 819             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 820
 821             /* Calculate squared distance and things based on it */
 822             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 823             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 824             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 825             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 826
 827             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 828             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 829             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 830
 831             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 832             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 833             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 834             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 835
 836             /* Load parameters for j particles */
 837             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 838             vdwjidx0A        = 2*vdwtype[jnrA+0];
 839
 840             fjx0             = _fjsp_setzero_v2r8();
 841             fjy0             = _fjsp_setzero_v2r8();
 842             fjz0             = _fjsp_setzero_v2r8();
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* Compute parameters for interactions between i and j atoms */
 849             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 850
 851             /* LENNARD-JONES DISPERSION/REPULSION */
 852
 853             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 854             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 855
 856             fscal            = fvdw;
 857
 858             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 859
 860             /* Update vectorial force */
 861             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 862             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 863             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 864
 865             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 866             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 867             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 868
 869             /**************************
 870              * CALCULATE INTERACTIONS *
 871              **************************/
 872
 873             /* Compute parameters for interactions between i and j atoms */
 874             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 875
 876             /* COULOMB ELECTROSTATICS */
 877             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 878             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 879
 880             fscal            = felec;
 881
 882             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 883
 884             /* Update vectorial force */
 885             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 886             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 887             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 888
 889             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 890             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 891             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 892
 893             /**************************
 894              * CALCULATE INTERACTIONS *
 895              **************************/
 896
 897             /* Compute parameters for interactions between i and j atoms */
 898             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 899
 900             /* COULOMB ELECTROSTATICS */
 901             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 902             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 903
 904             fscal            = felec;
 905
 906             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 907
 908             /* Update vectorial force */
 909             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 910             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 911             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 912
 913             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 914             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 915             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 916
 917             /**************************
 918              * CALCULATE INTERACTIONS *
 919              **************************/
 920
 921             /* Compute parameters for interactions between i and j atoms */
 922             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 923
 924             /* COULOMB ELECTROSTATICS */
 925             velec            = _fjsp_mul_v2r8(qq30,rinv30);
 926             felec            = _fjsp_mul_v2r8(velec,rinvsq30);
 927
 928             fscal            = felec;
 929
 930             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 931
 932             /* Update vectorial force */
 933             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 934             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 935             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 936
 937             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 938             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 939             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 940
 941             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 942
 943             /* Inner loop uses 123 flops */
 944         }
 945
 946         /* End of innermost loop */
 947
 948         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 949                                               f+i_coord_offset,fshift+i_shift_offset);
 950
 951         /* Increment number of inner iterations */
 952         inneriter                  += j_index_end - j_index_start;
 953
 954         /* Outer loop uses 24 flops */
 955     }
 956
 957     /* Increment number of outer iterations */
 958     outeriter        += nri;
 959
 960     /* Update outer/inner flops */
 961
 962     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*123);
 963 }