src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEw_VdwLJEw_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_ElecEw_VdwLJEw_GeomW3W3_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJEw_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           c6grid_00;
 109     _fjsp_v2r8           c6grid_01;
 110     _fjsp_v2r8           c6grid_02;
 111     _fjsp_v2r8           c6grid_10;
 112     _fjsp_v2r8           c6grid_11;
 113     _fjsp_v2r8           c6grid_12;
 114     _fjsp_v2r8           c6grid_20;
 115     _fjsp_v2r8           c6grid_21;
 116     _fjsp_v2r8           c6grid_22;
 117     real                 *vdwgridparam;
 118     _fjsp_v2r8           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 119     _fjsp_v2r8           one_half = gmx_fjsp_set1_v2r8(0.5);
 120     _fjsp_v2r8           minus_one = gmx_fjsp_set1_v2r8(-1.0);
 121     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 122     real             *ewtab;
 123     _fjsp_v2r8       itab_tmp;
 124     _fjsp_v2r8       dummy_mask,cutoff_mask;
 125     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 126     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 127     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 128
 129     x                = xx[0];
 130     f                = ff[0];
 131
 132     nri              = nlist->nri;
 133     iinr             = nlist->iinr;
 134     jindex           = nlist->jindex;
 135     jjnr             = nlist->jjnr;
 136     shiftidx         = nlist->shift;
 137     gid              = nlist->gid;
 138     shiftvec         = fr->shift_vec[0];
 139     fshift           = fr->fshift[0];
 140     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 141     charge           = mdatoms->chargeA;
 142     nvdwtype         = fr->ntype;
 143     vdwparam         = fr->nbfp;
 144     vdwtype          = mdatoms->typeA;
 145     vdwgridparam     = fr->ljpme_c6grid;
 146     sh_lj_ewald      = gmx_fjsp_set1_v2r8(fr->ic->sh_lj_ewald);
 147     ewclj            = gmx_fjsp_set1_v2r8(fr->ewaldcoeff_lj);
 148     ewclj2           = _fjsp_mul_v2r8(minus_one,_fjsp_mul_v2r8(ewclj,ewclj));
 149
 150     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 151     ewtab            = fr->ic->tabq_coul_FDV0;
 152     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 153     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 154
 155     /* Setup water-specific parameters */
 156     inr              = nlist->iinr[0];
 157     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 158     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 159     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 160     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 161
 162     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 163     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 164     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 165     vdwjidx0A        = 2*vdwtype[inr+0];
 166     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 167     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 168     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 169     c6grid_00        = gmx_fjsp_set1_v2r8(vdwgridparam[vdwioffset0+vdwjidx0A]);
 170     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 171     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 172     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 173     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 174     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 175     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 176     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 177     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 178
 179     /* Avoid stupid compiler warnings */
 180     jnrA = jnrB = 0;
 181     j_coord_offsetA = 0;
 182     j_coord_offsetB = 0;
 183
 184     outeriter        = 0;
 185     inneriter        = 0;
 186
 187     /* Start outer loop over neighborlists */
 188     for(iidx=0; iidx<nri; iidx++)
 189     {
 190         /* Load shift vector for this list */
 191         i_shift_offset   = DIM*shiftidx[iidx];
 192
 193         /* Load limits for loop over neighbors */
 194         j_index_start    = jindex[iidx];
 195         j_index_end      = jindex[iidx+1];
 196
 197         /* Get outer coordinate index */
 198         inr              = iinr[iidx];
 199         i_coord_offset   = DIM*inr;
 200
 201         /* Load i particle coords and add shift vector */
 202         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 203                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 204
 205         fix0             = _fjsp_setzero_v2r8();
 206         fiy0             = _fjsp_setzero_v2r8();
 207         fiz0             = _fjsp_setzero_v2r8();
 208         fix1             = _fjsp_setzero_v2r8();
 209         fiy1             = _fjsp_setzero_v2r8();
 210         fiz1             = _fjsp_setzero_v2r8();
 211         fix2             = _fjsp_setzero_v2r8();
 212         fiy2             = _fjsp_setzero_v2r8();
 213         fiz2             = _fjsp_setzero_v2r8();
 214
 215         /* Reset potential sums */
 216         velecsum         = _fjsp_setzero_v2r8();
 217         vvdwsum          = _fjsp_setzero_v2r8();
 218
 219         /* Start inner kernel loop */
 220         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 221         {
 222
 223             /* Get j neighbor index, and coordinate index */
 224             jnrA             = jjnr[jidx];
 225             jnrB             = jjnr[jidx+1];
 226             j_coord_offsetA  = DIM*jnrA;
 227             j_coord_offsetB  = DIM*jnrB;
 228
 229             /* load j atom coordinates */
 230             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 231                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 232
 233             /* Calculate displacement vector */
 234             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 235             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 236             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 237             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 238             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 239             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 240             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 241             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 242             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 243             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 244             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 245             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 246             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 247             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 248             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 249             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 250             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 251             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 252             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 253             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 254             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 255             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 256             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 257             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 258             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 259             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 260             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 261
 262             /* Calculate squared distance and things based on it */
 263             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 264             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 265             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 266             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 267             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 268             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 269             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 270             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 271             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 272
 273             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 274             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 275             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 276             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 277             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 278             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 279             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 280             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 281             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 282
 283             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 284             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 285             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 286             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 287             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 288             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 289             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 290             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 291             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 292
 293             fjx0             = _fjsp_setzero_v2r8();
 294             fjy0             = _fjsp_setzero_v2r8();
 295             fjz0             = _fjsp_setzero_v2r8();
 296             fjx1             = _fjsp_setzero_v2r8();
 297             fjy1             = _fjsp_setzero_v2r8();
 298             fjz1             = _fjsp_setzero_v2r8();
 299             fjx2             = _fjsp_setzero_v2r8();
 300             fjy2             = _fjsp_setzero_v2r8();
 301             fjz2             = _fjsp_setzero_v2r8();
 302
 303             /**************************
 304              * CALCULATE INTERACTIONS *
 305              **************************/
 306
 307             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 308
 309             /* EWALD ELECTROSTATICS */
 310
 311             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 312             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 313             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 314             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 315             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 316
 317             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 318             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 319             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 320             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 321             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 322             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 323             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 324             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 325             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 326             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 327
 328             /* Analytical LJ-PME */
 329             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 330             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
 331             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
 332             exponent         = gmx_simd_exp_d(-ewcljrsq);
 333             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 334             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
 335             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 336             vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(-c6grid_00,_fjsp_sub_v2r8(one,poly),c6_00),rinvsix);
 337             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 338             vvdw             = _fjsp_msub_v2r8(vvdw12,one_twelfth,_fjsp_mul_v2r8(vvdw6,one_sixth));
 339             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 340             fvdw             = _fjsp_mul_v2r8(_fjsp_add_v2r8(vvdw12,_fjsp_msub_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6),vvdw6)),rinvsq00);
 341
 342             /* Update potential sum for this i atom from the interaction with this j atom. */
 343             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 344             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 345
 346             fscal            = _fjsp_add_v2r8(felec,fvdw);
 347
 348             /* Update vectorial force */
 349             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 350             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 351             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 352
 353             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 354             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 355             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 362
 363             /* EWALD ELECTROSTATICS */
 364
 365             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 366             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 367             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 368             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 369             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 370
 371             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 372             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 373             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 374             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 375             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 376             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 377             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 378             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 379             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
 380             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387             /* Update vectorial force */
 388             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 389             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 390             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 391
 392             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 393             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 394             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 395
 396             /**************************
 397              * CALCULATE INTERACTIONS *
 398              **************************/
 399
 400             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
 401
 402             /* EWALD ELECTROSTATICS */
 403
 404             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 405             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
 406             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 407             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 408             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 409
 410             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 411             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 412             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 413             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 414             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 415             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 416             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 417             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 418             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
 419             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
 420
 421             /* Update potential sum for this i atom from the interaction with this j atom. */
 422             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             /* Update vectorial force */
 427             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 428             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 429             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 430
 431             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 432             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 433             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 440
 441             /* EWALD ELECTROSTATICS */
 442
 443             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 444             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 445             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 446             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 447             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 448
 449             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 450             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 451             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 452             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 453             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 454             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 455             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 456             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 457             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 458             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             /* Update vectorial force */
 466             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 467             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 468             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 469
 470             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 471             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 472             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 479
 480             /* EWALD ELECTROSTATICS */
 481
 482             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 483             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 484             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 485             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 486             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 487
 488             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 489             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 490             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 491             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 492             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 493             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 494             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 495             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 496             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
 497             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 498
 499             /* Update potential sum for this i atom from the interaction with this j atom. */
 500             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 501
 502             fscal            = felec;
 503
 504             /* Update vectorial force */
 505             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 506             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 507             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 508
 509             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 510             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 511             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 512
 513             /**************************
 514              * CALCULATE INTERACTIONS *
 515              **************************/
 516
 517             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 518
 519             /* EWALD ELECTROSTATICS */
 520
 521             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 522             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 523             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 524             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 525             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 526
 527             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 528             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 529             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 530             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 531             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 532             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 533             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 534             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 535             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
 536             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 540
 541             fscal            = felec;
 542
 543             /* Update vectorial force */
 544             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 545             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 546             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 547
 548             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 549             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 550             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 551
 552             /**************************
 553              * CALCULATE INTERACTIONS *
 554              **************************/
 555
 556             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 557
 558             /* EWALD ELECTROSTATICS */
 559
 560             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 561             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 562             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 563             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 564             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 565
 566             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 567             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 568             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 569             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 570             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 571             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 572             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 573             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 574             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 575             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 576
 577             /* Update potential sum for this i atom from the interaction with this j atom. */
 578             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 579
 580             fscal            = felec;
 581
 582             /* Update vectorial force */
 583             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 584             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 585             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 586
 587             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 588             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 589             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 590
 591             /**************************
 592              * CALCULATE INTERACTIONS *
 593              **************************/
 594
 595             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 596
 597             /* EWALD ELECTROSTATICS */
 598
 599             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 600             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 601             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 602             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 603             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 604
 605             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 606             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 607             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 608             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 609             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 610             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 611             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 612             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 613             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
 614             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 615
 616             /* Update potential sum for this i atom from the interaction with this j atom. */
 617             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 618
 619             fscal            = felec;
 620
 621             /* Update vectorial force */
 622             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 623             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 624             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 625
 626             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 627             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 628             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 629
 630             /**************************
 631              * CALCULATE INTERACTIONS *
 632              **************************/
 633
 634             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 635
 636             /* EWALD ELECTROSTATICS */
 637
 638             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 639             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
 640             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 641             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 642             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 643
 644             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 645             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 646             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 647             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 648             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 649             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 650             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 651             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 652             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
 653             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
 654
 655             /* Update potential sum for this i atom from the interaction with this j atom. */
 656             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 657
 658             fscal            = felec;
 659
 660             /* Update vectorial force */
 661             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 662             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 663             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 664
 665             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 666             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 667             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 668
 669             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 670
 671             /* Inner loop uses 420 flops */
 672         }
 673
 674         if(jidx<j_index_end)
 675         {
 676
 677             jnrA             = jjnr[jidx];
 678             j_coord_offsetA  = DIM*jnrA;
 679
 680             /* load j atom coordinates */
 681             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 682                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 683
 684             /* Calculate displacement vector */
 685             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 686             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 687             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 688             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 689             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 690             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 691             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 692             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 693             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 694             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 695             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 696             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 697             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 698             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 699             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 700             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 701             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 702             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 703             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 704             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 705             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 706             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 707             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 708             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 709             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 710             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 711             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 712
 713             /* Calculate squared distance and things based on it */
 714             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 715             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 716             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 717             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 718             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 719             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 720             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 721             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 722             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 723
 724             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 725             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 726             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 727             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 728             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 729             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 730             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 731             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 732             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 733
 734             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 735             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 736             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 737             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 738             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 739             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 740             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 741             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 742             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 743
 744             fjx0             = _fjsp_setzero_v2r8();
 745             fjy0             = _fjsp_setzero_v2r8();
 746             fjz0             = _fjsp_setzero_v2r8();
 747             fjx1             = _fjsp_setzero_v2r8();
 748             fjy1             = _fjsp_setzero_v2r8();
 749             fjz1             = _fjsp_setzero_v2r8();
 750             fjx2             = _fjsp_setzero_v2r8();
 751             fjy2             = _fjsp_setzero_v2r8();
 752             fjz2             = _fjsp_setzero_v2r8();
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 759
 760             /* EWALD ELECTROSTATICS */
 761
 762             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 763             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 764             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 765             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 766             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 767
 768             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 769             ewtabD           = _fjsp_setzero_v2r8();
 770             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 771             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 772             ewtabFn          = _fjsp_setzero_v2r8();
 773             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 774             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 775             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 776             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 777             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 778
 779             /* Analytical LJ-PME */
 780             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 781             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
 782             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
 783             exponent         = gmx_simd_exp_d(-ewcljrsq);
 784             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 785             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
 786             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 787             vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(-c6grid_00,_fjsp_sub_v2r8(one,poly),c6_00),rinvsix);
 788             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 789             vvdw             = _fjsp_msub_v2r8(vvdw12,one_twelfth,_fjsp_mul_v2r8(vvdw6,one_sixth));
 790             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 791             fvdw             = _fjsp_mul_v2r8(_fjsp_add_v2r8(vvdw12,_fjsp_msub_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6),vvdw6)),rinvsq00);
 792
 793             /* Update potential sum for this i atom from the interaction with this j atom. */
 794             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 795             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 796             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 797             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 798
 799             fscal            = _fjsp_add_v2r8(felec,fvdw);
 800
 801             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 802
 803             /* Update vectorial force */
 804             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 805             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 806             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 807
 808             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 809             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 810             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 811
 812             /**************************
 813              * CALCULATE INTERACTIONS *
 814              **************************/
 815
 816             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 817
 818             /* EWALD ELECTROSTATICS */
 819
 820             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 821             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 822             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 823             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 824             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 825
 826             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 827             ewtabD           = _fjsp_setzero_v2r8();
 828             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 829             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 830             ewtabFn          = _fjsp_setzero_v2r8();
 831             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 832             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 833             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 834             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
 835             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 839             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 840
 841             fscal            = felec;
 842
 843             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 844
 845             /* Update vectorial force */
 846             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 847             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 848             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 849
 850             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 851             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 852             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
 859
 860             /* EWALD ELECTROSTATICS */
 861
 862             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 863             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
 864             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 865             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 866             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 867
 868             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 869             ewtabD           = _fjsp_setzero_v2r8();
 870             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 871             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 872             ewtabFn          = _fjsp_setzero_v2r8();
 873             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 874             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 875             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 876             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
 877             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
 878
 879             /* Update potential sum for this i atom from the interaction with this j atom. */
 880             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 881             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 882
 883             fscal            = felec;
 884
 885             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 886
 887             /* Update vectorial force */
 888             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 889             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 890             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 891
 892             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 893             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 894             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 895
 896             /**************************
 897              * CALCULATE INTERACTIONS *
 898              **************************/
 899
 900             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 901
 902             /* EWALD ELECTROSTATICS */
 903
 904             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 905             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 906             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 907             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 908             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 909
 910             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 911             ewtabD           = _fjsp_setzero_v2r8();
 912             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 913             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 914             ewtabFn          = _fjsp_setzero_v2r8();
 915             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 916             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 917             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 918             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 919             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 920
 921             /* Update potential sum for this i atom from the interaction with this j atom. */
 922             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 923             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 924
 925             fscal            = felec;
 926
 927             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 928
 929             /* Update vectorial force */
 930             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 931             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 932             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 933
 934             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 935             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 936             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 937
 938             /**************************
 939              * CALCULATE INTERACTIONS *
 940              **************************/
 941
 942             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 943
 944             /* EWALD ELECTROSTATICS */
 945
 946             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 947             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 948             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 949             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 950             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 951
 952             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 953             ewtabD           = _fjsp_setzero_v2r8();
 954             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 955             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 956             ewtabFn          = _fjsp_setzero_v2r8();
 957             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 958             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 959             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 960             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
 961             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 962
 963             /* Update potential sum for this i atom from the interaction with this j atom. */
 964             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 965             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 966
 967             fscal            = felec;
 968
 969             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 970
 971             /* Update vectorial force */
 972             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 973             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 974             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 975
 976             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 977             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 978             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 979
 980             /**************************
 981              * CALCULATE INTERACTIONS *
 982              **************************/
 983
 984             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 985
 986             /* EWALD ELECTROSTATICS */
 987
 988             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 989             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 990             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 991             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 992             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 993
 994             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 995             ewtabD           = _fjsp_setzero_v2r8();
 996             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 997             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 998             ewtabFn          = _fjsp_setzero_v2r8();
 999             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1000             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1001             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1002             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
1003             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1004
1005             /* Update potential sum for this i atom from the interaction with this j atom. */
1006             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1007             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1008
1009             fscal            = felec;
1010
1011             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1012
1013             /* Update vectorial force */
1014             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1015             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1016             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1017
1018             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1019             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1020             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1021
1022             /**************************
1023              * CALCULATE INTERACTIONS *
1024              **************************/
1025
1026             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1027
1028             /* EWALD ELECTROSTATICS */
1029
1030             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1031             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1032             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1033             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1034             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1035
1036             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1037             ewtabD           = _fjsp_setzero_v2r8();
1038             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1039             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1040             ewtabFn          = _fjsp_setzero_v2r8();
1041             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1042             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1043             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1044             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
1045             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1046
1047             /* Update potential sum for this i atom from the interaction with this j atom. */
1048             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1049             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1050
1051             fscal            = felec;
1052
1053             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1054
1055             /* Update vectorial force */
1056             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1057             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1058             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1059
1060             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1061             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1062             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1063
1064             /**************************
1065              * CALCULATE INTERACTIONS *
1066              **************************/
1067
1068             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1069
1070             /* EWALD ELECTROSTATICS */
1071
1072             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1073             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1074             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1075             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1076             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1077
1078             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1079             ewtabD           = _fjsp_setzero_v2r8();
1080             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1081             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1082             ewtabFn          = _fjsp_setzero_v2r8();
1083             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1084             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1085             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1086             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
1087             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1088
1089             /* Update potential sum for this i atom from the interaction with this j atom. */
1090             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1091             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1092
1093             fscal            = felec;
1094
1095             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1096
1097             /* Update vectorial force */
1098             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1099             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1100             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1101
1102             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1103             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1104             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1105
1106             /**************************
1107              * CALCULATE INTERACTIONS *
1108              **************************/
1109
1110             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1111
1112             /* EWALD ELECTROSTATICS */
1113
1114             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1115             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1116             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1117             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1118             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1119
1120             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1121             ewtabD           = _fjsp_setzero_v2r8();
1122             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1123             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1124             ewtabFn          = _fjsp_setzero_v2r8();
1125             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1126             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1127             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1128             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
1129             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1130
1131             /* Update potential sum for this i atom from the interaction with this j atom. */
1132             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1133             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1134
1135             fscal            = felec;
1136
1137             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1138
1139             /* Update vectorial force */
1140             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1141             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1142             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1143
1144             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1145             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1146             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1147
1148             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1149
1150             /* Inner loop uses 420 flops */
1151         }
1152
1153         /* End of innermost loop */
1154
1155         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1156                                               f+i_coord_offset,fshift+i_shift_offset);
1157
1158         ggid                        = gid[iidx];
1159         /* Update potential energies */
1160         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1161         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
1162
1163         /* Increment number of inner iterations */
1164         inneriter                  += j_index_end - j_index_start;
1165
1166         /* Outer loop uses 20 flops */
1167     }
1168
1169     /* Increment number of outer iterations */
1170     outeriter        += nri;
1171
1172     /* Update outer/inner flops */
1173
1174     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*420);
1175 }
1176 /*
1177  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomW3W3_F_sparc64_hpc_ace_double
1178  * Electrostatics interaction: Ewald
1179  * VdW interaction:            LJEwald
1180  * Geometry:                   Water3-Water3
1181  * Calculate force/pot:        Force
1182  */
1183 void
1184 nb_kernel_ElecEw_VdwLJEw_GeomW3W3_F_sparc64_hpc_ace_double
1185                     (t_nblist                    * gmx_restrict       nlist,
1186                      rvec                        * gmx_restrict          xx,
1187                      rvec                        * gmx_restrict          ff,
1188                      t_forcerec                  * gmx_restrict          fr,
1189                      t_mdatoms                   * gmx_restrict     mdatoms,
1190                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1191                      t_nrnb                      * gmx_restrict        nrnb)
1192 {
1193     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1194      * just 0 for non-waters.
1195      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1196      * jnr indices corresponding to data put in the four positions in the SIMD register.
1197      */
1198     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1199     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1200     int              jnrA,jnrB;
1201     int              j_coord_offsetA,j_coord_offsetB;
1202     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1203     real             rcutoff_scalar;
1204     real             *shiftvec,*fshift,*x,*f;
1205     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1206     int              vdwioffset0;
1207     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1208     int              vdwioffset1;
1209     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1210     int              vdwioffset2;
1211     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1212     int              vdwjidx0A,vdwjidx0B;
1213     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1214     int              vdwjidx1A,vdwjidx1B;
1215     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1216     int              vdwjidx2A,vdwjidx2B;
1217     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1218     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1219     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1220     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1221     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1222     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1223     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1224     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1225     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1226     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1227     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1228     real             *charge;
1229     int              nvdwtype;
1230     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1231     int              *vdwtype;
1232     real             *vdwparam;
1233     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
1234     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
1235     _fjsp_v2r8           c6grid_00;
1236     _fjsp_v2r8           c6grid_01;
1237     _fjsp_v2r8           c6grid_02;
1238     _fjsp_v2r8           c6grid_10;
1239     _fjsp_v2r8           c6grid_11;
1240     _fjsp_v2r8           c6grid_12;
1241     _fjsp_v2r8           c6grid_20;
1242     _fjsp_v2r8           c6grid_21;
1243     _fjsp_v2r8           c6grid_22;
1244     real                 *vdwgridparam;
1245     _fjsp_v2r8           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
1246     _fjsp_v2r8           one_half = gmx_fjsp_set1_v2r8(0.5);
1247     _fjsp_v2r8           minus_one = gmx_fjsp_set1_v2r8(-1.0);
1248     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1249     real             *ewtab;
1250     _fjsp_v2r8       itab_tmp;
1251     _fjsp_v2r8       dummy_mask,cutoff_mask;
1252     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1253     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1254     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1255
1256     x                = xx[0];
1257     f                = ff[0];
1258
1259     nri              = nlist->nri;
1260     iinr             = nlist->iinr;
1261     jindex           = nlist->jindex;
1262     jjnr             = nlist->jjnr;
1263     shiftidx         = nlist->shift;
1264     gid              = nlist->gid;
1265     shiftvec         = fr->shift_vec[0];
1266     fshift           = fr->fshift[0];
1267     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1268     charge           = mdatoms->chargeA;
1269     nvdwtype         = fr->ntype;
1270     vdwparam         = fr->nbfp;
1271     vdwtype          = mdatoms->typeA;
1272     vdwgridparam     = fr->ljpme_c6grid;
1273     sh_lj_ewald      = gmx_fjsp_set1_v2r8(fr->ic->sh_lj_ewald);
1274     ewclj            = gmx_fjsp_set1_v2r8(fr->ewaldcoeff_lj);
1275     ewclj2           = _fjsp_mul_v2r8(minus_one,_fjsp_mul_v2r8(ewclj,ewclj));
1276
1277     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
1278     ewtab            = fr->ic->tabq_coul_F;
1279     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
1280     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
1281
1282     /* Setup water-specific parameters */
1283     inr              = nlist->iinr[0];
1284     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1285     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1286     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1287     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1288
1289     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1290     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1291     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1292     vdwjidx0A        = 2*vdwtype[inr+0];
1293     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1294     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
1295     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
1296     c6grid_00        = gmx_fjsp_set1_v2r8(vdwgridparam[vdwioffset0+vdwjidx0A]);
1297     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1298     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1299     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1300     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1301     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1302     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1303     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1304     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1305
1306     /* Avoid stupid compiler warnings */
1307     jnrA = jnrB = 0;
1308     j_coord_offsetA = 0;
1309     j_coord_offsetB = 0;
1310
1311     outeriter        = 0;
1312     inneriter        = 0;
1313
1314     /* Start outer loop over neighborlists */
1315     for(iidx=0; iidx<nri; iidx++)
1316     {
1317         /* Load shift vector for this list */
1318         i_shift_offset   = DIM*shiftidx[iidx];
1319
1320         /* Load limits for loop over neighbors */
1321         j_index_start    = jindex[iidx];
1322         j_index_end      = jindex[iidx+1];
1323
1324         /* Get outer coordinate index */
1325         inr              = iinr[iidx];
1326         i_coord_offset   = DIM*inr;
1327
1328         /* Load i particle coords and add shift vector */
1329         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1330                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1331
1332         fix0             = _fjsp_setzero_v2r8();
1333         fiy0             = _fjsp_setzero_v2r8();
1334         fiz0             = _fjsp_setzero_v2r8();
1335         fix1             = _fjsp_setzero_v2r8();
1336         fiy1             = _fjsp_setzero_v2r8();
1337         fiz1             = _fjsp_setzero_v2r8();
1338         fix2             = _fjsp_setzero_v2r8();
1339         fiy2             = _fjsp_setzero_v2r8();
1340         fiz2             = _fjsp_setzero_v2r8();
1341
1342         /* Start inner kernel loop */
1343         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1344         {
1345
1346             /* Get j neighbor index, and coordinate index */
1347             jnrA             = jjnr[jidx];
1348             jnrB             = jjnr[jidx+1];
1349             j_coord_offsetA  = DIM*jnrA;
1350             j_coord_offsetB  = DIM*jnrB;
1351
1352             /* load j atom coordinates */
1353             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1354                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1355
1356             /* Calculate displacement vector */
1357             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1358             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1359             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1360             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1361             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1362             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1363             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1364             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1365             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1366             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1367             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1368             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1369             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1370             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1371             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1372             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1373             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1374             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1375             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1376             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1377             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1378             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1379             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1380             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1381             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1382             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1383             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1384
1385             /* Calculate squared distance and things based on it */
1386             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1387             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1388             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1389             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1390             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1391             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1392             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1393             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1394             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1395
1396             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1397             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1398             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1399             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1400             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1401             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1402             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1403             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1404             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1405
1406             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1407             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1408             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1409             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1410             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1411             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1412             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1413             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1414             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1415
1416             fjx0             = _fjsp_setzero_v2r8();
1417             fjy0             = _fjsp_setzero_v2r8();
1418             fjz0             = _fjsp_setzero_v2r8();
1419             fjx1             = _fjsp_setzero_v2r8();
1420             fjy1             = _fjsp_setzero_v2r8();
1421             fjz1             = _fjsp_setzero_v2r8();
1422             fjx2             = _fjsp_setzero_v2r8();
1423             fjy2             = _fjsp_setzero_v2r8();
1424             fjz2             = _fjsp_setzero_v2r8();
1425
1426             /**************************
1427              * CALCULATE INTERACTIONS *
1428              **************************/
1429
1430             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1431
1432             /* EWALD ELECTROSTATICS */
1433
1434             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1435             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
1436             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1437             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1438             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1439
1440             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1441                                          &ewtabF,&ewtabFn);
1442             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1443             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
1444
1445             /* Analytical LJ-PME */
1446             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1447             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
1448             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
1449             exponent         = gmx_simd_exp_d(-ewcljrsq);
1450             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1451             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
1452             /* f6A = 6 * C6grid * (1 - poly) */
1453             f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_msub_v2r8(one,poly));
1454             /* f6B = C6grid * exponent * beta^6 */
1455             f6B              = _fjsp_mul_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6));
1456             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1457             fvdw              = _fjsp_mul_v2r8(_fjsp_madd_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,_fjsp_sub_v2r8(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1458
1459             fscal            = _fjsp_add_v2r8(felec,fvdw);
1460
1461             /* Update vectorial force */
1462             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1463             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1464             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1465
1466             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1467             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1468             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1469
1470             /**************************
1471              * CALCULATE INTERACTIONS *
1472              **************************/
1473
1474             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1475
1476             /* EWALD ELECTROSTATICS */
1477
1478             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1479             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1480             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1481             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1482             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1483
1484             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1485                                          &ewtabF,&ewtabFn);
1486             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1487             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1488
1489             fscal            = felec;
1490
1491             /* Update vectorial force */
1492             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1493             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1494             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1495
1496             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1497             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1498             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1499
1500             /**************************
1501              * CALCULATE INTERACTIONS *
1502              **************************/
1503
1504             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1505
1506             /* EWALD ELECTROSTATICS */
1507
1508             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1509             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1510             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1511             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1512             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1513
1514             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1515                                          &ewtabF,&ewtabFn);
1516             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1517             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1518
1519             fscal            = felec;
1520
1521             /* Update vectorial force */
1522             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1523             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1524             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1525
1526             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1527             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1528             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1529
1530             /**************************
1531              * CALCULATE INTERACTIONS *
1532              **************************/
1533
1534             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1535
1536             /* EWALD ELECTROSTATICS */
1537
1538             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1539             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1540             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1541             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1542             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1543
1544             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1545                                          &ewtabF,&ewtabFn);
1546             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1547             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1548
1549             fscal            = felec;
1550
1551             /* Update vectorial force */
1552             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1553             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1554             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1555
1556             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1557             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1558             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1559
1560             /**************************
1561              * CALCULATE INTERACTIONS *
1562              **************************/
1563
1564             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1565
1566             /* EWALD ELECTROSTATICS */
1567
1568             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1569             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1570             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1571             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1572             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1573
1574             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1575                                          &ewtabF,&ewtabFn);
1576             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1577             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1578
1579             fscal            = felec;
1580
1581             /* Update vectorial force */
1582             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1583             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1584             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1585
1586             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1587             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1588             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1595
1596             /* EWALD ELECTROSTATICS */
1597
1598             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1599             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1600             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1601             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1602             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1603
1604             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1605                                          &ewtabF,&ewtabFn);
1606             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1607             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1608
1609             fscal            = felec;
1610
1611             /* Update vectorial force */
1612             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1613             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1614             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1615
1616             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1617             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1618             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1619
1620             /**************************
1621              * CALCULATE INTERACTIONS *
1622              **************************/
1623
1624             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1625
1626             /* EWALD ELECTROSTATICS */
1627
1628             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1629             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1630             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1631             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1632             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1633
1634             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1635                                          &ewtabF,&ewtabFn);
1636             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1637             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1638
1639             fscal            = felec;
1640
1641             /* Update vectorial force */
1642             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1643             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1644             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1645
1646             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1647             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1648             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1649
1650             /**************************
1651              * CALCULATE INTERACTIONS *
1652              **************************/
1653
1654             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1655
1656             /* EWALD ELECTROSTATICS */
1657
1658             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1659             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1660             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1661             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1662             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1663
1664             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1665                                          &ewtabF,&ewtabFn);
1666             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1667             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1668
1669             fscal            = felec;
1670
1671             /* Update vectorial force */
1672             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1673             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1674             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1675
1676             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1677             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1678             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1685
1686             /* EWALD ELECTROSTATICS */
1687
1688             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1689             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1690             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1691             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1692             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1693
1694             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1695                                          &ewtabF,&ewtabFn);
1696             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1697             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1698
1699             fscal            = felec;
1700
1701             /* Update vectorial force */
1702             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1703             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1704             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1705
1706             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1707             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1708             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1709
1710             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1711
1712             /* Inner loop uses 373 flops */
1713         }
1714
1715         if(jidx<j_index_end)
1716         {
1717
1718             jnrA             = jjnr[jidx];
1719             j_coord_offsetA  = DIM*jnrA;
1720
1721             /* load j atom coordinates */
1722             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1723                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1724
1725             /* Calculate displacement vector */
1726             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1727             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1728             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1729             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1730             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1731             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1732             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1733             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1734             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1735             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1736             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1737             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1738             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1739             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1740             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1741             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1742             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1743             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1744             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1745             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1746             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1747             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1748             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1749             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1750             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1751             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1752             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1753
1754             /* Calculate squared distance and things based on it */
1755             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1756             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1757             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1758             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1759             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1760             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1761             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1762             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1763             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1764
1765             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1766             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1767             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1768             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1769             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1770             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1771             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1772             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1773             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1774
1775             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1776             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1777             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1778             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1779             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1780             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1781             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1782             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1783             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1784
1785             fjx0             = _fjsp_setzero_v2r8();
1786             fjy0             = _fjsp_setzero_v2r8();
1787             fjz0             = _fjsp_setzero_v2r8();
1788             fjx1             = _fjsp_setzero_v2r8();
1789             fjy1             = _fjsp_setzero_v2r8();
1790             fjz1             = _fjsp_setzero_v2r8();
1791             fjx2             = _fjsp_setzero_v2r8();
1792             fjy2             = _fjsp_setzero_v2r8();
1793             fjz2             = _fjsp_setzero_v2r8();
1794
1795             /**************************
1796              * CALCULATE INTERACTIONS *
1797              **************************/
1798
1799             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1800
1801             /* EWALD ELECTROSTATICS */
1802
1803             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1804             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
1805             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1806             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1807             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1808
1809             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1810             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1811             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
1812
1813             /* Analytical LJ-PME */
1814             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1815             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
1816             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
1817             exponent         = gmx_simd_exp_d(-ewcljrsq);
1818             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1819             poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
1820             /* f6A = 6 * C6grid * (1 - poly) */
1821             f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_msub_v2r8(one,poly));
1822             /* f6B = C6grid * exponent * beta^6 */
1823             f6B              = _fjsp_mul_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6));
1824             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1825             fvdw              = _fjsp_mul_v2r8(_fjsp_madd_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,_fjsp_sub_v2r8(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1826
1827             fscal            = _fjsp_add_v2r8(felec,fvdw);
1828
1829             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1830
1831             /* Update vectorial force */
1832             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1833             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1834             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1835
1836             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1837             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1838             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1839
1840             /**************************
1841              * CALCULATE INTERACTIONS *
1842              **************************/
1843
1844             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1845
1846             /* EWALD ELECTROSTATICS */
1847
1848             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1849             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1850             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1851             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1852             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1853
1854             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1855             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1856             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1857
1858             fscal            = felec;
1859
1860             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1861
1862             /* Update vectorial force */
1863             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1864             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1865             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1866
1867             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1868             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1869             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1870
1871             /**************************
1872              * CALCULATE INTERACTIONS *
1873              **************************/
1874
1875             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1876
1877             /* EWALD ELECTROSTATICS */
1878
1879             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1880             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1881             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1882             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1883             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1884
1885             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1886             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1887             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1888
1889             fscal            = felec;
1890
1891             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1892
1893             /* Update vectorial force */
1894             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1895             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1896             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1897
1898             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1899             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1900             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1901
1902             /**************************
1903              * CALCULATE INTERACTIONS *
1904              **************************/
1905
1906             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1907
1908             /* EWALD ELECTROSTATICS */
1909
1910             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1911             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1912             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1913             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1914             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1915
1916             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1917             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1918             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1919
1920             fscal            = felec;
1921
1922             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1923
1924             /* Update vectorial force */
1925             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1926             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1927             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1928
1929             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1930             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1931             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1932
1933             /**************************
1934              * CALCULATE INTERACTIONS *
1935              **************************/
1936
1937             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1938
1939             /* EWALD ELECTROSTATICS */
1940
1941             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1942             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1943             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1944             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1945             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1946
1947             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1948             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1949             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1950
1951             fscal            = felec;
1952
1953             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1954
1955             /* Update vectorial force */
1956             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1957             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1958             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1959
1960             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1961             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1962             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1963
1964             /**************************
1965              * CALCULATE INTERACTIONS *
1966              **************************/
1967
1968             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1969
1970             /* EWALD ELECTROSTATICS */
1971
1972             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1973             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1974             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1975             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1976             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1977
1978             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1979             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1980             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1981
1982             fscal            = felec;
1983
1984             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1985
1986             /* Update vectorial force */
1987             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1988             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1989             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1990
1991             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1992             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1993             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1994
1995             /**************************
1996              * CALCULATE INTERACTIONS *
1997              **************************/
1998
1999             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
2000
2001             /* EWALD ELECTROSTATICS */
2002
2003             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2004             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
2005             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2006             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2007             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2008
2009             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2010             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2011             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
2012
2013             fscal            = felec;
2014
2015             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2016
2017             /* Update vectorial force */
2018             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
2019             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
2020             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
2021
2022             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
2023             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
2024             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
2025
2026             /**************************
2027              * CALCULATE INTERACTIONS *
2028              **************************/
2029
2030             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2031
2032             /* EWALD ELECTROSTATICS */
2033
2034             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2035             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2036             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2037             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2038             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2039
2040             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2041             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2042             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2043
2044             fscal            = felec;
2045
2046             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2047
2048             /* Update vectorial force */
2049             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2050             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2051             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2052
2053             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2054             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2055             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2056
2057             /**************************
2058              * CALCULATE INTERACTIONS *
2059              **************************/
2060
2061             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2062
2063             /* EWALD ELECTROSTATICS */
2064
2065             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2066             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2067             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2068             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2069             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2070
2071             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2072             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2073             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2074
2075             fscal            = felec;
2076
2077             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2078
2079             /* Update vectorial force */
2080             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2081             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2082             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2083
2084             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2085             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2086             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2087
2088             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2089
2090             /* Inner loop uses 373 flops */
2091         }
2092
2093         /* End of innermost loop */
2094
2095         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2096                                               f+i_coord_offset,fshift+i_shift_offset);
2097
2098         /* Increment number of inner iterations */
2099         inneriter                  += j_index_end - j_index_start;
2100
2101         /* Outer loop uses 18 flops */
2102     }
2103
2104     /* Increment number of outer iterations */
2105     outeriter        += nri;
2106
2107     /* Update outer/inner flops */
2108
2109     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*373);
2110 }