src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRF_VdwLJ_GeomW4P1_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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_ElecRF_VdwLJ_GeomW4P1_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_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     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 121     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 122     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 130     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 131     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 132     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 133
 134     /* Avoid stupid compiler warnings */
 135     jnrA = jnrB = 0;
 136     j_coord_offsetA = 0;
 137     j_coord_offsetB = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     /* Start outer loop over neighborlists */
 143     for(iidx=0; iidx<nri; iidx++)
 144     {
 145         /* Load shift vector for this list */
 146         i_shift_offset   = DIM*shiftidx[iidx];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 158                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 159
 160         fix0             = _fjsp_setzero_v2r8();
 161         fiy0             = _fjsp_setzero_v2r8();
 162         fiz0             = _fjsp_setzero_v2r8();
 163         fix1             = _fjsp_setzero_v2r8();
 164         fiy1             = _fjsp_setzero_v2r8();
 165         fiz1             = _fjsp_setzero_v2r8();
 166         fix2             = _fjsp_setzero_v2r8();
 167         fiy2             = _fjsp_setzero_v2r8();
 168         fiz2             = _fjsp_setzero_v2r8();
 169         fix3             = _fjsp_setzero_v2r8();
 170         fiy3             = _fjsp_setzero_v2r8();
 171         fiz3             = _fjsp_setzero_v2r8();
 172
 173         /* Reset potential sums */
 174         velecsum         = _fjsp_setzero_v2r8();
 175         vvdwsum          = _fjsp_setzero_v2r8();
 176
 177         /* Start inner kernel loop */
 178         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 179         {
 180
 181             /* Get j neighbor index, and coordinate index */
 182             jnrA             = jjnr[jidx];
 183             jnrB             = jjnr[jidx+1];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186
 187             /* load j atom coordinates */
 188             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                               &jx0,&jy0,&jz0);
 190
 191             /* Calculate displacement vector */
 192             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 193             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 194             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 195             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 196             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 197             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 198             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 199             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 200             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 201             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 202             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 203             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 204
 205             /* Calculate squared distance and things based on it */
 206             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 207             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 208             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 209             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 210
 211             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 212             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 213             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 214
 215             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 216             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 217             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 218             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 219
 220             /* Load parameters for j particles */
 221             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 222             vdwjidx0A        = 2*vdwtype[jnrA+0];
 223             vdwjidx0B        = 2*vdwtype[jnrB+0];
 224
 225             fjx0             = _fjsp_setzero_v2r8();
 226             fjy0             = _fjsp_setzero_v2r8();
 227             fjz0             = _fjsp_setzero_v2r8();
 228
 229             /**************************
 230              * CALCULATE INTERACTIONS *
 231              **************************/
 232
 233             /* Compute parameters for interactions between i and j atoms */
 234             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 235                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 236
 237             /* LENNARD-JONES DISPERSION/REPULSION */
 238
 239             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 240             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 241             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 242             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 243             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 244
 245             /* Update potential sum for this i atom from the interaction with this j atom. */
 246             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 247
 248             fscal            = fvdw;
 249
 250             /* Update vectorial force */
 251             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 252             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 253             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 254
 255             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 256             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 257             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 258
 259             /**************************
 260              * CALCULATE INTERACTIONS *
 261              **************************/
 262
 263             /* Compute parameters for interactions between i and j atoms */
 264             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 265
 266             /* REACTION-FIELD ELECTROSTATICS */
 267             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 268             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 269
 270             /* Update potential sum for this i atom from the interaction with this j atom. */
 271             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 272
 273             fscal            = felec;
 274
 275             /* Update vectorial force */
 276             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 277             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 278             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 279
 280             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 281             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 282             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 283
 284             /**************************
 285              * CALCULATE INTERACTIONS *
 286              **************************/
 287
 288             /* Compute parameters for interactions between i and j atoms */
 289             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 290
 291             /* REACTION-FIELD ELECTROSTATICS */
 292             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 293             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 294
 295             /* Update potential sum for this i atom from the interaction with this j atom. */
 296             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 297
 298             fscal            = felec;
 299
 300             /* Update vectorial force */
 301             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 302             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 303             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 304
 305             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 306             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 307             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             /* Compute parameters for interactions between i and j atoms */
 314             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 315
 316             /* REACTION-FIELD ELECTROSTATICS */
 317             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 318             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 319
 320             /* Update potential sum for this i atom from the interaction with this j atom. */
 321             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 322
 323             fscal            = felec;
 324
 325             /* Update vectorial force */
 326             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 327             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 328             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 329
 330             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 331             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 332             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 333
 334             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 335
 336             /* Inner loop uses 143 flops */
 337         }
 338
 339         if(jidx<j_index_end)
 340         {
 341
 342             jnrA             = jjnr[jidx];
 343             j_coord_offsetA  = DIM*jnrA;
 344
 345             /* load j atom coordinates */
 346             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 347                                               &jx0,&jy0,&jz0);
 348
 349             /* Calculate displacement vector */
 350             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 351             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 352             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 353             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 354             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 355             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 356             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 357             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 358             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 359             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 360             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 361             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 362
 363             /* Calculate squared distance and things based on it */
 364             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 365             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 366             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 367             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 368
 369             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 370             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 371             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 372
 373             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 374             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 375             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 376             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 377
 378             /* Load parameters for j particles */
 379             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 380             vdwjidx0A        = 2*vdwtype[jnrA+0];
 381
 382             fjx0             = _fjsp_setzero_v2r8();
 383             fjy0             = _fjsp_setzero_v2r8();
 384             fjz0             = _fjsp_setzero_v2r8();
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             /* Compute parameters for interactions between i and j atoms */
 391             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 392                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 393
 394             /* LENNARD-JONES DISPERSION/REPULSION */
 395
 396             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 397             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 398             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 399             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 400             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 404             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 405
 406             fscal            = fvdw;
 407
 408             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 409
 410             /* Update vectorial force */
 411             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 412             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 413             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 414
 415             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 416             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 417             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             /* Compute parameters for interactions between i and j atoms */
 424             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 425
 426             /* REACTION-FIELD ELECTROSTATICS */
 427             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 428             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 432             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 437
 438             /* Update vectorial force */
 439             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 440             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 441             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 442
 443             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 444             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 445             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             /* Compute parameters for interactions between i and j atoms */
 452             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 453
 454             /* REACTION-FIELD ELECTROSTATICS */
 455             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 456             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 457
 458             /* Update potential sum for this i atom from the interaction with this j atom. */
 459             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 460             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 461
 462             fscal            = felec;
 463
 464             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 465
 466             /* Update vectorial force */
 467             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 468             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 469             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 470
 471             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 472             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 473             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 474
 475             /**************************
 476              * CALCULATE INTERACTIONS *
 477              **************************/
 478
 479             /* Compute parameters for interactions between i and j atoms */
 480             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 481
 482             /* REACTION-FIELD ELECTROSTATICS */
 483             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 484             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 485
 486             /* Update potential sum for this i atom from the interaction with this j atom. */
 487             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 488             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 493
 494             /* Update vectorial force */
 495             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 496             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 497             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 498
 499             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 500             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 501             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 502
 503             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 504
 505             /* Inner loop uses 143 flops */
 506         }
 507
 508         /* End of innermost loop */
 509
 510         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 511                                               f+i_coord_offset,fshift+i_shift_offset);
 512
 513         ggid                        = gid[iidx];
 514         /* Update potential energies */
 515         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 516         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 517
 518         /* Increment number of inner iterations */
 519         inneriter                  += j_index_end - j_index_start;
 520
 521         /* Outer loop uses 26 flops */
 522     }
 523
 524     /* Increment number of outer iterations */
 525     outeriter        += nri;
 526
 527     /* Update outer/inner flops */
 528
 529     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*143);
 530 }
 531 /*
 532  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 533  * Electrostatics interaction: ReactionField
 534  * VdW interaction:            LennardJones
 535  * Geometry:                   Water4-Particle
 536  * Calculate force/pot:        Force
 537  */
 538 void
 539 nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 540                     (t_nblist                    * gmx_restrict       nlist,
 541                      rvec                        * gmx_restrict          xx,
 542                      rvec                        * gmx_restrict          ff,
 543                      struct t_forcerec           * gmx_restrict          fr,
 544                      t_mdatoms                   * gmx_restrict     mdatoms,
 545                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 546                      t_nrnb                      * gmx_restrict        nrnb)
 547 {
 548     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 549      * just 0 for non-waters.
 550      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 551      * jnr indices corresponding to data put in the four positions in the SIMD register.
 552      */
 553     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 554     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 555     int              jnrA,jnrB;
 556     int              j_coord_offsetA,j_coord_offsetB;
 557     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 558     real             rcutoff_scalar;
 559     real             *shiftvec,*fshift,*x,*f;
 560     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 561     int              vdwioffset0;
 562     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 563     int              vdwioffset1;
 564     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 565     int              vdwioffset2;
 566     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 567     int              vdwioffset3;
 568     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 569     int              vdwjidx0A,vdwjidx0B;
 570     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 571     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 572     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 573     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 574     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 575     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 576     real             *charge;
 577     int              nvdwtype;
 578     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 579     int              *vdwtype;
 580     real             *vdwparam;
 581     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 582     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 583     _fjsp_v2r8       itab_tmp;
 584     _fjsp_v2r8       dummy_mask,cutoff_mask;
 585     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 586     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 587     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 588
 589     x                = xx[0];
 590     f                = ff[0];
 591
 592     nri              = nlist->nri;
 593     iinr             = nlist->iinr;
 594     jindex           = nlist->jindex;
 595     jjnr             = nlist->jjnr;
 596     shiftidx         = nlist->shift;
 597     gid              = nlist->gid;
 598     shiftvec         = fr->shift_vec[0];
 599     fshift           = fr->fshift[0];
 600     facel            = gmx_fjsp_set1_v2r8(fr->ic->epsfac);
 601     charge           = mdatoms->chargeA;
 602     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 603     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 604     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 605     nvdwtype         = fr->ntype;
 606     vdwparam         = fr->nbfp;
 607     vdwtype          = mdatoms->typeA;
 608
 609     /* Setup water-specific parameters */
 610     inr              = nlist->iinr[0];
 611     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 612     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 613     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 614     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 615
 616     /* Avoid stupid compiler warnings */
 617     jnrA = jnrB = 0;
 618     j_coord_offsetA = 0;
 619     j_coord_offsetB = 0;
 620
 621     outeriter        = 0;
 622     inneriter        = 0;
 623
 624     /* Start outer loop over neighborlists */
 625     for(iidx=0; iidx<nri; iidx++)
 626     {
 627         /* Load shift vector for this list */
 628         i_shift_offset   = DIM*shiftidx[iidx];
 629
 630         /* Load limits for loop over neighbors */
 631         j_index_start    = jindex[iidx];
 632         j_index_end      = jindex[iidx+1];
 633
 634         /* Get outer coordinate index */
 635         inr              = iinr[iidx];
 636         i_coord_offset   = DIM*inr;
 637
 638         /* Load i particle coords and add shift vector */
 639         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 640                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 641
 642         fix0             = _fjsp_setzero_v2r8();
 643         fiy0             = _fjsp_setzero_v2r8();
 644         fiz0             = _fjsp_setzero_v2r8();
 645         fix1             = _fjsp_setzero_v2r8();
 646         fiy1             = _fjsp_setzero_v2r8();
 647         fiz1             = _fjsp_setzero_v2r8();
 648         fix2             = _fjsp_setzero_v2r8();
 649         fiy2             = _fjsp_setzero_v2r8();
 650         fiz2             = _fjsp_setzero_v2r8();
 651         fix3             = _fjsp_setzero_v2r8();
 652         fiy3             = _fjsp_setzero_v2r8();
 653         fiz3             = _fjsp_setzero_v2r8();
 654
 655         /* Start inner kernel loop */
 656         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 657         {
 658
 659             /* Get j neighbor index, and coordinate index */
 660             jnrA             = jjnr[jidx];
 661             jnrB             = jjnr[jidx+1];
 662             j_coord_offsetA  = DIM*jnrA;
 663             j_coord_offsetB  = DIM*jnrB;
 664
 665             /* load j atom coordinates */
 666             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 667                                               &jx0,&jy0,&jz0);
 668
 669             /* Calculate displacement vector */
 670             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 671             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 672             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 673             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 674             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 675             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 676             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 677             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 678             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 679             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 680             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 681             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 682
 683             /* Calculate squared distance and things based on it */
 684             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 685             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 686             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 687             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 688
 689             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 690             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 691             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 692
 693             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 694             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 695             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 696             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 697
 698             /* Load parameters for j particles */
 699             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 700             vdwjidx0A        = 2*vdwtype[jnrA+0];
 701             vdwjidx0B        = 2*vdwtype[jnrB+0];
 702
 703             fjx0             = _fjsp_setzero_v2r8();
 704             fjy0             = _fjsp_setzero_v2r8();
 705             fjz0             = _fjsp_setzero_v2r8();
 706
 707             /**************************
 708              * CALCULATE INTERACTIONS *
 709              **************************/
 710
 711             /* Compute parameters for interactions between i and j atoms */
 712             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 713                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 714
 715             /* LENNARD-JONES DISPERSION/REPULSION */
 716
 717             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 718             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 719
 720             fscal            = fvdw;
 721
 722             /* Update vectorial force */
 723             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 724             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 725             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 726
 727             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 728             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 729             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 730
 731             /**************************
 732              * CALCULATE INTERACTIONS *
 733              **************************/
 734
 735             /* Compute parameters for interactions between i and j atoms */
 736             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 737
 738             /* REACTION-FIELD ELECTROSTATICS */
 739             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 740
 741             fscal            = felec;
 742
 743             /* Update vectorial force */
 744             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 745             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 746             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 747
 748             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 749             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 750             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             /* Compute parameters for interactions between i and j atoms */
 757             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 758
 759             /* REACTION-FIELD ELECTROSTATICS */
 760             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 761
 762             fscal            = felec;
 763
 764             /* Update vectorial force */
 765             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 766             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 767             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 768
 769             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 770             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 771             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 772
 773             /**************************
 774              * CALCULATE INTERACTIONS *
 775              **************************/
 776
 777             /* Compute parameters for interactions between i and j atoms */
 778             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 779
 780             /* REACTION-FIELD ELECTROSTATICS */
 781             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 782
 783             fscal            = felec;
 784
 785             /* Update vectorial force */
 786             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 787             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 788             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 789
 790             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 791             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 792             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 793
 794             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 795
 796             /* Inner loop uses 123 flops */
 797         }
 798
 799         if(jidx<j_index_end)
 800         {
 801
 802             jnrA             = jjnr[jidx];
 803             j_coord_offsetA  = DIM*jnrA;
 804
 805             /* load j atom coordinates */
 806             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 807                                               &jx0,&jy0,&jz0);
 808
 809             /* Calculate displacement vector */
 810             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 811             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 812             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 813             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 814             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 815             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 816             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 817             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 818             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 819             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 820             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 821             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 822
 823             /* Calculate squared distance and things based on it */
 824             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 825             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 826             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 827             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 828
 829             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 830             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 831             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 832
 833             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 834             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 835             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 836             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 837
 838             /* Load parameters for j particles */
 839             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 840             vdwjidx0A        = 2*vdwtype[jnrA+0];
 841
 842             fjx0             = _fjsp_setzero_v2r8();
 843             fjy0             = _fjsp_setzero_v2r8();
 844             fjz0             = _fjsp_setzero_v2r8();
 845
 846             /**************************
 847              * CALCULATE INTERACTIONS *
 848              **************************/
 849
 850             /* Compute parameters for interactions between i and j atoms */
 851             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 852                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 853
 854             /* LENNARD-JONES DISPERSION/REPULSION */
 855
 856             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 857             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 858
 859             fscal            = fvdw;
 860
 861             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 862
 863             /* Update vectorial force */
 864             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 865             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 866             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 867
 868             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 869             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 870             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 871
 872             /**************************
 873              * CALCULATE INTERACTIONS *
 874              **************************/
 875
 876             /* Compute parameters for interactions between i and j atoms */
 877             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 878
 879             /* REACTION-FIELD ELECTROSTATICS */
 880             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 881
 882             fscal            = felec;
 883
 884             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 885
 886             /* Update vectorial force */
 887             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 888             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 889             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 890
 891             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 892             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 893             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 894
 895             /**************************
 896              * CALCULATE INTERACTIONS *
 897              **************************/
 898
 899             /* Compute parameters for interactions between i and j atoms */
 900             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 901
 902             /* REACTION-FIELD ELECTROSTATICS */
 903             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 904
 905             fscal            = felec;
 906
 907             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 908
 909             /* Update vectorial force */
 910             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 911             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 912             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 913
 914             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 915             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 916             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 917
 918             /**************************
 919              * CALCULATE INTERACTIONS *
 920              **************************/
 921
 922             /* Compute parameters for interactions between i and j atoms */
 923             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 924
 925             /* REACTION-FIELD ELECTROSTATICS */
 926             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 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 }