src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEwSw_VdwNone_GeomW3W3_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomW3W3_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: Ewald
  54  * VdW interaction:            None
  55  * Geometry:                   Water3-Water3
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecEwSw_VdwNone_GeomW3W3_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  94     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  95     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  96     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  97     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 100     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 103     real             *charge;
 104     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 105     real             *ewtab;
 106     _fjsp_v2r8       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 107     real             rswitch_scalar,d_scalar;
 108     _fjsp_v2r8       itab_tmp;
 109     _fjsp_v2r8       dummy_mask,cutoff_mask;
 110     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 111     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 112     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 113
 114     x                = xx[0];
 115     f                = ff[0];
 116
 117     nri              = nlist->nri;
 118     iinr             = nlist->iinr;
 119     jindex           = nlist->jindex;
 120     jjnr             = nlist->jjnr;
 121     shiftidx         = nlist->shift;
 122     gid              = nlist->gid;
 123     shiftvec         = fr->shift_vec[0];
 124     fshift           = fr->fshift[0];
 125     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 126     charge           = mdatoms->chargeA;
 127
 128     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 129     ewtab            = fr->ic->tabq_coul_FDV0;
 130     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 131     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 132
 133     /* Setup water-specific parameters */
 134     inr              = nlist->iinr[0];
 135     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 136     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 137     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 138
 139     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 140     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 141     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 142     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 143     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 144     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 145     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 146     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 147     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 148     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 149     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 150     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 151
 152     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 153     rcutoff_scalar   = fr->rcoulomb;
 154     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 155     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 156
 157     rswitch_scalar   = fr->rcoulomb_switch;
 158     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
 159     /* Setup switch parameters */
 160     d_scalar         = rcutoff_scalar-rswitch_scalar;
 161     d                = gmx_fjsp_set1_v2r8(d_scalar);
 162     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
 163     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 164     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 165     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
 166     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 167     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 168
 169     /* Avoid stupid compiler warnings */
 170     jnrA = jnrB = 0;
 171     j_coord_offsetA = 0;
 172     j_coord_offsetB = 0;
 173
 174     outeriter        = 0;
 175     inneriter        = 0;
 176
 177     /* Start outer loop over neighborlists */
 178     for(iidx=0; iidx<nri; iidx++)
 179     {
 180         /* Load shift vector for this list */
 181         i_shift_offset   = DIM*shiftidx[iidx];
 182
 183         /* Load limits for loop over neighbors */
 184         j_index_start    = jindex[iidx];
 185         j_index_end      = jindex[iidx+1];
 186
 187         /* Get outer coordinate index */
 188         inr              = iinr[iidx];
 189         i_coord_offset   = DIM*inr;
 190
 191         /* Load i particle coords and add shift vector */
 192         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 193                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 194
 195         fix0             = _fjsp_setzero_v2r8();
 196         fiy0             = _fjsp_setzero_v2r8();
 197         fiz0             = _fjsp_setzero_v2r8();
 198         fix1             = _fjsp_setzero_v2r8();
 199         fiy1             = _fjsp_setzero_v2r8();
 200         fiz1             = _fjsp_setzero_v2r8();
 201         fix2             = _fjsp_setzero_v2r8();
 202         fiy2             = _fjsp_setzero_v2r8();
 203         fiz2             = _fjsp_setzero_v2r8();
 204
 205         /* Reset potential sums */
 206         velecsum         = _fjsp_setzero_v2r8();
 207
 208         /* Start inner kernel loop */
 209         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 210         {
 211
 212             /* Get j neighbor index, and coordinate index */
 213             jnrA             = jjnr[jidx];
 214             jnrB             = jjnr[jidx+1];
 215             j_coord_offsetA  = DIM*jnrA;
 216             j_coord_offsetB  = DIM*jnrB;
 217
 218             /* load j atom coordinates */
 219             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 220                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 221
 222             /* Calculate displacement vector */
 223             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 224             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 225             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 226             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 227             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 228             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 229             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 230             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 231             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 232             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 233             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 234             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 235             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 236             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 237             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 238             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 239             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 240             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 241             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 242             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 243             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 244             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 245             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 246             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 247             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 248             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 249             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 250
 251             /* Calculate squared distance and things based on it */
 252             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 253             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 254             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 255             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 256             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 257             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 258             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 259             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 260             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 261
 262             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 263             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 264             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 265             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 266             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 267             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 268             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 269             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 270             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 271
 272             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 273             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 274             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 275             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 276             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 277             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 278             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 279             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 280             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 281
 282             fjx0             = _fjsp_setzero_v2r8();
 283             fjy0             = _fjsp_setzero_v2r8();
 284             fjz0             = _fjsp_setzero_v2r8();
 285             fjx1             = _fjsp_setzero_v2r8();
 286             fjy1             = _fjsp_setzero_v2r8();
 287             fjz1             = _fjsp_setzero_v2r8();
 288             fjx2             = _fjsp_setzero_v2r8();
 289             fjy2             = _fjsp_setzero_v2r8();
 290             fjz2             = _fjsp_setzero_v2r8();
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 297             {
 298
 299             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 300
 301             /* EWALD ELECTROSTATICS */
 302
 303             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 304             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 305             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 306             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 307             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 308
 309             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 310             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 311             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 312             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 313             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 314             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 315             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 316             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 317             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 318             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 319
 320             d                = _fjsp_sub_v2r8(r00,rswitch);
 321             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 322             d2               = _fjsp_mul_v2r8(d,d);
 323             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 324
 325             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 326
 327             /* Evaluate switch function */
 328             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 329             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
 330             velec            = _fjsp_mul_v2r8(velec,sw);
 331             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 335             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 336
 337             fscal            = felec;
 338
 339             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 340
 341             /* Update vectorial force */
 342             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 343             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 344             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 345
 346             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 347             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 348             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 349
 350             }
 351
 352             /**************************
 353              * CALCULATE INTERACTIONS *
 354              **************************/
 355
 356             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 357             {
 358
 359             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 360
 361             /* EWALD ELECTROSTATICS */
 362
 363             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 364             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 365             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 366             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 367             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 368
 369             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 370             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 371             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 372             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 373             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 374             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 375             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 376             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 377             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
 378             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
 379
 380             d                = _fjsp_sub_v2r8(r01,rswitch);
 381             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 382             d2               = _fjsp_mul_v2r8(d,d);
 383             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 384
 385             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 386
 387             /* Evaluate switch function */
 388             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 389             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
 390             velec            = _fjsp_mul_v2r8(velec,sw);
 391             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 392
 393             /* Update potential sum for this i atom from the interaction with this j atom. */
 394             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 395             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 396
 397             fscal            = felec;
 398
 399             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 400
 401             /* Update vectorial force */
 402             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 403             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 404             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 405
 406             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 407             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 408             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 409
 410             }
 411
 412             /**************************
 413              * CALCULATE INTERACTIONS *
 414              **************************/
 415
 416             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 417             {
 418
 419             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
 420
 421             /* EWALD ELECTROSTATICS */
 422
 423             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 424             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
 425             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 426             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 427             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 428
 429             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 430             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 431             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 432             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 433             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 434             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 435             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 436             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 437             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
 438             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
 439
 440             d                = _fjsp_sub_v2r8(r02,rswitch);
 441             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 442             d2               = _fjsp_mul_v2r8(d,d);
 443             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 444
 445             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 446
 447             /* Evaluate switch function */
 448             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 449             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
 450             velec            = _fjsp_mul_v2r8(velec,sw);
 451             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 452
 453             /* Update potential sum for this i atom from the interaction with this j atom. */
 454             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 455             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 456
 457             fscal            = felec;
 458
 459             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 460
 461             /* Update vectorial force */
 462             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 463             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 464             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 465
 466             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 467             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 468             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 469
 470             }
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 477             {
 478
 479             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 480
 481             /* EWALD ELECTROSTATICS */
 482
 483             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 484             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 485             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 486             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 487             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 488
 489             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 490             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 491             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 492             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 493             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 494             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 495             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 496             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 497             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 498             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 499
 500             d                = _fjsp_sub_v2r8(r10,rswitch);
 501             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 502             d2               = _fjsp_mul_v2r8(d,d);
 503             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 504
 505             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 506
 507             /* Evaluate switch function */
 508             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 509             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
 510             velec            = _fjsp_mul_v2r8(velec,sw);
 511             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 512
 513             /* Update potential sum for this i atom from the interaction with this j atom. */
 514             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 515             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 516
 517             fscal            = felec;
 518
 519             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 520
 521             /* Update vectorial force */
 522             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 523             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 524             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 525
 526             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 527             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 528             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 529
 530             }
 531
 532             /**************************
 533              * CALCULATE INTERACTIONS *
 534              **************************/
 535
 536             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 537             {
 538
 539             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 540
 541             /* EWALD ELECTROSTATICS */
 542
 543             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 544             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 545             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 546             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 547             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 548
 549             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 550             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 551             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 552             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 553             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 554             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 555             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 556             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 557             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
 558             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 559
 560             d                = _fjsp_sub_v2r8(r11,rswitch);
 561             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 562             d2               = _fjsp_mul_v2r8(d,d);
 563             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 564
 565             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 566
 567             /* Evaluate switch function */
 568             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 569             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
 570             velec            = _fjsp_mul_v2r8(velec,sw);
 571             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 572
 573             /* Update potential sum for this i atom from the interaction with this j atom. */
 574             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 575             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 576
 577             fscal            = felec;
 578
 579             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 580
 581             /* Update vectorial force */
 582             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 583             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 584             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 585
 586             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 587             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 588             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 589
 590             }
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 597             {
 598
 599             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 600
 601             /* EWALD ELECTROSTATICS */
 602
 603             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 604             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 605             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 606             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 607             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 608
 609             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 610             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 611             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 612             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 613             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 614             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 615             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 616             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 617             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
 618             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 619
 620             d                = _fjsp_sub_v2r8(r12,rswitch);
 621             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 622             d2               = _fjsp_mul_v2r8(d,d);
 623             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 624
 625             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 626
 627             /* Evaluate switch function */
 628             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 629             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
 630             velec            = _fjsp_mul_v2r8(velec,sw);
 631             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 632
 633             /* Update potential sum for this i atom from the interaction with this j atom. */
 634             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 635             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 636
 637             fscal            = felec;
 638
 639             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 640
 641             /* Update vectorial force */
 642             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 643             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 644             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 645
 646             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 647             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 648             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 649
 650             }
 651
 652             /**************************
 653              * CALCULATE INTERACTIONS *
 654              **************************/
 655
 656             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 657             {
 658
 659             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 660
 661             /* EWALD ELECTROSTATICS */
 662
 663             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 664             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 665             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 666             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 667             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 668
 669             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 670             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 671             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 672             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 673             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 674             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 675             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 676             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 677             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 678             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 679
 680             d                = _fjsp_sub_v2r8(r20,rswitch);
 681             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 682             d2               = _fjsp_mul_v2r8(d,d);
 683             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 684
 685             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 686
 687             /* Evaluate switch function */
 688             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 689             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
 690             velec            = _fjsp_mul_v2r8(velec,sw);
 691             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 692
 693             /* Update potential sum for this i atom from the interaction with this j atom. */
 694             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 695             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 696
 697             fscal            = felec;
 698
 699             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 700
 701             /* Update vectorial force */
 702             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 703             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 704             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 705
 706             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 707             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 708             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 709
 710             }
 711
 712             /**************************
 713              * CALCULATE INTERACTIONS *
 714              **************************/
 715
 716             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 717             {
 718
 719             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 720
 721             /* EWALD ELECTROSTATICS */
 722
 723             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 724             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 725             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 726             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 727             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 728
 729             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 730             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 731             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 732             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 733             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 734             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 735             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 736             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 737             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
 738             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 739
 740             d                = _fjsp_sub_v2r8(r21,rswitch);
 741             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 742             d2               = _fjsp_mul_v2r8(d,d);
 743             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 744
 745             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 746
 747             /* Evaluate switch function */
 748             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 749             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
 750             velec            = _fjsp_mul_v2r8(velec,sw);
 751             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 752
 753             /* Update potential sum for this i atom from the interaction with this j atom. */
 754             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 755             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 756
 757             fscal            = felec;
 758
 759             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 760
 761             /* Update vectorial force */
 762             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 763             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 764             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 765
 766             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 767             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 768             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 769
 770             }
 771
 772             /**************************
 773              * CALCULATE INTERACTIONS *
 774              **************************/
 775
 776             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 777             {
 778
 779             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 780
 781             /* EWALD ELECTROSTATICS */
 782
 783             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 784             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
 785             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 786             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 787             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 788
 789             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 790             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 791             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 792             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 793             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 794             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 795             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 796             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 797             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
 798             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
 799
 800             d                = _fjsp_sub_v2r8(r22,rswitch);
 801             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 802             d2               = _fjsp_mul_v2r8(d,d);
 803             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 804
 805             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 806
 807             /* Evaluate switch function */
 808             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 809             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
 810             velec            = _fjsp_mul_v2r8(velec,sw);
 811             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 815             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 816
 817             fscal            = felec;
 818
 819             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 820
 821             /* Update vectorial force */
 822             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 823             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 824             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 825
 826             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 827             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 828             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 829
 830             }
 831
 832             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 833
 834             /* Inner loop uses 612 flops */
 835         }
 836
 837         if(jidx<j_index_end)
 838         {
 839
 840             jnrA             = jjnr[jidx];
 841             j_coord_offsetA  = DIM*jnrA;
 842
 843             /* load j atom coordinates */
 844             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 845                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 846
 847             /* Calculate displacement vector */
 848             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 849             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 850             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 851             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 852             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 853             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 854             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 855             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 856             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 857             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 858             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 859             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 860             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 861             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 862             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 863             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 864             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 865             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 866             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 867             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 868             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 869             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 870             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 871             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 872             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 873             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 874             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 875
 876             /* Calculate squared distance and things based on it */
 877             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 878             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 879             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 880             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 881             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 882             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 883             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 884             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 885             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 886
 887             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 888             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 889             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 890             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 891             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 892             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 893             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 894             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 895             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 896
 897             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 898             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 899             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 900             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 901             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 902             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 903             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 904             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 905             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 906
 907             fjx0             = _fjsp_setzero_v2r8();
 908             fjy0             = _fjsp_setzero_v2r8();
 909             fjz0             = _fjsp_setzero_v2r8();
 910             fjx1             = _fjsp_setzero_v2r8();
 911             fjy1             = _fjsp_setzero_v2r8();
 912             fjz1             = _fjsp_setzero_v2r8();
 913             fjx2             = _fjsp_setzero_v2r8();
 914             fjy2             = _fjsp_setzero_v2r8();
 915             fjz2             = _fjsp_setzero_v2r8();
 916
 917             /**************************
 918              * CALCULATE INTERACTIONS *
 919              **************************/
 920
 921             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 922             {
 923
 924             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 925
 926             /* EWALD ELECTROSTATICS */
 927
 928             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 929             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 930             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 931             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 932             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 933
 934             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 935             ewtabD           = _fjsp_setzero_v2r8();
 936             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 937             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 938             ewtabFn          = _fjsp_setzero_v2r8();
 939             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 940             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 941             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 942             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 943             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 944
 945             d                = _fjsp_sub_v2r8(r00,rswitch);
 946             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
 947             d2               = _fjsp_mul_v2r8(d,d);
 948             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
 949
 950             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
 951
 952             /* Evaluate switch function */
 953             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 954             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
 955             velec            = _fjsp_mul_v2r8(velec,sw);
 956             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 957
 958             /* Update potential sum for this i atom from the interaction with this j atom. */
 959             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 960             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 961             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 962
 963             fscal            = felec;
 964
 965             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 966
 967             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 968
 969             /* Update vectorial force */
 970             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 971             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 972             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 973
 974             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 975             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 976             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 977
 978             }
 979
 980             /**************************
 981              * CALCULATE INTERACTIONS *
 982              **************************/
 983
 984             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 985             {
 986
 987             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
 988
 989             /* EWALD ELECTROSTATICS */
 990
 991             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 992             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
 993             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 994             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 995             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 996
 997             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 998             ewtabD           = _fjsp_setzero_v2r8();
 999             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1000             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1001             ewtabFn          = _fjsp_setzero_v2r8();
1002             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1003             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1004             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1005             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
1006             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1007
1008             d                = _fjsp_sub_v2r8(r01,rswitch);
1009             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1010             d2               = _fjsp_mul_v2r8(d,d);
1011             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1012
1013             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1014
1015             /* Evaluate switch function */
1016             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1017             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
1018             velec            = _fjsp_mul_v2r8(velec,sw);
1019             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1020
1021             /* Update potential sum for this i atom from the interaction with this j atom. */
1022             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1023             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1024             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1025
1026             fscal            = felec;
1027
1028             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1029
1030             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1031
1032             /* Update vectorial force */
1033             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1034             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1035             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1036
1037             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1038             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1039             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1040
1041             }
1042
1043             /**************************
1044              * CALCULATE INTERACTIONS *
1045              **************************/
1046
1047             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1048             {
1049
1050             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1051
1052             /* EWALD ELECTROSTATICS */
1053
1054             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1055             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1056             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1057             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1058             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1059
1060             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1061             ewtabD           = _fjsp_setzero_v2r8();
1062             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1063             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1064             ewtabFn          = _fjsp_setzero_v2r8();
1065             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1066             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1067             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1068             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
1069             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1070
1071             d                = _fjsp_sub_v2r8(r02,rswitch);
1072             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1073             d2               = _fjsp_mul_v2r8(d,d);
1074             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1075
1076             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1077
1078             /* Evaluate switch function */
1079             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1080             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
1081             velec            = _fjsp_mul_v2r8(velec,sw);
1082             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1083
1084             /* Update potential sum for this i atom from the interaction with this j atom. */
1085             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1086             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1087             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1088
1089             fscal            = felec;
1090
1091             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1092
1093             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1094
1095             /* Update vectorial force */
1096             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1097             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1098             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1099
1100             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1101             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1102             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1103
1104             }
1105
1106             /**************************
1107              * CALCULATE INTERACTIONS *
1108              **************************/
1109
1110             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1111             {
1112
1113             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1114
1115             /* EWALD ELECTROSTATICS */
1116
1117             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1118             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1119             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1120             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1121             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1122
1123             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1124             ewtabD           = _fjsp_setzero_v2r8();
1125             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1126             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1127             ewtabFn          = _fjsp_setzero_v2r8();
1128             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1129             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1130             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1131             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
1132             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1133
1134             d                = _fjsp_sub_v2r8(r10,rswitch);
1135             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1136             d2               = _fjsp_mul_v2r8(d,d);
1137             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1138
1139             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1140
1141             /* Evaluate switch function */
1142             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1143             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
1144             velec            = _fjsp_mul_v2r8(velec,sw);
1145             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1146
1147             /* Update potential sum for this i atom from the interaction with this j atom. */
1148             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1149             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1150             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1151
1152             fscal            = felec;
1153
1154             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1155
1156             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1157
1158             /* Update vectorial force */
1159             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1160             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1161             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1162
1163             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1164             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1165             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1166
1167             }
1168
1169             /**************************
1170              * CALCULATE INTERACTIONS *
1171              **************************/
1172
1173             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1174             {
1175
1176             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1177
1178             /* EWALD ELECTROSTATICS */
1179
1180             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1181             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1182             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1183             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1184             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1185
1186             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1187             ewtabD           = _fjsp_setzero_v2r8();
1188             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1189             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1190             ewtabFn          = _fjsp_setzero_v2r8();
1191             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1192             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1193             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1194             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
1195             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1196
1197             d                = _fjsp_sub_v2r8(r11,rswitch);
1198             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1199             d2               = _fjsp_mul_v2r8(d,d);
1200             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1201
1202             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1203
1204             /* Evaluate switch function */
1205             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1206             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
1207             velec            = _fjsp_mul_v2r8(velec,sw);
1208             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1209
1210             /* Update potential sum for this i atom from the interaction with this j atom. */
1211             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1212             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1213             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1214
1215             fscal            = felec;
1216
1217             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1218
1219             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1220
1221             /* Update vectorial force */
1222             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1223             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1224             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1225
1226             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1227             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1228             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1229
1230             }
1231
1232             /**************************
1233              * CALCULATE INTERACTIONS *
1234              **************************/
1235
1236             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1237             {
1238
1239             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1240
1241             /* EWALD ELECTROSTATICS */
1242
1243             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1244             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1245             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1246             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1247             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1248
1249             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1250             ewtabD           = _fjsp_setzero_v2r8();
1251             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1252             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1253             ewtabFn          = _fjsp_setzero_v2r8();
1254             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1255             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1256             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1257             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
1258             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1259
1260             d                = _fjsp_sub_v2r8(r12,rswitch);
1261             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1262             d2               = _fjsp_mul_v2r8(d,d);
1263             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1264
1265             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1266
1267             /* Evaluate switch function */
1268             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1269             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
1270             velec            = _fjsp_mul_v2r8(velec,sw);
1271             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1272
1273             /* Update potential sum for this i atom from the interaction with this j atom. */
1274             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1275             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1276             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1277
1278             fscal            = felec;
1279
1280             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1281
1282             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1283
1284             /* Update vectorial force */
1285             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1286             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1287             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1288
1289             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1290             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1291             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1292
1293             }
1294
1295             /**************************
1296              * CALCULATE INTERACTIONS *
1297              **************************/
1298
1299             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1300             {
1301
1302             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1303
1304             /* EWALD ELECTROSTATICS */
1305
1306             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1307             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1308             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1309             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1310             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1311
1312             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1313             ewtabD           = _fjsp_setzero_v2r8();
1314             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1315             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1316             ewtabFn          = _fjsp_setzero_v2r8();
1317             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1318             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1319             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1320             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
1321             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1322
1323             d                = _fjsp_sub_v2r8(r20,rswitch);
1324             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1325             d2               = _fjsp_mul_v2r8(d,d);
1326             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1327
1328             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1329
1330             /* Evaluate switch function */
1331             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1332             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
1333             velec            = _fjsp_mul_v2r8(velec,sw);
1334             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1335
1336             /* Update potential sum for this i atom from the interaction with this j atom. */
1337             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1338             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1339             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1340
1341             fscal            = felec;
1342
1343             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1344
1345             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1346
1347             /* Update vectorial force */
1348             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1349             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1350             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1351
1352             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1353             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1354             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1355
1356             }
1357
1358             /**************************
1359              * CALCULATE INTERACTIONS *
1360              **************************/
1361
1362             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1363             {
1364
1365             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1366
1367             /* EWALD ELECTROSTATICS */
1368
1369             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1370             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1371             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1372             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1373             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1374
1375             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1376             ewtabD           = _fjsp_setzero_v2r8();
1377             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1378             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1379             ewtabFn          = _fjsp_setzero_v2r8();
1380             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1381             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1382             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1383             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
1384             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1385
1386             d                = _fjsp_sub_v2r8(r21,rswitch);
1387             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1388             d2               = _fjsp_mul_v2r8(d,d);
1389             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1390
1391             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1392
1393             /* Evaluate switch function */
1394             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1395             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
1396             velec            = _fjsp_mul_v2r8(velec,sw);
1397             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1398
1399             /* Update potential sum for this i atom from the interaction with this j atom. */
1400             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1401             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1402             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1403
1404             fscal            = felec;
1405
1406             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1407
1408             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1409
1410             /* Update vectorial force */
1411             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1412             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1413             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1414
1415             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1416             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1417             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1418
1419             }
1420
1421             /**************************
1422              * CALCULATE INTERACTIONS *
1423              **************************/
1424
1425             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1426             {
1427
1428             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1429
1430             /* EWALD ELECTROSTATICS */
1431
1432             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1433             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1434             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1435             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1436             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1437
1438             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1439             ewtabD           = _fjsp_setzero_v2r8();
1440             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1441             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1442             ewtabFn          = _fjsp_setzero_v2r8();
1443             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1444             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1445             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1446             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
1447             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1448
1449             d                = _fjsp_sub_v2r8(r22,rswitch);
1450             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1451             d2               = _fjsp_mul_v2r8(d,d);
1452             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1453
1454             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1455
1456             /* Evaluate switch function */
1457             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1458             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
1459             velec            = _fjsp_mul_v2r8(velec,sw);
1460             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1461
1462             /* Update potential sum for this i atom from the interaction with this j atom. */
1463             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1464             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1465             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1466
1467             fscal            = felec;
1468
1469             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1470
1471             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1472
1473             /* Update vectorial force */
1474             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1475             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1476             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1477
1478             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1479             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1480             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1481
1482             }
1483
1484             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1485
1486             /* Inner loop uses 612 flops */
1487         }
1488
1489         /* End of innermost loop */
1490
1491         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1492                                               f+i_coord_offset,fshift+i_shift_offset);
1493
1494         ggid                        = gid[iidx];
1495         /* Update potential energies */
1496         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1497
1498         /* Increment number of inner iterations */
1499         inneriter                  += j_index_end - j_index_start;
1500
1501         /* Outer loop uses 19 flops */
1502     }
1503
1504     /* Increment number of outer iterations */
1505     outeriter        += nri;
1506
1507     /* Update outer/inner flops */
1508
1509     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*612);
1510 }
1511 /*
1512  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomW3W3_F_sparc64_hpc_ace_double
1513  * Electrostatics interaction: Ewald
1514  * VdW interaction:            None
1515  * Geometry:                   Water3-Water3
1516  * Calculate force/pot:        Force
1517  */
1518 void
1519 nb_kernel_ElecEwSw_VdwNone_GeomW3W3_F_sparc64_hpc_ace_double
1520                     (t_nblist                    * gmx_restrict       nlist,
1521                      rvec                        * gmx_restrict          xx,
1522                      rvec                        * gmx_restrict          ff,
1523                      t_forcerec                  * gmx_restrict          fr,
1524                      t_mdatoms                   * gmx_restrict     mdatoms,
1525                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1526                      t_nrnb                      * gmx_restrict        nrnb)
1527 {
1528     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1529      * just 0 for non-waters.
1530      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1531      * jnr indices corresponding to data put in the four positions in the SIMD register.
1532      */
1533     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1534     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1535     int              jnrA,jnrB;
1536     int              j_coord_offsetA,j_coord_offsetB;
1537     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1538     real             rcutoff_scalar;
1539     real             *shiftvec,*fshift,*x,*f;
1540     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1541     int              vdwioffset0;
1542     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1543     int              vdwioffset1;
1544     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1545     int              vdwioffset2;
1546     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1547     int              vdwjidx0A,vdwjidx0B;
1548     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1549     int              vdwjidx1A,vdwjidx1B;
1550     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1551     int              vdwjidx2A,vdwjidx2B;
1552     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1553     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1554     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1555     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1556     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1557     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1558     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1559     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1560     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1561     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1562     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1563     real             *charge;
1564     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1565     real             *ewtab;
1566     _fjsp_v2r8       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1567     real             rswitch_scalar,d_scalar;
1568     _fjsp_v2r8       itab_tmp;
1569     _fjsp_v2r8       dummy_mask,cutoff_mask;
1570     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1571     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1572     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1573
1574     x                = xx[0];
1575     f                = ff[0];
1576
1577     nri              = nlist->nri;
1578     iinr             = nlist->iinr;
1579     jindex           = nlist->jindex;
1580     jjnr             = nlist->jjnr;
1581     shiftidx         = nlist->shift;
1582     gid              = nlist->gid;
1583     shiftvec         = fr->shift_vec[0];
1584     fshift           = fr->fshift[0];
1585     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1586     charge           = mdatoms->chargeA;
1587
1588     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
1589     ewtab            = fr->ic->tabq_coul_FDV0;
1590     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
1591     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
1592
1593     /* Setup water-specific parameters */
1594     inr              = nlist->iinr[0];
1595     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1596     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1597     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1598
1599     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1600     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1601     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1602     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1603     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1604     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1605     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1606     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1607     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1608     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1609     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1610     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1611
1612     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1613     rcutoff_scalar   = fr->rcoulomb;
1614     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1615     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1616
1617     rswitch_scalar   = fr->rcoulomb_switch;
1618     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
1619     /* Setup switch parameters */
1620     d_scalar         = rcutoff_scalar-rswitch_scalar;
1621     d                = gmx_fjsp_set1_v2r8(d_scalar);
1622     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
1623     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1624     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1625     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
1626     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1627     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1628
1629     /* Avoid stupid compiler warnings */
1630     jnrA = jnrB = 0;
1631     j_coord_offsetA = 0;
1632     j_coord_offsetB = 0;
1633
1634     outeriter        = 0;
1635     inneriter        = 0;
1636
1637     /* Start outer loop over neighborlists */
1638     for(iidx=0; iidx<nri; iidx++)
1639     {
1640         /* Load shift vector for this list */
1641         i_shift_offset   = DIM*shiftidx[iidx];
1642
1643         /* Load limits for loop over neighbors */
1644         j_index_start    = jindex[iidx];
1645         j_index_end      = jindex[iidx+1];
1646
1647         /* Get outer coordinate index */
1648         inr              = iinr[iidx];
1649         i_coord_offset   = DIM*inr;
1650
1651         /* Load i particle coords and add shift vector */
1652         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1653                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1654
1655         fix0             = _fjsp_setzero_v2r8();
1656         fiy0             = _fjsp_setzero_v2r8();
1657         fiz0             = _fjsp_setzero_v2r8();
1658         fix1             = _fjsp_setzero_v2r8();
1659         fiy1             = _fjsp_setzero_v2r8();
1660         fiz1             = _fjsp_setzero_v2r8();
1661         fix2             = _fjsp_setzero_v2r8();
1662         fiy2             = _fjsp_setzero_v2r8();
1663         fiz2             = _fjsp_setzero_v2r8();
1664
1665         /* Start inner kernel loop */
1666         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1667         {
1668
1669             /* Get j neighbor index, and coordinate index */
1670             jnrA             = jjnr[jidx];
1671             jnrB             = jjnr[jidx+1];
1672             j_coord_offsetA  = DIM*jnrA;
1673             j_coord_offsetB  = DIM*jnrB;
1674
1675             /* load j atom coordinates */
1676             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1677                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1678
1679             /* Calculate displacement vector */
1680             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1681             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1682             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1683             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1684             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1685             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1686             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1687             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1688             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1689             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1690             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1691             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1692             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1693             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1694             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1695             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1696             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1697             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1698             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1699             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1700             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1701             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1702             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1703             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1704             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1705             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1706             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1707
1708             /* Calculate squared distance and things based on it */
1709             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1710             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1711             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1712             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1713             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1714             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1715             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1716             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1717             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1718
1719             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1720             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1721             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1722             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1723             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1724             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1725             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1726             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1727             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1728
1729             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1730             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1731             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1732             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1733             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1734             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1735             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1736             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1737             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1738
1739             fjx0             = _fjsp_setzero_v2r8();
1740             fjy0             = _fjsp_setzero_v2r8();
1741             fjz0             = _fjsp_setzero_v2r8();
1742             fjx1             = _fjsp_setzero_v2r8();
1743             fjy1             = _fjsp_setzero_v2r8();
1744             fjz1             = _fjsp_setzero_v2r8();
1745             fjx2             = _fjsp_setzero_v2r8();
1746             fjy2             = _fjsp_setzero_v2r8();
1747             fjz2             = _fjsp_setzero_v2r8();
1748
1749             /**************************
1750              * CALCULATE INTERACTIONS *
1751              **************************/
1752
1753             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1754             {
1755
1756             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1757
1758             /* EWALD ELECTROSTATICS */
1759
1760             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1761             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
1762             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1763             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1764             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1765
1766             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1767             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1768             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1769             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1770             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1771             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1772             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1773             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1774             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
1775             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
1776
1777             d                = _fjsp_sub_v2r8(r00,rswitch);
1778             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1779             d2               = _fjsp_mul_v2r8(d,d);
1780             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1781
1782             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1783
1784             /* Evaluate switch function */
1785             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1786             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
1787             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1788
1789             fscal            = felec;
1790
1791             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1792
1793             /* Update vectorial force */
1794             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1795             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1796             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1797
1798             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1799             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1800             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1801
1802             }
1803
1804             /**************************
1805              * CALCULATE INTERACTIONS *
1806              **************************/
1807
1808             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1809             {
1810
1811             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
1812
1813             /* EWALD ELECTROSTATICS */
1814
1815             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1816             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
1817             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1818             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1819             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1820
1821             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1822             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1823             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1824             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1825             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1826             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1827             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1828             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1829             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
1830             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
1831
1832             d                = _fjsp_sub_v2r8(r01,rswitch);
1833             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1834             d2               = _fjsp_mul_v2r8(d,d);
1835             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1836
1837             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1838
1839             /* Evaluate switch function */
1840             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1841             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
1842             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1843
1844             fscal            = felec;
1845
1846             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1847
1848             /* Update vectorial force */
1849             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1850             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1851             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1852
1853             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1854             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1855             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1856
1857             }
1858
1859             /**************************
1860              * CALCULATE INTERACTIONS *
1861              **************************/
1862
1863             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1864             {
1865
1866             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
1867
1868             /* EWALD ELECTROSTATICS */
1869
1870             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1871             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
1872             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1873             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1874             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1875
1876             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1877             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1878             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1879             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1880             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1881             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1882             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1883             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1884             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
1885             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
1886
1887             d                = _fjsp_sub_v2r8(r02,rswitch);
1888             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1889             d2               = _fjsp_mul_v2r8(d,d);
1890             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1891
1892             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1893
1894             /* Evaluate switch function */
1895             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1896             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
1897             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1898
1899             fscal            = felec;
1900
1901             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1902
1903             /* Update vectorial force */
1904             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1905             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1906             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1907
1908             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1909             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1910             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1911
1912             }
1913
1914             /**************************
1915              * CALCULATE INTERACTIONS *
1916              **************************/
1917
1918             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1919             {
1920
1921             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1922
1923             /* EWALD ELECTROSTATICS */
1924
1925             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1926             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1927             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1928             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1929             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1930
1931             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1932             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1933             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1934             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1935             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1936             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1937             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1938             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1939             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
1940             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1941
1942             d                = _fjsp_sub_v2r8(r10,rswitch);
1943             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1944             d2               = _fjsp_mul_v2r8(d,d);
1945             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1946
1947             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1948
1949             /* Evaluate switch function */
1950             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1951             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
1952             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1953
1954             fscal            = felec;
1955
1956             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1957
1958             /* Update vectorial force */
1959             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1960             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1961             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1962
1963             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1964             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1965             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1966
1967             }
1968
1969             /**************************
1970              * CALCULATE INTERACTIONS *
1971              **************************/
1972
1973             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1974             {
1975
1976             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1977
1978             /* EWALD ELECTROSTATICS */
1979
1980             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1981             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1982             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1983             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1984             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1985
1986             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1987             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
1988             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1989             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1990             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
1991             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1992             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1993             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1994             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
1995             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1996
1997             d                = _fjsp_sub_v2r8(r11,rswitch);
1998             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1999             d2               = _fjsp_mul_v2r8(d,d);
2000             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2001
2002             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2003
2004             /* Evaluate switch function */
2005             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2006             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
2007             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
2008
2009             fscal            = felec;
2010
2011             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2012
2013             /* Update vectorial force */
2014             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
2015             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
2016             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
2017
2018             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
2019             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
2020             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
2021
2022             }
2023
2024             /**************************
2025              * CALCULATE INTERACTIONS *
2026              **************************/
2027
2028             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2029             {
2030
2031             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2032
2033             /* EWALD ELECTROSTATICS */
2034
2035             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2036             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2037             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2038             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2039             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2040
2041             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2042             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2043             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2044             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2045             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2046             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2047             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2048             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2049             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
2050             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2051
2052             d                = _fjsp_sub_v2r8(r12,rswitch);
2053             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2054             d2               = _fjsp_mul_v2r8(d,d);
2055             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2056
2057             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2058
2059             /* Evaluate switch function */
2060             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2061             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
2062             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2063
2064             fscal            = felec;
2065
2066             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2067
2068             /* Update vectorial force */
2069             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2070             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2071             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2072
2073             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2074             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2075             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2076
2077             }
2078
2079             /**************************
2080              * CALCULATE INTERACTIONS *
2081              **************************/
2082
2083             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
2084             {
2085
2086             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
2087
2088             /* EWALD ELECTROSTATICS */
2089
2090             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2091             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
2092             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2093             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2094             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2095
2096             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2097             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2098             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2099             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2100             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2101             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2102             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2103             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2104             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
2105             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
2106
2107             d                = _fjsp_sub_v2r8(r20,rswitch);
2108             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2109             d2               = _fjsp_mul_v2r8(d,d);
2110             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2111
2112             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2113
2114             /* Evaluate switch function */
2115             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2116             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
2117             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
2118
2119             fscal            = felec;
2120
2121             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2122
2123             /* Update vectorial force */
2124             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
2125             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
2126             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
2127
2128             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
2129             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
2130             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
2131
2132             }
2133
2134             /**************************
2135              * CALCULATE INTERACTIONS *
2136              **************************/
2137
2138             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2139             {
2140
2141             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2142
2143             /* EWALD ELECTROSTATICS */
2144
2145             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2146             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2147             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2148             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2149             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2150
2151             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2152             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2153             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2154             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2155             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2156             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2157             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2158             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2159             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
2160             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2161
2162             d                = _fjsp_sub_v2r8(r21,rswitch);
2163             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2164             d2               = _fjsp_mul_v2r8(d,d);
2165             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2166
2167             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2168
2169             /* Evaluate switch function */
2170             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2171             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
2172             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2173
2174             fscal            = felec;
2175
2176             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2177
2178             /* Update vectorial force */
2179             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2180             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2181             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2182
2183             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2184             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2185             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2186
2187             }
2188
2189             /**************************
2190              * CALCULATE INTERACTIONS *
2191              **************************/
2192
2193             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2194             {
2195
2196             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2197
2198             /* EWALD ELECTROSTATICS */
2199
2200             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2201             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2202             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2203             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2204             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2205
2206             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2207             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
2208             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2209             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2210             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
2211             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2212             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2213             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2214             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
2215             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2216
2217             d                = _fjsp_sub_v2r8(r22,rswitch);
2218             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2219             d2               = _fjsp_mul_v2r8(d,d);
2220             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2221
2222             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2223
2224             /* Evaluate switch function */
2225             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2226             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
2227             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2228
2229             fscal            = felec;
2230
2231             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2232
2233             /* Update vectorial force */
2234             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2235             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2236             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2237
2238             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2239             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2240             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2241
2242             }
2243
2244             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2245
2246             /* Inner loop uses 585 flops */
2247         }
2248
2249         if(jidx<j_index_end)
2250         {
2251
2252             jnrA             = jjnr[jidx];
2253             j_coord_offsetA  = DIM*jnrA;
2254
2255             /* load j atom coordinates */
2256             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
2257                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
2258
2259             /* Calculate displacement vector */
2260             dx00             = _fjsp_sub_v2r8(ix0,jx0);
2261             dy00             = _fjsp_sub_v2r8(iy0,jy0);
2262             dz00             = _fjsp_sub_v2r8(iz0,jz0);
2263             dx01             = _fjsp_sub_v2r8(ix0,jx1);
2264             dy01             = _fjsp_sub_v2r8(iy0,jy1);
2265             dz01             = _fjsp_sub_v2r8(iz0,jz1);
2266             dx02             = _fjsp_sub_v2r8(ix0,jx2);
2267             dy02             = _fjsp_sub_v2r8(iy0,jy2);
2268             dz02             = _fjsp_sub_v2r8(iz0,jz2);
2269             dx10             = _fjsp_sub_v2r8(ix1,jx0);
2270             dy10             = _fjsp_sub_v2r8(iy1,jy0);
2271             dz10             = _fjsp_sub_v2r8(iz1,jz0);
2272             dx11             = _fjsp_sub_v2r8(ix1,jx1);
2273             dy11             = _fjsp_sub_v2r8(iy1,jy1);
2274             dz11             = _fjsp_sub_v2r8(iz1,jz1);
2275             dx12             = _fjsp_sub_v2r8(ix1,jx2);
2276             dy12             = _fjsp_sub_v2r8(iy1,jy2);
2277             dz12             = _fjsp_sub_v2r8(iz1,jz2);
2278             dx20             = _fjsp_sub_v2r8(ix2,jx0);
2279             dy20             = _fjsp_sub_v2r8(iy2,jy0);
2280             dz20             = _fjsp_sub_v2r8(iz2,jz0);
2281             dx21             = _fjsp_sub_v2r8(ix2,jx1);
2282             dy21             = _fjsp_sub_v2r8(iy2,jy1);
2283             dz21             = _fjsp_sub_v2r8(iz2,jz1);
2284             dx22             = _fjsp_sub_v2r8(ix2,jx2);
2285             dy22             = _fjsp_sub_v2r8(iy2,jy2);
2286             dz22             = _fjsp_sub_v2r8(iz2,jz2);
2287
2288             /* Calculate squared distance and things based on it */
2289             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
2290             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
2291             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
2292             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
2293             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
2294             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
2295             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
2296             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
2297             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
2298
2299             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
2300             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
2301             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
2302             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
2303             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
2304             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
2305             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
2306             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
2307             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
2308
2309             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
2310             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
2311             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
2312             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
2313             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
2314             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
2315             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
2316             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
2317             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
2318
2319             fjx0             = _fjsp_setzero_v2r8();
2320             fjy0             = _fjsp_setzero_v2r8();
2321             fjz0             = _fjsp_setzero_v2r8();
2322             fjx1             = _fjsp_setzero_v2r8();
2323             fjy1             = _fjsp_setzero_v2r8();
2324             fjz1             = _fjsp_setzero_v2r8();
2325             fjx2             = _fjsp_setzero_v2r8();
2326             fjy2             = _fjsp_setzero_v2r8();
2327             fjz2             = _fjsp_setzero_v2r8();
2328
2329             /**************************
2330              * CALCULATE INTERACTIONS *
2331              **************************/
2332
2333             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
2334             {
2335
2336             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
2337
2338             /* EWALD ELECTROSTATICS */
2339
2340             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2341             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
2342             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2343             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2344             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2345
2346             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2347             ewtabD           = _fjsp_setzero_v2r8();
2348             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2349             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2350             ewtabFn          = _fjsp_setzero_v2r8();
2351             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2352             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2353             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2354             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
2355             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
2356
2357             d                = _fjsp_sub_v2r8(r00,rswitch);
2358             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2359             d2               = _fjsp_mul_v2r8(d,d);
2360             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2361
2362             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2363
2364             /* Evaluate switch function */
2365             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2366             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(velec,dsw)) );
2367             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
2368
2369             fscal            = felec;
2370
2371             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2372
2373             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2374
2375             /* Update vectorial force */
2376             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
2377             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
2378             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
2379
2380             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
2381             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
2382             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
2383
2384             }
2385
2386             /**************************
2387              * CALCULATE INTERACTIONS *
2388              **************************/
2389
2390             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
2391             {
2392
2393             r01              = _fjsp_mul_v2r8(rsq01,rinv01);
2394
2395             /* EWALD ELECTROSTATICS */
2396
2397             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2398             ewrt             = _fjsp_mul_v2r8(r01,ewtabscale);
2399             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2400             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2401             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2402
2403             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2404             ewtabD           = _fjsp_setzero_v2r8();
2405             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2406             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2407             ewtabFn          = _fjsp_setzero_v2r8();
2408             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2409             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2410             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2411             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(rinv01,velec));
2412             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq01,rinv01),_fjsp_sub_v2r8(rinvsq01,felec));
2413
2414             d                = _fjsp_sub_v2r8(r01,rswitch);
2415             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2416             d2               = _fjsp_mul_v2r8(d,d);
2417             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2418
2419             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2420
2421             /* Evaluate switch function */
2422             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2423             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv01,_fjsp_mul_v2r8(velec,dsw)) );
2424             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
2425
2426             fscal            = felec;
2427
2428             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2429
2430             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2431
2432             /* Update vectorial force */
2433             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
2434             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
2435             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
2436
2437             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
2438             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
2439             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
2440
2441             }
2442
2443             /**************************
2444              * CALCULATE INTERACTIONS *
2445              **************************/
2446
2447             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
2448             {
2449
2450             r02              = _fjsp_mul_v2r8(rsq02,rinv02);
2451
2452             /* EWALD ELECTROSTATICS */
2453
2454             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2455             ewrt             = _fjsp_mul_v2r8(r02,ewtabscale);
2456             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2457             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2458             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2459
2460             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2461             ewtabD           = _fjsp_setzero_v2r8();
2462             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2463             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2464             ewtabFn          = _fjsp_setzero_v2r8();
2465             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2466             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2467             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2468             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(rinv02,velec));
2469             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq02,rinv02),_fjsp_sub_v2r8(rinvsq02,felec));
2470
2471             d                = _fjsp_sub_v2r8(r02,rswitch);
2472             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2473             d2               = _fjsp_mul_v2r8(d,d);
2474             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2475
2476             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2477
2478             /* Evaluate switch function */
2479             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2480             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv02,_fjsp_mul_v2r8(velec,dsw)) );
2481             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
2482
2483             fscal            = felec;
2484
2485             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2486
2487             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2488
2489             /* Update vectorial force */
2490             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
2491             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
2492             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
2493
2494             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
2495             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
2496             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
2497
2498             }
2499
2500             /**************************
2501              * CALCULATE INTERACTIONS *
2502              **************************/
2503
2504             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
2505             {
2506
2507             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
2508
2509             /* EWALD ELECTROSTATICS */
2510
2511             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2512             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
2513             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2514             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2515             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2516
2517             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2518             ewtabD           = _fjsp_setzero_v2r8();
2519             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2520             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2521             ewtabFn          = _fjsp_setzero_v2r8();
2522             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2523             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2524             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2525             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
2526             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
2527
2528             d                = _fjsp_sub_v2r8(r10,rswitch);
2529             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2530             d2               = _fjsp_mul_v2r8(d,d);
2531             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2532
2533             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2534
2535             /* Evaluate switch function */
2536             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2537             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv10,_fjsp_mul_v2r8(velec,dsw)) );
2538             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
2539
2540             fscal            = felec;
2541
2542             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2543
2544             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2545
2546             /* Update vectorial force */
2547             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
2548             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
2549             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
2550
2551             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
2552             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
2553             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
2554
2555             }
2556
2557             /**************************
2558              * CALCULATE INTERACTIONS *
2559              **************************/
2560
2561             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
2562             {
2563
2564             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
2565
2566             /* EWALD ELECTROSTATICS */
2567
2568             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2569             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
2570             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2571             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2572             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2573
2574             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2575             ewtabD           = _fjsp_setzero_v2r8();
2576             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2577             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2578             ewtabFn          = _fjsp_setzero_v2r8();
2579             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2580             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2581             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2582             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(rinv11,velec));
2583             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
2584
2585             d                = _fjsp_sub_v2r8(r11,rswitch);
2586             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2587             d2               = _fjsp_mul_v2r8(d,d);
2588             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2589
2590             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2591
2592             /* Evaluate switch function */
2593             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2594             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv11,_fjsp_mul_v2r8(velec,dsw)) );
2595             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
2596
2597             fscal            = felec;
2598
2599             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2600
2601             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2602
2603             /* Update vectorial force */
2604             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
2605             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
2606             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
2607
2608             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
2609             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
2610             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
2611
2612             }
2613
2614             /**************************
2615              * CALCULATE INTERACTIONS *
2616              **************************/
2617
2618             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2619             {
2620
2621             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2622
2623             /* EWALD ELECTROSTATICS */
2624
2625             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2626             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2627             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2628             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2629             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2630
2631             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2632             ewtabD           = _fjsp_setzero_v2r8();
2633             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2634             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2635             ewtabFn          = _fjsp_setzero_v2r8();
2636             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2637             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2638             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2639             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(rinv12,velec));
2640             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2641
2642             d                = _fjsp_sub_v2r8(r12,rswitch);
2643             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2644             d2               = _fjsp_mul_v2r8(d,d);
2645             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2646
2647             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2648
2649             /* Evaluate switch function */
2650             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2651             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv12,_fjsp_mul_v2r8(velec,dsw)) );
2652             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2653
2654             fscal            = felec;
2655
2656             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2657
2658             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2659
2660             /* Update vectorial force */
2661             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2662             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2663             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2664
2665             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2666             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2667             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2668
2669             }
2670
2671             /**************************
2672              * CALCULATE INTERACTIONS *
2673              **************************/
2674
2675             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
2676             {
2677
2678             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
2679
2680             /* EWALD ELECTROSTATICS */
2681
2682             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2683             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
2684             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2685             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2686             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2687
2688             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2689             ewtabD           = _fjsp_setzero_v2r8();
2690             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2691             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2692             ewtabFn          = _fjsp_setzero_v2r8();
2693             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2694             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2695             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2696             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
2697             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
2698
2699             d                = _fjsp_sub_v2r8(r20,rswitch);
2700             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2701             d2               = _fjsp_mul_v2r8(d,d);
2702             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2703
2704             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2705
2706             /* Evaluate switch function */
2707             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2708             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv20,_fjsp_mul_v2r8(velec,dsw)) );
2709             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
2710
2711             fscal            = felec;
2712
2713             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2714
2715             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2716
2717             /* Update vectorial force */
2718             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
2719             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
2720             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
2721
2722             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
2723             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
2724             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
2725
2726             }
2727
2728             /**************************
2729              * CALCULATE INTERACTIONS *
2730              **************************/
2731
2732             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2733             {
2734
2735             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2736
2737             /* EWALD ELECTROSTATICS */
2738
2739             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2740             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2741             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2742             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2743             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2744
2745             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2746             ewtabD           = _fjsp_setzero_v2r8();
2747             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2748             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2749             ewtabFn          = _fjsp_setzero_v2r8();
2750             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2751             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2752             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2753             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(rinv21,velec));
2754             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2755
2756             d                = _fjsp_sub_v2r8(r21,rswitch);
2757             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2758             d2               = _fjsp_mul_v2r8(d,d);
2759             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2760
2761             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2762
2763             /* Evaluate switch function */
2764             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2765             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv21,_fjsp_mul_v2r8(velec,dsw)) );
2766             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2767
2768             fscal            = felec;
2769
2770             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2771
2772             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2773
2774             /* Update vectorial force */
2775             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2776             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2777             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2778
2779             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2780             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2781             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2782
2783             }
2784
2785             /**************************
2786              * CALCULATE INTERACTIONS *
2787              **************************/
2788
2789             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2790             {
2791
2792             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2793
2794             /* EWALD ELECTROSTATICS */
2795
2796             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2797             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2798             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2799             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2800             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2801
2802             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
2803             ewtabD           = _fjsp_setzero_v2r8();
2804             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
2805             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
2806             ewtabFn          = _fjsp_setzero_v2r8();
2807             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
2808             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
2809             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
2810             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(rinv22,velec));
2811             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2812
2813             d                = _fjsp_sub_v2r8(r22,rswitch);
2814             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
2815             d2               = _fjsp_mul_v2r8(d,d);
2816             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
2817
2818             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
2819
2820             /* Evaluate switch function */
2821             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2822             felec            = _fjsp_msub_v2r8( felec,sw , _fjsp_mul_v2r8(rinv22,_fjsp_mul_v2r8(velec,dsw)) );
2823             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2824
2825             fscal            = felec;
2826
2827             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2828
2829             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2830
2831             /* Update vectorial force */
2832             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2833             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2834             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2835
2836             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2837             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2838             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2839
2840             }
2841
2842             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2843
2844             /* Inner loop uses 585 flops */
2845         }
2846
2847         /* End of innermost loop */
2848
2849         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2850                                               f+i_coord_offset,fshift+i_shift_offset);
2851
2852         /* Increment number of inner iterations */
2853         inneriter                  += j_index_end - j_index_start;
2854
2855         /* Outer loop uses 18 flops */
2856     }
2857
2858     /* Increment number of outer iterations */
2859     outeriter        += nri;
2860
2861     /* Update outer/inner flops */
2862
2863     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*585);
2864 }