src/gromacs/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) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sparc64_hpc_ace_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwjidx0A,vdwjidx0B;
  86     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     int              vdwjidx1A,vdwjidx1B;
  88     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     int              nvdwtype;
 103     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 104     int              *vdwtype;
 105     real             *vdwparam;
 106     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 107     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 108     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 109     real             *ewtab;
 110     _fjsp_v2r8       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 111     real             rswitch_scalar,d_scalar;
 112     _fjsp_v2r8       itab_tmp;
 113     _fjsp_v2r8       dummy_mask,cutoff_mask;
 114     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 115     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 116     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 117
 118     x                = xx[0];
 119     f                = ff[0];
 120
 121     nri              = nlist->nri;
 122     iinr             = nlist->iinr;
 123     jindex           = nlist->jindex;
 124     jjnr             = nlist->jjnr;
 125     shiftidx         = nlist->shift;
 126     gid              = nlist->gid;
 127     shiftvec         = fr->shift_vec[0];
 128     fshift           = fr->fshift[0];
 129     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 130     charge           = mdatoms->chargeA;
 131     nvdwtype         = fr->ntype;
 132     vdwparam         = fr->nbfp;
 133     vdwtype          = mdatoms->typeA;
 134
 135     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 136     ewtab            = fr->ic->tabq_coul_FDV0;
 137     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 138     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 139
 140     /* Setup water-specific parameters */
 141     inr              = nlist->iinr[0];
 142     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 143     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 144     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 148     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 149     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 152     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 153     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 154     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 155     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 156     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 157     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 158     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 159     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 160     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 161     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 162
 163     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 164     rcutoff_scalar   = fr->rcoulomb;
 165     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 166     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 167
 168     rswitch_scalar   = fr->rcoulomb_switch;
 169     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
 170     /* Setup switch parameters */
 171     d_scalar         = rcutoff_scalar-rswitch_scalar;
 172     d                = gmx_fjsp_set1_v2r8(d_scalar);
 173     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
 174     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 175     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 176     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
 177     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 178     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 179
 180     /* Avoid stupid compiler warnings */
 181     jnrA = jnrB = 0;
 182     j_coord_offsetA = 0;
 183     j_coord_offsetB = 0;
 184
 185     outeriter        = 0;
 186     inneriter        = 0;
 187
 188     /* Start outer loop over neighborlists */
 189     for(iidx=0; iidx<nri; iidx++)
 190     {
 191         /* Load shift vector for this list */
 192         i_shift_offset   = DIM*shiftidx[iidx];
 193
 194         /* Load limits for loop over neighbors */
 195         j_index_start    = jindex[iidx];
 196         j_index_end      = jindex[iidx+1];
 197
 198         /* Get outer coordinate index */
 199         inr              = iinr[iidx];
 200         i_coord_offset   = DIM*inr;
 201
 202         /* Load i particle coords and add shift vector */
 203         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 204                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 205
 206         fix0             = _fjsp_setzero_v2r8();
 207         fiy0             = _fjsp_setzero_v2r8();
 208         fiz0             = _fjsp_setzero_v2r8();
 209         fix1             = _fjsp_setzero_v2r8();
 210         fiy1             = _fjsp_setzero_v2r8();
 211         fiz1             = _fjsp_setzero_v2r8();
 212         fix2             = _fjsp_setzero_v2r8();
 213         fiy2             = _fjsp_setzero_v2r8();
 214         fiz2             = _fjsp_setzero_v2r8();
 215
 216         /* Reset potential sums */
 217         velecsum         = _fjsp_setzero_v2r8();
 218         vvdwsum          = _fjsp_setzero_v2r8();
 219
 220         /* Start inner kernel loop */
 221         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 222         {
 223
 224             /* Get j neighbor index, and coordinate index */
 225             jnrA             = jjnr[jidx];
 226             jnrB             = jjnr[jidx+1];
 227             j_coord_offsetA  = DIM*jnrA;
 228             j_coord_offsetB  = DIM*jnrB;
 229
 230             /* load j atom coordinates */
 231             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 232                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 233
 234             /* Calculate displacement vector */
 235             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 236             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 237             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 238             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 239             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 240             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 241             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 242             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 243             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 244             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 245             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 246             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 247             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 248             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 249             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 250             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 251             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 252             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 253             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 254             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 255             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 256             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 257             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 258             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 259             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 260             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 261             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 265             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 266             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 267             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 268             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 269             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 270             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 271             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 272             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 273
 274             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 275             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 276             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 277             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 278             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 279             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 280             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 281             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 282             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 283
 284             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 285             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 286             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 287             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 288             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 289             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 290             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 291             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 292             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 293
 294             fjx0             = _fjsp_setzero_v2r8();
 295             fjy0             = _fjsp_setzero_v2r8();
 296             fjz0             = _fjsp_setzero_v2r8();
 297             fjx1             = _fjsp_setzero_v2r8();
 298             fjy1             = _fjsp_setzero_v2r8();
 299             fjz1             = _fjsp_setzero_v2r8();
 300             fjx2             = _fjsp_setzero_v2r8();
 301             fjy2             = _fjsp_setzero_v2r8();
 302             fjz2             = _fjsp_setzero_v2r8();
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 309             {
 310
 311             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 312
 313             /* EWALD ELECTROSTATICS */
 314
 315             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 316             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 317             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 318             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 319             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 320
 321             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 322             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 323             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 324             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 325             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 326             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 327             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 328             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 329             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 330             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 331
 332             /* LENNARD-JONES DISPERSION/REPULSION */
 333
 334             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 335             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 336             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 337             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 338             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 339
 340             d                = _fjsp_sub_v2r8(r00,rswitch);
 341             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 342             d2               = _fjsp_mul_v2r8(d,d);
 343             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 344
 345             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 346
 347             /* Evaluate switch function */
 348             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 349             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
 350             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
 351             velec            = _fjsp_mul_v2r8(velec,sw);
 352             vvdw             = _fjsp_mul_v2r8(vvdw,sw);
 353             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 357             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 358             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 359             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 360
 361             fscal            = _fjsp_add_v2r8(felec,fvdw);
 362
 363             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 364
 365             /* Update vectorial force */
 366             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 367             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 368             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 369
 370             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 371             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 372             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 373
 374             }
 375
 376             /**************************
 377              * CALCULATE INTERACTIONS *
 378              **************************/
 379
 380             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 381             {
 382
 383             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 384
 385             /* EWALD ELECTROSTATICS */
 386
 387             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 388             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 389             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 390             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 391             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 392
 393             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 394             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 395             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 396             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 397             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 398             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 399             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 400             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 401             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
 402             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
 403
 404             d                = _fjsp_sub_v2r8(r01,rswitch);
 405             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 406             d2               = _fjsp_mul_v2r8(d,d);
 407             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 408
 409             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 410
 411             /* Evaluate switch function */
 412             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 413             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
 414             velec            = _fjsp_mul_v2r8(velec,sw);
 415             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 419             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 424
 425             /* Update vectorial force */
 426             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 427             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 428             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 429
 430             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 431             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 432             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 433
 434             }
 435
 436             /**************************
 437              * CALCULATE INTERACTIONS *
 438              **************************/
 439
 440             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 441             {
 442
 443             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
 444
 445             /* EWALD ELECTROSTATICS */
 446
 447             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 448             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
 449             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 450             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 451             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 452
 453             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 454             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 455             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 456             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 457             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 458             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 459             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 460             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 461             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
 462             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
 463
 464             d                = _fjsp_sub_v2r8(r02,rswitch);
 465             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 466             d2               = _fjsp_mul_v2r8(d,d);
 467             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 468
 469             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 470
 471             /* Evaluate switch function */
 472             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 473             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
 474             velec            = _fjsp_mul_v2r8(velec,sw);
 475             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 476
 477             /* Update potential sum for this i atom from the interaction with this j atom. */
 478             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 479             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 480
 481             fscal            = felec;
 482
 483             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 484
 485             /* Update vectorial force */
 486             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 487             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 488             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 489
 490             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 491             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 492             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 493
 494             }
 495
 496             /**************************
 497              * CALCULATE INTERACTIONS *
 498              **************************/
 499
 500             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 501             {
 502
 503             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 504
 505             /* EWALD ELECTROSTATICS */
 506
 507             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 508             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 509             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 510             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 511             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 512
 513             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 514             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 515             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 516             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 517             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 518             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 519             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 520             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 521             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 522             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 523
 524             d                = _fjsp_sub_v2r8(r10,rswitch);
 525             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 526             d2               = _fjsp_mul_v2r8(d,d);
 527             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 528
 529             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 530
 531             /* Evaluate switch function */
 532             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 533             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
 534             velec            = _fjsp_mul_v2r8(velec,sw);
 535             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 536
 537             /* Update potential sum for this i atom from the interaction with this j atom. */
 538             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 539             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 540
 541             fscal            = felec;
 542
 543             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 544
 545             /* Update vectorial force */
 546             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 547             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 548             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 549
 550             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 551             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 552             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 553
 554             }
 555
 556             /**************************
 557              * CALCULATE INTERACTIONS *
 558              **************************/
 559
 560             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 561             {
 562
 563             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 564
 565             /* EWALD ELECTROSTATICS */
 566
 567             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 568             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 569             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 570             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 571             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 572
 573             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 574             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 575             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 576             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 577             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 578             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 579             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 580             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 581             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
 582             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 583
 584             d                = _fjsp_sub_v2r8(r11,rswitch);
 585             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 586             d2               = _fjsp_mul_v2r8(d,d);
 587             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 588
 589             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 590
 591             /* Evaluate switch function */
 592             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 593             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
 594             velec            = _fjsp_mul_v2r8(velec,sw);
 595             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 596
 597             /* Update potential sum for this i atom from the interaction with this j atom. */
 598             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 599             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 600
 601             fscal            = felec;
 602
 603             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 604
 605             /* Update vectorial force */
 606             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 607             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 608             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 609
 610             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 611             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 612             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 613
 614             }
 615
 616             /**************************
 617              * CALCULATE INTERACTIONS *
 618              **************************/
 619
 620             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 621             {
 622
 623             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 624
 625             /* EWALD ELECTROSTATICS */
 626
 627             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 628             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 629             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 630             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 631             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 632
 633             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 634             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 635             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 636             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 637             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 638             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 639             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 640             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 641             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
 642             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 643
 644             d                = _fjsp_sub_v2r8(r12,rswitch);
 645             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 646             d2               = _fjsp_mul_v2r8(d,d);
 647             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 648
 649             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 650
 651             /* Evaluate switch function */
 652             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 653             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
 654             velec            = _fjsp_mul_v2r8(velec,sw);
 655             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 656
 657             /* Update potential sum for this i atom from the interaction with this j atom. */
 658             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 659             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 660
 661             fscal            = felec;
 662
 663             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 664
 665             /* Update vectorial force */
 666             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 667             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 668             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 669
 670             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 671             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 672             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 673
 674             }
 675
 676             /**************************
 677              * CALCULATE INTERACTIONS *
 678              **************************/
 679
 680             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 681             {
 682
 683             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 684
 685             /* EWALD ELECTROSTATICS */
 686
 687             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 688             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 689             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 690             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 691             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 692
 693             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 694             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 695             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 696             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 697             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 698             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 699             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 700             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 701             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 702             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 703
 704             d                = _fjsp_sub_v2r8(r20,rswitch);
 705             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 706             d2               = _fjsp_mul_v2r8(d,d);
 707             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 708
 709             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 710
 711             /* Evaluate switch function */
 712             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 713             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
 714             velec            = _fjsp_mul_v2r8(velec,sw);
 715             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 716
 717             /* Update potential sum for this i atom from the interaction with this j atom. */
 718             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 719             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 720
 721             fscal            = felec;
 722
 723             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 724
 725             /* Update vectorial force */
 726             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 727             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 728             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 729
 730             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 731             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 732             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 733
 734             }
 735
 736             /**************************
 737              * CALCULATE INTERACTIONS *
 738              **************************/
 739
 740             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 741             {
 742
 743             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 744
 745             /* EWALD ELECTROSTATICS */
 746
 747             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 748             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 749             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 750             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 751             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 752
 753             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 754             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 755             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 756             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 757             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 758             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 759             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 760             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 761             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
 762             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 763
 764             d                = _fjsp_sub_v2r8(r21,rswitch);
 765             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 766             d2               = _fjsp_mul_v2r8(d,d);
 767             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 768
 769             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 770
 771             /* Evaluate switch function */
 772             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 773             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
 774             velec            = _fjsp_mul_v2r8(velec,sw);
 775             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 779             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 784
 785             /* Update vectorial force */
 786             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 787             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 788             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 789
 790             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 791             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 792             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 793
 794             }
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 801             {
 802
 803             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 804
 805             /* EWALD ELECTROSTATICS */
 806
 807             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 808             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
 809             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 810             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 811             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 812
 813             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 814             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 815             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 816             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 817             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 818             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 819             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 820             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 821             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
 822             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
 823
 824             d                = _fjsp_sub_v2r8(r22,rswitch);
 825             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 826             d2               = _fjsp_mul_v2r8(d,d);
 827             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 828
 829             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 830
 831             /* Evaluate switch function */
 832             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 833             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
 834             velec            = _fjsp_mul_v2r8(velec,sw);
 835             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 839             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 840
 841             fscal            = felec;
 842
 843             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 844
 845             /* Update vectorial force */
 846             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 847             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 848             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 849
 850             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 851             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 852             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 853
 854             }
 855
 856             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 857
 858             /* Inner loop uses 630 flops */
 859         }
 860
 861         if(jidx<j_index_end)
 862         {
 863
 864             jnrA             = jjnr[jidx];
 865             j_coord_offsetA  = DIM*jnrA;
 866
 867             /* load j atom coordinates */
 868             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 869                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 870
 871             /* Calculate displacement vector */
 872             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 873             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 874             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 875             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 876             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 877             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 878             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 879             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 880             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 881             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 882             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 883             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 884             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 885             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 886             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 887             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 888             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 889             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 890             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 891             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 892             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 893             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 894             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 895             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 896             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 897             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 898             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 899
 900             /* Calculate squared distance and things based on it */
 901             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 902             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 903             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 904             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 905             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 906             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 907             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 908             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 909             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 910
 911             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 912             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 913             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 914             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 915             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 916             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 917             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 918             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 919             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 920
 921             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 922             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 923             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 924             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 925             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 926             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 927             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 928             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 929             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 930
 931             fjx0             = _fjsp_setzero_v2r8();
 932             fjy0             = _fjsp_setzero_v2r8();
 933             fjz0             = _fjsp_setzero_v2r8();
 934             fjx1             = _fjsp_setzero_v2r8();
 935             fjy1             = _fjsp_setzero_v2r8();
 936             fjz1             = _fjsp_setzero_v2r8();
 937             fjx2             = _fjsp_setzero_v2r8();
 938             fjy2             = _fjsp_setzero_v2r8();
 939             fjz2             = _fjsp_setzero_v2r8();
 940
 941             /**************************
 942              * CALCULATE INTERACTIONS *
 943              **************************/
 944
 945             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 946             {
 947
 948             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 949
 950             /* EWALD ELECTROSTATICS */
 951
 952             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 953             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 954             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 955             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 956             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 957
 958             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 959             ewtabD           = _fjsp_setzero_v2r8();
 960             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 961             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 962             ewtabFn          = _fjsp_setzero_v2r8();
 963             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 964             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 965             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 966             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 967             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 968
 969             /* LENNARD-JONES DISPERSION/REPULSION */
 970
 971             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 972             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 973             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 974             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 975             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 976
 977             d                = _fjsp_sub_v2r8(r00,rswitch);
 978             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 979             d2               = _fjsp_mul_v2r8(d,d);
 980             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 981
 982             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 983
 984             /* Evaluate switch function */
 985             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 986             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
 987             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
 988             velec            = _fjsp_mul_v2r8(velec,sw);
 989             vvdw             = _fjsp_mul_v2r8(vvdw,sw);
 990             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 991
 992             /* Update potential sum for this i atom from the interaction with this j atom. */
 993             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 994             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 995             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 996             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 997             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 998             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 999
1000             fscal            = _fjsp_add_v2r8(felec,fvdw);
1001
1002             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1003
1004             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1005
1006             /* Update vectorial force */
1007             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1008             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1009             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1010
1011             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1012             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1013             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1014
1015             }
1016
1017             /**************************
1018              * CALCULATE INTERACTIONS *
1019              **************************/
1020
1021             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1022             {
1023
1024             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1025
1026             /* EWALD ELECTROSTATICS */
1027
1028             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1029             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1030             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1031             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1032             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1033
1034             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1035             ewtabD           = _fjsp_setzero_v2r8();
1036             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1037             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1038             ewtabFn          = _fjsp_setzero_v2r8();
1039             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1040             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1041             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1042             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
1043             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1044
1045             d                = _fjsp_sub_v2r8(r01,rswitch);
1046             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1047             d2               = _fjsp_mul_v2r8(d,d);
1048             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1049
1050             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1051
1052             /* Evaluate switch function */
1053             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1054             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
1055             velec            = _fjsp_mul_v2r8(velec,sw);
1056             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1057
1058             /* Update potential sum for this i atom from the interaction with this j atom. */
1059             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1060             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1061             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1062
1063             fscal            = felec;
1064
1065             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1066
1067             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1068
1069             /* Update vectorial force */
1070             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1071             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1072             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1073
1074             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1075             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1076             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1077
1078             }
1079
1080             /**************************
1081              * CALCULATE INTERACTIONS *
1082              **************************/
1083
1084             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1085             {
1086
1087             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1088
1089             /* EWALD ELECTROSTATICS */
1090
1091             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1092             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1093             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1094             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1095             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1096
1097             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1098             ewtabD           = _fjsp_setzero_v2r8();
1099             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1100             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1101             ewtabFn          = _fjsp_setzero_v2r8();
1102             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1103             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1104             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1105             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
1106             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1107
1108             d                = _fjsp_sub_v2r8(r02,rswitch);
1109             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1110             d2               = _fjsp_mul_v2r8(d,d);
1111             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1112
1113             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1114
1115             /* Evaluate switch function */
1116             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1117             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
1118             velec            = _fjsp_mul_v2r8(velec,sw);
1119             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1120
1121             /* Update potential sum for this i atom from the interaction with this j atom. */
1122             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1123             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1124             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1125
1126             fscal            = felec;
1127
1128             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1129
1130             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1131
1132             /* Update vectorial force */
1133             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1134             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1135             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1136
1137             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1138             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1139             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1140
1141             }
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1148             {
1149
1150             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1151
1152             /* EWALD ELECTROSTATICS */
1153
1154             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1155             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1156             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1157             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1158             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1159
1160             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1161             ewtabD           = _fjsp_setzero_v2r8();
1162             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1163             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1164             ewtabFn          = _fjsp_setzero_v2r8();
1165             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1166             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1167             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1168             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
1169             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1170
1171             d                = _fjsp_sub_v2r8(r10,rswitch);
1172             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1173             d2               = _fjsp_mul_v2r8(d,d);
1174             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1175
1176             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1177
1178             /* Evaluate switch function */
1179             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1180             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
1181             velec            = _fjsp_mul_v2r8(velec,sw);
1182             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1183
1184             /* Update potential sum for this i atom from the interaction with this j atom. */
1185             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1186             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1187             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1188
1189             fscal            = felec;
1190
1191             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1192
1193             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1194
1195             /* Update vectorial force */
1196             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1197             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1198             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1199
1200             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1201             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1202             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1203
1204             }
1205
1206             /**************************
1207              * CALCULATE INTERACTIONS *
1208              **************************/
1209
1210             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1211             {
1212
1213             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1214
1215             /* EWALD ELECTROSTATICS */
1216
1217             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1218             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1219             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1220             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1221             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1222
1223             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1224             ewtabD           = _fjsp_setzero_v2r8();
1225             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1226             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1227             ewtabFn          = _fjsp_setzero_v2r8();
1228             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1229             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1230             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1231             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
1232             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1233
1234             d                = _fjsp_sub_v2r8(r11,rswitch);
1235             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1236             d2               = _fjsp_mul_v2r8(d,d);
1237             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1238
1239             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1240
1241             /* Evaluate switch function */
1242             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1243             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
1244             velec            = _fjsp_mul_v2r8(velec,sw);
1245             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1246
1247             /* Update potential sum for this i atom from the interaction with this j atom. */
1248             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1249             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1250             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1251
1252             fscal            = felec;
1253
1254             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1255
1256             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1257
1258             /* Update vectorial force */
1259             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1260             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1261             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1262
1263             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1264             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1265             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1266
1267             }
1268
1269             /**************************
1270              * CALCULATE INTERACTIONS *
1271              **************************/
1272
1273             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1274             {
1275
1276             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1277
1278             /* EWALD ELECTROSTATICS */
1279
1280             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1281             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1282             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1283             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1284             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1285
1286             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1287             ewtabD           = _fjsp_setzero_v2r8();
1288             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1289             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1290             ewtabFn          = _fjsp_setzero_v2r8();
1291             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1292             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1293             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1294             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
1295             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1296
1297             d                = _fjsp_sub_v2r8(r12,rswitch);
1298             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1299             d2               = _fjsp_mul_v2r8(d,d);
1300             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1301
1302             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1303
1304             /* Evaluate switch function */
1305             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1306             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
1307             velec            = _fjsp_mul_v2r8(velec,sw);
1308             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1309
1310             /* Update potential sum for this i atom from the interaction with this j atom. */
1311             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1312             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1313             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1314
1315             fscal            = felec;
1316
1317             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1318
1319             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1320
1321             /* Update vectorial force */
1322             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1323             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1324             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1325
1326             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1327             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1328             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1329
1330             }
1331
1332             /**************************
1333              * CALCULATE INTERACTIONS *
1334              **************************/
1335
1336             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1337             {
1338
1339             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1340
1341             /* EWALD ELECTROSTATICS */
1342
1343             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1344             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1345             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1346             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1347             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1348
1349             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1350             ewtabD           = _fjsp_setzero_v2r8();
1351             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1352             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1353             ewtabFn          = _fjsp_setzero_v2r8();
1354             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1355             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1356             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1357             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
1358             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1359
1360             d                = _fjsp_sub_v2r8(r20,rswitch);
1361             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1362             d2               = _fjsp_mul_v2r8(d,d);
1363             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1364
1365             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1366
1367             /* Evaluate switch function */
1368             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1369             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
1370             velec            = _fjsp_mul_v2r8(velec,sw);
1371             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1372
1373             /* Update potential sum for this i atom from the interaction with this j atom. */
1374             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1375             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1376             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1377
1378             fscal            = felec;
1379
1380             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1381
1382             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1383
1384             /* Update vectorial force */
1385             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1386             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1387             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1388
1389             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1390             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1391             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1392
1393             }
1394
1395             /**************************
1396              * CALCULATE INTERACTIONS *
1397              **************************/
1398
1399             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1400             {
1401
1402             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1403
1404             /* EWALD ELECTROSTATICS */
1405
1406             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1407             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1408             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1409             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1410             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1411
1412             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1413             ewtabD           = _fjsp_setzero_v2r8();
1414             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1415             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1416             ewtabFn          = _fjsp_setzero_v2r8();
1417             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1418             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1419             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1420             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
1421             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1422
1423             d                = _fjsp_sub_v2r8(r21,rswitch);
1424             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1425             d2               = _fjsp_mul_v2r8(d,d);
1426             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1427
1428             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1429
1430             /* Evaluate switch function */
1431             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1432             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
1433             velec            = _fjsp_mul_v2r8(velec,sw);
1434             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1435
1436             /* Update potential sum for this i atom from the interaction with this j atom. */
1437             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1438             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1439             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1440
1441             fscal            = felec;
1442
1443             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1444
1445             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1446
1447             /* Update vectorial force */
1448             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1449             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1450             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1451
1452             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1453             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1454             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1455
1456             }
1457
1458             /**************************
1459              * CALCULATE INTERACTIONS *
1460              **************************/
1461
1462             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1463             {
1464
1465             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1466
1467             /* EWALD ELECTROSTATICS */
1468
1469             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1470             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1471             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1472             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1473             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1474
1475             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1476             ewtabD           = _fjsp_setzero_v2r8();
1477             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1478             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1479             ewtabFn          = _fjsp_setzero_v2r8();
1480             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1481             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1482             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1483             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
1484             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1485
1486             d                = _fjsp_sub_v2r8(r22,rswitch);
1487             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1488             d2               = _fjsp_mul_v2r8(d,d);
1489             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1490
1491             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1492
1493             /* Evaluate switch function */
1494             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1495             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
1496             velec            = _fjsp_mul_v2r8(velec,sw);
1497             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1498
1499             /* Update potential sum for this i atom from the interaction with this j atom. */
1500             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1501             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1502             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1503
1504             fscal            = felec;
1505
1506             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1507
1508             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1509
1510             /* Update vectorial force */
1511             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1512             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1513             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1514
1515             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1516             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1517             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1518
1519             }
1520
1521             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1522
1523             /* Inner loop uses 630 flops */
1524         }
1525
1526         /* End of innermost loop */
1527
1528         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1529                                               f+i_coord_offset,fshift+i_shift_offset);
1530
1531         ggid                        = gid[iidx];
1532         /* Update potential energies */
1533         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1534         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
1535
1536         /* Increment number of inner iterations */
1537         inneriter                  += j_index_end - j_index_start;
1538
1539         /* Outer loop uses 20 flops */
1540     }
1541
1542     /* Increment number of outer iterations */
1543     outeriter        += nri;
1544
1545     /* Update outer/inner flops */
1546
1547     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*630);
1548 }
1549 /*
1550  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_F_sparc64_hpc_ace_double
1551  * Electrostatics interaction: Ewald
1552  * VdW interaction:            LennardJones
1553  * Geometry:                   Water3-Water3
1554  * Calculate force/pot:        Force
1555  */
1556 void
1557 nb_kernel_ElecEwSw_VdwLJSw_GeomW3W3_F_sparc64_hpc_ace_double
1558                     (t_nblist                    * gmx_restrict       nlist,
1559                      rvec                        * gmx_restrict          xx,
1560                      rvec                        * gmx_restrict          ff,
1561                      t_forcerec                  * gmx_restrict          fr,
1562                      t_mdatoms                   * gmx_restrict     mdatoms,
1563                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1564                      t_nrnb                      * gmx_restrict        nrnb)
1565 {
1566     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1567      * just 0 for non-waters.
1568      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1569      * jnr indices corresponding to data put in the four positions in the SIMD register.
1570      */
1571     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1572     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1573     int              jnrA,jnrB;
1574     int              j_coord_offsetA,j_coord_offsetB;
1575     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1576     real             rcutoff_scalar;
1577     real             *shiftvec,*fshift,*x,*f;
1578     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1579     int              vdwioffset0;
1580     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1581     int              vdwioffset1;
1582     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1583     int              vdwioffset2;
1584     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1585     int              vdwjidx0A,vdwjidx0B;
1586     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1587     int              vdwjidx1A,vdwjidx1B;
1588     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1589     int              vdwjidx2A,vdwjidx2B;
1590     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1591     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1592     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1593     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1594     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1595     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1596     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1597     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1598     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1599     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1600     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1601     real             *charge;
1602     int              nvdwtype;
1603     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1604     int              *vdwtype;
1605     real             *vdwparam;
1606     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
1607     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
1608     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1609     real             *ewtab;
1610     _fjsp_v2r8       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1611     real             rswitch_scalar,d_scalar;
1612     _fjsp_v2r8       itab_tmp;
1613     _fjsp_v2r8       dummy_mask,cutoff_mask;
1614     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1615     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1616     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1617
1618     x                = xx[0];
1619     f                = ff[0];
1620
1621     nri              = nlist->nri;
1622     iinr             = nlist->iinr;
1623     jindex           = nlist->jindex;
1624     jjnr             = nlist->jjnr;
1625     shiftidx         = nlist->shift;
1626     gid              = nlist->gid;
1627     shiftvec         = fr->shift_vec[0];
1628     fshift           = fr->fshift[0];
1629     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1630     charge           = mdatoms->chargeA;
1631     nvdwtype         = fr->ntype;
1632     vdwparam         = fr->nbfp;
1633     vdwtype          = mdatoms->typeA;
1634
1635     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
1636     ewtab            = fr->ic->tabq_coul_FDV0;
1637     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
1638     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
1639
1640     /* Setup water-specific parameters */
1641     inr              = nlist->iinr[0];
1642     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1643     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1644     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1645     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1646
1647     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1648     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1649     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1650     vdwjidx0A        = 2*vdwtype[inr+0];
1651     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1652     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
1653     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
1654     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1655     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1656     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1657     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1658     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1659     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1660     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1661     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1662
1663     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1664     rcutoff_scalar   = fr->rcoulomb;
1665     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1666     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1667
1668     rswitch_scalar   = fr->rcoulomb_switch;
1669     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
1670     /* Setup switch parameters */
1671     d_scalar         = rcutoff_scalar-rswitch_scalar;
1672     d                = gmx_fjsp_set1_v2r8(d_scalar);
1673     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
1674     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1675     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1676     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
1677     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1678     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1679
1680     /* Avoid stupid compiler warnings */
1681     jnrA = jnrB = 0;
1682     j_coord_offsetA = 0;
1683     j_coord_offsetB = 0;
1684
1685     outeriter        = 0;
1686     inneriter        = 0;
1687
1688     /* Start outer loop over neighborlists */
1689     for(iidx=0; iidx<nri; iidx++)
1690     {
1691         /* Load shift vector for this list */
1692         i_shift_offset   = DIM*shiftidx[iidx];
1693
1694         /* Load limits for loop over neighbors */
1695         j_index_start    = jindex[iidx];
1696         j_index_end      = jindex[iidx+1];
1697
1698         /* Get outer coordinate index */
1699         inr              = iinr[iidx];
1700         i_coord_offset   = DIM*inr;
1701
1702         /* Load i particle coords and add shift vector */
1703         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1704                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1705
1706         fix0             = _fjsp_setzero_v2r8();
1707         fiy0             = _fjsp_setzero_v2r8();
1708         fiz0             = _fjsp_setzero_v2r8();
1709         fix1             = _fjsp_setzero_v2r8();
1710         fiy1             = _fjsp_setzero_v2r8();
1711         fiz1             = _fjsp_setzero_v2r8();
1712         fix2             = _fjsp_setzero_v2r8();
1713         fiy2             = _fjsp_setzero_v2r8();
1714         fiz2             = _fjsp_setzero_v2r8();
1715
1716         /* Start inner kernel loop */
1717         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1718         {
1719
1720             /* Get j neighbor index, and coordinate index */
1721             jnrA             = jjnr[jidx];
1722             jnrB             = jjnr[jidx+1];
1723             j_coord_offsetA  = DIM*jnrA;
1724             j_coord_offsetB  = DIM*jnrB;
1725
1726             /* load j atom coordinates */
1727             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1728                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1729
1730             /* Calculate displacement vector */
1731             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1732             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1733             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1734             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1735             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1736             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1737             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1738             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1739             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1740             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1741             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1742             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1743             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1744             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1745             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1746             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1747             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1748             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1749             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1750             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1751             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1752             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1753             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1754             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1755             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1756             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1757             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1758
1759             /* Calculate squared distance and things based on it */
1760             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1761             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1762             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1763             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1764             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1765             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1766             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1767             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1768             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1769
1770             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1771             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1772             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1773             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1774             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1775             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1776             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1777             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1778             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1779
1780             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1781             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1782             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1783             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1784             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1785             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1786             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1787             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1788             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1789
1790             fjx0             = _fjsp_setzero_v2r8();
1791             fjy0             = _fjsp_setzero_v2r8();
1792             fjz0             = _fjsp_setzero_v2r8();
1793             fjx1             = _fjsp_setzero_v2r8();
1794             fjy1             = _fjsp_setzero_v2r8();
1795             fjz1             = _fjsp_setzero_v2r8();
1796             fjx2             = _fjsp_setzero_v2r8();
1797             fjy2             = _fjsp_setzero_v2r8();
1798             fjz2             = _fjsp_setzero_v2r8();
1799
1800             /**************************
1801              * CALCULATE INTERACTIONS *
1802              **************************/
1803
1804             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1805             {
1806
1807             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1808
1809             /* EWALD ELECTROSTATICS */
1810
1811             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1812             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
1813             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1814             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1815             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1816
1817             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1818             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1819             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1820             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1821             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1822             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1823             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1824             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1825             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
1826             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
1827
1828             /* LENNARD-JONES DISPERSION/REPULSION */
1829
1830             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1831             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
1832             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
1833             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
1834             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
1835
1836             d                = _fjsp_sub_v2r8(r00,rswitch);
1837             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1838             d2               = _fjsp_mul_v2r8(d,d);
1839             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1840
1841             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1842
1843             /* Evaluate switch function */
1844             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1845             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
1846             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
1847             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1848
1849             fscal            = _fjsp_add_v2r8(felec,fvdw);
1850
1851             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1852
1853             /* Update vectorial force */
1854             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1855             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1856             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1857
1858             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1859             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1860             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1861
1862             }
1863
1864             /**************************
1865              * CALCULATE INTERACTIONS *
1866              **************************/
1867
1868             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1869             {
1870
1871             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1872
1873             /* EWALD ELECTROSTATICS */
1874
1875             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1876             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1877             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1878             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1879             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1880
1881             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1882             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1883             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1884             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1885             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1886             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1887             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1888             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1889             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
1890             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1891
1892             d                = _fjsp_sub_v2r8(r01,rswitch);
1893             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1894             d2               = _fjsp_mul_v2r8(d,d);
1895             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1896
1897             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1898
1899             /* Evaluate switch function */
1900             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1901             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
1902             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1903
1904             fscal            = felec;
1905
1906             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1907
1908             /* Update vectorial force */
1909             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1910             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1911             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1912
1913             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1914             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1915             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1916
1917             }
1918
1919             /**************************
1920              * CALCULATE INTERACTIONS *
1921              **************************/
1922
1923             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1924             {
1925
1926             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1927
1928             /* EWALD ELECTROSTATICS */
1929
1930             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1931             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1932             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1933             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1934             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1935
1936             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1937             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1938             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1939             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1940             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1941             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1942             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1943             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1944             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
1945             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1946
1947             d                = _fjsp_sub_v2r8(r02,rswitch);
1948             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1949             d2               = _fjsp_mul_v2r8(d,d);
1950             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1951
1952             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1953
1954             /* Evaluate switch function */
1955             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1956             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
1957             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1958
1959             fscal            = felec;
1960
1961             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1962
1963             /* Update vectorial force */
1964             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1965             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1966             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1967
1968             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1969             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1970             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1971
1972             }
1973
1974             /**************************
1975              * CALCULATE INTERACTIONS *
1976              **************************/
1977
1978             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1979             {
1980
1981             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1982
1983             /* EWALD ELECTROSTATICS */
1984
1985             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1986             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1987             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1988             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1989             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1990
1991             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1992             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1993             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1994             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1995             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1996             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1997             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1998             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1999             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
2000             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
2001
2002             d                = _fjsp_sub_v2r8(r10,rswitch);
2003             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2004             d2               = _fjsp_mul_v2r8(d,d);
2005             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2006
2007             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2008
2009             /* Evaluate switch function */
2010             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2011             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
2012             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
2013
2014             fscal            = felec;
2015
2016             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2017
2018             /* Update vectorial force */
2019             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
2020             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
2021             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
2022
2023             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
2024             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
2025             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
2026
2027             }
2028
2029             /**************************
2030              * CALCULATE INTERACTIONS *
2031              **************************/
2032
2033             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
2034             {
2035
2036             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
2037
2038             /* EWALD ELECTROSTATICS */
2039
2040             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2041             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
2042             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2043             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2044             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2045
2046             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2047             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2048             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2049             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2050             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2051             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2052             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2053             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2054             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
2055             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
2056
2057             d                = _fjsp_sub_v2r8(r11,rswitch);
2058             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2059             d2               = _fjsp_mul_v2r8(d,d);
2060             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2061
2062             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2063
2064             /* Evaluate switch function */
2065             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2066             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
2067             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
2068
2069             fscal            = felec;
2070
2071             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2072
2073             /* Update vectorial force */
2074             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
2075             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
2076             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
2077
2078             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
2079             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
2080             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
2081
2082             }
2083
2084             /**************************
2085              * CALCULATE INTERACTIONS *
2086              **************************/
2087
2088             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2089             {
2090
2091             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2092
2093             /* EWALD ELECTROSTATICS */
2094
2095             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2096             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2097             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2098             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2099             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2100
2101             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2102             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2103             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2104             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2105             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2106             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2107             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2108             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2109             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
2110             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2111
2112             d                = _fjsp_sub_v2r8(r12,rswitch);
2113             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2114             d2               = _fjsp_mul_v2r8(d,d);
2115             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2116
2117             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2118
2119             /* Evaluate switch function */
2120             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2121             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
2122             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2123
2124             fscal            = felec;
2125
2126             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2127
2128             /* Update vectorial force */
2129             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2130             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2131             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2132
2133             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2134             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2135             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2136
2137             }
2138
2139             /**************************
2140              * CALCULATE INTERACTIONS *
2141              **************************/
2142
2143             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
2144             {
2145
2146             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
2147
2148             /* EWALD ELECTROSTATICS */
2149
2150             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2151             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
2152             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2153             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2154             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2155
2156             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2157             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2158             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2159             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2160             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2161             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2162             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2163             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2164             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
2165             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
2166
2167             d                = _fjsp_sub_v2r8(r20,rswitch);
2168             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2169             d2               = _fjsp_mul_v2r8(d,d);
2170             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2171
2172             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2173
2174             /* Evaluate switch function */
2175             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2176             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
2177             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
2178
2179             fscal            = felec;
2180
2181             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2182
2183             /* Update vectorial force */
2184             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
2185             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
2186             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
2187
2188             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
2189             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
2190             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
2191
2192             }
2193
2194             /**************************
2195              * CALCULATE INTERACTIONS *
2196              **************************/
2197
2198             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2199             {
2200
2201             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2202
2203             /* EWALD ELECTROSTATICS */
2204
2205             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2206             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2207             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2208             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2209             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2210
2211             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2212             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2213             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2214             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2215             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2216             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2217             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2218             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2219             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
2220             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2221
2222             d                = _fjsp_sub_v2r8(r21,rswitch);
2223             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2224             d2               = _fjsp_mul_v2r8(d,d);
2225             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2226
2227             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2228
2229             /* Evaluate switch function */
2230             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2231             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
2232             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2233
2234             fscal            = felec;
2235
2236             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2237
2238             /* Update vectorial force */
2239             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2240             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2241             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2242
2243             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2244             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2245             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2246
2247             }
2248
2249             /**************************
2250              * CALCULATE INTERACTIONS *
2251              **************************/
2252
2253             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2254             {
2255
2256             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2257
2258             /* EWALD ELECTROSTATICS */
2259
2260             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2261             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2262             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2263             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2264             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2265
2266             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2267             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2268             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2269             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2270             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2271             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2272             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2273             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2274             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
2275             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2276
2277             d                = _fjsp_sub_v2r8(r22,rswitch);
2278             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2279             d2               = _fjsp_mul_v2r8(d,d);
2280             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2281
2282             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2283
2284             /* Evaluate switch function */
2285             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2286             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
2287             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2288
2289             fscal            = felec;
2290
2291             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2292
2293             /* Update vectorial force */
2294             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2295             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2296             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2297
2298             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2299             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2300             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2301
2302             }
2303
2304             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2305
2306             /* Inner loop uses 600 flops */
2307         }
2308
2309         if(jidx<j_index_end)
2310         {
2311
2312             jnrA             = jjnr[jidx];
2313             j_coord_offsetA  = DIM*jnrA;
2314
2315             /* load j atom coordinates */
2316             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
2317                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
2318
2319             /* Calculate displacement vector */
2320             dx00             = _fjsp_sub_v2r8(ix0,jx0);
2321             dy00             = _fjsp_sub_v2r8(iy0,jy0);
2322             dz00             = _fjsp_sub_v2r8(iz0,jz0);
2323             dx01             = _fjsp_sub_v2r8(ix0,jx1);
2324             dy01             = _fjsp_sub_v2r8(iy0,jy1);
2325             dz01             = _fjsp_sub_v2r8(iz0,jz1);
2326             dx02             = _fjsp_sub_v2r8(ix0,jx2);
2327             dy02             = _fjsp_sub_v2r8(iy0,jy2);
2328             dz02             = _fjsp_sub_v2r8(iz0,jz2);
2329             dx10             = _fjsp_sub_v2r8(ix1,jx0);
2330             dy10             = _fjsp_sub_v2r8(iy1,jy0);
2331             dz10             = _fjsp_sub_v2r8(iz1,jz0);
2332             dx11             = _fjsp_sub_v2r8(ix1,jx1);
2333             dy11             = _fjsp_sub_v2r8(iy1,jy1);
2334             dz11             = _fjsp_sub_v2r8(iz1,jz1);
2335             dx12             = _fjsp_sub_v2r8(ix1,jx2);
2336             dy12             = _fjsp_sub_v2r8(iy1,jy2);
2337             dz12             = _fjsp_sub_v2r8(iz1,jz2);
2338             dx20             = _fjsp_sub_v2r8(ix2,jx0);
2339             dy20             = _fjsp_sub_v2r8(iy2,jy0);
2340             dz20             = _fjsp_sub_v2r8(iz2,jz0);
2341             dx21             = _fjsp_sub_v2r8(ix2,jx1);
2342             dy21             = _fjsp_sub_v2r8(iy2,jy1);
2343             dz21             = _fjsp_sub_v2r8(iz2,jz1);
2344             dx22             = _fjsp_sub_v2r8(ix2,jx2);
2345             dy22             = _fjsp_sub_v2r8(iy2,jy2);
2346             dz22             = _fjsp_sub_v2r8(iz2,jz2);
2347
2348             /* Calculate squared distance and things based on it */
2349             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
2350             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
2351             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
2352             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
2353             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
2354             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
2355             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
2356             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
2357             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
2358
2359             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
2360             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
2361             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
2362             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
2363             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
2364             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
2365             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
2366             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
2367             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
2368
2369             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
2370             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
2371             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
2372             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
2373             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
2374             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
2375             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
2376             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
2377             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
2378
2379             fjx0             = _fjsp_setzero_v2r8();
2380             fjy0             = _fjsp_setzero_v2r8();
2381             fjz0             = _fjsp_setzero_v2r8();
2382             fjx1             = _fjsp_setzero_v2r8();
2383             fjy1             = _fjsp_setzero_v2r8();
2384             fjz1             = _fjsp_setzero_v2r8();
2385             fjx2             = _fjsp_setzero_v2r8();
2386             fjy2             = _fjsp_setzero_v2r8();
2387             fjz2             = _fjsp_setzero_v2r8();
2388
2389             /**************************
2390              * CALCULATE INTERACTIONS *
2391              **************************/
2392
2393             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
2394             {
2395
2396             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
2397
2398             /* EWALD ELECTROSTATICS */
2399
2400             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2401             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
2402             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2403             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2404             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2405
2406             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2407             ewtabD           = _fjsp_setzero_v2r8();
2408             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2409             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2410             ewtabFn          = _fjsp_setzero_v2r8();
2411             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2412             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2413             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2414             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
2415             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
2416
2417             /* LENNARD-JONES DISPERSION/REPULSION */
2418
2419             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
2420             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
2421             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
2422             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
2423             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
2424
2425             d                = _fjsp_sub_v2r8(r00,rswitch);
2426             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2427             d2               = _fjsp_mul_v2r8(d,d);
2428             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2429
2430             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2431
2432             /* Evaluate switch function */
2433             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2434             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
2435             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
2436             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
2437
2438             fscal            = _fjsp_add_v2r8(felec,fvdw);
2439
2440             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2441
2442             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2443
2444             /* Update vectorial force */
2445             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
2446             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
2447             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
2448
2449             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
2450             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
2451             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
2452
2453             }
2454
2455             /**************************
2456              * CALCULATE INTERACTIONS *
2457              **************************/
2458
2459             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
2460             {
2461
2462             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
2463
2464             /* EWALD ELECTROSTATICS */
2465
2466             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2467             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
2468             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2469             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2470             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2471
2472             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2473             ewtabD           = _fjsp_setzero_v2r8();
2474             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2475             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2476             ewtabFn          = _fjsp_setzero_v2r8();
2477             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2478             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2479             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2480             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
2481             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
2482
2483             d                = _fjsp_sub_v2r8(r01,rswitch);
2484             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2485             d2               = _fjsp_mul_v2r8(d,d);
2486             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2487
2488             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2489
2490             /* Evaluate switch function */
2491             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2492             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
2493             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
2494
2495             fscal            = felec;
2496
2497             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2498
2499             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2500
2501             /* Update vectorial force */
2502             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
2503             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
2504             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
2505
2506             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
2507             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
2508             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
2509
2510             }
2511
2512             /**************************
2513              * CALCULATE INTERACTIONS *
2514              **************************/
2515
2516             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
2517             {
2518
2519             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
2520
2521             /* EWALD ELECTROSTATICS */
2522
2523             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2524             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
2525             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2526             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2527             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2528
2529             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2530             ewtabD           = _fjsp_setzero_v2r8();
2531             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2532             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2533             ewtabFn          = _fjsp_setzero_v2r8();
2534             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2535             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2536             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2537             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
2538             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
2539
2540             d                = _fjsp_sub_v2r8(r02,rswitch);
2541             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2542             d2               = _fjsp_mul_v2r8(d,d);
2543             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2544
2545             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2546
2547             /* Evaluate switch function */
2548             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2549             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
2550             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
2551
2552             fscal            = felec;
2553
2554             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2555
2556             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2557
2558             /* Update vectorial force */
2559             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
2560             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
2561             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
2562
2563             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
2564             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
2565             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
2566
2567             }
2568
2569             /**************************
2570              * CALCULATE INTERACTIONS *
2571              **************************/
2572
2573             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
2574             {
2575
2576             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
2577
2578             /* EWALD ELECTROSTATICS */
2579
2580             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2581             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
2582             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2583             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2584             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2585
2586             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2587             ewtabD           = _fjsp_setzero_v2r8();
2588             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2589             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2590             ewtabFn          = _fjsp_setzero_v2r8();
2591             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2592             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2593             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2594             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
2595             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
2596
2597             d                = _fjsp_sub_v2r8(r10,rswitch);
2598             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2599             d2               = _fjsp_mul_v2r8(d,d);
2600             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2601
2602             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2603
2604             /* Evaluate switch function */
2605             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2606             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
2607             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
2608
2609             fscal            = felec;
2610
2611             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2612
2613             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2614
2615             /* Update vectorial force */
2616             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
2617             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
2618             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
2619
2620             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
2621             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
2622             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
2623
2624             }
2625
2626             /**************************
2627              * CALCULATE INTERACTIONS *
2628              **************************/
2629
2630             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
2631             {
2632
2633             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
2634
2635             /* EWALD ELECTROSTATICS */
2636
2637             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2638             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
2639             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2640             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2641             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2642
2643             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2644             ewtabD           = _fjsp_setzero_v2r8();
2645             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2646             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2647             ewtabFn          = _fjsp_setzero_v2r8();
2648             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2649             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2650             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2651             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
2652             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
2653
2654             d                = _fjsp_sub_v2r8(r11,rswitch);
2655             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2656             d2               = _fjsp_mul_v2r8(d,d);
2657             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2658
2659             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2660
2661             /* Evaluate switch function */
2662             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2663             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
2664             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
2665
2666             fscal            = felec;
2667
2668             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2669
2670             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2671
2672             /* Update vectorial force */
2673             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
2674             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
2675             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
2676
2677             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
2678             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
2679             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
2680
2681             }
2682
2683             /**************************
2684              * CALCULATE INTERACTIONS *
2685              **************************/
2686
2687             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2688             {
2689
2690             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2691
2692             /* EWALD ELECTROSTATICS */
2693
2694             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2695             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2696             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2697             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2698             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2699
2700             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2701             ewtabD           = _fjsp_setzero_v2r8();
2702             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2703             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2704             ewtabFn          = _fjsp_setzero_v2r8();
2705             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2706             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2707             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2708             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
2709             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2710
2711             d                = _fjsp_sub_v2r8(r12,rswitch);
2712             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2713             d2               = _fjsp_mul_v2r8(d,d);
2714             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2715
2716             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2717
2718             /* Evaluate switch function */
2719             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2720             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
2721             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2722
2723             fscal            = felec;
2724
2725             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2726
2727             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2728
2729             /* Update vectorial force */
2730             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2731             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2732             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2733
2734             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2735             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2736             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2737
2738             }
2739
2740             /**************************
2741              * CALCULATE INTERACTIONS *
2742              **************************/
2743
2744             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
2745             {
2746
2747             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
2748
2749             /* EWALD ELECTROSTATICS */
2750
2751             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2752             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
2753             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2754             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2755             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2756
2757             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2758             ewtabD           = _fjsp_setzero_v2r8();
2759             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2760             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2761             ewtabFn          = _fjsp_setzero_v2r8();
2762             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2763             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2764             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2765             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
2766             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
2767
2768             d                = _fjsp_sub_v2r8(r20,rswitch);
2769             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2770             d2               = _fjsp_mul_v2r8(d,d);
2771             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2772
2773             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2774
2775             /* Evaluate switch function */
2776             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2777             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
2778             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
2779
2780             fscal            = felec;
2781
2782             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2783
2784             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2785
2786             /* Update vectorial force */
2787             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
2788             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
2789             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
2790
2791             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
2792             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
2793             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
2794
2795             }
2796
2797             /**************************
2798              * CALCULATE INTERACTIONS *
2799              **************************/
2800
2801             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2802             {
2803
2804             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2805
2806             /* EWALD ELECTROSTATICS */
2807
2808             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2809             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2810             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2811             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2812             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2813
2814             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2815             ewtabD           = _fjsp_setzero_v2r8();
2816             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2817             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2818             ewtabFn          = _fjsp_setzero_v2r8();
2819             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2820             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2821             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2822             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
2823             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2824
2825             d                = _fjsp_sub_v2r8(r21,rswitch);
2826             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2827             d2               = _fjsp_mul_v2r8(d,d);
2828             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2829
2830             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2831
2832             /* Evaluate switch function */
2833             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2834             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
2835             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2836
2837             fscal            = felec;
2838
2839             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2840
2841             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2842
2843             /* Update vectorial force */
2844             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2845             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2846             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2847
2848             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2849             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2850             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2851
2852             }
2853
2854             /**************************
2855              * CALCULATE INTERACTIONS *
2856              **************************/
2857
2858             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2859             {
2860
2861             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2862
2863             /* EWALD ELECTROSTATICS */
2864
2865             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2866             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2867             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2868             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2869             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2870
2871             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2872             ewtabD           = _fjsp_setzero_v2r8();
2873             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2874             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2875             ewtabFn          = _fjsp_setzero_v2r8();
2876             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2877             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2878             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2879             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
2880             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2881
2882             d                = _fjsp_sub_v2r8(r22,rswitch);
2883             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2884             d2               = _fjsp_mul_v2r8(d,d);
2885             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2886
2887             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2888
2889             /* Evaluate switch function */
2890             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2891             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
2892             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2893
2894             fscal            = felec;
2895
2896             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2897
2898             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2899
2900             /* Update vectorial force */
2901             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2902             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2903             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2904
2905             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2906             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2907             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2908
2909             }
2910
2911             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2912
2913             /* Inner loop uses 600 flops */
2914         }
2915
2916         /* End of innermost loop */
2917
2918         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2919                                               f+i_coord_offset,fshift+i_shift_offset);
2920
2921         /* Increment number of inner iterations */
2922         inneriter                  += j_index_end - j_index_start;
2923
2924         /* Outer loop uses 18 flops */
2925     }
2926
2927     /* Increment number of outer iterations */
2928     outeriter        += nri;
2929
2930     /* Update outer/inner flops */
2931
2932     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*600);
2933 }