src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEwSh_VdwNone_GeomW4W4_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sparc64_hpc_ace_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: Ewald
  54  * VdW interaction:            None
  55  * Geometry:                   Water4-Water4
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecEwSh_VdwNone_GeomW4W4_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              vdwioffset1;
  82     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwioffset3;
  86     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  87     int              vdwjidx1A,vdwjidx1B;
  88     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     int              vdwjidx3A,vdwjidx3B;
  92     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  93     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  94     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  95     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  96     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  97     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  98     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  99     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 100     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 101     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 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       itab_tmp;
 107     _fjsp_v2r8       dummy_mask,cutoff_mask;
 108     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 109     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 110     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 111
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 124     charge           = mdatoms->chargeA;
 125
 126     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 127     ewtab            = fr->ic->tabq_coul_FDV0;
 128     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 129     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 130
 131     /* Setup water-specific parameters */
 132     inr              = nlist->iinr[0];
 133     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 134     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 135     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 136
 137     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 138     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 139     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 140     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 141     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 142     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 143     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 144     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 145     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 146     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 147     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 148     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 149
 150     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 151     rcutoff_scalar   = fr->rcoulomb;
 152     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 153     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 154
 155     /* Avoid stupid compiler warnings */
 156     jnrA = jnrB = 0;
 157     j_coord_offsetA = 0;
 158     j_coord_offsetB = 0;
 159
 160     outeriter        = 0;
 161     inneriter        = 0;
 162
 163     /* Start outer loop over neighborlists */
 164     for(iidx=0; iidx<nri; iidx++)
 165     {
 166         /* Load shift vector for this list */
 167         i_shift_offset   = DIM*shiftidx[iidx];
 168
 169         /* Load limits for loop over neighbors */
 170         j_index_start    = jindex[iidx];
 171         j_index_end      = jindex[iidx+1];
 172
 173         /* Get outer coordinate index */
 174         inr              = iinr[iidx];
 175         i_coord_offset   = DIM*inr;
 176
 177         /* Load i particle coords and add shift vector */
 178         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 179                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 180
 181         fix1             = _fjsp_setzero_v2r8();
 182         fiy1             = _fjsp_setzero_v2r8();
 183         fiz1             = _fjsp_setzero_v2r8();
 184         fix2             = _fjsp_setzero_v2r8();
 185         fiy2             = _fjsp_setzero_v2r8();
 186         fiz2             = _fjsp_setzero_v2r8();
 187         fix3             = _fjsp_setzero_v2r8();
 188         fiy3             = _fjsp_setzero_v2r8();
 189         fiz3             = _fjsp_setzero_v2r8();
 190
 191         /* Reset potential sums */
 192         velecsum         = _fjsp_setzero_v2r8();
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 196         {
 197
 198             /* Get j neighbor index, and coordinate index */
 199             jnrA             = jjnr[jidx];
 200             jnrB             = jjnr[jidx+1];
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203
 204             /* load j atom coordinates */
 205             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 206                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 207
 208             /* Calculate displacement vector */
 209             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 210             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 211             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 212             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 213             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 214             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 215             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 216             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 217             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 218             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 219             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 220             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 221             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 222             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 223             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 224             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 225             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 226             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 227             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 228             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 229             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 230             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 231             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 232             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 233             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 234             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 235             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 236
 237             /* Calculate squared distance and things based on it */
 238             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 239             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 240             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 241             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 242             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 243             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 244             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 245             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 246             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 247
 248             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 249             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 250             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 251             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 252             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 253             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 254             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 255             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 256             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 257
 258             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 259             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 260             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 261             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 262             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 263             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 264             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 265             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 266             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 267
 268             fjx1             = _fjsp_setzero_v2r8();
 269             fjy1             = _fjsp_setzero_v2r8();
 270             fjz1             = _fjsp_setzero_v2r8();
 271             fjx2             = _fjsp_setzero_v2r8();
 272             fjy2             = _fjsp_setzero_v2r8();
 273             fjz2             = _fjsp_setzero_v2r8();
 274             fjx3             = _fjsp_setzero_v2r8();
 275             fjy3             = _fjsp_setzero_v2r8();
 276             fjz3             = _fjsp_setzero_v2r8();
 277
 278             /**************************
 279              * CALCULATE INTERACTIONS *
 280              **************************/
 281
 282             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 283             {
 284
 285             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 286
 287             /* EWALD ELECTROSTATICS */
 288
 289             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 290             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 291             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 292             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 293             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 294
 295             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 296             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 297             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 298             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 299             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 300             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 301             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 302             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 303             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv11,sh_ewald),velec));
 304             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 305
 306             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 307
 308             /* Update potential sum for this i atom from the interaction with this j atom. */
 309             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 310             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 311
 312             fscal            = felec;
 313
 314             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 315
 316             /* Update vectorial force */
 317             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 318             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 319             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 320
 321             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 322             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 323             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 324
 325             }
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 332             {
 333
 334             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 335
 336             /* EWALD ELECTROSTATICS */
 337
 338             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 339             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 340             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 341             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 342             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 343
 344             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 345             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 346             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 347             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 348             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 349             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 350             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 351             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 352             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv12,sh_ewald),velec));
 353             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 354
 355             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 356
 357             /* Update potential sum for this i atom from the interaction with this j atom. */
 358             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 359             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 360
 361             fscal            = felec;
 362
 363             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 364
 365             /* Update vectorial force */
 366             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 367             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 368             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 369
 370             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 371             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 372             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 373
 374             }
 375
 376             /**************************
 377              * CALCULATE INTERACTIONS *
 378              **************************/
 379
 380             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
 381             {
 382
 383             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 384
 385             /* EWALD ELECTROSTATICS */
 386
 387             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 388             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
 389             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 390             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 391             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 392
 393             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 394             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 395             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 396             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 397             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 398             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 399             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 400             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 401             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv13,sh_ewald),velec));
 402             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
 403
 404             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 408             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 413
 414             /* Update vectorial force */
 415             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 416             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 417             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 418
 419             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 420             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 421             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 422
 423             }
 424
 425             /**************************
 426              * CALCULATE INTERACTIONS *
 427              **************************/
 428
 429             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 430             {
 431
 432             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 433
 434             /* EWALD ELECTROSTATICS */
 435
 436             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 437             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 438             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 439             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 440             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 441
 442             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 443             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 444             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 445             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 446             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 447             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 448             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 449             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 450             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv21,sh_ewald),velec));
 451             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 452
 453             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 457             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 462
 463             /* Update vectorial force */
 464             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 465             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 466             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 467
 468             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 469             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 470             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 471
 472             }
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 479             {
 480
 481             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 482
 483             /* EWALD ELECTROSTATICS */
 484
 485             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 486             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
 487             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 488             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 489             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 490
 491             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 492             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 493             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 494             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 495             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 496             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 497             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 498             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 499             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv22,sh_ewald),velec));
 500             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
 501
 502             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 503
 504             /* Update potential sum for this i atom from the interaction with this j atom. */
 505             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 506             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 507
 508             fscal            = felec;
 509
 510             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 511
 512             /* Update vectorial force */
 513             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 514             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 515             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 516
 517             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 518             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 519             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 520
 521             }
 522
 523             /**************************
 524              * CALCULATE INTERACTIONS *
 525              **************************/
 526
 527             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
 528             {
 529
 530             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
 531
 532             /* EWALD ELECTROSTATICS */
 533
 534             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 535             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
 536             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 537             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 538             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 539
 540             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 541             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 542             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 543             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 544             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 545             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 546             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 547             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 548             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv23,sh_ewald),velec));
 549             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
 550
 551             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
 552
 553             /* Update potential sum for this i atom from the interaction with this j atom. */
 554             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 555             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 556
 557             fscal            = felec;
 558
 559             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 560
 561             /* Update vectorial force */
 562             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 563             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 564             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 565
 566             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 567             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 568             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 569
 570             }
 571
 572             /**************************
 573              * CALCULATE INTERACTIONS *
 574              **************************/
 575
 576             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
 577             {
 578
 579             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
 580
 581             /* EWALD ELECTROSTATICS */
 582
 583             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 584             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
 585             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 586             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 587             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 588
 589             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 590             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 591             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 592             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 593             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 594             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 595             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 596             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 597             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv31,sh_ewald),velec));
 598             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
 599
 600             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
 601
 602             /* Update potential sum for this i atom from the interaction with this j atom. */
 603             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 604             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 605
 606             fscal            = felec;
 607
 608             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 609
 610             /* Update vectorial force */
 611             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 612             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 613             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 614
 615             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 616             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 617             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 618
 619             }
 620
 621             /**************************
 622              * CALCULATE INTERACTIONS *
 623              **************************/
 624
 625             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
 626             {
 627
 628             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
 629
 630             /* EWALD ELECTROSTATICS */
 631
 632             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 633             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
 634             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 635             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 636             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 637
 638             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 639             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 640             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 641             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 642             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 643             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 644             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 645             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 646             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv32,sh_ewald),velec));
 647             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
 648
 649             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
 650
 651             /* Update potential sum for this i atom from the interaction with this j atom. */
 652             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 653             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 654
 655             fscal            = felec;
 656
 657             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 658
 659             /* Update vectorial force */
 660             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 661             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 662             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 663
 664             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 665             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 666             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 667
 668             }
 669
 670             /**************************
 671              * CALCULATE INTERACTIONS *
 672              **************************/
 673
 674             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
 675             {
 676
 677             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
 678
 679             /* EWALD ELECTROSTATICS */
 680
 681             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 682             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
 683             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 684             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 685             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 686
 687             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 688             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 689             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 690             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 691             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 692             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 693             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 694             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 695             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv33,sh_ewald),velec));
 696             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
 697
 698             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
 699
 700             /* Update potential sum for this i atom from the interaction with this j atom. */
 701             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 702             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 703
 704             fscal            = felec;
 705
 706             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 707
 708             /* Update vectorial force */
 709             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 710             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 711             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 712
 713             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 714             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 715             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 716
 717             }
 718
 719             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 720
 721             /* Inner loop uses 441 flops */
 722         }
 723
 724         if(jidx<j_index_end)
 725         {
 726
 727             jnrA             = jjnr[jidx];
 728             j_coord_offsetA  = DIM*jnrA;
 729
 730             /* load j atom coordinates */
 731             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
 732                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 733
 734             /* Calculate displacement vector */
 735             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 736             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 737             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 738             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 739             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 740             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 741             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 742             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 743             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 744             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 745             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 746             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 747             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 748             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 749             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 750             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 751             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 752             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 753             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 754             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 755             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 756             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 757             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 758             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 759             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 760             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 761             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 762
 763             /* Calculate squared distance and things based on it */
 764             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 765             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 766             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 767             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 768             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 769             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 770             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 771             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 772             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 773
 774             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 775             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 776             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 777             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 778             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 779             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 780             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 781             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 782             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 783
 784             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 785             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 786             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 787             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 788             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 789             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 790             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 791             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 792             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 793
 794             fjx1             = _fjsp_setzero_v2r8();
 795             fjy1             = _fjsp_setzero_v2r8();
 796             fjz1             = _fjsp_setzero_v2r8();
 797             fjx2             = _fjsp_setzero_v2r8();
 798             fjy2             = _fjsp_setzero_v2r8();
 799             fjz2             = _fjsp_setzero_v2r8();
 800             fjx3             = _fjsp_setzero_v2r8();
 801             fjy3             = _fjsp_setzero_v2r8();
 802             fjz3             = _fjsp_setzero_v2r8();
 803
 804             /**************************
 805              * CALCULATE INTERACTIONS *
 806              **************************/
 807
 808             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 809             {
 810
 811             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 812
 813             /* EWALD ELECTROSTATICS */
 814
 815             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 816             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 817             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 818             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 819             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 820
 821             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 822             ewtabD           = _fjsp_setzero_v2r8();
 823             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 824             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 825             ewtabFn          = _fjsp_setzero_v2r8();
 826             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 827             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 828             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 829             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv11,sh_ewald),velec));
 830             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 831
 832             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 833
 834             /* Update potential sum for this i atom from the interaction with this j atom. */
 835             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 836             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 837             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 842
 843             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 844
 845             /* Update vectorial force */
 846             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 847             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 848             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 849
 850             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 851             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 852             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 853
 854             }
 855
 856             /**************************
 857              * CALCULATE INTERACTIONS *
 858              **************************/
 859
 860             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 861             {
 862
 863             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 864
 865             /* EWALD ELECTROSTATICS */
 866
 867             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 868             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 869             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 870             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 871             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 872
 873             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 874             ewtabD           = _fjsp_setzero_v2r8();
 875             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 876             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 877             ewtabFn          = _fjsp_setzero_v2r8();
 878             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 879             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 880             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 881             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv12,sh_ewald),velec));
 882             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 883
 884             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 885
 886             /* Update potential sum for this i atom from the interaction with this j atom. */
 887             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 888             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 889             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 894
 895             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 896
 897             /* Update vectorial force */
 898             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 899             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 900             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 901
 902             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 903             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 904             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 905
 906             }
 907
 908             /**************************
 909              * CALCULATE INTERACTIONS *
 910              **************************/
 911
 912             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
 913             {
 914
 915             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 916
 917             /* EWALD ELECTROSTATICS */
 918
 919             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 920             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
 921             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 922             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 923             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 924
 925             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 926             ewtabD           = _fjsp_setzero_v2r8();
 927             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 928             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 929             ewtabFn          = _fjsp_setzero_v2r8();
 930             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 931             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 932             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 933             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv13,sh_ewald),velec));
 934             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
 935
 936             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
 937
 938             /* Update potential sum for this i atom from the interaction with this j atom. */
 939             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 940             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 941             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 942
 943             fscal            = felec;
 944
 945             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 946
 947             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 948
 949             /* Update vectorial force */
 950             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 951             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 952             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 953
 954             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 955             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 956             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 957
 958             }
 959
 960             /**************************
 961              * CALCULATE INTERACTIONS *
 962              **************************/
 963
 964             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 965             {
 966
 967             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 968
 969             /* EWALD ELECTROSTATICS */
 970
 971             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 972             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 973             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 974             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 975             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 976
 977             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 978             ewtabD           = _fjsp_setzero_v2r8();
 979             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 980             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 981             ewtabFn          = _fjsp_setzero_v2r8();
 982             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 983             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 984             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 985             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv21,sh_ewald),velec));
 986             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 987
 988             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 989
 990             /* Update potential sum for this i atom from the interaction with this j atom. */
 991             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 992             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 993             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 994
 995             fscal            = felec;
 996
 997             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 998
 999             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1000
1001             /* Update vectorial force */
1002             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1003             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1004             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1005
1006             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1007             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1008             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1009
1010             }
1011
1012             /**************************
1013              * CALCULATE INTERACTIONS *
1014              **************************/
1015
1016             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1017             {
1018
1019             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1020
1021             /* EWALD ELECTROSTATICS */
1022
1023             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1024             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1025             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1026             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1027             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1028
1029             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1030             ewtabD           = _fjsp_setzero_v2r8();
1031             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1032             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1033             ewtabFn          = _fjsp_setzero_v2r8();
1034             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1035             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1036             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1037             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv22,sh_ewald),velec));
1038             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1039
1040             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1041
1042             /* Update potential sum for this i atom from the interaction with this j atom. */
1043             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1044             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1045             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1046
1047             fscal            = felec;
1048
1049             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1050
1051             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1052
1053             /* Update vectorial force */
1054             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1055             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1056             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1057
1058             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1059             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1060             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1061
1062             }
1063
1064             /**************************
1065              * CALCULATE INTERACTIONS *
1066              **************************/
1067
1068             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
1069             {
1070
1071             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1072
1073             /* EWALD ELECTROSTATICS */
1074
1075             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1076             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
1077             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1078             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1079             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1080
1081             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1082             ewtabD           = _fjsp_setzero_v2r8();
1083             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1084             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1085             ewtabFn          = _fjsp_setzero_v2r8();
1086             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1087             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1088             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1089             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv23,sh_ewald),velec));
1090             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
1091
1092             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
1093
1094             /* Update potential sum for this i atom from the interaction with this j atom. */
1095             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1096             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1097             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1098
1099             fscal            = felec;
1100
1101             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1102
1103             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1104
1105             /* Update vectorial force */
1106             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1107             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1108             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1109
1110             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1111             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1112             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1113
1114             }
1115
1116             /**************************
1117              * CALCULATE INTERACTIONS *
1118              **************************/
1119
1120             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
1121             {
1122
1123             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1124
1125             /* EWALD ELECTROSTATICS */
1126
1127             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1128             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
1129             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1130             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1131             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1132
1133             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1134             ewtabD           = _fjsp_setzero_v2r8();
1135             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1136             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1137             ewtabFn          = _fjsp_setzero_v2r8();
1138             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1139             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1140             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1141             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv31,sh_ewald),velec));
1142             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
1143
1144             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
1145
1146             /* Update potential sum for this i atom from the interaction with this j atom. */
1147             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1148             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1149             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1150
1151             fscal            = felec;
1152
1153             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1154
1155             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1156
1157             /* Update vectorial force */
1158             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1159             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1160             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1161
1162             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1163             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1164             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1165
1166             }
1167
1168             /**************************
1169              * CALCULATE INTERACTIONS *
1170              **************************/
1171
1172             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
1173             {
1174
1175             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1176
1177             /* EWALD ELECTROSTATICS */
1178
1179             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1180             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
1181             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1182             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1183             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1184
1185             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1186             ewtabD           = _fjsp_setzero_v2r8();
1187             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1188             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1189             ewtabFn          = _fjsp_setzero_v2r8();
1190             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1191             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1192             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1193             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv32,sh_ewald),velec));
1194             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
1195
1196             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
1197
1198             /* Update potential sum for this i atom from the interaction with this j atom. */
1199             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1200             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1201             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1202
1203             fscal            = felec;
1204
1205             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1206
1207             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1208
1209             /* Update vectorial force */
1210             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1211             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1212             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1213
1214             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1215             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1216             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1217
1218             }
1219
1220             /**************************
1221              * CALCULATE INTERACTIONS *
1222              **************************/
1223
1224             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
1225             {
1226
1227             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1228
1229             /* EWALD ELECTROSTATICS */
1230
1231             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1232             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
1233             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1234             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1235             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1236
1237             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1238             ewtabD           = _fjsp_setzero_v2r8();
1239             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1240             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1241             ewtabFn          = _fjsp_setzero_v2r8();
1242             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1243             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1244             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1245             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv33,sh_ewald),velec));
1246             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
1247
1248             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
1249
1250             /* Update potential sum for this i atom from the interaction with this j atom. */
1251             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1252             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1253             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1254
1255             fscal            = felec;
1256
1257             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1258
1259             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1260
1261             /* Update vectorial force */
1262             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1263             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1264             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1265
1266             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1267             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1268             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1269
1270             }
1271
1272             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1273
1274             /* Inner loop uses 441 flops */
1275         }
1276
1277         /* End of innermost loop */
1278
1279         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1280                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1281
1282         ggid                        = gid[iidx];
1283         /* Update potential energies */
1284         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1285
1286         /* Increment number of inner iterations */
1287         inneriter                  += j_index_end - j_index_start;
1288
1289         /* Outer loop uses 19 flops */
1290     }
1291
1292     /* Increment number of outer iterations */
1293     outeriter        += nri;
1294
1295     /* Update outer/inner flops */
1296
1297     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*441);
1298 }
1299 /*
1300  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1301  * Electrostatics interaction: Ewald
1302  * VdW interaction:            None
1303  * Geometry:                   Water4-Water4
1304  * Calculate force/pot:        Force
1305  */
1306 void
1307 nb_kernel_ElecEwSh_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1308                     (t_nblist                    * gmx_restrict       nlist,
1309                      rvec                        * gmx_restrict          xx,
1310                      rvec                        * gmx_restrict          ff,
1311                      t_forcerec                  * gmx_restrict          fr,
1312                      t_mdatoms                   * gmx_restrict     mdatoms,
1313                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1314                      t_nrnb                      * gmx_restrict        nrnb)
1315 {
1316     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1317      * just 0 for non-waters.
1318      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1319      * jnr indices corresponding to data put in the four positions in the SIMD register.
1320      */
1321     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1322     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1323     int              jnrA,jnrB;
1324     int              j_coord_offsetA,j_coord_offsetB;
1325     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1326     real             rcutoff_scalar;
1327     real             *shiftvec,*fshift,*x,*f;
1328     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1329     int              vdwioffset1;
1330     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1331     int              vdwioffset2;
1332     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1333     int              vdwioffset3;
1334     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1335     int              vdwjidx1A,vdwjidx1B;
1336     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1337     int              vdwjidx2A,vdwjidx2B;
1338     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1339     int              vdwjidx3A,vdwjidx3B;
1340     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1341     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1342     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1343     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1344     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1345     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1346     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1347     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1348     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1349     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1350     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1351     real             *charge;
1352     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1353     real             *ewtab;
1354     _fjsp_v2r8       itab_tmp;
1355     _fjsp_v2r8       dummy_mask,cutoff_mask;
1356     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1357     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1358     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1359
1360     x                = xx[0];
1361     f                = ff[0];
1362
1363     nri              = nlist->nri;
1364     iinr             = nlist->iinr;
1365     jindex           = nlist->jindex;
1366     jjnr             = nlist->jjnr;
1367     shiftidx         = nlist->shift;
1368     gid              = nlist->gid;
1369     shiftvec         = fr->shift_vec[0];
1370     fshift           = fr->fshift[0];
1371     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1372     charge           = mdatoms->chargeA;
1373
1374     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
1375     ewtab            = fr->ic->tabq_coul_F;
1376     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
1377     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
1378
1379     /* Setup water-specific parameters */
1380     inr              = nlist->iinr[0];
1381     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1382     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1383     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
1384
1385     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1386     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1387     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
1388     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1389     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1390     qq13             = _fjsp_mul_v2r8(iq1,jq3);
1391     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1392     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1393     qq23             = _fjsp_mul_v2r8(iq2,jq3);
1394     qq31             = _fjsp_mul_v2r8(iq3,jq1);
1395     qq32             = _fjsp_mul_v2r8(iq3,jq2);
1396     qq33             = _fjsp_mul_v2r8(iq3,jq3);
1397
1398     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1399     rcutoff_scalar   = fr->rcoulomb;
1400     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1401     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1402
1403     /* Avoid stupid compiler warnings */
1404     jnrA = jnrB = 0;
1405     j_coord_offsetA = 0;
1406     j_coord_offsetB = 0;
1407
1408     outeriter        = 0;
1409     inneriter        = 0;
1410
1411     /* Start outer loop over neighborlists */
1412     for(iidx=0; iidx<nri; iidx++)
1413     {
1414         /* Load shift vector for this list */
1415         i_shift_offset   = DIM*shiftidx[iidx];
1416
1417         /* Load limits for loop over neighbors */
1418         j_index_start    = jindex[iidx];
1419         j_index_end      = jindex[iidx+1];
1420
1421         /* Get outer coordinate index */
1422         inr              = iinr[iidx];
1423         i_coord_offset   = DIM*inr;
1424
1425         /* Load i particle coords and add shift vector */
1426         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1427                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1428
1429         fix1             = _fjsp_setzero_v2r8();
1430         fiy1             = _fjsp_setzero_v2r8();
1431         fiz1             = _fjsp_setzero_v2r8();
1432         fix2             = _fjsp_setzero_v2r8();
1433         fiy2             = _fjsp_setzero_v2r8();
1434         fiz2             = _fjsp_setzero_v2r8();
1435         fix3             = _fjsp_setzero_v2r8();
1436         fiy3             = _fjsp_setzero_v2r8();
1437         fiz3             = _fjsp_setzero_v2r8();
1438
1439         /* Start inner kernel loop */
1440         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1441         {
1442
1443             /* Get j neighbor index, and coordinate index */
1444             jnrA             = jjnr[jidx];
1445             jnrB             = jjnr[jidx+1];
1446             j_coord_offsetA  = DIM*jnrA;
1447             j_coord_offsetB  = DIM*jnrB;
1448
1449             /* load j atom coordinates */
1450             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1451                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1452
1453             /* Calculate displacement vector */
1454             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1455             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1456             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1457             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1458             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1459             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1460             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1461             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1462             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1463             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1464             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1465             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1466             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1467             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1468             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1469             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1470             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1471             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1472             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1473             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1474             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1475             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1476             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1477             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1478             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1479             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1480             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1481
1482             /* Calculate squared distance and things based on it */
1483             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1484             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1485             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1486             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1487             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1488             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1489             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1490             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1491             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1492
1493             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1494             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1495             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1496             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1497             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1498             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1499             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1500             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1501             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1502
1503             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1504             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1505             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1506             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1507             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1508             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1509             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1510             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1511             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1512
1513             fjx1             = _fjsp_setzero_v2r8();
1514             fjy1             = _fjsp_setzero_v2r8();
1515             fjz1             = _fjsp_setzero_v2r8();
1516             fjx2             = _fjsp_setzero_v2r8();
1517             fjy2             = _fjsp_setzero_v2r8();
1518             fjz2             = _fjsp_setzero_v2r8();
1519             fjx3             = _fjsp_setzero_v2r8();
1520             fjy3             = _fjsp_setzero_v2r8();
1521             fjz3             = _fjsp_setzero_v2r8();
1522
1523             /**************************
1524              * CALCULATE INTERACTIONS *
1525              **************************/
1526
1527             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1528             {
1529
1530             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1531
1532             /* EWALD ELECTROSTATICS */
1533
1534             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1535             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1536             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1537             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1538             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1539
1540             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1541                                          &ewtabF,&ewtabFn);
1542             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1543             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1544
1545             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1546
1547             fscal            = felec;
1548
1549             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1550
1551             /* Update vectorial force */
1552             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1553             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1554             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1555
1556             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1557             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1558             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1559
1560             }
1561
1562             /**************************
1563              * CALCULATE INTERACTIONS *
1564              **************************/
1565
1566             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1567             {
1568
1569             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1570
1571             /* EWALD ELECTROSTATICS */
1572
1573             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1574             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1575             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1576             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1577             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1578
1579             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1580                                          &ewtabF,&ewtabFn);
1581             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1582             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1583
1584             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1585
1586             fscal            = felec;
1587
1588             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1589
1590             /* Update vectorial force */
1591             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1592             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1593             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1594
1595             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1596             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1597             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1598
1599             }
1600
1601             /**************************
1602              * CALCULATE INTERACTIONS *
1603              **************************/
1604
1605             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
1606             {
1607
1608             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
1609
1610             /* EWALD ELECTROSTATICS */
1611
1612             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1613             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
1614             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1615             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1616             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1617
1618             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1619                                          &ewtabF,&ewtabFn);
1620             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1621             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
1622
1623             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
1624
1625             fscal            = felec;
1626
1627             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1628
1629             /* Update vectorial force */
1630             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1631             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1632             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1633
1634             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1635             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1636             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1637
1638             }
1639
1640             /**************************
1641              * CALCULATE INTERACTIONS *
1642              **************************/
1643
1644             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1645             {
1646
1647             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1648
1649             /* EWALD ELECTROSTATICS */
1650
1651             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1652             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1653             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1654             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1655             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1656
1657             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1658                                          &ewtabF,&ewtabFn);
1659             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1660             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1661
1662             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1663
1664             fscal            = felec;
1665
1666             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1667
1668             /* Update vectorial force */
1669             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1670             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1671             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1672
1673             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1674             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1675             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1676
1677             }
1678
1679             /**************************
1680              * CALCULATE INTERACTIONS *
1681              **************************/
1682
1683             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1684             {
1685
1686             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1687
1688             /* EWALD ELECTROSTATICS */
1689
1690             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1691             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1692             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1693             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1694             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1695
1696             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1697                                          &ewtabF,&ewtabFn);
1698             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1699             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1700
1701             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1702
1703             fscal            = felec;
1704
1705             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1706
1707             /* Update vectorial force */
1708             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1709             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1710             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1711
1712             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1713             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1714             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1715
1716             }
1717
1718             /**************************
1719              * CALCULATE INTERACTIONS *
1720              **************************/
1721
1722             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
1723             {
1724
1725             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1726
1727             /* EWALD ELECTROSTATICS */
1728
1729             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1730             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
1731             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1732             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1733             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1734
1735             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1736                                          &ewtabF,&ewtabFn);
1737             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1738             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
1739
1740             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
1741
1742             fscal            = felec;
1743
1744             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1745
1746             /* Update vectorial force */
1747             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1748             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1749             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1750
1751             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1752             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1753             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1754
1755             }
1756
1757             /**************************
1758              * CALCULATE INTERACTIONS *
1759              **************************/
1760
1761             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
1762             {
1763
1764             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1765
1766             /* EWALD ELECTROSTATICS */
1767
1768             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1769             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
1770             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1771             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1772             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1773
1774             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1775                                          &ewtabF,&ewtabFn);
1776             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1777             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
1778
1779             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
1780
1781             fscal            = felec;
1782
1783             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1784
1785             /* Update vectorial force */
1786             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1787             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1788             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1789
1790             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1791             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1792             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1793
1794             }
1795
1796             /**************************
1797              * CALCULATE INTERACTIONS *
1798              **************************/
1799
1800             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
1801             {
1802
1803             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1804
1805             /* EWALD ELECTROSTATICS */
1806
1807             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1808             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
1809             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1810             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1811             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1812
1813             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1814                                          &ewtabF,&ewtabFn);
1815             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1816             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
1817
1818             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
1819
1820             fscal            = felec;
1821
1822             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1823
1824             /* Update vectorial force */
1825             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1826             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1827             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1828
1829             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1830             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1831             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1832
1833             }
1834
1835             /**************************
1836              * CALCULATE INTERACTIONS *
1837              **************************/
1838
1839             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
1840             {
1841
1842             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1843
1844             /* EWALD ELECTROSTATICS */
1845
1846             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1847             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
1848             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1849             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1850             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1851
1852             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1853                                          &ewtabF,&ewtabFn);
1854             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1855             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
1856
1857             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
1858
1859             fscal            = felec;
1860
1861             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1862
1863             /* Update vectorial force */
1864             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1865             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1866             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1867
1868             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1869             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1870             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1871
1872             }
1873
1874             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1875
1876             /* Inner loop uses 378 flops */
1877         }
1878
1879         if(jidx<j_index_end)
1880         {
1881
1882             jnrA             = jjnr[jidx];
1883             j_coord_offsetA  = DIM*jnrA;
1884
1885             /* load j atom coordinates */
1886             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1887                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1888
1889             /* Calculate displacement vector */
1890             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1891             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1892             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1893             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1894             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1895             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1896             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1897             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1898             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1899             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1900             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1901             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1902             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1903             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1904             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1905             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1906             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1907             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1908             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1909             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1910             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1911             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1912             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1913             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1914             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1915             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1916             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1917
1918             /* Calculate squared distance and things based on it */
1919             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1920             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1921             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1922             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1923             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1924             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1925             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1926             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1927             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1928
1929             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1930             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1931             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1932             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1933             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1934             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1935             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1936             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1937             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1938
1939             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1940             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1941             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1942             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1943             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1944             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1945             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1946             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1947             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1948
1949             fjx1             = _fjsp_setzero_v2r8();
1950             fjy1             = _fjsp_setzero_v2r8();
1951             fjz1             = _fjsp_setzero_v2r8();
1952             fjx2             = _fjsp_setzero_v2r8();
1953             fjy2             = _fjsp_setzero_v2r8();
1954             fjz2             = _fjsp_setzero_v2r8();
1955             fjx3             = _fjsp_setzero_v2r8();
1956             fjy3             = _fjsp_setzero_v2r8();
1957             fjz3             = _fjsp_setzero_v2r8();
1958
1959             /**************************
1960              * CALCULATE INTERACTIONS *
1961              **************************/
1962
1963             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1964             {
1965
1966             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1967
1968             /* EWALD ELECTROSTATICS */
1969
1970             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1971             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1972             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1973             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1974             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1975
1976             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1977             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1978             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1979
1980             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1981
1982             fscal            = felec;
1983
1984             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1985
1986             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1987
1988             /* Update vectorial force */
1989             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1990             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1991             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1992
1993             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1994             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1995             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1996
1997             }
1998
1999             /**************************
2000              * CALCULATE INTERACTIONS *
2001              **************************/
2002
2003             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2004             {
2005
2006             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2007
2008             /* EWALD ELECTROSTATICS */
2009
2010             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2011             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2012             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2013             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2014             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2015
2016             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2017             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2018             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2019
2020             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2021
2022             fscal            = felec;
2023
2024             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2025
2026             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2027
2028             /* Update vectorial force */
2029             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2030             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2031             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2032
2033             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2034             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2035             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2036
2037             }
2038
2039             /**************************
2040              * CALCULATE INTERACTIONS *
2041              **************************/
2042
2043             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
2044             {
2045
2046             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
2047
2048             /* EWALD ELECTROSTATICS */
2049
2050             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2051             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
2052             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2053             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2054             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2055
2056             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2057             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2058             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
2059
2060             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
2061
2062             fscal            = felec;
2063
2064             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2065
2066             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2067
2068             /* Update vectorial force */
2069             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
2070             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
2071             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
2072
2073             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
2074             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
2075             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
2076
2077             }
2078
2079             /**************************
2080              * CALCULATE INTERACTIONS *
2081              **************************/
2082
2083             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2084             {
2085
2086             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2087
2088             /* EWALD ELECTROSTATICS */
2089
2090             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2091             ewrt             = _fjsp_mul_v2r8(r21,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             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2097             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2098             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2099
2100             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2101
2102             fscal            = felec;
2103
2104             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2105
2106             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2107
2108             /* Update vectorial force */
2109             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2110             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2111             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2112
2113             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2114             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2115             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2116
2117             }
2118
2119             /**************************
2120              * CALCULATE INTERACTIONS *
2121              **************************/
2122
2123             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2124             {
2125
2126             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2127
2128             /* EWALD ELECTROSTATICS */
2129
2130             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2131             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2132             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2133             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2134             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2135
2136             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2137             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2138             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2139
2140             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2141
2142             fscal            = felec;
2143
2144             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2145
2146             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2147
2148             /* Update vectorial force */
2149             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2150             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2151             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2152
2153             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2154             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2155             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2156
2157             }
2158
2159             /**************************
2160              * CALCULATE INTERACTIONS *
2161              **************************/
2162
2163             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
2164             {
2165
2166             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
2167
2168             /* EWALD ELECTROSTATICS */
2169
2170             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2171             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
2172             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2173             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2174             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2175
2176             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2177             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2178             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
2179
2180             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
2181
2182             fscal            = felec;
2183
2184             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2185
2186             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2187
2188             /* Update vectorial force */
2189             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
2190             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
2191             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
2192
2193             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
2194             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
2195             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
2196
2197             }
2198
2199             /**************************
2200              * CALCULATE INTERACTIONS *
2201              **************************/
2202
2203             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
2204             {
2205
2206             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
2207
2208             /* EWALD ELECTROSTATICS */
2209
2210             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2211             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
2212             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2213             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2214             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2215
2216             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2217             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2218             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
2219
2220             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
2221
2222             fscal            = felec;
2223
2224             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2225
2226             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2227
2228             /* Update vectorial force */
2229             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
2230             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
2231             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
2232
2233             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
2234             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
2235             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
2236
2237             }
2238
2239             /**************************
2240              * CALCULATE INTERACTIONS *
2241              **************************/
2242
2243             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
2244             {
2245
2246             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
2247
2248             /* EWALD ELECTROSTATICS */
2249
2250             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2251             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
2252             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2253             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2254             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2255
2256             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2257             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2258             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
2259
2260             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
2261
2262             fscal            = felec;
2263
2264             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2265
2266             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2267
2268             /* Update vectorial force */
2269             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
2270             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
2271             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
2272
2273             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
2274             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
2275             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
2276
2277             }
2278
2279             /**************************
2280              * CALCULATE INTERACTIONS *
2281              **************************/
2282
2283             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
2284             {
2285
2286             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
2287
2288             /* EWALD ELECTROSTATICS */
2289
2290             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2291             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
2292             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2293             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2294             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2295
2296             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2297             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2298             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
2299
2300             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
2301
2302             fscal            = felec;
2303
2304             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2305
2306             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2307
2308             /* Update vectorial force */
2309             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
2310             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
2311             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
2312
2313             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
2314             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
2315             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
2316
2317             }
2318
2319             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2320
2321             /* Inner loop uses 378 flops */
2322         }
2323
2324         /* End of innermost loop */
2325
2326         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2327                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2328
2329         /* Increment number of inner iterations */
2330         inneriter                  += j_index_end - j_index_start;
2331
2332         /* Outer loop uses 18 flops */
2333     }
2334
2335     /* Increment number of outer iterations */
2336     outeriter        += nri;
2337
2338     /* Update outer/inner flops */
2339
2340     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*378);
2341 }