src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEw_VdwLJEw_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_ElecEw_VdwLJEw_GeomW4P1_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJEw_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           c6grid_00;
 102     _fjsp_v2r8           c6grid_10;
 103     _fjsp_v2r8           c6grid_20;
 104     _fjsp_v2r8           c6grid_30;
 105     real                 *vdwgridparam;
 106     _fjsp_v2r8           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 107     _fjsp_v2r8           one_half = gmx_fjsp_set1_v2r8(0.5);
 108     _fjsp_v2r8           minus_one = gmx_fjsp_set1_v2r8(-1.0);
 109     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 110     real             *ewtab;
 111     _fjsp_v2r8       itab_tmp;
 112     _fjsp_v2r8       dummy_mask,cutoff_mask;
 113     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 114     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 115     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 116
 117     x                = xx[0];
 118     f                = ff[0];
 119
 120     nri              = nlist->nri;
 121     iinr             = nlist->iinr;
 122     jindex           = nlist->jindex;
 123     jjnr             = nlist->jjnr;
 124     shiftidx         = nlist->shift;
 125     gid              = nlist->gid;
 126     shiftvec         = fr->shift_vec[0];
 127     fshift           = fr->fshift[0];
 128     facel            = gmx_fjsp_set1_v2r8(fr->ic->epsfac);
 129     charge           = mdatoms->chargeA;
 130     nvdwtype         = fr->ntype;
 131     vdwparam         = fr->nbfp;
 132     vdwtype          = mdatoms->typeA;
 133     vdwgridparam     = fr->ljpme_c6grid;
 134     sh_lj_ewald      = gmx_fjsp_set1_v2r8(fr->ic->sh_lj_ewald);
 135     ewclj            = gmx_fjsp_set1_v2r8(fr->ic->ewaldcoeff_lj);
 136     ewclj2           = _fjsp_mul_v2r8(minus_one,_fjsp_mul_v2r8(ewclj,ewclj));
 137
 138     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 139     ewtab            = fr->ic->tabq_coul_FDV0;
 140     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 141     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 142
 143     /* Setup water-specific parameters */
 144     inr              = nlist->iinr[0];
 145     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 146     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 147     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 148     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 149
 150     /* Avoid stupid compiler warnings */
 151     jnrA = jnrB = 0;
 152     j_coord_offsetA = 0;
 153     j_coord_offsetB = 0;
 154
 155     outeriter        = 0;
 156     inneriter        = 0;
 157
 158     /* Start outer loop over neighborlists */
 159     for(iidx=0; iidx<nri; iidx++)
 160     {
 161         /* Load shift vector for this list */
 162         i_shift_offset   = DIM*shiftidx[iidx];
 163
 164         /* Load limits for loop over neighbors */
 165         j_index_start    = jindex[iidx];
 166         j_index_end      = jindex[iidx+1];
 167
 168         /* Get outer coordinate index */
 169         inr              = iinr[iidx];
 170         i_coord_offset   = DIM*inr;
 171
 172         /* Load i particle coords and add shift vector */
 173         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 174                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 175
 176         fix0             = _fjsp_setzero_v2r8();
 177         fiy0             = _fjsp_setzero_v2r8();
 178         fiz0             = _fjsp_setzero_v2r8();
 179         fix1             = _fjsp_setzero_v2r8();
 180         fiy1             = _fjsp_setzero_v2r8();
 181         fiz1             = _fjsp_setzero_v2r8();
 182         fix2             = _fjsp_setzero_v2r8();
 183         fiy2             = _fjsp_setzero_v2r8();
 184         fiz2             = _fjsp_setzero_v2r8();
 185         fix3             = _fjsp_setzero_v2r8();
 186         fiy3             = _fjsp_setzero_v2r8();
 187         fiz3             = _fjsp_setzero_v2r8();
 188
 189         /* Reset potential sums */
 190         velecsum         = _fjsp_setzero_v2r8();
 191         vvdwsum          = _fjsp_setzero_v2r8();
 192
 193         /* Start inner kernel loop */
 194         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 195         {
 196
 197             /* Get j neighbor index, and coordinate index */
 198             jnrA             = jjnr[jidx];
 199             jnrB             = jjnr[jidx+1];
 200             j_coord_offsetA  = DIM*jnrA;
 201             j_coord_offsetB  = DIM*jnrB;
 202
 203             /* load j atom coordinates */
 204             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 205                                               &jx0,&jy0,&jz0);
 206
 207             /* Calculate displacement vector */
 208             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 209             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 210             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 211             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 212             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 213             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 214             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 215             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 216             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 217             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 218             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 219             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 220
 221             /* Calculate squared distance and things based on it */
 222             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 223             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 224             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 225             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 226
 227             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 228             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 229             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 230             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 231
 232             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 233             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 234             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 235             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 236
 237             /* Load parameters for j particles */
 238             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 239             vdwjidx0A        = 2*vdwtype[jnrA+0];
 240             vdwjidx0B        = 2*vdwtype[jnrB+0];
 241
 242             fjx0             = _fjsp_setzero_v2r8();
 243             fjy0             = _fjsp_setzero_v2r8();
 244             fjz0             = _fjsp_setzero_v2r8();
 245
 246             /**************************
 247              * CALCULATE INTERACTIONS *
 248              **************************/
 249
 250             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 251
 252             /* Compute parameters for interactions between i and j atoms */
 253             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 254                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 255
 256             c6grid_00       = gmx_fjsp_load_2real_swizzle_v2r8(vdwgridparam+vdwioffset0+vdwjidx0A,
 257                                                                    vdwgridparam+vdwioffset0+vdwjidx0B);
 258
 259             /* Analytical LJ-PME */
 260             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 261             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
 262             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
 263             exponent         = gmx_simd_exp_d(ewcljrsq);
 264             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 265             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
 266             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 267             vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(c6grid_00,_fjsp_sub_v2r8(poly,one),c6_00),rinvsix);
 268             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 269             vvdw             = _fjsp_msub_v2r8(vvdw12,one_twelfth,_fjsp_mul_v2r8(vvdw6,one_sixth));
 270             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 271             fvdw             = _fjsp_mul_v2r8(_fjsp_add_v2r8(vvdw12,_fjsp_msub_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6),vvdw6)),rinvsq00);
 272
 273             /* Update potential sum for this i atom from the interaction with this j atom. */
 274             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 275
 276             fscal            = fvdw;
 277
 278             /* Update vectorial force */
 279             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 280             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 281             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 282
 283             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 284             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 285             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 286
 287             /**************************
 288              * CALCULATE INTERACTIONS *
 289              **************************/
 290
 291             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 292
 293             /* Compute parameters for interactions between i and j atoms */
 294             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 295
 296             /* EWALD ELECTROSTATICS */
 297
 298             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 299             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 300             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 301             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 302             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 303
 304             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 305             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 306             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 307             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 308             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 309             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 310             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 311             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 312             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 313             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 314
 315             /* Update potential sum for this i atom from the interaction with this j atom. */
 316             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 317
 318             fscal            = felec;
 319
 320             /* Update vectorial force */
 321             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 322             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 323             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 324
 325             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 326             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 327             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 334
 335             /* Compute parameters for interactions between i and j atoms */
 336             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 337
 338             /* EWALD ELECTROSTATICS */
 339
 340             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 341             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 342             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 343             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 344             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 345
 346             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 347             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 348             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 349             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 350             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 351             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 352             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 353             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 354             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 355             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 356
 357             /* Update potential sum for this i atom from the interaction with this j atom. */
 358             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 359
 360             fscal            = felec;
 361
 362             /* Update vectorial force */
 363             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 364             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 365             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 366
 367             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 368             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 369             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 370
 371             /**************************
 372              * CALCULATE INTERACTIONS *
 373              **************************/
 374
 375             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
 376
 377             /* Compute parameters for interactions between i and j atoms */
 378             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 379
 380             /* EWALD ELECTROSTATICS */
 381
 382             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 383             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
 384             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 385             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 386             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 387
 388             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 389             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 390             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 391             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 392             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 393             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 394             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 395             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 396             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(rinv30,velec));
 397             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 401
 402             fscal            = felec;
 403
 404             /* Update vectorial force */
 405             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 406             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 407             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 408
 409             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 410             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 411             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 412
 413             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 414
 415             /* Inner loop uses 185 flops */
 416         }
 417
 418         if(jidx<j_index_end)
 419         {
 420
 421             jnrA             = jjnr[jidx];
 422             j_coord_offsetA  = DIM*jnrA;
 423
 424             /* load j atom coordinates */
 425             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 426                                               &jx0,&jy0,&jz0);
 427
 428             /* Calculate displacement vector */
 429             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 430             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 431             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 432             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 433             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 434             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 435             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 436             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 437             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 438             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 439             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 440             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 441
 442             /* Calculate squared distance and things based on it */
 443             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 444             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 445             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 446             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 447
 448             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 449             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 450             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 451             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 452
 453             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 454             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 455             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 456             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 457
 458             /* Load parameters for j particles */
 459             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 460             vdwjidx0A        = 2*vdwtype[jnrA+0];
 461
 462             fjx0             = _fjsp_setzero_v2r8();
 463             fjy0             = _fjsp_setzero_v2r8();
 464             fjz0             = _fjsp_setzero_v2r8();
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 471
 472             /* Compute parameters for interactions between i and j atoms */
 473             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 474                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 475
 476             c6grid_00       = gmx_fjsp_load_2real_swizzle_v2r8(vdwgridparam+vdwioffset0+vdwjidx0A,
 477                                                                    vdwgridparam+vdwioffset0+vdwjidx0B);
 478
 479             /* Analytical LJ-PME */
 480             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 481             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
 482             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
 483             exponent         = gmx_simd_exp_d(ewcljrsq);
 484             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 485             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
 486             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 487             vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(c6grid_00,_fjsp_sub_v2r8(poly,one),c6_00),rinvsix);
 488             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 489             vvdw             = _fjsp_msub_v2r8(vvdw12,one_twelfth,_fjsp_mul_v2r8(vvdw6,one_sixth));
 490             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 491             fvdw             = _fjsp_mul_v2r8(_fjsp_add_v2r8(vvdw12,_fjsp_msub_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6),vvdw6)),rinvsq00);
 492
 493             /* Update potential sum for this i atom from the interaction with this j atom. */
 494             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 495             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 496
 497             fscal            = fvdw;
 498
 499             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 500
 501             /* Update vectorial force */
 502             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 503             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 504             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 505
 506             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 507             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 508             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 509
 510             /**************************
 511              * CALCULATE INTERACTIONS *
 512              **************************/
 513
 514             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 515
 516             /* Compute parameters for interactions between i and j atoms */
 517             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 518
 519             /* EWALD ELECTROSTATICS */
 520
 521             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 522             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 523             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 524             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 525             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 526
 527             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 528             ewtabD           = _fjsp_setzero_v2r8();
 529             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 530             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 531             ewtabFn          = _fjsp_setzero_v2r8();
 532             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 533             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 534             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 535             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 536             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 540             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 541
 542             fscal            = felec;
 543
 544             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 545
 546             /* Update vectorial force */
 547             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 548             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 549             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 550
 551             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 552             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 553             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 560
 561             /* Compute parameters for interactions between i and j atoms */
 562             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 563
 564             /* EWALD ELECTROSTATICS */
 565
 566             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 567             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 568             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 569             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 570             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 571
 572             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 573             ewtabD           = _fjsp_setzero_v2r8();
 574             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 575             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 576             ewtabFn          = _fjsp_setzero_v2r8();
 577             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 578             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 579             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 580             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 581             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 582
 583             /* Update potential sum for this i atom from the interaction with this j atom. */
 584             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 585             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 586
 587             fscal            = felec;
 588
 589             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 590
 591             /* Update vectorial force */
 592             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 593             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 594             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 595
 596             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 597             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 598             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 599
 600             /**************************
 601              * CALCULATE INTERACTIONS *
 602              **************************/
 603
 604             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
 605
 606             /* Compute parameters for interactions between i and j atoms */
 607             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 608
 609             /* EWALD ELECTROSTATICS */
 610
 611             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 612             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
 613             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 614             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 615             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 616
 617             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 618             ewtabD           = _fjsp_setzero_v2r8();
 619             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 620             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 621             ewtabFn          = _fjsp_setzero_v2r8();
 622             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 623             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 624             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 625             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(rinv30,velec));
 626             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
 627
 628             /* Update potential sum for this i atom from the interaction with this j atom. */
 629             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 630             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 631
 632             fscal            = felec;
 633
 634             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 635
 636             /* Update vectorial force */
 637             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 638             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 639             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 640
 641             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 642             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 643             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 644
 645             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 646
 647             /* Inner loop uses 185 flops */
 648         }
 649
 650         /* End of innermost loop */
 651
 652         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 653                                               f+i_coord_offset,fshift+i_shift_offset);
 654
 655         ggid                        = gid[iidx];
 656         /* Update potential energies */
 657         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 658         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 659
 660         /* Increment number of inner iterations */
 661         inneriter                  += j_index_end - j_index_start;
 662
 663         /* Outer loop uses 26 flops */
 664     }
 665
 666     /* Increment number of outer iterations */
 667     outeriter        += nri;
 668
 669     /* Update outer/inner flops */
 670
 671     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*185);
 672 }
 673 /*
 674  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomW4P1_F_sparc64_hpc_ace_double
 675  * Electrostatics interaction: Ewald
 676  * VdW interaction:            LJEwald
 677  * Geometry:                   Water4-Particle
 678  * Calculate force/pot:        Force
 679  */
 680 void
 681 nb_kernel_ElecEw_VdwLJEw_GeomW4P1_F_sparc64_hpc_ace_double
 682                     (t_nblist                    * gmx_restrict       nlist,
 683                      rvec                        * gmx_restrict          xx,
 684                      rvec                        * gmx_restrict          ff,
 685                      struct t_forcerec           * gmx_restrict          fr,
 686                      t_mdatoms                   * gmx_restrict     mdatoms,
 687                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 688                      t_nrnb                      * gmx_restrict        nrnb)
 689 {
 690     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 691      * just 0 for non-waters.
 692      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 693      * jnr indices corresponding to data put in the four positions in the SIMD register.
 694      */
 695     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 696     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 697     int              jnrA,jnrB;
 698     int              j_coord_offsetA,j_coord_offsetB;
 699     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 700     real             rcutoff_scalar;
 701     real             *shiftvec,*fshift,*x,*f;
 702     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 703     int              vdwioffset0;
 704     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 705     int              vdwioffset1;
 706     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 707     int              vdwioffset2;
 708     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 709     int              vdwioffset3;
 710     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 711     int              vdwjidx0A,vdwjidx0B;
 712     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 713     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 714     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 715     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 716     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 717     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 718     real             *charge;
 719     int              nvdwtype;
 720     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 721     int              *vdwtype;
 722     real             *vdwparam;
 723     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 724     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 725     _fjsp_v2r8           c6grid_00;
 726     _fjsp_v2r8           c6grid_10;
 727     _fjsp_v2r8           c6grid_20;
 728     _fjsp_v2r8           c6grid_30;
 729     real                 *vdwgridparam;
 730     _fjsp_v2r8           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 731     _fjsp_v2r8           one_half = gmx_fjsp_set1_v2r8(0.5);
 732     _fjsp_v2r8           minus_one = gmx_fjsp_set1_v2r8(-1.0);
 733     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 734     real             *ewtab;
 735     _fjsp_v2r8       itab_tmp;
 736     _fjsp_v2r8       dummy_mask,cutoff_mask;
 737     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 738     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 739     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 740
 741     x                = xx[0];
 742     f                = ff[0];
 743
 744     nri              = nlist->nri;
 745     iinr             = nlist->iinr;
 746     jindex           = nlist->jindex;
 747     jjnr             = nlist->jjnr;
 748     shiftidx         = nlist->shift;
 749     gid              = nlist->gid;
 750     shiftvec         = fr->shift_vec[0];
 751     fshift           = fr->fshift[0];
 752     facel            = gmx_fjsp_set1_v2r8(fr->ic->epsfac);
 753     charge           = mdatoms->chargeA;
 754     nvdwtype         = fr->ntype;
 755     vdwparam         = fr->nbfp;
 756     vdwtype          = mdatoms->typeA;
 757     vdwgridparam     = fr->ljpme_c6grid;
 758     sh_lj_ewald      = gmx_fjsp_set1_v2r8(fr->ic->sh_lj_ewald);
 759     ewclj            = gmx_fjsp_set1_v2r8(fr->ic->ewaldcoeff_lj);
 760     ewclj2           = _fjsp_mul_v2r8(minus_one,_fjsp_mul_v2r8(ewclj,ewclj));
 761
 762     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 763     ewtab            = fr->ic->tabq_coul_F;
 764     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 765     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 766
 767     /* Setup water-specific parameters */
 768     inr              = nlist->iinr[0];
 769     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 770     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 771     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 772     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 773
 774     /* Avoid stupid compiler warnings */
 775     jnrA = jnrB = 0;
 776     j_coord_offsetA = 0;
 777     j_coord_offsetB = 0;
 778
 779     outeriter        = 0;
 780     inneriter        = 0;
 781
 782     /* Start outer loop over neighborlists */
 783     for(iidx=0; iidx<nri; iidx++)
 784     {
 785         /* Load shift vector for this list */
 786         i_shift_offset   = DIM*shiftidx[iidx];
 787
 788         /* Load limits for loop over neighbors */
 789         j_index_start    = jindex[iidx];
 790         j_index_end      = jindex[iidx+1];
 791
 792         /* Get outer coordinate index */
 793         inr              = iinr[iidx];
 794         i_coord_offset   = DIM*inr;
 795
 796         /* Load i particle coords and add shift vector */
 797         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 798                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 799
 800         fix0             = _fjsp_setzero_v2r8();
 801         fiy0             = _fjsp_setzero_v2r8();
 802         fiz0             = _fjsp_setzero_v2r8();
 803         fix1             = _fjsp_setzero_v2r8();
 804         fiy1             = _fjsp_setzero_v2r8();
 805         fiz1             = _fjsp_setzero_v2r8();
 806         fix2             = _fjsp_setzero_v2r8();
 807         fiy2             = _fjsp_setzero_v2r8();
 808         fiz2             = _fjsp_setzero_v2r8();
 809         fix3             = _fjsp_setzero_v2r8();
 810         fiy3             = _fjsp_setzero_v2r8();
 811         fiz3             = _fjsp_setzero_v2r8();
 812
 813         /* Start inner kernel loop */
 814         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 815         {
 816
 817             /* Get j neighbor index, and coordinate index */
 818             jnrA             = jjnr[jidx];
 819             jnrB             = jjnr[jidx+1];
 820             j_coord_offsetA  = DIM*jnrA;
 821             j_coord_offsetB  = DIM*jnrB;
 822
 823             /* load j atom coordinates */
 824             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 825                                               &jx0,&jy0,&jz0);
 826
 827             /* Calculate displacement vector */
 828             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 829             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 830             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 831             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 832             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 833             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 834             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 835             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 836             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 837             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 838             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 839             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 840
 841             /* Calculate squared distance and things based on it */
 842             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 843             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 844             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 845             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 846
 847             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 848             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 849             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 850             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 851
 852             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 853             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 854             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 855             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 856
 857             /* Load parameters for j particles */
 858             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 859             vdwjidx0A        = 2*vdwtype[jnrA+0];
 860             vdwjidx0B        = 2*vdwtype[jnrB+0];
 861
 862             fjx0             = _fjsp_setzero_v2r8();
 863             fjy0             = _fjsp_setzero_v2r8();
 864             fjz0             = _fjsp_setzero_v2r8();
 865
 866             /**************************
 867              * CALCULATE INTERACTIONS *
 868              **************************/
 869
 870             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 871
 872             /* Compute parameters for interactions between i and j atoms */
 873             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 874                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 875
 876             c6grid_00       = gmx_fjsp_load_2real_swizzle_v2r8(vdwgridparam+vdwioffset0+vdwjidx0A,
 877                                                                    vdwgridparam+vdwioffset0+vdwjidx0B);
 878
 879             /* Analytical LJ-PME */
 880             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 881             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
 882             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
 883             exponent         = gmx_simd_exp_d(ewcljrsq);
 884             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 885             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
 886             /* f6A = 6 * C6grid * (1 - poly) */
 887             f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_sub_v2r8(one,poly));
 888             /* f6B = C6grid * exponent * beta^6 */
 889             f6B              = _fjsp_mul_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6));
 890             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 891             fvdw              = _fjsp_mul_v2r8(_fjsp_madd_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,_fjsp_sub_v2r8(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 892
 893             fscal            = fvdw;
 894
 895             /* Update vectorial force */
 896             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 897             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 898             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 899
 900             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 901             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 902             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 909
 910             /* Compute parameters for interactions between i and j atoms */
 911             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 912
 913             /* EWALD ELECTROSTATICS */
 914
 915             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 916             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 917             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 918             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 919             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 920
 921             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 922                                          &ewtabF,&ewtabFn);
 923             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 924             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 925
 926             fscal            = felec;
 927
 928             /* Update vectorial force */
 929             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 930             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 931             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 932
 933             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 934             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 935             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 936
 937             /**************************
 938              * CALCULATE INTERACTIONS *
 939              **************************/
 940
 941             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 942
 943             /* Compute parameters for interactions between i and j atoms */
 944             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 945
 946             /* EWALD ELECTROSTATICS */
 947
 948             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 949             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 950             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 951             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 952             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 953
 954             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 955                                          &ewtabF,&ewtabFn);
 956             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 957             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 958
 959             fscal            = felec;
 960
 961             /* Update vectorial force */
 962             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 963             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 964             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 965
 966             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 967             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 968             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 969
 970             /**************************
 971              * CALCULATE INTERACTIONS *
 972              **************************/
 973
 974             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
 975
 976             /* Compute parameters for interactions between i and j atoms */
 977             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 978
 979             /* EWALD ELECTROSTATICS */
 980
 981             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 982             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
 983             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 984             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 985             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 986
 987             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 988                                          &ewtabF,&ewtabFn);
 989             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 990             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
 991
 992             fscal            = felec;
 993
 994             /* Update vectorial force */
 995             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 996             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 997             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 998
 999             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
1000             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
1001             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
1002
1003             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
1004
1005             /* Inner loop uses 168 flops */
1006         }
1007
1008         if(jidx<j_index_end)
1009         {
1010
1011             jnrA             = jjnr[jidx];
1012             j_coord_offsetA  = DIM*jnrA;
1013
1014             /* load j atom coordinates */
1015             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1016                                               &jx0,&jy0,&jz0);
1017
1018             /* Calculate displacement vector */
1019             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1020             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1021             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1022             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1023             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1024             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1025             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1026             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1027             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1028             dx30             = _fjsp_sub_v2r8(ix3,jx0);
1029             dy30             = _fjsp_sub_v2r8(iy3,jy0);
1030             dz30             = _fjsp_sub_v2r8(iz3,jz0);
1031
1032             /* Calculate squared distance and things based on it */
1033             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1034             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1035             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1036             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
1037
1038             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1039             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1040             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1041             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
1042
1043             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1044             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1045             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1046             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
1047
1048             /* Load parameters for j particles */
1049             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
1050             vdwjidx0A        = 2*vdwtype[jnrA+0];
1051
1052             fjx0             = _fjsp_setzero_v2r8();
1053             fjy0             = _fjsp_setzero_v2r8();
1054             fjz0             = _fjsp_setzero_v2r8();
1055
1056             /**************************
1057              * CALCULATE INTERACTIONS *
1058              **************************/
1059
1060             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1061
1062             /* Compute parameters for interactions between i and j atoms */
1063             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
1064                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
1065
1066             c6grid_00       = gmx_fjsp_load_2real_swizzle_v2r8(vdwgridparam+vdwioffset0+vdwjidx0A,
1067                                                                    vdwgridparam+vdwioffset0+vdwjidx0B);
1068
1069             /* Analytical LJ-PME */
1070             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1071             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
1072             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
1073             exponent         = gmx_simd_exp_d(ewcljrsq);
1074             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1075             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
1076             /* f6A = 6 * C6grid * (1 - poly) */
1077             f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_sub_v2r8(one,poly));
1078             /* f6B = C6grid * exponent * beta^6 */
1079             f6B              = _fjsp_mul_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6));
1080             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1081             fvdw              = _fjsp_mul_v2r8(_fjsp_madd_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,_fjsp_sub_v2r8(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1082
1083             fscal            = fvdw;
1084
1085             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1086
1087             /* Update vectorial force */
1088             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1089             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1090             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1091
1092             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1093             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1094             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1095
1096             /**************************
1097              * CALCULATE INTERACTIONS *
1098              **************************/
1099
1100             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1101
1102             /* Compute parameters for interactions between i and j atoms */
1103             qq10             = _fjsp_mul_v2r8(iq1,jq0);
1104
1105             /* EWALD ELECTROSTATICS */
1106
1107             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1108             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1109             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1110             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1111             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1112
1113             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1114             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1115             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1116
1117             fscal            = felec;
1118
1119             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1120
1121             /* Update vectorial force */
1122             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1123             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1124             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1125
1126             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1127             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1128             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1129
1130             /**************************
1131              * CALCULATE INTERACTIONS *
1132              **************************/
1133
1134             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1135
1136             /* Compute parameters for interactions between i and j atoms */
1137             qq20             = _fjsp_mul_v2r8(iq2,jq0);
1138
1139             /* EWALD ELECTROSTATICS */
1140
1141             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1142             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1143             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1144             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1145             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1146
1147             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1148             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1149             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1150
1151             fscal            = felec;
1152
1153             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1154
1155             /* Update vectorial force */
1156             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1157             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1158             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1159
1160             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1161             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1162             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1163
1164             /**************************
1165              * CALCULATE INTERACTIONS *
1166              **************************/
1167
1168             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
1169
1170             /* Compute parameters for interactions between i and j atoms */
1171             qq30             = _fjsp_mul_v2r8(iq3,jq0);
1172
1173             /* EWALD ELECTROSTATICS */
1174
1175             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1176             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
1177             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1178             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1179             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1180
1181             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1182             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1183             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
1184
1185             fscal            = felec;
1186
1187             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1188
1189             /* Update vectorial force */
1190             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
1191             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
1192             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
1193
1194             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
1195             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
1196             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
1197
1198             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1199
1200             /* Inner loop uses 168 flops */
1201         }
1202
1203         /* End of innermost loop */
1204
1205         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1206                                               f+i_coord_offset,fshift+i_shift_offset);
1207
1208         /* Increment number of inner iterations */
1209         inneriter                  += j_index_end - j_index_start;
1210
1211         /* Outer loop uses 24 flops */
1212     }
1213
1214     /* Increment number of outer iterations */
1215     outeriter        += nri;
1216
1217     /* Update outer/inner flops */
1218
1219     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*168);
1220 }