src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecCoul_VdwLJ_GeomW4P1_sparc64_hpc_ace_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4P1_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_GeomW4P1_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwioffset3;
  86     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  91     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  92     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  93     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 100     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 101     _fjsp_v2r8       itab_tmp;
 102     _fjsp_v2r8       dummy_mask,cutoff_mask;
 103     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 104     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 105     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 106
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = gmx_fjsp_set1_v2r8(fr->ic->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 127     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 128     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 129     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 130
 131     /* Avoid stupid compiler warnings */
 132     jnrA = jnrB = 0;
 133     j_coord_offsetA = 0;
 134     j_coord_offsetB = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 155                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 156
 157         fix0             = _fjsp_setzero_v2r8();
 158         fiy0             = _fjsp_setzero_v2r8();
 159         fiz0             = _fjsp_setzero_v2r8();
 160         fix1             = _fjsp_setzero_v2r8();
 161         fiy1             = _fjsp_setzero_v2r8();
 162         fiz1             = _fjsp_setzero_v2r8();
 163         fix2             = _fjsp_setzero_v2r8();
 164         fiy2             = _fjsp_setzero_v2r8();
 165         fiz2             = _fjsp_setzero_v2r8();
 166         fix3             = _fjsp_setzero_v2r8();
 167         fiy3             = _fjsp_setzero_v2r8();
 168         fiz3             = _fjsp_setzero_v2r8();
 169
 170         /* Reset potential sums */
 171         velecsum         = _fjsp_setzero_v2r8();
 172         vvdwsum          = _fjsp_setzero_v2r8();
 173
 174         /* Start inner kernel loop */
 175         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 176         {
 177
 178             /* Get j neighbor index, and coordinate index */
 179             jnrA             = jjnr[jidx];
 180             jnrB             = jjnr[jidx+1];
 181             j_coord_offsetA  = DIM*jnrA;
 182             j_coord_offsetB  = DIM*jnrB;
 183
 184             /* load j atom coordinates */
 185             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 186                                               &jx0,&jy0,&jz0);
 187
 188             /* Calculate displacement vector */
 189             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 190             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 191             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 192             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 193             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 194             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 195             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 196             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 197             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 198             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 199             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 200             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 204             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 205             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 206             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 207
 208             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 209             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 210             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 211
 212             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 213             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 214             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 215             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 216
 217             /* Load parameters for j particles */
 218             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 219             vdwjidx0A        = 2*vdwtype[jnrA+0];
 220             vdwjidx0B        = 2*vdwtype[jnrB+0];
 221
 222             fjx0             = _fjsp_setzero_v2r8();
 223             fjy0             = _fjsp_setzero_v2r8();
 224             fjz0             = _fjsp_setzero_v2r8();
 225
 226             /**************************
 227              * CALCULATE INTERACTIONS *
 228              **************************/
 229
 230             /* Compute parameters for interactions between i and j atoms */
 231             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 232                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 233
 234             /* LENNARD-JONES DISPERSION/REPULSION */
 235
 236             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 237             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 238             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 239             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 240             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 241
 242             /* Update potential sum for this i atom from the interaction with this j atom. */
 243             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 244
 245             fscal            = fvdw;
 246
 247             /* Update vectorial force */
 248             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 249             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 250             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 251
 252             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 253             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 254             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 255
 256             /**************************
 257              * CALCULATE INTERACTIONS *
 258              **************************/
 259
 260             /* Compute parameters for interactions between i and j atoms */
 261             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 262
 263             /* COULOMB ELECTROSTATICS */
 264             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 265             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 266
 267             /* Update potential sum for this i atom from the interaction with this j atom. */
 268             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 269
 270             fscal            = felec;
 271
 272             /* Update vectorial force */
 273             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 274             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 275             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 276
 277             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 278             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 279             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 280
 281             /**************************
 282              * CALCULATE INTERACTIONS *
 283              **************************/
 284
 285             /* Compute parameters for interactions between i and j atoms */
 286             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 287
 288             /* COULOMB ELECTROSTATICS */
 289             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 290             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 291
 292             /* Update potential sum for this i atom from the interaction with this j atom. */
 293             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 294
 295             fscal            = felec;
 296
 297             /* Update vectorial force */
 298             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 299             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 300             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 301
 302             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 303             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 304             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             /* Compute parameters for interactions between i and j atoms */
 311             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 312
 313             /* COULOMB ELECTROSTATICS */
 314             velec            = _fjsp_mul_v2r8(qq30,rinv30);
 315             felec            = _fjsp_mul_v2r8(velec,rinvsq30);
 316
 317             /* Update potential sum for this i atom from the interaction with this j atom. */
 318             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 319
 320             fscal            = felec;
 321
 322             /* Update vectorial force */
 323             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 324             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 325             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 326
 327             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 328             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 329             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 330
 331             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 332
 333             /* Inner loop uses 131 flops */
 334         }
 335
 336         if(jidx<j_index_end)
 337         {
 338
 339             jnrA             = jjnr[jidx];
 340             j_coord_offsetA  = DIM*jnrA;
 341
 342             /* load j atom coordinates */
 343             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 344                                               &jx0,&jy0,&jz0);
 345
 346             /* Calculate displacement vector */
 347             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 348             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 349             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 350             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 351             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 352             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 353             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 354             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 355             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 356             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 357             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 358             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 359
 360             /* Calculate squared distance and things based on it */
 361             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 362             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 363             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 364             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 365
 366             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 367             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 368             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 369
 370             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 371             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 372             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 373             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 374
 375             /* Load parameters for j particles */
 376             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 377             vdwjidx0A        = 2*vdwtype[jnrA+0];
 378
 379             fjx0             = _fjsp_setzero_v2r8();
 380             fjy0             = _fjsp_setzero_v2r8();
 381             fjz0             = _fjsp_setzero_v2r8();
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             /* Compute parameters for interactions between i and j atoms */
 388             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 389                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 390
 391             /* LENNARD-JONES DISPERSION/REPULSION */
 392
 393             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 394             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 395             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 396             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 397             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 401             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 402
 403             fscal            = fvdw;
 404
 405             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 406
 407             /* Update vectorial force */
 408             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 409             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 410             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 411
 412             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 413             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 414             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             /* Compute parameters for interactions between i and j atoms */
 421             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 422
 423             /* COULOMB ELECTROSTATICS */
 424             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 425             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 426
 427             /* Update potential sum for this i atom from the interaction with this j atom. */
 428             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 429             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 430
 431             fscal            = felec;
 432
 433             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 434
 435             /* Update vectorial force */
 436             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 437             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 438             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 439
 440             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 441             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 442             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             /* Compute parameters for interactions between i and j atoms */
 449             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 450
 451             /* COULOMB ELECTROSTATICS */
 452             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 453             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 457             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 462
 463             /* Update vectorial force */
 464             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 465             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 466             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 467
 468             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 469             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 470             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             /* Compute parameters for interactions between i and j atoms */
 477             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 478
 479             /* COULOMB ELECTROSTATICS */
 480             velec            = _fjsp_mul_v2r8(qq30,rinv30);
 481             felec            = _fjsp_mul_v2r8(velec,rinvsq30);
 482
 483             /* Update potential sum for this i atom from the interaction with this j atom. */
 484             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 485             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 490
 491             /* Update vectorial force */
 492             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 493             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 494             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 495
 496             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 497             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 498             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 499
 500             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 501
 502             /* Inner loop uses 131 flops */
 503         }
 504
 505         /* End of innermost loop */
 506
 507         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 508                                               f+i_coord_offset,fshift+i_shift_offset);
 509
 510         ggid                        = gid[iidx];
 511         /* Update potential energies */
 512         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 513         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 514
 515         /* Increment number of inner iterations */
 516         inneriter                  += j_index_end - j_index_start;
 517
 518         /* Outer loop uses 26 flops */
 519     }
 520
 521     /* Increment number of outer iterations */
 522     outeriter        += nri;
 523
 524     /* Update outer/inner flops */
 525
 526     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*131);
 527 }
 528 /*
 529  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 530  * Electrostatics interaction: Coulomb
 531  * VdW interaction:            LennardJones
 532  * Geometry:                   Water4-Particle
 533  * Calculate force/pot:        Force
 534  */
 535 void
 536 nb_kernel_ElecCoul_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 537                     (t_nblist                    * gmx_restrict       nlist,
 538                      rvec                        * gmx_restrict          xx,
 539                      rvec                        * gmx_restrict          ff,
 540                      struct t_forcerec           * gmx_restrict          fr,
 541                      t_mdatoms                   * gmx_restrict     mdatoms,
 542                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 543                      t_nrnb                      * gmx_restrict        nrnb)
 544 {
 545     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 546      * just 0 for non-waters.
 547      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 548      * jnr indices corresponding to data put in the four positions in the SIMD register.
 549      */
 550     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 551     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 552     int              jnrA,jnrB;
 553     int              j_coord_offsetA,j_coord_offsetB;
 554     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 555     real             rcutoff_scalar;
 556     real             *shiftvec,*fshift,*x,*f;
 557     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 558     int              vdwioffset0;
 559     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 560     int              vdwioffset1;
 561     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 562     int              vdwioffset2;
 563     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 564     int              vdwioffset3;
 565     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 566     int              vdwjidx0A,vdwjidx0B;
 567     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 568     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 569     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 570     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 571     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 572     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 573     real             *charge;
 574     int              nvdwtype;
 575     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 576     int              *vdwtype;
 577     real             *vdwparam;
 578     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 579     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 580     _fjsp_v2r8       itab_tmp;
 581     _fjsp_v2r8       dummy_mask,cutoff_mask;
 582     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 583     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 584     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 585
 586     x                = xx[0];
 587     f                = ff[0];
 588
 589     nri              = nlist->nri;
 590     iinr             = nlist->iinr;
 591     jindex           = nlist->jindex;
 592     jjnr             = nlist->jjnr;
 593     shiftidx         = nlist->shift;
 594     gid              = nlist->gid;
 595     shiftvec         = fr->shift_vec[0];
 596     fshift           = fr->fshift[0];
 597     facel            = gmx_fjsp_set1_v2r8(fr->ic->epsfac);
 598     charge           = mdatoms->chargeA;
 599     nvdwtype         = fr->ntype;
 600     vdwparam         = fr->nbfp;
 601     vdwtype          = mdatoms->typeA;
 602
 603     /* Setup water-specific parameters */
 604     inr              = nlist->iinr[0];
 605     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 606     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 607     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 608     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 609
 610     /* Avoid stupid compiler warnings */
 611     jnrA = jnrB = 0;
 612     j_coord_offsetA = 0;
 613     j_coord_offsetB = 0;
 614
 615     outeriter        = 0;
 616     inneriter        = 0;
 617
 618     /* Start outer loop over neighborlists */
 619     for(iidx=0; iidx<nri; iidx++)
 620     {
 621         /* Load shift vector for this list */
 622         i_shift_offset   = DIM*shiftidx[iidx];
 623
 624         /* Load limits for loop over neighbors */
 625         j_index_start    = jindex[iidx];
 626         j_index_end      = jindex[iidx+1];
 627
 628         /* Get outer coordinate index */
 629         inr              = iinr[iidx];
 630         i_coord_offset   = DIM*inr;
 631
 632         /* Load i particle coords and add shift vector */
 633         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 634                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 635
 636         fix0             = _fjsp_setzero_v2r8();
 637         fiy0             = _fjsp_setzero_v2r8();
 638         fiz0             = _fjsp_setzero_v2r8();
 639         fix1             = _fjsp_setzero_v2r8();
 640         fiy1             = _fjsp_setzero_v2r8();
 641         fiz1             = _fjsp_setzero_v2r8();
 642         fix2             = _fjsp_setzero_v2r8();
 643         fiy2             = _fjsp_setzero_v2r8();
 644         fiz2             = _fjsp_setzero_v2r8();
 645         fix3             = _fjsp_setzero_v2r8();
 646         fiy3             = _fjsp_setzero_v2r8();
 647         fiz3             = _fjsp_setzero_v2r8();
 648
 649         /* Start inner kernel loop */
 650         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 651         {
 652
 653             /* Get j neighbor index, and coordinate index */
 654             jnrA             = jjnr[jidx];
 655             jnrB             = jjnr[jidx+1];
 656             j_coord_offsetA  = DIM*jnrA;
 657             j_coord_offsetB  = DIM*jnrB;
 658
 659             /* load j atom coordinates */
 660             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 661                                               &jx0,&jy0,&jz0);
 662
 663             /* Calculate displacement vector */
 664             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 665             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 666             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 667             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 668             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 669             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 670             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 671             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 672             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 673             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 674             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 675             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 676
 677             /* Calculate squared distance and things based on it */
 678             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 679             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 680             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 681             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 682
 683             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 684             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 685             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 686
 687             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 688             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 689             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 690             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 691
 692             /* Load parameters for j particles */
 693             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 694             vdwjidx0A        = 2*vdwtype[jnrA+0];
 695             vdwjidx0B        = 2*vdwtype[jnrB+0];
 696
 697             fjx0             = _fjsp_setzero_v2r8();
 698             fjy0             = _fjsp_setzero_v2r8();
 699             fjz0             = _fjsp_setzero_v2r8();
 700
 701             /**************************
 702              * CALCULATE INTERACTIONS *
 703              **************************/
 704
 705             /* Compute parameters for interactions between i and j atoms */
 706             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 707                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 708
 709             /* LENNARD-JONES DISPERSION/REPULSION */
 710
 711             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 712             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 713
 714             fscal            = fvdw;
 715
 716             /* Update vectorial force */
 717             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 718             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 719             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 720
 721             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 722             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 723             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 724
 725             /**************************
 726              * CALCULATE INTERACTIONS *
 727              **************************/
 728
 729             /* Compute parameters for interactions between i and j atoms */
 730             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 731
 732             /* COULOMB ELECTROSTATICS */
 733             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 734             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 735
 736             fscal            = felec;
 737
 738             /* Update vectorial force */
 739             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 740             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 741             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 742
 743             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 744             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 745             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 746
 747             /**************************
 748              * CALCULATE INTERACTIONS *
 749              **************************/
 750
 751             /* Compute parameters for interactions between i and j atoms */
 752             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 753
 754             /* COULOMB ELECTROSTATICS */
 755             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 756             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 757
 758             fscal            = felec;
 759
 760             /* Update vectorial force */
 761             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 762             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 763             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 764
 765             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 766             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 767             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* Compute parameters for interactions between i and j atoms */
 774             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 775
 776             /* COULOMB ELECTROSTATICS */
 777             velec            = _fjsp_mul_v2r8(qq30,rinv30);
 778             felec            = _fjsp_mul_v2r8(velec,rinvsq30);
 779
 780             fscal            = felec;
 781
 782             /* Update vectorial force */
 783             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 784             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 785             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 786
 787             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 788             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 789             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 790
 791             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 792
 793             /* Inner loop uses 123 flops */
 794         }
 795
 796         if(jidx<j_index_end)
 797         {
 798
 799             jnrA             = jjnr[jidx];
 800             j_coord_offsetA  = DIM*jnrA;
 801
 802             /* load j atom coordinates */
 803             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 804                                               &jx0,&jy0,&jz0);
 805
 806             /* Calculate displacement vector */
 807             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 808             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 809             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 810             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 811             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 812             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 813             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 814             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 815             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 816             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 817             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 818             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 819
 820             /* Calculate squared distance and things based on it */
 821             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 822             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 823             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 824             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 825
 826             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 827             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 828             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 829
 830             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 831             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 832             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 833             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 834
 835             /* Load parameters for j particles */
 836             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 837             vdwjidx0A        = 2*vdwtype[jnrA+0];
 838
 839             fjx0             = _fjsp_setzero_v2r8();
 840             fjy0             = _fjsp_setzero_v2r8();
 841             fjz0             = _fjsp_setzero_v2r8();
 842
 843             /**************************
 844              * CALCULATE INTERACTIONS *
 845              **************************/
 846
 847             /* Compute parameters for interactions between i and j atoms */
 848             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 849                                          vdwparam+vdwioffset0+vdwjidx0B,&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 }