src/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012, by the GROMACS development team, led by
   5  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   6  * others, as listed in the AUTHORS file in the top-level source
   7  * 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: Ewald
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water3-Water3
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_VF_sparc64_hpc_ace_double
  60                     (t_nblist * gmx_restrict                nlist,
  61                      rvec * gmx_restrict                    xx,
  62                      rvec * gmx_restrict                    ff,
  63                      t_forcerec * gmx_restrict              fr,
  64                      t_mdatoms * gmx_restrict               mdatoms,
  65                      nb_kernel_data_t * gmx_restrict        kernel_data,
  66                      t_nrnb * gmx_restrict                  nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  94     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  95     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  96     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  97     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 100     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 103     real             *charge;
 104     int              nvdwtype;
 105     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 106     int              *vdwtype;
 107     real             *vdwparam;
 108     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 109     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 110     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 111     real             *ewtab;
 112     _fjsp_v2r8       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 113     real             rswitch_scalar,d_scalar;
 114     _fjsp_v2r8       itab_tmp;
 115     _fjsp_v2r8       dummy_mask,cutoff_mask;
 116     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 117     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 118     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 119
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 132     charge           = mdatoms->chargeA;
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 138     ewtab            = fr->ic->tabq_coul_FDV0;
 139     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 140     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 141
 142     /* Setup water-specific parameters */
 143     inr              = nlist->iinr[0];
 144     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 145     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 146     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 147     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 148
 149     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 150     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 151     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 152     vdwjidx0A        = 2*vdwtype[inr+0];
 153     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 154     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 155     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 156     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 157     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 158     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 159     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 160     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 161     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 162     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 163     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 164
 165     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 166     rcutoff_scalar   = fr->rcoulomb;
 167     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 168     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 169
 170     rswitch_scalar   = fr->rcoulomb_switch;
 171     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
 172     /* Setup switch parameters */
 173     d_scalar         = rcutoff_scalar-rswitch_scalar;
 174     d                = gmx_fjsp_set1_v2r8(d_scalar);
 175     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
 176     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 177     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 178     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
 179     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 180     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 181
 182     /* Avoid stupid compiler warnings */
 183     jnrA = jnrB = 0;
 184     j_coord_offsetA = 0;
 185     j_coord_offsetB = 0;
 186
 187     outeriter        = 0;
 188     inneriter        = 0;
 189
 190     /* Start outer loop over neighborlists */
 191     for(iidx=0; iidx<nri; iidx++)
 192     {
 193         /* Load shift vector for this list */
 194         i_shift_offset   = DIM*shiftidx[iidx];
 195
 196         /* Load limits for loop over neighbors */
 197         j_index_start    = jindex[iidx];
 198         j_index_end      = jindex[iidx+1];
 199
 200         /* Get outer coordinate index */
 201         inr              = iinr[iidx];
 202         i_coord_offset   = DIM*inr;
 203
 204         /* Load i particle coords and add shift vector */
 205         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 206                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 207
 208         fix0             = _fjsp_setzero_v2r8();
 209         fiy0             = _fjsp_setzero_v2r8();
 210         fiz0             = _fjsp_setzero_v2r8();
 211         fix1             = _fjsp_setzero_v2r8();
 212         fiy1             = _fjsp_setzero_v2r8();
 213         fiz1             = _fjsp_setzero_v2r8();
 214         fix2             = _fjsp_setzero_v2r8();
 215         fiy2             = _fjsp_setzero_v2r8();
 216         fiz2             = _fjsp_setzero_v2r8();
 217
 218         /* Reset potential sums */
 219         velecsum         = _fjsp_setzero_v2r8();
 220         vvdwsum          = _fjsp_setzero_v2r8();
 221
 222         /* Start inner kernel loop */
 223         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 224         {
 225
 226             /* Get j neighbor index, and coordinate index */
 227             jnrA             = jjnr[jidx];
 228             jnrB             = jjnr[jidx+1];
 229             j_coord_offsetA  = DIM*jnrA;
 230             j_coord_offsetB  = DIM*jnrB;
 231
 232             /* load j atom coordinates */
 233             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 234                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 235
 236             /* Calculate displacement vector */
 237             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 238             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 239             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 240             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 241             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 242             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 243             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 244             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 245             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 246             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 247             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 248             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 249             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 250             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 251             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 252             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 253             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 254             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 255             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 256             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 257             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 258             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 259             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 260             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 261             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 262             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 263             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 264
 265             /* Calculate squared distance and things based on it */
 266             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 267             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 268             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 269             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 270             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 271             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 272             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 273             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 274             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 275
 276             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 277             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 278             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 279             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 280             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 281             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 282             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 283             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 284             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 285
 286             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 287             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 288             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 289             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 290             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 291             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 292             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 293             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 294             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 295
 296             fjx0             = _fjsp_setzero_v2r8();
 297             fjy0             = _fjsp_setzero_v2r8();
 298             fjz0             = _fjsp_setzero_v2r8();
 299             fjx1             = _fjsp_setzero_v2r8();
 300             fjy1             = _fjsp_setzero_v2r8();
 301             fjz1             = _fjsp_setzero_v2r8();
 302             fjx2             = _fjsp_setzero_v2r8();
 303             fjy2             = _fjsp_setzero_v2r8();
 304             fjz2             = _fjsp_setzero_v2r8();
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 311             {
 312
 313             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 314
 315             /* EWALD ELECTROSTATICS */
 316
 317             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 318             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 319             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 320             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 321             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 322
 323             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 324             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 325             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 326             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 327             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 328             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 329             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 330             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 331             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 332             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 333
 334             /* LENNARD-JONES DISPERSION/REPULSION */
 335
 336             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 337             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 338             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 339             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 340             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 341
 342             d                = _fjsp_sub_v2r8(r00,rswitch);
 343             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 344             d2               = _fjsp_mul_v2r8(d,d);
 345             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 346
 347             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 348
 349             /* Evaluate switch function */
 350             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 351             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
 352             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
 353             velec            = _fjsp_mul_v2r8(velec,sw);
 354             vvdw             = _fjsp_mul_v2r8(vvdw,sw);
 355             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 356
 357             /* Update potential sum for this i atom from the interaction with this j atom. */
 358             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 359             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 360             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 361             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 362
 363             fscal            = _fjsp_add_v2r8(felec,fvdw);
 364
 365             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 366
 367             /* Update vectorial force */
 368             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 369             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 370             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 371
 372             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 373             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 374             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 375
 376             }
 377
 378             /**************************
 379              * CALCULATE INTERACTIONS *
 380              **************************/
 381
 382             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 383             {
 384
 385             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 386
 387             /* EWALD ELECTROSTATICS */
 388
 389             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 390             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 391             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 392             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 393             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 394
 395             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 396             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 397             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 398             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 399             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 400             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 401             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 402             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 403             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
 404             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
 405
 406             d                = _fjsp_sub_v2r8(r01,rswitch);
 407             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 408             d2               = _fjsp_mul_v2r8(d,d);
 409             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 410
 411             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 412
 413             /* Evaluate switch function */
 414             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 415             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
 416             velec            = _fjsp_mul_v2r8(velec,sw);
 417             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 421             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 422
 423             fscal            = felec;
 424
 425             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 426
 427             /* Update vectorial force */
 428             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 429             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 430             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 431
 432             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 433             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 434             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 435
 436             }
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 443             {
 444
 445             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
 446
 447             /* EWALD ELECTROSTATICS */
 448
 449             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 450             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
 451             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 452             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 453             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 454
 455             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 456             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 457             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 458             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 459             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 460             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 461             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 462             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 463             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
 464             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
 465
 466             d                = _fjsp_sub_v2r8(r02,rswitch);
 467             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 468             d2               = _fjsp_mul_v2r8(d,d);
 469             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 470
 471             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 472
 473             /* Evaluate switch function */
 474             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 475             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
 476             velec            = _fjsp_mul_v2r8(velec,sw);
 477             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 478
 479             /* Update potential sum for this i atom from the interaction with this j atom. */
 480             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 481             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 482
 483             fscal            = felec;
 484
 485             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 486
 487             /* Update vectorial force */
 488             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 489             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 490             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 491
 492             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 493             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 494             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 495
 496             }
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 503             {
 504
 505             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 506
 507             /* EWALD ELECTROSTATICS */
 508
 509             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 510             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 511             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 512             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 513             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 514
 515             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 516             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 517             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 518             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 519             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 520             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 521             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 522             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 523             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 524             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 525
 526             d                = _fjsp_sub_v2r8(r10,rswitch);
 527             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 528             d2               = _fjsp_mul_v2r8(d,d);
 529             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 530
 531             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 532
 533             /* Evaluate switch function */
 534             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 535             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
 536             velec            = _fjsp_mul_v2r8(velec,sw);
 537             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 538
 539             /* Update potential sum for this i atom from the interaction with this j atom. */
 540             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 541             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 542
 543             fscal            = felec;
 544
 545             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 546
 547             /* Update vectorial force */
 548             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 549             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 550             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 551
 552             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 553             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 554             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 555
 556             }
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 563             {
 564
 565             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 566
 567             /* EWALD ELECTROSTATICS */
 568
 569             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 570             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 571             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 572             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 573             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 574
 575             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 576             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 577             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 578             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 579             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 580             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 581             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 582             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 583             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
 584             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 585
 586             d                = _fjsp_sub_v2r8(r11,rswitch);
 587             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 588             d2               = _fjsp_mul_v2r8(d,d);
 589             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 590
 591             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 592
 593             /* Evaluate switch function */
 594             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 595             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
 596             velec            = _fjsp_mul_v2r8(velec,sw);
 597             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 598
 599             /* Update potential sum for this i atom from the interaction with this j atom. */
 600             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 601             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 602
 603             fscal            = felec;
 604
 605             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 606
 607             /* Update vectorial force */
 608             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 609             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 610             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 611
 612             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 613             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 614             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 615
 616             }
 617
 618             /**************************
 619              * CALCULATE INTERACTIONS *
 620              **************************/
 621
 622             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 623             {
 624
 625             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 626
 627             /* EWALD ELECTROSTATICS */
 628
 629             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 630             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 631             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 632             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 633             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 634
 635             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 636             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 637             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 638             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 639             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 640             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 641             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 642             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 643             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
 644             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 645
 646             d                = _fjsp_sub_v2r8(r12,rswitch);
 647             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 648             d2               = _fjsp_mul_v2r8(d,d);
 649             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 650
 651             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 652
 653             /* Evaluate switch function */
 654             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 655             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
 656             velec            = _fjsp_mul_v2r8(velec,sw);
 657             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 658
 659             /* Update potential sum for this i atom from the interaction with this j atom. */
 660             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 661             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 662
 663             fscal            = felec;
 664
 665             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 666
 667             /* Update vectorial force */
 668             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 669             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 670             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 671
 672             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 673             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 674             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 675
 676             }
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 683             {
 684
 685             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 686
 687             /* EWALD ELECTROSTATICS */
 688
 689             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 690             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 691             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 692             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 693             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 694
 695             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 696             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 697             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 698             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 699             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 700             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 701             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 702             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 703             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 704             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 705
 706             d                = _fjsp_sub_v2r8(r20,rswitch);
 707             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 708             d2               = _fjsp_mul_v2r8(d,d);
 709             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 710
 711             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 712
 713             /* Evaluate switch function */
 714             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 715             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
 716             velec            = _fjsp_mul_v2r8(velec,sw);
 717             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 718
 719             /* Update potential sum for this i atom from the interaction with this j atom. */
 720             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 721             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 722
 723             fscal            = felec;
 724
 725             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 726
 727             /* Update vectorial force */
 728             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 729             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 730             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 731
 732             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 733             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 734             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 735
 736             }
 737
 738             /**************************
 739              * CALCULATE INTERACTIONS *
 740              **************************/
 741
 742             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 743             {
 744
 745             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 746
 747             /* EWALD ELECTROSTATICS */
 748
 749             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 750             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 751             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 752             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 753             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 754
 755             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 756             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 757             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 758             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 759             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 760             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 761             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 762             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 763             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
 764             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 765
 766             d                = _fjsp_sub_v2r8(r21,rswitch);
 767             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 768             d2               = _fjsp_mul_v2r8(d,d);
 769             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 770
 771             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 772
 773             /* Evaluate switch function */
 774             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 775             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
 776             velec            = _fjsp_mul_v2r8(velec,sw);
 777             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 781             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 782
 783             fscal            = felec;
 784
 785             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 786
 787             /* Update vectorial force */
 788             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 789             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 790             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 791
 792             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 793             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 794             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 795
 796             }
 797
 798             /**************************
 799              * CALCULATE INTERACTIONS *
 800              **************************/
 801
 802             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 803             {
 804
 805             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 806
 807             /* EWALD ELECTROSTATICS */
 808
 809             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 810             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
 811             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 812             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 813             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 814
 815             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 816             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 817             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 818             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 819             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 820             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 821             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 822             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 823             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
 824             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
 825
 826             d                = _fjsp_sub_v2r8(r22,rswitch);
 827             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 828             d2               = _fjsp_mul_v2r8(d,d);
 829             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 830
 831             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 832
 833             /* Evaluate switch function */
 834             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 835             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
 836             velec            = _fjsp_mul_v2r8(velec,sw);
 837             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 838
 839             /* Update potential sum for this i atom from the interaction with this j atom. */
 840             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 841             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 842
 843             fscal            = felec;
 844
 845             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 846
 847             /* Update vectorial force */
 848             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 849             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 850             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 851
 852             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 853             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 854             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 855
 856             }
 857
 858             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 859
 860             /* Inner loop uses 630 flops */
 861         }
 862
 863         if(jidx<j_index_end)
 864         {
 865
 866             jnrA             = jjnr[jidx];
 867             j_coord_offsetA  = DIM*jnrA;
 868
 869             /* load j atom coordinates */
 870             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 871                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 872
 873             /* Calculate displacement vector */
 874             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 875             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 876             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 877             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 878             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 879             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 880             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 881             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 882             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 883             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 884             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 885             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 886             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 887             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 888             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 889             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 890             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 891             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 892             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 893             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 894             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 895             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 896             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 897             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 898             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 899             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 900             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 901
 902             /* Calculate squared distance and things based on it */
 903             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 904             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 905             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 906             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 907             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 908             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 909             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 910             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 911             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 912
 913             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 914             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 915             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 916             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 917             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 918             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 919             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 920             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 921             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 922
 923             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 924             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 925             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 926             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 927             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 928             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 929             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 930             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 931             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 932
 933             fjx0             = _fjsp_setzero_v2r8();
 934             fjy0             = _fjsp_setzero_v2r8();
 935             fjz0             = _fjsp_setzero_v2r8();
 936             fjx1             = _fjsp_setzero_v2r8();
 937             fjy1             = _fjsp_setzero_v2r8();
 938             fjz1             = _fjsp_setzero_v2r8();
 939             fjx2             = _fjsp_setzero_v2r8();
 940             fjy2             = _fjsp_setzero_v2r8();
 941             fjz2             = _fjsp_setzero_v2r8();
 942
 943             /**************************
 944              * CALCULATE INTERACTIONS *
 945              **************************/
 946
 947             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 948             {
 949
 950             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 951
 952             /* EWALD ELECTROSTATICS */
 953
 954             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 955             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 956             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 957             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 958             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 959
 960             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 961             ewtabD           = _fjsp_setzero_v2r8();
 962             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 963             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 964             ewtabFn          = _fjsp_setzero_v2r8();
 965             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 966             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 967             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 968             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 969             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 970
 971             /* LENNARD-JONES DISPERSION/REPULSION */
 972
 973             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 974             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 975             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 976             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 977             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 978
 979             d                = _fjsp_sub_v2r8(r00,rswitch);
 980             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 981             d2               = _fjsp_mul_v2r8(d,d);
 982             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 983
 984             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 985
 986             /* Evaluate switch function */
 987             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 988             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
 989             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
 990             velec            = _fjsp_mul_v2r8(velec,sw);
 991             vvdw             = _fjsp_mul_v2r8(vvdw,sw);
 992             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 993
 994             /* Update potential sum for this i atom from the interaction with this j atom. */
 995             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 996             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 997             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 998             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 999             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
1000             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
1001
1002             fscal            = _fjsp_add_v2r8(felec,fvdw);
1003
1004             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1005
1006             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1007
1008             /* Update vectorial force */
1009             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1010             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1011             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1012
1013             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1014             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1015             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1016
1017             }
1018
1019             /**************************
1020              * CALCULATE INTERACTIONS *
1021              **************************/
1022
1023             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1024             {
1025
1026             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1027
1028             /* EWALD ELECTROSTATICS */
1029
1030             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1031             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1032             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1033             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1034             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1035
1036             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1037             ewtabD           = _fjsp_setzero_v2r8();
1038             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1039             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1040             ewtabFn          = _fjsp_setzero_v2r8();
1041             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1042             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1043             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1044             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
1045             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1046
1047             d                = _fjsp_sub_v2r8(r01,rswitch);
1048             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1049             d2               = _fjsp_mul_v2r8(d,d);
1050             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1051
1052             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1053
1054             /* Evaluate switch function */
1055             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1056             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
1057             velec            = _fjsp_mul_v2r8(velec,sw);
1058             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1059
1060             /* Update potential sum for this i atom from the interaction with this j atom. */
1061             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1062             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1063             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1064
1065             fscal            = felec;
1066
1067             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1068
1069             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1070
1071             /* Update vectorial force */
1072             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1073             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1074             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1075
1076             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1077             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1078             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1079
1080             }
1081
1082             /**************************
1083              * CALCULATE INTERACTIONS *
1084              **************************/
1085
1086             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1087             {
1088
1089             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1090
1091             /* EWALD ELECTROSTATICS */
1092
1093             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1094             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1095             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1096             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1097             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1098
1099             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1100             ewtabD           = _fjsp_setzero_v2r8();
1101             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1102             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1103             ewtabFn          = _fjsp_setzero_v2r8();
1104             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1105             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1106             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1107             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
1108             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1109
1110             d                = _fjsp_sub_v2r8(r02,rswitch);
1111             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1112             d2               = _fjsp_mul_v2r8(d,d);
1113             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1114
1115             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1116
1117             /* Evaluate switch function */
1118             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1119             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
1120             velec            = _fjsp_mul_v2r8(velec,sw);
1121             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1122
1123             /* Update potential sum for this i atom from the interaction with this j atom. */
1124             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1125             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1126             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1127
1128             fscal            = felec;
1129
1130             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1131
1132             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1133
1134             /* Update vectorial force */
1135             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1136             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1137             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1138
1139             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1140             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1141             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1142
1143             }
1144
1145             /**************************
1146              * CALCULATE INTERACTIONS *
1147              **************************/
1148
1149             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1150             {
1151
1152             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1153
1154             /* EWALD ELECTROSTATICS */
1155
1156             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1157             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1158             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1159             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1160             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1161
1162             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1163             ewtabD           = _fjsp_setzero_v2r8();
1164             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1165             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1166             ewtabFn          = _fjsp_setzero_v2r8();
1167             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1168             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1169             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1170             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
1171             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1172
1173             d                = _fjsp_sub_v2r8(r10,rswitch);
1174             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1175             d2               = _fjsp_mul_v2r8(d,d);
1176             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1177
1178             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1179
1180             /* Evaluate switch function */
1181             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1182             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
1183             velec            = _fjsp_mul_v2r8(velec,sw);
1184             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1185
1186             /* Update potential sum for this i atom from the interaction with this j atom. */
1187             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1188             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1189             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1190
1191             fscal            = felec;
1192
1193             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1194
1195             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1196
1197             /* Update vectorial force */
1198             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1199             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1200             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1201
1202             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1203             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1204             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1205
1206             }
1207
1208             /**************************
1209              * CALCULATE INTERACTIONS *
1210              **************************/
1211
1212             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1213             {
1214
1215             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1216
1217             /* EWALD ELECTROSTATICS */
1218
1219             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1220             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1221             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1222             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1223             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1224
1225             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1226             ewtabD           = _fjsp_setzero_v2r8();
1227             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1228             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1229             ewtabFn          = _fjsp_setzero_v2r8();
1230             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1231             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1232             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1233             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
1234             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1235
1236             d                = _fjsp_sub_v2r8(r11,rswitch);
1237             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1238             d2               = _fjsp_mul_v2r8(d,d);
1239             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1240
1241             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1242
1243             /* Evaluate switch function */
1244             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1245             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
1246             velec            = _fjsp_mul_v2r8(velec,sw);
1247             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1248
1249             /* Update potential sum for this i atom from the interaction with this j atom. */
1250             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1251             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1252             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1253
1254             fscal            = felec;
1255
1256             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1257
1258             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1259
1260             /* Update vectorial force */
1261             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1262             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1263             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1264
1265             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1266             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1267             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1268
1269             }
1270
1271             /**************************
1272              * CALCULATE INTERACTIONS *
1273              **************************/
1274
1275             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1276             {
1277
1278             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1279
1280             /* EWALD ELECTROSTATICS */
1281
1282             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1283             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1284             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1285             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1286             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1287
1288             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1289             ewtabD           = _fjsp_setzero_v2r8();
1290             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1291             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1292             ewtabFn          = _fjsp_setzero_v2r8();
1293             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1294             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1295             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1296             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
1297             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1298
1299             d                = _fjsp_sub_v2r8(r12,rswitch);
1300             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1301             d2               = _fjsp_mul_v2r8(d,d);
1302             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1303
1304             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1305
1306             /* Evaluate switch function */
1307             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1308             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
1309             velec            = _fjsp_mul_v2r8(velec,sw);
1310             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1311
1312             /* Update potential sum for this i atom from the interaction with this j atom. */
1313             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1314             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1315             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1316
1317             fscal            = felec;
1318
1319             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1320
1321             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1322
1323             /* Update vectorial force */
1324             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1325             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1326             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1327
1328             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1329             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1330             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1331
1332             }
1333
1334             /**************************
1335              * CALCULATE INTERACTIONS *
1336              **************************/
1337
1338             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1339             {
1340
1341             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1342
1343             /* EWALD ELECTROSTATICS */
1344
1345             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1346             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1347             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1348             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1349             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1350
1351             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1352             ewtabD           = _fjsp_setzero_v2r8();
1353             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1354             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1355             ewtabFn          = _fjsp_setzero_v2r8();
1356             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1357             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1358             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1359             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
1360             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1361
1362             d                = _fjsp_sub_v2r8(r20,rswitch);
1363             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1364             d2               = _fjsp_mul_v2r8(d,d);
1365             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1366
1367             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1368
1369             /* Evaluate switch function */
1370             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1371             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
1372             velec            = _fjsp_mul_v2r8(velec,sw);
1373             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1374
1375             /* Update potential sum for this i atom from the interaction with this j atom. */
1376             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1377             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1378             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1379
1380             fscal            = felec;
1381
1382             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1383
1384             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1385
1386             /* Update vectorial force */
1387             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1388             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1389             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1390
1391             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1392             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1393             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1394
1395             }
1396
1397             /**************************
1398              * CALCULATE INTERACTIONS *
1399              **************************/
1400
1401             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1402             {
1403
1404             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1405
1406             /* EWALD ELECTROSTATICS */
1407
1408             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1409             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1410             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1411             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1412             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1413
1414             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1415             ewtabD           = _fjsp_setzero_v2r8();
1416             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1417             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1418             ewtabFn          = _fjsp_setzero_v2r8();
1419             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1420             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1421             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1422             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
1423             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1424
1425             d                = _fjsp_sub_v2r8(r21,rswitch);
1426             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1427             d2               = _fjsp_mul_v2r8(d,d);
1428             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1429
1430             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1431
1432             /* Evaluate switch function */
1433             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1434             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
1435             velec            = _fjsp_mul_v2r8(velec,sw);
1436             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1437
1438             /* Update potential sum for this i atom from the interaction with this j atom. */
1439             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1440             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1441             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1442
1443             fscal            = felec;
1444
1445             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1446
1447             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1448
1449             /* Update vectorial force */
1450             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1451             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1452             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1453
1454             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1455             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1456             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1457
1458             }
1459
1460             /**************************
1461              * CALCULATE INTERACTIONS *
1462              **************************/
1463
1464             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1465             {
1466
1467             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1468
1469             /* EWALD ELECTROSTATICS */
1470
1471             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1472             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1473             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1474             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1475             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1476
1477             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1478             ewtabD           = _fjsp_setzero_v2r8();
1479             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1480             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1481             ewtabFn          = _fjsp_setzero_v2r8();
1482             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1483             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1484             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1485             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
1486             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1487
1488             d                = _fjsp_sub_v2r8(r22,rswitch);
1489             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1490             d2               = _fjsp_mul_v2r8(d,d);
1491             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1492
1493             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1494
1495             /* Evaluate switch function */
1496             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1497             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
1498             velec            = _fjsp_mul_v2r8(velec,sw);
1499             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1500
1501             /* Update potential sum for this i atom from the interaction with this j atom. */
1502             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1503             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1504             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1505
1506             fscal            = felec;
1507
1508             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1509
1510             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1511
1512             /* Update vectorial force */
1513             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1514             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1515             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1516
1517             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1518             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1519             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1520
1521             }
1522
1523             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1524
1525             /* Inner loop uses 630 flops */
1526         }
1527
1528         /* End of innermost loop */
1529
1530         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1531                                               f+i_coord_offset,fshift+i_shift_offset);
1532
1533         ggid                        = gid[iidx];
1534         /* Update potential energies */
1535         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1536         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
1537
1538         /* Increment number of inner iterations */
1539         inneriter                  += j_index_end - j_index_start;
1540
1541         /* Outer loop uses 20 flops */
1542     }
1543
1544     /* Increment number of outer iterations */
1545     outeriter        += nri;
1546
1547     /* Update outer/inner flops */
1548
1549     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*630);
1550 }
1551 /*
1552  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_F_sparc64_hpc_ace_double
1553  * Electrostatics interaction: Ewald
1554  * VdW interaction:            LennardJones
1555  * Geometry:                   Water3-Water3
1556  * Calculate force/pot:        Force
1557  */
1558 void
1559 nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_F_sparc64_hpc_ace_double
1560                     (t_nblist * gmx_restrict                nlist,
1561                      rvec * gmx_restrict                    xx,
1562                      rvec * gmx_restrict                    ff,
1563                      t_forcerec * gmx_restrict              fr,
1564                      t_mdatoms * gmx_restrict               mdatoms,
1565                      nb_kernel_data_t * gmx_restrict        kernel_data,
1566                      t_nrnb * gmx_restrict                  nrnb)
1567 {
1568     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1569      * just 0 for non-waters.
1570      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1571      * jnr indices corresponding to data put in the four positions in the SIMD register.
1572      */
1573     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1574     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1575     int              jnrA,jnrB;
1576     int              j_coord_offsetA,j_coord_offsetB;
1577     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1578     real             rcutoff_scalar;
1579     real             *shiftvec,*fshift,*x,*f;
1580     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1581     int              vdwioffset0;
1582     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1583     int              vdwioffset1;
1584     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1585     int              vdwioffset2;
1586     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1587     int              vdwjidx0A,vdwjidx0B;
1588     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1589     int              vdwjidx1A,vdwjidx1B;
1590     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1591     int              vdwjidx2A,vdwjidx2B;
1592     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1593     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1594     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1595     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1596     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1597     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1598     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1599     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1600     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1601     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1602     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1603     real             *charge;
1604     int              nvdwtype;
1605     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1606     int              *vdwtype;
1607     real             *vdwparam;
1608     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
1609     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
1610     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1611     real             *ewtab;
1612     _fjsp_v2r8       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1613     real             rswitch_scalar,d_scalar;
1614     _fjsp_v2r8       itab_tmp;
1615     _fjsp_v2r8       dummy_mask,cutoff_mask;
1616     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1617     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1618     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1619
1620     x                = xx[0];
1621     f                = ff[0];
1622
1623     nri              = nlist->nri;
1624     iinr             = nlist->iinr;
1625     jindex           = nlist->jindex;
1626     jjnr             = nlist->jjnr;
1627     shiftidx         = nlist->shift;
1628     gid              = nlist->gid;
1629     shiftvec         = fr->shift_vec[0];
1630     fshift           = fr->fshift[0];
1631     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1632     charge           = mdatoms->chargeA;
1633     nvdwtype         = fr->ntype;
1634     vdwparam         = fr->nbfp;
1635     vdwtype          = mdatoms->typeA;
1636
1637     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
1638     ewtab            = fr->ic->tabq_coul_FDV0;
1639     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
1640     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
1641
1642     /* Setup water-specific parameters */
1643     inr              = nlist->iinr[0];
1644     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1645     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1646     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1647     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1648
1649     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1650     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1651     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1652     vdwjidx0A        = 2*vdwtype[inr+0];
1653     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1654     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
1655     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
1656     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1657     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1658     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1659     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1660     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1661     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1662     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1663     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1664
1665     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1666     rcutoff_scalar   = fr->rcoulomb;
1667     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1668     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1669
1670     rswitch_scalar   = fr->rcoulomb_switch;
1671     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
1672     /* Setup switch parameters */
1673     d_scalar         = rcutoff_scalar-rswitch_scalar;
1674     d                = gmx_fjsp_set1_v2r8(d_scalar);
1675     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
1676     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1677     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1678     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
1679     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1680     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1681
1682     /* Avoid stupid compiler warnings */
1683     jnrA = jnrB = 0;
1684     j_coord_offsetA = 0;
1685     j_coord_offsetB = 0;
1686
1687     outeriter        = 0;
1688     inneriter        = 0;
1689
1690     /* Start outer loop over neighborlists */
1691     for(iidx=0; iidx<nri; iidx++)
1692     {
1693         /* Load shift vector for this list */
1694         i_shift_offset   = DIM*shiftidx[iidx];
1695
1696         /* Load limits for loop over neighbors */
1697         j_index_start    = jindex[iidx];
1698         j_index_end      = jindex[iidx+1];
1699
1700         /* Get outer coordinate index */
1701         inr              = iinr[iidx];
1702         i_coord_offset   = DIM*inr;
1703
1704         /* Load i particle coords and add shift vector */
1705         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1706                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1707
1708         fix0             = _fjsp_setzero_v2r8();
1709         fiy0             = _fjsp_setzero_v2r8();
1710         fiz0             = _fjsp_setzero_v2r8();
1711         fix1             = _fjsp_setzero_v2r8();
1712         fiy1             = _fjsp_setzero_v2r8();
1713         fiz1             = _fjsp_setzero_v2r8();
1714         fix2             = _fjsp_setzero_v2r8();
1715         fiy2             = _fjsp_setzero_v2r8();
1716         fiz2             = _fjsp_setzero_v2r8();
1717
1718         /* Start inner kernel loop */
1719         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1720         {
1721
1722             /* Get j neighbor index, and coordinate index */
1723             jnrA             = jjnr[jidx];
1724             jnrB             = jjnr[jidx+1];
1725             j_coord_offsetA  = DIM*jnrA;
1726             j_coord_offsetB  = DIM*jnrB;
1727
1728             /* load j atom coordinates */
1729             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1730                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1731
1732             /* Calculate displacement vector */
1733             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1734             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1735             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1736             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1737             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1738             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1739             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1740             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1741             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1742             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1743             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1744             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1745             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1746             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1747             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1748             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1749             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1750             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1751             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1752             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1753             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1754             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1755             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1756             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1757             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1758             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1759             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1760
1761             /* Calculate squared distance and things based on it */
1762             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1763             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1764             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1765             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1766             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1767             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1768             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1769             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1770             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1771
1772             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1773             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1774             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1775             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1776             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1777             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1778             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1779             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1780             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1781
1782             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1783             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1784             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1785             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1786             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1787             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1788             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1789             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1790             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1791
1792             fjx0             = _fjsp_setzero_v2r8();
1793             fjy0             = _fjsp_setzero_v2r8();
1794             fjz0             = _fjsp_setzero_v2r8();
1795             fjx1             = _fjsp_setzero_v2r8();
1796             fjy1             = _fjsp_setzero_v2r8();
1797             fjz1             = _fjsp_setzero_v2r8();
1798             fjx2             = _fjsp_setzero_v2r8();
1799             fjy2             = _fjsp_setzero_v2r8();
1800             fjz2             = _fjsp_setzero_v2r8();
1801
1802             /**************************
1803              * CALCULATE INTERACTIONS *
1804              **************************/
1805
1806             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1807             {
1808
1809             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1810
1811             /* EWALD ELECTROSTATICS */
1812
1813             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1814             ewrt             = _fjsp_mul_v2r8(r00,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             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1820             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1821             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1822             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1823             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1824             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1825             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1826             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1827             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
1828             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
1829
1830             /* LENNARD-JONES DISPERSION/REPULSION */
1831
1832             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1833             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
1834             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
1835             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
1836             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
1837
1838             d                = _fjsp_sub_v2r8(r00,rswitch);
1839             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1840             d2               = _fjsp_mul_v2r8(d,d);
1841             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1842
1843             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1844
1845             /* Evaluate switch function */
1846             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1847             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
1848             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
1849             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1850
1851             fscal            = _fjsp_add_v2r8(felec,fvdw);
1852
1853             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1854
1855             /* Update vectorial force */
1856             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1857             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1858             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1859
1860             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1861             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1862             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1863
1864             }
1865
1866             /**************************
1867              * CALCULATE INTERACTIONS *
1868              **************************/
1869
1870             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1871             {
1872
1873             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1874
1875             /* EWALD ELECTROSTATICS */
1876
1877             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1878             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1879             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1880             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1881             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1882
1883             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1884             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1885             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1886             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1887             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1888             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1889             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1890             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1891             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
1892             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1893
1894             d                = _fjsp_sub_v2r8(r01,rswitch);
1895             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1896             d2               = _fjsp_mul_v2r8(d,d);
1897             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1898
1899             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1900
1901             /* Evaluate switch function */
1902             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1903             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
1904             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1905
1906             fscal            = felec;
1907
1908             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1909
1910             /* Update vectorial force */
1911             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1912             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1913             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1914
1915             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1916             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1917             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1918
1919             }
1920
1921             /**************************
1922              * CALCULATE INTERACTIONS *
1923              **************************/
1924
1925             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1926             {
1927
1928             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1929
1930             /* EWALD ELECTROSTATICS */
1931
1932             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1933             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1934             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1935             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1936             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1937
1938             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1939             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1940             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1941             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1942             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1943             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1944             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1945             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1946             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
1947             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1948
1949             d                = _fjsp_sub_v2r8(r02,rswitch);
1950             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1951             d2               = _fjsp_mul_v2r8(d,d);
1952             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1953
1954             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1955
1956             /* Evaluate switch function */
1957             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1958             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
1959             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1960
1961             fscal            = felec;
1962
1963             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1964
1965             /* Update vectorial force */
1966             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1967             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1968             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1969
1970             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1971             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1972             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1973
1974             }
1975
1976             /**************************
1977              * CALCULATE INTERACTIONS *
1978              **************************/
1979
1980             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1981             {
1982
1983             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1984
1985             /* EWALD ELECTROSTATICS */
1986
1987             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1988             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1989             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1990             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1991             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1992
1993             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1994             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1995             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1996             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1997             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1998             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1999             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2000             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2001             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
2002             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
2003
2004             d                = _fjsp_sub_v2r8(r10,rswitch);
2005             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2006             d2               = _fjsp_mul_v2r8(d,d);
2007             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2008
2009             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2010
2011             /* Evaluate switch function */
2012             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2013             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
2014             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
2015
2016             fscal            = felec;
2017
2018             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2019
2020             /* Update vectorial force */
2021             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
2022             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
2023             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
2024
2025             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
2026             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
2027             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
2028
2029             }
2030
2031             /**************************
2032              * CALCULATE INTERACTIONS *
2033              **************************/
2034
2035             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
2036             {
2037
2038             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
2039
2040             /* EWALD ELECTROSTATICS */
2041
2042             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2043             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
2044             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2045             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2046             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2047
2048             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2049             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2050             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2051             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2052             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2053             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2054             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2055             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2056             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
2057             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
2058
2059             d                = _fjsp_sub_v2r8(r11,rswitch);
2060             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2061             d2               = _fjsp_mul_v2r8(d,d);
2062             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2063
2064             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2065
2066             /* Evaluate switch function */
2067             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2068             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
2069             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
2070
2071             fscal            = felec;
2072
2073             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2074
2075             /* Update vectorial force */
2076             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
2077             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
2078             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
2079
2080             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
2081             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
2082             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
2083
2084             }
2085
2086             /**************************
2087              * CALCULATE INTERACTIONS *
2088              **************************/
2089
2090             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2091             {
2092
2093             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2094
2095             /* EWALD ELECTROSTATICS */
2096
2097             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2098             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2099             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2100             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2101             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2102
2103             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2104             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2105             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2106             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2107             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2108             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2109             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2110             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2111             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
2112             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2113
2114             d                = _fjsp_sub_v2r8(r12,rswitch);
2115             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2116             d2               = _fjsp_mul_v2r8(d,d);
2117             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2118
2119             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2120
2121             /* Evaluate switch function */
2122             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2123             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
2124             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2125
2126             fscal            = felec;
2127
2128             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2129
2130             /* Update vectorial force */
2131             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2132             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2133             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2134
2135             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2136             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2137             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2138
2139             }
2140
2141             /**************************
2142              * CALCULATE INTERACTIONS *
2143              **************************/
2144
2145             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
2146             {
2147
2148             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
2149
2150             /* EWALD ELECTROSTATICS */
2151
2152             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2153             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
2154             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2155             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2156             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2157
2158             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2159             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2160             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2161             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2162             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2163             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2164             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2165             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2166             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
2167             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
2168
2169             d                = _fjsp_sub_v2r8(r20,rswitch);
2170             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2171             d2               = _fjsp_mul_v2r8(d,d);
2172             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2173
2174             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2175
2176             /* Evaluate switch function */
2177             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2178             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
2179             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
2180
2181             fscal            = felec;
2182
2183             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2184
2185             /* Update vectorial force */
2186             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
2187             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
2188             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
2189
2190             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
2191             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
2192             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
2193
2194             }
2195
2196             /**************************
2197              * CALCULATE INTERACTIONS *
2198              **************************/
2199
2200             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2201             {
2202
2203             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2204
2205             /* EWALD ELECTROSTATICS */
2206
2207             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2208             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2209             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2210             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2211             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2212
2213             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2214             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2215             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2216             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2217             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2218             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2219             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2220             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2221             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
2222             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2223
2224             d                = _fjsp_sub_v2r8(r21,rswitch);
2225             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2226             d2               = _fjsp_mul_v2r8(d,d);
2227             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2228
2229             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2230
2231             /* Evaluate switch function */
2232             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2233             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
2234             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2235
2236             fscal            = felec;
2237
2238             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2239
2240             /* Update vectorial force */
2241             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2242             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2243             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2244
2245             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2246             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2247             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2248
2249             }
2250
2251             /**************************
2252              * CALCULATE INTERACTIONS *
2253              **************************/
2254
2255             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2256             {
2257
2258             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2259
2260             /* EWALD ELECTROSTATICS */
2261
2262             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2263             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2264             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2265             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2266             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2267
2268             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2269             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2270             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2271             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2272             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2273             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2274             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2275             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2276             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
2277             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2278
2279             d                = _fjsp_sub_v2r8(r22,rswitch);
2280             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2281             d2               = _fjsp_mul_v2r8(d,d);
2282             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2283
2284             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2285
2286             /* Evaluate switch function */
2287             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2288             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
2289             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2290
2291             fscal            = felec;
2292
2293             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2294
2295             /* Update vectorial force */
2296             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2297             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2298             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2299
2300             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2301             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2302             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2303
2304             }
2305
2306             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2307
2308             /* Inner loop uses 600 flops */
2309         }
2310
2311         if(jidx<j_index_end)
2312         {
2313
2314             jnrA             = jjnr[jidx];
2315             j_coord_offsetA  = DIM*jnrA;
2316
2317             /* load j atom coordinates */
2318             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
2319                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
2320
2321             /* Calculate displacement vector */
2322             dx00             = _fjsp_sub_v2r8(ix0,jx0);
2323             dy00             = _fjsp_sub_v2r8(iy0,jy0);
2324             dz00             = _fjsp_sub_v2r8(iz0,jz0);
2325             dx01             = _fjsp_sub_v2r8(ix0,jx1);
2326             dy01             = _fjsp_sub_v2r8(iy0,jy1);
2327             dz01             = _fjsp_sub_v2r8(iz0,jz1);
2328             dx02             = _fjsp_sub_v2r8(ix0,jx2);
2329             dy02             = _fjsp_sub_v2r8(iy0,jy2);
2330             dz02             = _fjsp_sub_v2r8(iz0,jz2);
2331             dx10             = _fjsp_sub_v2r8(ix1,jx0);
2332             dy10             = _fjsp_sub_v2r8(iy1,jy0);
2333             dz10             = _fjsp_sub_v2r8(iz1,jz0);
2334             dx11             = _fjsp_sub_v2r8(ix1,jx1);
2335             dy11             = _fjsp_sub_v2r8(iy1,jy1);
2336             dz11             = _fjsp_sub_v2r8(iz1,jz1);
2337             dx12             = _fjsp_sub_v2r8(ix1,jx2);
2338             dy12             = _fjsp_sub_v2r8(iy1,jy2);
2339             dz12             = _fjsp_sub_v2r8(iz1,jz2);
2340             dx20             = _fjsp_sub_v2r8(ix2,jx0);
2341             dy20             = _fjsp_sub_v2r8(iy2,jy0);
2342             dz20             = _fjsp_sub_v2r8(iz2,jz0);
2343             dx21             = _fjsp_sub_v2r8(ix2,jx1);
2344             dy21             = _fjsp_sub_v2r8(iy2,jy1);
2345             dz21             = _fjsp_sub_v2r8(iz2,jz1);
2346             dx22             = _fjsp_sub_v2r8(ix2,jx2);
2347             dy22             = _fjsp_sub_v2r8(iy2,jy2);
2348             dz22             = _fjsp_sub_v2r8(iz2,jz2);
2349
2350             /* Calculate squared distance and things based on it */
2351             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
2352             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
2353             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
2354             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
2355             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
2356             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
2357             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
2358             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
2359             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
2360
2361             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
2362             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
2363             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
2364             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
2365             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
2366             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
2367             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
2368             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
2369             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
2370
2371             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
2372             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
2373             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
2374             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
2375             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
2376             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
2377             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
2378             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
2379             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
2380
2381             fjx0             = _fjsp_setzero_v2r8();
2382             fjy0             = _fjsp_setzero_v2r8();
2383             fjz0             = _fjsp_setzero_v2r8();
2384             fjx1             = _fjsp_setzero_v2r8();
2385             fjy1             = _fjsp_setzero_v2r8();
2386             fjz1             = _fjsp_setzero_v2r8();
2387             fjx2             = _fjsp_setzero_v2r8();
2388             fjy2             = _fjsp_setzero_v2r8();
2389             fjz2             = _fjsp_setzero_v2r8();
2390
2391             /**************************
2392              * CALCULATE INTERACTIONS *
2393              **************************/
2394
2395             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
2396             {
2397
2398             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
2399
2400             /* EWALD ELECTROSTATICS */
2401
2402             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2403             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
2404             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2405             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2406             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2407
2408             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2409             ewtabD           = _fjsp_setzero_v2r8();
2410             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2411             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2412             ewtabFn          = _fjsp_setzero_v2r8();
2413             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2414             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2415             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2416             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
2417             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
2418
2419             /* LENNARD-JONES DISPERSION/REPULSION */
2420
2421             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
2422             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
2423             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
2424             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
2425             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
2426
2427             d                = _fjsp_sub_v2r8(r00,rswitch);
2428             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2429             d2               = _fjsp_mul_v2r8(d,d);
2430             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2431
2432             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2433
2434             /* Evaluate switch function */
2435             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2436             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
2437             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
2438             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
2439
2440             fscal            = _fjsp_add_v2r8(felec,fvdw);
2441
2442             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2443
2444             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2445
2446             /* Update vectorial force */
2447             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
2448             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
2449             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
2450
2451             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
2452             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
2453             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
2454
2455             }
2456
2457             /**************************
2458              * CALCULATE INTERACTIONS *
2459              **************************/
2460
2461             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
2462             {
2463
2464             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
2465
2466             /* EWALD ELECTROSTATICS */
2467
2468             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2469             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
2470             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2471             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2472             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2473
2474             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2475             ewtabD           = _fjsp_setzero_v2r8();
2476             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2477             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2478             ewtabFn          = _fjsp_setzero_v2r8();
2479             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2480             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2481             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2482             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
2483             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
2484
2485             d                = _fjsp_sub_v2r8(r01,rswitch);
2486             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2487             d2               = _fjsp_mul_v2r8(d,d);
2488             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2489
2490             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2491
2492             /* Evaluate switch function */
2493             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2494             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
2495             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
2496
2497             fscal            = felec;
2498
2499             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2500
2501             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2502
2503             /* Update vectorial force */
2504             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
2505             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
2506             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
2507
2508             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
2509             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
2510             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
2511
2512             }
2513
2514             /**************************
2515              * CALCULATE INTERACTIONS *
2516              **************************/
2517
2518             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
2519             {
2520
2521             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
2522
2523             /* EWALD ELECTROSTATICS */
2524
2525             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2526             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
2527             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2528             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2529             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2530
2531             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2532             ewtabD           = _fjsp_setzero_v2r8();
2533             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2534             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2535             ewtabFn          = _fjsp_setzero_v2r8();
2536             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2537             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2538             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2539             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
2540             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
2541
2542             d                = _fjsp_sub_v2r8(r02,rswitch);
2543             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2544             d2               = _fjsp_mul_v2r8(d,d);
2545             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2546
2547             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2548
2549             /* Evaluate switch function */
2550             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2551             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
2552             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
2553
2554             fscal            = felec;
2555
2556             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2557
2558             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2559
2560             /* Update vectorial force */
2561             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
2562             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
2563             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
2564
2565             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
2566             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
2567             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
2568
2569             }
2570
2571             /**************************
2572              * CALCULATE INTERACTIONS *
2573              **************************/
2574
2575             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
2576             {
2577
2578             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
2579
2580             /* EWALD ELECTROSTATICS */
2581
2582             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2583             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
2584             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2585             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2586             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2587
2588             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2589             ewtabD           = _fjsp_setzero_v2r8();
2590             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2591             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2592             ewtabFn          = _fjsp_setzero_v2r8();
2593             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2594             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2595             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2596             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
2597             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
2598
2599             d                = _fjsp_sub_v2r8(r10,rswitch);
2600             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2601             d2               = _fjsp_mul_v2r8(d,d);
2602             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2603
2604             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2605
2606             /* Evaluate switch function */
2607             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2608             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
2609             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
2610
2611             fscal            = felec;
2612
2613             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2614
2615             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2616
2617             /* Update vectorial force */
2618             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
2619             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
2620             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
2621
2622             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
2623             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
2624             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
2625
2626             }
2627
2628             /**************************
2629              * CALCULATE INTERACTIONS *
2630              **************************/
2631
2632             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
2633             {
2634
2635             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
2636
2637             /* EWALD ELECTROSTATICS */
2638
2639             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2640             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
2641             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2642             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2643             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2644
2645             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2646             ewtabD           = _fjsp_setzero_v2r8();
2647             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2648             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2649             ewtabFn          = _fjsp_setzero_v2r8();
2650             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2651             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2652             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2653             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
2654             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
2655
2656             d                = _fjsp_sub_v2r8(r11,rswitch);
2657             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2658             d2               = _fjsp_mul_v2r8(d,d);
2659             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2660
2661             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2662
2663             /* Evaluate switch function */
2664             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2665             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
2666             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
2667
2668             fscal            = felec;
2669
2670             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2671
2672             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2673
2674             /* Update vectorial force */
2675             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
2676             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
2677             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
2678
2679             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
2680             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
2681             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
2682
2683             }
2684
2685             /**************************
2686              * CALCULATE INTERACTIONS *
2687              **************************/
2688
2689             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2690             {
2691
2692             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2693
2694             /* EWALD ELECTROSTATICS */
2695
2696             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2697             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2698             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2699             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2700             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2701
2702             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2703             ewtabD           = _fjsp_setzero_v2r8();
2704             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2705             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2706             ewtabFn          = _fjsp_setzero_v2r8();
2707             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2708             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2709             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2710             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
2711             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2712
2713             d                = _fjsp_sub_v2r8(r12,rswitch);
2714             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2715             d2               = _fjsp_mul_v2r8(d,d);
2716             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2717
2718             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2719
2720             /* Evaluate switch function */
2721             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2722             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
2723             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2724
2725             fscal            = felec;
2726
2727             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2728
2729             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2730
2731             /* Update vectorial force */
2732             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2733             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2734             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2735
2736             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2737             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2738             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2739
2740             }
2741
2742             /**************************
2743              * CALCULATE INTERACTIONS *
2744              **************************/
2745
2746             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
2747             {
2748
2749             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
2750
2751             /* EWALD ELECTROSTATICS */
2752
2753             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2754             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
2755             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2756             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2757             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2758
2759             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2760             ewtabD           = _fjsp_setzero_v2r8();
2761             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2762             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2763             ewtabFn          = _fjsp_setzero_v2r8();
2764             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2765             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2766             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2767             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
2768             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
2769
2770             d                = _fjsp_sub_v2r8(r20,rswitch);
2771             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2772             d2               = _fjsp_mul_v2r8(d,d);
2773             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2774
2775             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2776
2777             /* Evaluate switch function */
2778             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2779             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
2780             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
2781
2782             fscal            = felec;
2783
2784             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2785
2786             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2787
2788             /* Update vectorial force */
2789             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
2790             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
2791             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
2792
2793             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
2794             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
2795             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
2796
2797             }
2798
2799             /**************************
2800              * CALCULATE INTERACTIONS *
2801              **************************/
2802
2803             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2804             {
2805
2806             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2807
2808             /* EWALD ELECTROSTATICS */
2809
2810             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2811             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2812             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2813             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2814             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2815
2816             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2817             ewtabD           = _fjsp_setzero_v2r8();
2818             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2819             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2820             ewtabFn          = _fjsp_setzero_v2r8();
2821             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2822             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2823             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2824             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
2825             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2826
2827             d                = _fjsp_sub_v2r8(r21,rswitch);
2828             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2829             d2               = _fjsp_mul_v2r8(d,d);
2830             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2831
2832             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2833
2834             /* Evaluate switch function */
2835             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2836             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
2837             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2838
2839             fscal            = felec;
2840
2841             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2842
2843             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2844
2845             /* Update vectorial force */
2846             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2847             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2848             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2849
2850             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2851             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2852             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2853
2854             }
2855
2856             /**************************
2857              * CALCULATE INTERACTIONS *
2858              **************************/
2859
2860             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2861             {
2862
2863             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2864
2865             /* EWALD ELECTROSTATICS */
2866
2867             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2868             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2869             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2870             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2871             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2872
2873             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2874             ewtabD           = _fjsp_setzero_v2r8();
2875             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2876             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2877             ewtabFn          = _fjsp_setzero_v2r8();
2878             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2879             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2880             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2881             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
2882             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2883
2884             d                = _fjsp_sub_v2r8(r22,rswitch);
2885             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2886             d2               = _fjsp_mul_v2r8(d,d);
2887             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2888
2889             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2890
2891             /* Evaluate switch function */
2892             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2893             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
2894             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2895
2896             fscal            = felec;
2897
2898             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2899
2900             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2901
2902             /* Update vectorial force */
2903             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2904             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2905             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2906
2907             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2908             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2909             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2910
2911             }
2912
2913             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2914
2915             /* Inner loop uses 600 flops */
2916         }
2917
2918         /* End of innermost loop */
2919
2920         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2921                                               f+i_coord_offset,fshift+i_shift_offset);
2922
2923         /* Increment number of inner iterations */
2924         inneriter                  += j_index_end - j_index_start;
2925
2926         /* Outer loop uses 18 flops */
2927     }
2928
2929     /* Increment number of outer iterations */
2930     outeriter        += nri;
2931
2932     /* Update outer/inner flops */
2933
2934     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*600);
2935 }