src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEw_VdwLJ_GeomW3W3_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sparc64_hpc_ace_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      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              vdwjidx0A,vdwjidx0B;
  86     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     int              vdwjidx1A,vdwjidx1B;
  88     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     int              nvdwtype;
 103     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 104     int              *vdwtype;
 105     real             *vdwparam;
 106     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 107     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 108     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 109     real             *ewtab;
 110     _fjsp_v2r8       itab_tmp;
 111     _fjsp_v2r8       dummy_mask,cutoff_mask;
 112     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 113     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 114     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 115
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 128     charge           = mdatoms->chargeA;
 129     nvdwtype         = fr->ntype;
 130     vdwparam         = fr->nbfp;
 131     vdwtype          = mdatoms->typeA;
 132
 133     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 136     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 141     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 142     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 146     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 147     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 148     vdwjidx0A        = 2*vdwtype[inr+0];
 149     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 150     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 151     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 152     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 153     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 154     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 155     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 156     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 157     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 158     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 159     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 160
 161     /* Avoid stupid compiler warnings */
 162     jnrA = jnrB = 0;
 163     j_coord_offsetA = 0;
 164     j_coord_offsetB = 0;
 165
 166     outeriter        = 0;
 167     inneriter        = 0;
 168
 169     /* Start outer loop over neighborlists */
 170     for(iidx=0; iidx<nri; iidx++)
 171     {
 172         /* Load shift vector for this list */
 173         i_shift_offset   = DIM*shiftidx[iidx];
 174
 175         /* Load limits for loop over neighbors */
 176         j_index_start    = jindex[iidx];
 177         j_index_end      = jindex[iidx+1];
 178
 179         /* Get outer coordinate index */
 180         inr              = iinr[iidx];
 181         i_coord_offset   = DIM*inr;
 182
 183         /* Load i particle coords and add shift vector */
 184         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 185                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 186
 187         fix0             = _fjsp_setzero_v2r8();
 188         fiy0             = _fjsp_setzero_v2r8();
 189         fiz0             = _fjsp_setzero_v2r8();
 190         fix1             = _fjsp_setzero_v2r8();
 191         fiy1             = _fjsp_setzero_v2r8();
 192         fiz1             = _fjsp_setzero_v2r8();
 193         fix2             = _fjsp_setzero_v2r8();
 194         fiy2             = _fjsp_setzero_v2r8();
 195         fiz2             = _fjsp_setzero_v2r8();
 196
 197         /* Reset potential sums */
 198         velecsum         = _fjsp_setzero_v2r8();
 199         vvdwsum          = _fjsp_setzero_v2r8();
 200
 201         /* Start inner kernel loop */
 202         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 203         {
 204
 205             /* Get j neighbor index, and coordinate index */
 206             jnrA             = jjnr[jidx];
 207             jnrB             = jjnr[jidx+1];
 208             j_coord_offsetA  = DIM*jnrA;
 209             j_coord_offsetB  = DIM*jnrB;
 210
 211             /* load j atom coordinates */
 212             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 213                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 214
 215             /* Calculate displacement vector */
 216             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 217             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 218             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 219             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 220             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 221             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 222             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 223             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 224             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 225             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 226             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 227             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 228             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 229             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 230             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 231             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 232             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 233             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 234             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 235             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 236             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 237             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 238             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 239             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 240             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 241             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 242             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 243
 244             /* Calculate squared distance and things based on it */
 245             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 246             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 247             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 248             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 249             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 250             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 251             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 252             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 253             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 254
 255             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 256             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 257             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 258             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 259             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 260             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 261             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 262             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 263             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 264
 265             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 266             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 267             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 268             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 269             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 270             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 271             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 272             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 273             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 274
 275             fjx0             = _fjsp_setzero_v2r8();
 276             fjy0             = _fjsp_setzero_v2r8();
 277             fjz0             = _fjsp_setzero_v2r8();
 278             fjx1             = _fjsp_setzero_v2r8();
 279             fjy1             = _fjsp_setzero_v2r8();
 280             fjz1             = _fjsp_setzero_v2r8();
 281             fjx2             = _fjsp_setzero_v2r8();
 282             fjy2             = _fjsp_setzero_v2r8();
 283             fjz2             = _fjsp_setzero_v2r8();
 284
 285             /**************************
 286              * CALCULATE INTERACTIONS *
 287              **************************/
 288
 289             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 290
 291             /* EWALD ELECTROSTATICS */
 292
 293             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 294             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 295             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 296             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 297             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 298
 299             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 300             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 301             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 302             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 303             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 304             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 305             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 306             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 307             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 308             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 309
 310             /* LENNARD-JONES DISPERSION/REPULSION */
 311
 312             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 313             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 314             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 315             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 316             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 320             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 321
 322             fscal            = _fjsp_add_v2r8(felec,fvdw);
 323
 324             /* Update vectorial force */
 325             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 326             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 327             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 328
 329             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 330             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 331             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 338
 339             /* EWALD ELECTROSTATICS */
 340
 341             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 342             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 343             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 344             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 345             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 346
 347             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 348             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 349             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 350             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 351             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 352             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 353             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 354             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 355             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
 356             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 360
 361             fscal            = felec;
 362
 363             /* Update vectorial force */
 364             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 365             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 366             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 367
 368             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 369             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 370             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
 377
 378             /* EWALD ELECTROSTATICS */
 379
 380             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 381             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
 382             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 383             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 384             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 385
 386             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 387             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 388             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 389             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 390             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 391             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 392             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 393             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 394             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
 395             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402             /* Update vectorial force */
 403             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 404             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 405             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 406
 407             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 408             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 409             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 416
 417             /* EWALD ELECTROSTATICS */
 418
 419             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 420             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 421             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 422             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 423             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 424
 425             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 426             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 427             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 428             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 429             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 430             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 431             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 432             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 433             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 434             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 435
 436             /* Update potential sum for this i atom from the interaction with this j atom. */
 437             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 438
 439             fscal            = felec;
 440
 441             /* Update vectorial force */
 442             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 443             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 444             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 445
 446             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 447             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 448             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 449
 450             /**************************
 451              * CALCULATE INTERACTIONS *
 452              **************************/
 453
 454             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 455
 456             /* EWALD ELECTROSTATICS */
 457
 458             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 459             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 460             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 461             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 462             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 463
 464             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 465             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 466             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 467             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 468             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 469             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 470             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 471             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 472             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
 473             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480             /* Update vectorial force */
 481             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 482             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 483             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 484
 485             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 486             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 487             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 488
 489             /**************************
 490              * CALCULATE INTERACTIONS *
 491              **************************/
 492
 493             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 494
 495             /* EWALD ELECTROSTATICS */
 496
 497             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 498             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 499             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 500             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 501             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 502
 503             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 504             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 505             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 506             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 507             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 508             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 509             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 510             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 511             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
 512             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 513
 514             /* Update potential sum for this i atom from the interaction with this j atom. */
 515             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 516
 517             fscal            = felec;
 518
 519             /* Update vectorial force */
 520             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 521             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 522             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 523
 524             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 525             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 526             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 533
 534             /* EWALD ELECTROSTATICS */
 535
 536             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 537             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 538             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 539             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 540             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 541
 542             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 543             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 544             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 545             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 546             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 547             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 548             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 549             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 550             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 551             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 552
 553             /* Update potential sum for this i atom from the interaction with this j atom. */
 554             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 555
 556             fscal            = felec;
 557
 558             /* Update vectorial force */
 559             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 560             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 561             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 562
 563             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 564             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 565             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 566
 567             /**************************
 568              * CALCULATE INTERACTIONS *
 569              **************************/
 570
 571             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 572
 573             /* EWALD ELECTROSTATICS */
 574
 575             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 576             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 577             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 578             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 579             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 580
 581             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 582             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 583             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 584             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 585             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 586             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 587             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 588             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 589             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
 590             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 591
 592             /* Update potential sum for this i atom from the interaction with this j atom. */
 593             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 594
 595             fscal            = felec;
 596
 597             /* Update vectorial force */
 598             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 599             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 600             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 601
 602             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 603             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 604             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 605
 606             /**************************
 607              * CALCULATE INTERACTIONS *
 608              **************************/
 609
 610             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 611
 612             /* EWALD ELECTROSTATICS */
 613
 614             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 615             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
 616             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 617             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 618             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 619
 620             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 621             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 622             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 623             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 624             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 625             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 626             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 627             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 628             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
 629             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
 630
 631             /* Update potential sum for this i atom from the interaction with this j atom. */
 632             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 633
 634             fscal            = felec;
 635
 636             /* Update vectorial force */
 637             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 638             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 639             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 640
 641             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 642             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 643             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 644
 645             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 646
 647             /* Inner loop uses 408 flops */
 648         }
 649
 650         if(jidx<j_index_end)
 651         {
 652
 653             jnrA             = jjnr[jidx];
 654             j_coord_offsetA  = DIM*jnrA;
 655
 656             /* load j atom coordinates */
 657             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 658                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 659
 660             /* Calculate displacement vector */
 661             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 662             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 663             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 664             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 665             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 666             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 667             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 668             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 669             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 670             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 671             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 672             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 673             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 674             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 675             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 676             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 677             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 678             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 679             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 680             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 681             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 682             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 683             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 684             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 685             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 686             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 687             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 688
 689             /* Calculate squared distance and things based on it */
 690             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 691             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 692             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 693             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 694             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 695             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 696             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 697             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 698             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 699
 700             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 701             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 702             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 703             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 704             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 705             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 706             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 707             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 708             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 709
 710             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 711             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 712             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 713             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 714             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 715             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 716             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 717             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 718             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 719
 720             fjx0             = _fjsp_setzero_v2r8();
 721             fjy0             = _fjsp_setzero_v2r8();
 722             fjz0             = _fjsp_setzero_v2r8();
 723             fjx1             = _fjsp_setzero_v2r8();
 724             fjy1             = _fjsp_setzero_v2r8();
 725             fjz1             = _fjsp_setzero_v2r8();
 726             fjx2             = _fjsp_setzero_v2r8();
 727             fjy2             = _fjsp_setzero_v2r8();
 728             fjz2             = _fjsp_setzero_v2r8();
 729
 730             /**************************
 731              * CALCULATE INTERACTIONS *
 732              **************************/
 733
 734             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 735
 736             /* EWALD ELECTROSTATICS */
 737
 738             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 739             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 740             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 741             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 742             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 743
 744             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 745             ewtabD           = _fjsp_setzero_v2r8();
 746             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 747             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 748             ewtabFn          = _fjsp_setzero_v2r8();
 749             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 750             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 751             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 752             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 753             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 754
 755             /* LENNARD-JONES DISPERSION/REPULSION */
 756
 757             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 758             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 759             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 760             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 761             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 765             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 766             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 767             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 768
 769             fscal            = _fjsp_add_v2r8(felec,fvdw);
 770
 771             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 772
 773             /* Update vectorial force */
 774             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 775             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 776             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 777
 778             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 779             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 780             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 781
 782             /**************************
 783              * CALCULATE INTERACTIONS *
 784              **************************/
 785
 786             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 787
 788             /* EWALD ELECTROSTATICS */
 789
 790             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 791             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 792             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 793             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 794             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 795
 796             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 797             ewtabD           = _fjsp_setzero_v2r8();
 798             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 799             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 800             ewtabFn          = _fjsp_setzero_v2r8();
 801             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 802             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 803             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 804             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
 805             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
 806
 807             /* Update potential sum for this i atom from the interaction with this j atom. */
 808             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 809             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 810
 811             fscal            = felec;
 812
 813             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 814
 815             /* Update vectorial force */
 816             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 817             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 818             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 819
 820             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 821             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 822             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
 829
 830             /* EWALD ELECTROSTATICS */
 831
 832             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 833             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
 834             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 835             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 836             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 837
 838             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 839             ewtabD           = _fjsp_setzero_v2r8();
 840             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 841             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 842             ewtabFn          = _fjsp_setzero_v2r8();
 843             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 844             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 845             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 846             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
 847             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
 848
 849             /* Update potential sum for this i atom from the interaction with this j atom. */
 850             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 851             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 856
 857             /* Update vectorial force */
 858             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 859             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 860             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 861
 862             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 863             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 864             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 865
 866             /**************************
 867              * CALCULATE INTERACTIONS *
 868              **************************/
 869
 870             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 871
 872             /* EWALD ELECTROSTATICS */
 873
 874             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 875             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 876             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 877             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 878             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 879
 880             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 881             ewtabD           = _fjsp_setzero_v2r8();
 882             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 883             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 884             ewtabFn          = _fjsp_setzero_v2r8();
 885             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 886             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 887             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 888             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 889             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 890
 891             /* Update potential sum for this i atom from the interaction with this j atom. */
 892             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 893             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 894
 895             fscal            = felec;
 896
 897             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 898
 899             /* Update vectorial force */
 900             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 901             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 902             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 903
 904             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 905             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 906             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 907
 908             /**************************
 909              * CALCULATE INTERACTIONS *
 910              **************************/
 911
 912             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 913
 914             /* EWALD ELECTROSTATICS */
 915
 916             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 917             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 918             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 919             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 920             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 921
 922             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 923             ewtabD           = _fjsp_setzero_v2r8();
 924             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 925             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 926             ewtabFn          = _fjsp_setzero_v2r8();
 927             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 928             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 929             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 930             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
 931             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 932
 933             /* Update potential sum for this i atom from the interaction with this j atom. */
 934             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 935             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 936
 937             fscal            = felec;
 938
 939             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 940
 941             /* Update vectorial force */
 942             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 943             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 944             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 945
 946             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 947             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 948             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 949
 950             /**************************
 951              * CALCULATE INTERACTIONS *
 952              **************************/
 953
 954             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 955
 956             /* EWALD ELECTROSTATICS */
 957
 958             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 959             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 960             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 961             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 962             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 963
 964             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 965             ewtabD           = _fjsp_setzero_v2r8();
 966             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 967             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 968             ewtabFn          = _fjsp_setzero_v2r8();
 969             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 970             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 971             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 972             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
 973             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 974
 975             /* Update potential sum for this i atom from the interaction with this j atom. */
 976             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 977             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 978
 979             fscal            = felec;
 980
 981             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 982
 983             /* Update vectorial force */
 984             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 985             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 986             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 987
 988             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 989             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 990             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 991
 992             /**************************
 993              * CALCULATE INTERACTIONS *
 994              **************************/
 995
 996             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 997
 998             /* EWALD ELECTROSTATICS */
 999
1000             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1001             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1002             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1003             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1004             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1005
1006             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1007             ewtabD           = _fjsp_setzero_v2r8();
1008             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1009             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1010             ewtabFn          = _fjsp_setzero_v2r8();
1011             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1012             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1013             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1014             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
1015             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1016
1017             /* Update potential sum for this i atom from the interaction with this j atom. */
1018             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1019             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1020
1021             fscal            = felec;
1022
1023             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1024
1025             /* Update vectorial force */
1026             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1027             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1028             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1029
1030             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1031             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1032             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1039
1040             /* EWALD ELECTROSTATICS */
1041
1042             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1043             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1044             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1045             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1046             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1047
1048             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1049             ewtabD           = _fjsp_setzero_v2r8();
1050             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1051             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1052             ewtabFn          = _fjsp_setzero_v2r8();
1053             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1054             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1055             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1056             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
1057             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1058
1059             /* Update potential sum for this i atom from the interaction with this j atom. */
1060             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1061             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1062
1063             fscal            = felec;
1064
1065             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1066
1067             /* Update vectorial force */
1068             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1069             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1070             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1071
1072             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1073             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1074             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1075
1076             /**************************
1077              * CALCULATE INTERACTIONS *
1078              **************************/
1079
1080             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1081
1082             /* EWALD ELECTROSTATICS */
1083
1084             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1085             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1086             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1087             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1088             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1089
1090             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1091             ewtabD           = _fjsp_setzero_v2r8();
1092             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1093             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1094             ewtabFn          = _fjsp_setzero_v2r8();
1095             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1096             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1097             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1098             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
1099             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1100
1101             /* Update potential sum for this i atom from the interaction with this j atom. */
1102             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1103             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1104
1105             fscal            = felec;
1106
1107             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1108
1109             /* Update vectorial force */
1110             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1111             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1112             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1113
1114             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1115             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1116             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1117
1118             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1119
1120             /* Inner loop uses 408 flops */
1121         }
1122
1123         /* End of innermost loop */
1124
1125         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1126                                               f+i_coord_offset,fshift+i_shift_offset);
1127
1128         ggid                        = gid[iidx];
1129         /* Update potential energies */
1130         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1131         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
1132
1133         /* Increment number of inner iterations */
1134         inneriter                  += j_index_end - j_index_start;
1135
1136         /* Outer loop uses 20 flops */
1137     }
1138
1139     /* Increment number of outer iterations */
1140     outeriter        += nri;
1141
1142     /* Update outer/inner flops */
1143
1144     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*408);
1145 }
1146 /*
1147  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
1148  * Electrostatics interaction: Ewald
1149  * VdW interaction:            LennardJones
1150  * Geometry:                   Water3-Water3
1151  * Calculate force/pot:        Force
1152  */
1153 void
1154 nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
1155                     (t_nblist                    * gmx_restrict       nlist,
1156                      rvec                        * gmx_restrict          xx,
1157                      rvec                        * gmx_restrict          ff,
1158                      t_forcerec                  * gmx_restrict          fr,
1159                      t_mdatoms                   * gmx_restrict     mdatoms,
1160                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1161                      t_nrnb                      * gmx_restrict        nrnb)
1162 {
1163     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1164      * just 0 for non-waters.
1165      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1166      * jnr indices corresponding to data put in the four positions in the SIMD register.
1167      */
1168     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1169     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1170     int              jnrA,jnrB;
1171     int              j_coord_offsetA,j_coord_offsetB;
1172     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1173     real             rcutoff_scalar;
1174     real             *shiftvec,*fshift,*x,*f;
1175     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1176     int              vdwioffset0;
1177     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1178     int              vdwioffset1;
1179     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1180     int              vdwioffset2;
1181     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1182     int              vdwjidx0A,vdwjidx0B;
1183     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1184     int              vdwjidx1A,vdwjidx1B;
1185     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1186     int              vdwjidx2A,vdwjidx2B;
1187     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1188     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1189     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1190     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1191     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1192     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1193     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1194     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1195     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1196     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1197     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1198     real             *charge;
1199     int              nvdwtype;
1200     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1201     int              *vdwtype;
1202     real             *vdwparam;
1203     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
1204     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
1205     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1206     real             *ewtab;
1207     _fjsp_v2r8       itab_tmp;
1208     _fjsp_v2r8       dummy_mask,cutoff_mask;
1209     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1210     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1211     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1212
1213     x                = xx[0];
1214     f                = ff[0];
1215
1216     nri              = nlist->nri;
1217     iinr             = nlist->iinr;
1218     jindex           = nlist->jindex;
1219     jjnr             = nlist->jjnr;
1220     shiftidx         = nlist->shift;
1221     gid              = nlist->gid;
1222     shiftvec         = fr->shift_vec[0];
1223     fshift           = fr->fshift[0];
1224     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1225     charge           = mdatoms->chargeA;
1226     nvdwtype         = fr->ntype;
1227     vdwparam         = fr->nbfp;
1228     vdwtype          = mdatoms->typeA;
1229
1230     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
1231     ewtab            = fr->ic->tabq_coul_F;
1232     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
1233     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
1234
1235     /* Setup water-specific parameters */
1236     inr              = nlist->iinr[0];
1237     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1238     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1239     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1240     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1241
1242     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1243     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1244     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1245     vdwjidx0A        = 2*vdwtype[inr+0];
1246     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1247     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
1248     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
1249     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1250     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1251     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1252     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1253     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1254     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1255     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1256     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1257
1258     /* Avoid stupid compiler warnings */
1259     jnrA = jnrB = 0;
1260     j_coord_offsetA = 0;
1261     j_coord_offsetB = 0;
1262
1263     outeriter        = 0;
1264     inneriter        = 0;
1265
1266     /* Start outer loop over neighborlists */
1267     for(iidx=0; iidx<nri; iidx++)
1268     {
1269         /* Load shift vector for this list */
1270         i_shift_offset   = DIM*shiftidx[iidx];
1271
1272         /* Load limits for loop over neighbors */
1273         j_index_start    = jindex[iidx];
1274         j_index_end      = jindex[iidx+1];
1275
1276         /* Get outer coordinate index */
1277         inr              = iinr[iidx];
1278         i_coord_offset   = DIM*inr;
1279
1280         /* Load i particle coords and add shift vector */
1281         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1282                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1283
1284         fix0             = _fjsp_setzero_v2r8();
1285         fiy0             = _fjsp_setzero_v2r8();
1286         fiz0             = _fjsp_setzero_v2r8();
1287         fix1             = _fjsp_setzero_v2r8();
1288         fiy1             = _fjsp_setzero_v2r8();
1289         fiz1             = _fjsp_setzero_v2r8();
1290         fix2             = _fjsp_setzero_v2r8();
1291         fiy2             = _fjsp_setzero_v2r8();
1292         fiz2             = _fjsp_setzero_v2r8();
1293
1294         /* Start inner kernel loop */
1295         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1296         {
1297
1298             /* Get j neighbor index, and coordinate index */
1299             jnrA             = jjnr[jidx];
1300             jnrB             = jjnr[jidx+1];
1301             j_coord_offsetA  = DIM*jnrA;
1302             j_coord_offsetB  = DIM*jnrB;
1303
1304             /* load j atom coordinates */
1305             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1306                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1307
1308             /* Calculate displacement vector */
1309             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1310             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1311             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1312             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1313             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1314             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1315             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1316             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1317             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1318             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1319             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1320             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1321             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1322             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1323             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1324             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1325             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1326             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1327             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1328             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1329             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1330             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1331             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1332             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1333             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1334             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1335             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1336
1337             /* Calculate squared distance and things based on it */
1338             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1339             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1340             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1341             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1342             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1343             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1344             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1345             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1346             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1347
1348             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1349             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1350             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1351             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1352             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1353             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1354             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1355             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1356             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1357
1358             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1359             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1360             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1361             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1362             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1363             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1364             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1365             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1366             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1367
1368             fjx0             = _fjsp_setzero_v2r8();
1369             fjy0             = _fjsp_setzero_v2r8();
1370             fjz0             = _fjsp_setzero_v2r8();
1371             fjx1             = _fjsp_setzero_v2r8();
1372             fjy1             = _fjsp_setzero_v2r8();
1373             fjz1             = _fjsp_setzero_v2r8();
1374             fjx2             = _fjsp_setzero_v2r8();
1375             fjy2             = _fjsp_setzero_v2r8();
1376             fjz2             = _fjsp_setzero_v2r8();
1377
1378             /**************************
1379              * CALCULATE INTERACTIONS *
1380              **************************/
1381
1382             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1383
1384             /* EWALD ELECTROSTATICS */
1385
1386             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1387             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
1388             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1389             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1390             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1391
1392             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1393                                          &ewtabF,&ewtabFn);
1394             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1395             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
1396
1397             /* LENNARD-JONES DISPERSION/REPULSION */
1398
1399             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1400             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1401
1402             fscal            = _fjsp_add_v2r8(felec,fvdw);
1403
1404             /* Update vectorial force */
1405             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1406             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1407             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1408
1409             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1410             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1411             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1412
1413             /**************************
1414              * CALCULATE INTERACTIONS *
1415              **************************/
1416
1417             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1418
1419             /* EWALD ELECTROSTATICS */
1420
1421             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1422             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1423             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1424             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1425             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1426
1427             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1428                                          &ewtabF,&ewtabFn);
1429             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1430             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1431
1432             fscal            = felec;
1433
1434             /* Update vectorial force */
1435             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1436             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1437             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1438
1439             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1440             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1441             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1442
1443             /**************************
1444              * CALCULATE INTERACTIONS *
1445              **************************/
1446
1447             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1448
1449             /* EWALD ELECTROSTATICS */
1450
1451             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1452             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1453             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1454             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1455             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1456
1457             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1458                                          &ewtabF,&ewtabFn);
1459             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1460             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1461
1462             fscal            = felec;
1463
1464             /* Update vectorial force */
1465             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1466             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1467             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1468
1469             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1470             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1471             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1472
1473             /**************************
1474              * CALCULATE INTERACTIONS *
1475              **************************/
1476
1477             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1478
1479             /* EWALD ELECTROSTATICS */
1480
1481             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1482             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1483             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1484             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1485             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1486
1487             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1488                                          &ewtabF,&ewtabFn);
1489             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1490             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1491
1492             fscal            = felec;
1493
1494             /* Update vectorial force */
1495             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1496             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1497             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1498
1499             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1500             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1501             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1502
1503             /**************************
1504              * CALCULATE INTERACTIONS *
1505              **************************/
1506
1507             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1508
1509             /* EWALD ELECTROSTATICS */
1510
1511             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1512             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1513             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1514             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1515             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1516
1517             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1518                                          &ewtabF,&ewtabFn);
1519             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1520             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1521
1522             fscal            = felec;
1523
1524             /* Update vectorial force */
1525             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1526             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1527             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1528
1529             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1530             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1531             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1532
1533             /**************************
1534              * CALCULATE INTERACTIONS *
1535              **************************/
1536
1537             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1538
1539             /* EWALD ELECTROSTATICS */
1540
1541             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1542             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1543             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1544             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1545             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1546
1547             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1548                                          &ewtabF,&ewtabFn);
1549             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1550             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1551
1552             fscal            = felec;
1553
1554             /* Update vectorial force */
1555             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1556             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1557             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1558
1559             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1560             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1561             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1562
1563             /**************************
1564              * CALCULATE INTERACTIONS *
1565              **************************/
1566
1567             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1568
1569             /* EWALD ELECTROSTATICS */
1570
1571             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1572             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1573             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1574             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1575             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1576
1577             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1578                                          &ewtabF,&ewtabFn);
1579             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1580             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1581
1582             fscal            = felec;
1583
1584             /* Update vectorial force */
1585             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1586             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1587             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1588
1589             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1590             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1591             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1592
1593             /**************************
1594              * CALCULATE INTERACTIONS *
1595              **************************/
1596
1597             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1598
1599             /* EWALD ELECTROSTATICS */
1600
1601             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1602             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1603             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1604             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1605             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1606
1607             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1608                                          &ewtabF,&ewtabFn);
1609             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1610             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1611
1612             fscal            = felec;
1613
1614             /* Update vectorial force */
1615             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1616             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1617             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1618
1619             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1620             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1621             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1622
1623             /**************************
1624              * CALCULATE INTERACTIONS *
1625              **************************/
1626
1627             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1628
1629             /* EWALD ELECTROSTATICS */
1630
1631             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1632             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1633             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1634             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1635             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1636
1637             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1638                                          &ewtabF,&ewtabFn);
1639             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1640             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1641
1642             fscal            = felec;
1643
1644             /* Update vectorial force */
1645             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1646             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1647             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1648
1649             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1650             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1651             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1652
1653             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1654
1655             /* Inner loop uses 358 flops */
1656         }
1657
1658         if(jidx<j_index_end)
1659         {
1660
1661             jnrA             = jjnr[jidx];
1662             j_coord_offsetA  = DIM*jnrA;
1663
1664             /* load j atom coordinates */
1665             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1666                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1667
1668             /* Calculate displacement vector */
1669             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1670             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1671             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1672             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1673             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1674             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1675             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1676             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1677             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1678             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1679             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1680             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1681             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1682             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1683             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1684             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1685             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1686             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1687             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1688             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1689             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1690             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1691             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1692             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1693             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1694             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1695             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1696
1697             /* Calculate squared distance and things based on it */
1698             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1699             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1700             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1701             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1702             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1703             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1704             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1705             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1706             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1707
1708             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1709             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1710             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1711             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1712             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1713             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1714             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1715             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1716             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1717
1718             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1719             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1720             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1721             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1722             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1723             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1724             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1725             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1726             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1727
1728             fjx0             = _fjsp_setzero_v2r8();
1729             fjy0             = _fjsp_setzero_v2r8();
1730             fjz0             = _fjsp_setzero_v2r8();
1731             fjx1             = _fjsp_setzero_v2r8();
1732             fjy1             = _fjsp_setzero_v2r8();
1733             fjz1             = _fjsp_setzero_v2r8();
1734             fjx2             = _fjsp_setzero_v2r8();
1735             fjy2             = _fjsp_setzero_v2r8();
1736             fjz2             = _fjsp_setzero_v2r8();
1737
1738             /**************************
1739              * CALCULATE INTERACTIONS *
1740              **************************/
1741
1742             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1743
1744             /* EWALD ELECTROSTATICS */
1745
1746             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1747             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
1748             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1749             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1750             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1751
1752             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1753             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1754             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
1755
1756             /* LENNARD-JONES DISPERSION/REPULSION */
1757
1758             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1759             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1760
1761             fscal            = _fjsp_add_v2r8(felec,fvdw);
1762
1763             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1764
1765             /* Update vectorial force */
1766             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1767             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1768             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1769
1770             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1771             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1772             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1773
1774             /**************************
1775              * CALCULATE INTERACTIONS *
1776              **************************/
1777
1778             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1779
1780             /* EWALD ELECTROSTATICS */
1781
1782             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1783             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1784             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1785             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1786             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1787
1788             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1789             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1790             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1791
1792             fscal            = felec;
1793
1794             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1795
1796             /* Update vectorial force */
1797             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1798             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1799             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1800
1801             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1802             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1803             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1804
1805             /**************************
1806              * CALCULATE INTERACTIONS *
1807              **************************/
1808
1809             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1810
1811             /* EWALD ELECTROSTATICS */
1812
1813             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1814             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1815             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1816             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1817             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1818
1819             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1820             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1821             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1822
1823             fscal            = felec;
1824
1825             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1826
1827             /* Update vectorial force */
1828             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1829             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1830             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1831
1832             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1833             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1834             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1835
1836             /**************************
1837              * CALCULATE INTERACTIONS *
1838              **************************/
1839
1840             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1841
1842             /* EWALD ELECTROSTATICS */
1843
1844             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1845             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1846             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1847             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1848             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1849
1850             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1851             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1852             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1853
1854             fscal            = felec;
1855
1856             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1857
1858             /* Update vectorial force */
1859             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1860             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1861             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1862
1863             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1864             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1865             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1866
1867             /**************************
1868              * CALCULATE INTERACTIONS *
1869              **************************/
1870
1871             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1872
1873             /* EWALD ELECTROSTATICS */
1874
1875             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1876             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1877             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1878             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1879             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1880
1881             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1882             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1883             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1884
1885             fscal            = felec;
1886
1887             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1888
1889             /* Update vectorial force */
1890             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1891             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1892             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1893
1894             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1895             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1896             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1897
1898             /**************************
1899              * CALCULATE INTERACTIONS *
1900              **************************/
1901
1902             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1903
1904             /* EWALD ELECTROSTATICS */
1905
1906             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1907             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1908             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1909             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1910             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1911
1912             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1913             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1914             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1915
1916             fscal            = felec;
1917
1918             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1919
1920             /* Update vectorial force */
1921             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1922             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1923             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1924
1925             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1926             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1927             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1928
1929             /**************************
1930              * CALCULATE INTERACTIONS *
1931              **************************/
1932
1933             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1934
1935             /* EWALD ELECTROSTATICS */
1936
1937             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1938             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1939             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1940             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1941             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1942
1943             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1944             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1945             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1946
1947             fscal            = felec;
1948
1949             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1950
1951             /* Update vectorial force */
1952             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1953             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1954             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1955
1956             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1957             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1958             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1959
1960             /**************************
1961              * CALCULATE INTERACTIONS *
1962              **************************/
1963
1964             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1965
1966             /* EWALD ELECTROSTATICS */
1967
1968             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1969             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1970             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1971             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1972             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1973
1974             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1975             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1976             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1977
1978             fscal            = felec;
1979
1980             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1981
1982             /* Update vectorial force */
1983             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1984             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1985             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1986
1987             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1988             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1989             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1990
1991             /**************************
1992              * CALCULATE INTERACTIONS *
1993              **************************/
1994
1995             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1996
1997             /* EWALD ELECTROSTATICS */
1998
1999             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2000             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2001             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2002             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2003             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2004
2005             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2006             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2007             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2008
2009             fscal            = felec;
2010
2011             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2012
2013             /* Update vectorial force */
2014             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2015             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2016             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2017
2018             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2019             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2020             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2021
2022             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2023
2024             /* Inner loop uses 358 flops */
2025         }
2026
2027         /* End of innermost loop */
2028
2029         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2030                                               f+i_coord_offset,fshift+i_shift_offset);
2031
2032         /* Increment number of inner iterations */
2033         inneriter                  += j_index_end - j_index_start;
2034
2035         /* Outer loop uses 18 flops */
2036     }
2037
2038     /* Increment number of outer iterations */
2039     outeriter        += nri;
2040
2041     /* Update outer/inner flops */
2042
2043     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*358);
2044 }