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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            None
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSh_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset1;
  80     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  81     int              vdwioffset2;
  82     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  83     int              vdwioffset3;
  84     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  85     int              vdwjidx1A,vdwjidx1B;
  86     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  87     int              vdwjidx2A,vdwjidx2B;
  88     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  89     int              vdwjidx3A,vdwjidx3B;
  90     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  91     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  92     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  93     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  94     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  95     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  96     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  97     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  98     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  99     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 100     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 103     real             *ewtab;
 104     _fjsp_v2r8       itab_tmp;
 105     _fjsp_v2r8       dummy_mask,cutoff_mask;
 106     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 107     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 108     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 109
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123
 124     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 125     ewtab            = fr->ic->tabq_coul_FDV0;
 126     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 127     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 132     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 133     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 134
 135     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 136     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 137     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 138     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 139     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 140     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 141     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 142     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 143     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 144     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 145     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 146     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 147
 148     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 149     rcutoff_scalar   = fr->rcoulomb;
 150     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 151     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 152
 153     /* Avoid stupid compiler warnings */
 154     jnrA = jnrB = 0;
 155     j_coord_offsetA = 0;
 156     j_coord_offsetB = 0;
 157
 158     outeriter        = 0;
 159     inneriter        = 0;
 160
 161     /* Start outer loop over neighborlists */
 162     for(iidx=0; iidx<nri; iidx++)
 163     {
 164         /* Load shift vector for this list */
 165         i_shift_offset   = DIM*shiftidx[iidx];
 166
 167         /* Load limits for loop over neighbors */
 168         j_index_start    = jindex[iidx];
 169         j_index_end      = jindex[iidx+1];
 170
 171         /* Get outer coordinate index */
 172         inr              = iinr[iidx];
 173         i_coord_offset   = DIM*inr;
 174
 175         /* Load i particle coords and add shift vector */
 176         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 177                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 178
 179         fix1             = _fjsp_setzero_v2r8();
 180         fiy1             = _fjsp_setzero_v2r8();
 181         fiz1             = _fjsp_setzero_v2r8();
 182         fix2             = _fjsp_setzero_v2r8();
 183         fiy2             = _fjsp_setzero_v2r8();
 184         fiz2             = _fjsp_setzero_v2r8();
 185         fix3             = _fjsp_setzero_v2r8();
 186         fiy3             = _fjsp_setzero_v2r8();
 187         fiz3             = _fjsp_setzero_v2r8();
 188
 189         /* Reset potential sums */
 190         velecsum         = _fjsp_setzero_v2r8();
 191
 192         /* Start inner kernel loop */
 193         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 194         {
 195
 196             /* Get j neighbor index, and coordinate index */
 197             jnrA             = jjnr[jidx];
 198             jnrB             = jjnr[jidx+1];
 199             j_coord_offsetA  = DIM*jnrA;
 200             j_coord_offsetB  = DIM*jnrB;
 201
 202             /* load j atom coordinates */
 203             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 204                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 205
 206             /* Calculate displacement vector */
 207             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 208             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 209             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 210             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 211             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 212             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 213             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 214             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 215             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 216             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 217             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 218             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 219             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 220             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 221             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 222             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 223             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 224             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 225             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 226             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 227             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 228             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 229             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 230             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 231             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 232             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 233             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 234
 235             /* Calculate squared distance and things based on it */
 236             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 237             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 238             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 239             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 240             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 241             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 242             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 243             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 244             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 245
 246             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 247             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 248             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 249             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 250             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 251             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 252             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 253             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 254             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 255
 256             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 257             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 258             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 259             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 260             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 261             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 262             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 263             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 264             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 265
 266             fjx1             = _fjsp_setzero_v2r8();
 267             fjy1             = _fjsp_setzero_v2r8();
 268             fjz1             = _fjsp_setzero_v2r8();
 269             fjx2             = _fjsp_setzero_v2r8();
 270             fjy2             = _fjsp_setzero_v2r8();
 271             fjz2             = _fjsp_setzero_v2r8();
 272             fjx3             = _fjsp_setzero_v2r8();
 273             fjy3             = _fjsp_setzero_v2r8();
 274             fjz3             = _fjsp_setzero_v2r8();
 275
 276             /**************************
 277              * CALCULATE INTERACTIONS *
 278              **************************/
 279
 280             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 281             {
 282
 283             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 284
 285             /* EWALD ELECTROSTATICS */
 286
 287             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 288             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 289             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 290             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 291             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 292
 293             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 294             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 295             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 296             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 297             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 298             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 299             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 300             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 301             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv11,sh_ewald),velec));
 302             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 303
 304             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 305
 306             /* Update potential sum for this i atom from the interaction with this j atom. */
 307             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 308             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 309
 310             fscal            = felec;
 311
 312             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 313
 314             /* Update vectorial force */
 315             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 316             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 317             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 318
 319             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 320             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 321             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 322
 323             }
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 330             {
 331
 332             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 333
 334             /* EWALD ELECTROSTATICS */
 335
 336             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 337             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 338             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 339             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 340             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 341
 342             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 343             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 344             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 345             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 346             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 347             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 348             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 349             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 350             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv12,sh_ewald),velec));
 351             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 352
 353             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 357             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 358
 359             fscal            = felec;
 360
 361             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 362
 363             /* Update vectorial force */
 364             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 365             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 366             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 367
 368             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 369             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 370             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 371
 372             }
 373
 374             /**************************
 375              * CALCULATE INTERACTIONS *
 376              **************************/
 377
 378             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
 379             {
 380
 381             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 382
 383             /* EWALD ELECTROSTATICS */
 384
 385             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 386             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
 387             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 388             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 389             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 390
 391             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 392             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 393             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 394             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 395             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 396             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 397             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 398             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 399             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv13,sh_ewald),velec));
 400             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
 401
 402             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 406             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 411
 412             /* Update vectorial force */
 413             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 414             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 415             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 416
 417             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 418             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 419             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 420
 421             }
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 428             {
 429
 430             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 431
 432             /* EWALD ELECTROSTATICS */
 433
 434             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 435             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 436             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 437             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 438             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 439
 440             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 441             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 442             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 443             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 444             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 445             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 446             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 447             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 448             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv21,sh_ewald),velec));
 449             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 450
 451             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 452
 453             /* Update potential sum for this i atom from the interaction with this j atom. */
 454             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 455             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 456
 457             fscal            = felec;
 458
 459             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 460
 461             /* Update vectorial force */
 462             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 463             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 464             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 465
 466             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 467             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 468             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 469
 470             }
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 477             {
 478
 479             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 480
 481             /* EWALD ELECTROSTATICS */
 482
 483             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 484             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
 485             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 486             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 487             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 488
 489             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 490             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 491             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 492             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 493             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 494             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 495             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 496             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 497             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv22,sh_ewald),velec));
 498             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
 499
 500             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 501
 502             /* Update potential sum for this i atom from the interaction with this j atom. */
 503             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 504             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 505
 506             fscal            = felec;
 507
 508             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 509
 510             /* Update vectorial force */
 511             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 512             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 513             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 514
 515             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 516             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 517             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 518
 519             }
 520
 521             /**************************
 522              * CALCULATE INTERACTIONS *
 523              **************************/
 524
 525             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
 526             {
 527
 528             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
 529
 530             /* EWALD ELECTROSTATICS */
 531
 532             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 533             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
 534             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 535             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 536             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 537
 538             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 539             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 540             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 541             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 542             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 543             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 544             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 545             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 546             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv23,sh_ewald),velec));
 547             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
 548
 549             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
 550
 551             /* Update potential sum for this i atom from the interaction with this j atom. */
 552             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 553             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 554
 555             fscal            = felec;
 556
 557             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 558
 559             /* Update vectorial force */
 560             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 561             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 562             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 563
 564             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 565             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 566             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 567
 568             }
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
 575             {
 576
 577             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
 578
 579             /* EWALD ELECTROSTATICS */
 580
 581             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 582             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
 583             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 584             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 585             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 586
 587             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 588             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 589             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 590             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 591             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 592             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 593             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 594             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 595             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv31,sh_ewald),velec));
 596             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
 597
 598             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
 599
 600             /* Update potential sum for this i atom from the interaction with this j atom. */
 601             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 602             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 603
 604             fscal            = felec;
 605
 606             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 607
 608             /* Update vectorial force */
 609             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 610             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 611             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 612
 613             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 614             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 615             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 616
 617             }
 618
 619             /**************************
 620              * CALCULATE INTERACTIONS *
 621              **************************/
 622
 623             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
 624             {
 625
 626             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
 627
 628             /* EWALD ELECTROSTATICS */
 629
 630             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 631             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
 632             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 633             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 634             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 635
 636             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 637             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 638             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 639             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 640             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 641             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 642             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 643             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 644             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv32,sh_ewald),velec));
 645             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
 646
 647             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
 648
 649             /* Update potential sum for this i atom from the interaction with this j atom. */
 650             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 651             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 652
 653             fscal            = felec;
 654
 655             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 656
 657             /* Update vectorial force */
 658             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 659             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 660             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 661
 662             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 663             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 664             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 665
 666             }
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
 673             {
 674
 675             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
 676
 677             /* EWALD ELECTROSTATICS */
 678
 679             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 680             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
 681             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 682             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 683             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 684
 685             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 686             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 687             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 688             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 689             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 690             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 691             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 692             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 693             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv33,sh_ewald),velec));
 694             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
 695
 696             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
 697
 698             /* Update potential sum for this i atom from the interaction with this j atom. */
 699             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 700             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 701
 702             fscal            = felec;
 703
 704             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 705
 706             /* Update vectorial force */
 707             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 708             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 709             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 710
 711             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 712             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 713             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 714
 715             }
 716
 717             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);
 718
 719             /* Inner loop uses 441 flops */
 720         }
 721
 722         if(jidx<j_index_end)
 723         {
 724
 725             jnrA             = jjnr[jidx];
 726             j_coord_offsetA  = DIM*jnrA;
 727
 728             /* load j atom coordinates */
 729             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
 730                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 731
 732             /* Calculate displacement vector */
 733             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 734             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 735             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 736             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 737             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 738             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 739             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 740             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 741             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 742             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 743             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 744             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 745             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 746             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 747             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 748             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 749             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 750             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 751             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 752             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 753             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 754             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 755             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 756             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 757             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 758             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 759             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 760
 761             /* Calculate squared distance and things based on it */
 762             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 763             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 764             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 765             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 766             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 767             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 768             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 769             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 770             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 771
 772             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 773             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 774             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 775             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 776             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 777             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 778             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 779             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 780             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 781
 782             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 783             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 784             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 785             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 786             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 787             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 788             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 789             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 790             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 791
 792             fjx1             = _fjsp_setzero_v2r8();
 793             fjy1             = _fjsp_setzero_v2r8();
 794             fjz1             = _fjsp_setzero_v2r8();
 795             fjx2             = _fjsp_setzero_v2r8();
 796             fjy2             = _fjsp_setzero_v2r8();
 797             fjz2             = _fjsp_setzero_v2r8();
 798             fjx3             = _fjsp_setzero_v2r8();
 799             fjy3             = _fjsp_setzero_v2r8();
 800             fjz3             = _fjsp_setzero_v2r8();
 801
 802             /**************************
 803              * CALCULATE INTERACTIONS *
 804              **************************/
 805
 806             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 807             {
 808
 809             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 810
 811             /* EWALD ELECTROSTATICS */
 812
 813             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 814             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
 815             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 816             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 817             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 818
 819             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 820             ewtabD           = _fjsp_setzero_v2r8();
 821             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 822             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 823             ewtabFn          = _fjsp_setzero_v2r8();
 824             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 825             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 826             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 827             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv11,sh_ewald),velec));
 828             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
 829
 830             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 831
 832             /* Update potential sum for this i atom from the interaction with this j atom. */
 833             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 834             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 835             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 836
 837             fscal            = felec;
 838
 839             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 840
 841             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 842
 843             /* Update vectorial force */
 844             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 845             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 846             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 847
 848             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 849             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 850             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 851
 852             }
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 859             {
 860
 861             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 862
 863             /* EWALD ELECTROSTATICS */
 864
 865             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 866             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
 867             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 868             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 869             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 870
 871             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 872             ewtabD           = _fjsp_setzero_v2r8();
 873             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 874             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 875             ewtabFn          = _fjsp_setzero_v2r8();
 876             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 877             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 878             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 879             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv12,sh_ewald),velec));
 880             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
 881
 882             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 883
 884             /* Update potential sum for this i atom from the interaction with this j atom. */
 885             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 886             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 887             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 888
 889             fscal            = felec;
 890
 891             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 892
 893             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 894
 895             /* Update vectorial force */
 896             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 897             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 898             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 899
 900             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 901             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 902             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 903
 904             }
 905
 906             /**************************
 907              * CALCULATE INTERACTIONS *
 908              **************************/
 909
 910             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
 911             {
 912
 913             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 914
 915             /* EWALD ELECTROSTATICS */
 916
 917             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 918             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
 919             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 920             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 921             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 922
 923             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 924             ewtabD           = _fjsp_setzero_v2r8();
 925             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 926             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 927             ewtabFn          = _fjsp_setzero_v2r8();
 928             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 929             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 930             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 931             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv13,sh_ewald),velec));
 932             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
 933
 934             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
 935
 936             /* Update potential sum for this i atom from the interaction with this j atom. */
 937             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 938             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 939             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 940
 941             fscal            = felec;
 942
 943             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 944
 945             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 946
 947             /* Update vectorial force */
 948             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 949             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 950             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 951
 952             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 953             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 954             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 955
 956             }
 957
 958             /**************************
 959              * CALCULATE INTERACTIONS *
 960              **************************/
 961
 962             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 963             {
 964
 965             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 966
 967             /* EWALD ELECTROSTATICS */
 968
 969             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 970             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
 971             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 972             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 973             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 974
 975             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 976             ewtabD           = _fjsp_setzero_v2r8();
 977             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 978             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 979             ewtabFn          = _fjsp_setzero_v2r8();
 980             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 981             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 982             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 983             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv21,sh_ewald),velec));
 984             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
 985
 986             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 987
 988             /* Update potential sum for this i atom from the interaction with this j atom. */
 989             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 990             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 991             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 992
 993             fscal            = felec;
 994
 995             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 996
 997             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 998
 999             /* Update vectorial force */
1000             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1001             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1002             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1003
1004             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1005             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1006             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1007
1008             }
1009
1010             /**************************
1011              * CALCULATE INTERACTIONS *
1012              **************************/
1013
1014             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1015             {
1016
1017             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1018
1019             /* EWALD ELECTROSTATICS */
1020
1021             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1022             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1023             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1024             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1025             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1026
1027             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1028             ewtabD           = _fjsp_setzero_v2r8();
1029             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1030             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1031             ewtabFn          = _fjsp_setzero_v2r8();
1032             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1033             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1034             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1035             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv22,sh_ewald),velec));
1036             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1037
1038             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1039
1040             /* Update potential sum for this i atom from the interaction with this j atom. */
1041             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1042             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1043             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1044
1045             fscal            = felec;
1046
1047             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1048
1049             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1050
1051             /* Update vectorial force */
1052             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1053             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1054             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1055
1056             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1057             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1058             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1059
1060             }
1061
1062             /**************************
1063              * CALCULATE INTERACTIONS *
1064              **************************/
1065
1066             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
1067             {
1068
1069             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1070
1071             /* EWALD ELECTROSTATICS */
1072
1073             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1074             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
1075             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1076             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1077             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1078
1079             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1080             ewtabD           = _fjsp_setzero_v2r8();
1081             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1082             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1083             ewtabFn          = _fjsp_setzero_v2r8();
1084             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1085             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1086             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1087             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv23,sh_ewald),velec));
1088             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
1089
1090             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
1091
1092             /* Update potential sum for this i atom from the interaction with this j atom. */
1093             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1094             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1095             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1096
1097             fscal            = felec;
1098
1099             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1100
1101             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1102
1103             /* Update vectorial force */
1104             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1105             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1106             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1107
1108             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1109             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1110             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1111
1112             }
1113
1114             /**************************
1115              * CALCULATE INTERACTIONS *
1116              **************************/
1117
1118             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
1119             {
1120
1121             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1122
1123             /* EWALD ELECTROSTATICS */
1124
1125             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1126             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
1127             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1128             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1129             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1130
1131             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1132             ewtabD           = _fjsp_setzero_v2r8();
1133             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1134             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1135             ewtabFn          = _fjsp_setzero_v2r8();
1136             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1137             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1138             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1139             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv31,sh_ewald),velec));
1140             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
1141
1142             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
1143
1144             /* Update potential sum for this i atom from the interaction with this j atom. */
1145             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1146             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1147             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1148
1149             fscal            = felec;
1150
1151             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1152
1153             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1154
1155             /* Update vectorial force */
1156             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1157             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1158             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1159
1160             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1161             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1162             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1163
1164             }
1165
1166             /**************************
1167              * CALCULATE INTERACTIONS *
1168              **************************/
1169
1170             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
1171             {
1172
1173             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1174
1175             /* EWALD ELECTROSTATICS */
1176
1177             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1178             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
1179             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1180             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1181             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1182
1183             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1184             ewtabD           = _fjsp_setzero_v2r8();
1185             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1186             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1187             ewtabFn          = _fjsp_setzero_v2r8();
1188             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1189             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1190             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1191             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv32,sh_ewald),velec));
1192             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
1193
1194             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
1195
1196             /* Update potential sum for this i atom from the interaction with this j atom. */
1197             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1198             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1199             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1200
1201             fscal            = felec;
1202
1203             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1204
1205             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1206
1207             /* Update vectorial force */
1208             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1209             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1210             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1211
1212             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1213             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1214             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1215
1216             }
1217
1218             /**************************
1219              * CALCULATE INTERACTIONS *
1220              **************************/
1221
1222             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
1223             {
1224
1225             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1226
1227             /* EWALD ELECTROSTATICS */
1228
1229             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1230             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
1231             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1232             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1233             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1234
1235             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
1236             ewtabD           = _fjsp_setzero_v2r8();
1237             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
1238             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
1239             ewtabFn          = _fjsp_setzero_v2r8();
1240             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
1241             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
1242             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
1243             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_sub_v2r8(rinv33,sh_ewald),velec));
1244             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
1245
1246             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
1247
1248             /* Update potential sum for this i atom from the interaction with this j atom. */
1249             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1250             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1251             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1252
1253             fscal            = felec;
1254
1255             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1256
1257             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1258
1259             /* Update vectorial force */
1260             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1261             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1262             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1263
1264             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1265             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1266             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1267
1268             }
1269
1270             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1271
1272             /* Inner loop uses 441 flops */
1273         }
1274
1275         /* End of innermost loop */
1276
1277         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1278                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1279
1280         ggid                        = gid[iidx];
1281         /* Update potential energies */
1282         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1283
1284         /* Increment number of inner iterations */
1285         inneriter                  += j_index_end - j_index_start;
1286
1287         /* Outer loop uses 19 flops */
1288     }
1289
1290     /* Increment number of outer iterations */
1291     outeriter        += nri;
1292
1293     /* Update outer/inner flops */
1294
1295     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*441);
1296 }
1297 /*
1298  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1299  * Electrostatics interaction: Ewald
1300  * VdW interaction:            None
1301  * Geometry:                   Water4-Water4
1302  * Calculate force/pot:        Force
1303  */
1304 void
1305 nb_kernel_ElecEwSh_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1306                     (t_nblist                    * gmx_restrict       nlist,
1307                      rvec                        * gmx_restrict          xx,
1308                      rvec                        * gmx_restrict          ff,
1309                      t_forcerec                  * gmx_restrict          fr,
1310                      t_mdatoms                   * gmx_restrict     mdatoms,
1311                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1312                      t_nrnb                      * gmx_restrict        nrnb)
1313 {
1314     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1315      * just 0 for non-waters.
1316      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1317      * jnr indices corresponding to data put in the four positions in the SIMD register.
1318      */
1319     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1320     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1321     int              jnrA,jnrB;
1322     int              j_coord_offsetA,j_coord_offsetB;
1323     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1324     real             rcutoff_scalar;
1325     real             *shiftvec,*fshift,*x,*f;
1326     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1327     int              vdwioffset1;
1328     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1329     int              vdwioffset2;
1330     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1331     int              vdwioffset3;
1332     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1333     int              vdwjidx1A,vdwjidx1B;
1334     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1335     int              vdwjidx2A,vdwjidx2B;
1336     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1337     int              vdwjidx3A,vdwjidx3B;
1338     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1339     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1340     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1341     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1342     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1343     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1344     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1345     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1346     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1347     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1348     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1349     real             *charge;
1350     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1351     real             *ewtab;
1352     _fjsp_v2r8       itab_tmp;
1353     _fjsp_v2r8       dummy_mask,cutoff_mask;
1354     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1355     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1356     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1357
1358     x                = xx[0];
1359     f                = ff[0];
1360
1361     nri              = nlist->nri;
1362     iinr             = nlist->iinr;
1363     jindex           = nlist->jindex;
1364     jjnr             = nlist->jjnr;
1365     shiftidx         = nlist->shift;
1366     gid              = nlist->gid;
1367     shiftvec         = fr->shift_vec[0];
1368     fshift           = fr->fshift[0];
1369     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1370     charge           = mdatoms->chargeA;
1371
1372     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
1373     ewtab            = fr->ic->tabq_coul_F;
1374     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
1375     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
1376
1377     /* Setup water-specific parameters */
1378     inr              = nlist->iinr[0];
1379     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1380     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1381     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
1382
1383     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1384     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1385     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
1386     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1387     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1388     qq13             = _fjsp_mul_v2r8(iq1,jq3);
1389     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1390     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1391     qq23             = _fjsp_mul_v2r8(iq2,jq3);
1392     qq31             = _fjsp_mul_v2r8(iq3,jq1);
1393     qq32             = _fjsp_mul_v2r8(iq3,jq2);
1394     qq33             = _fjsp_mul_v2r8(iq3,jq3);
1395
1396     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1397     rcutoff_scalar   = fr->rcoulomb;
1398     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1399     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1400
1401     /* Avoid stupid compiler warnings */
1402     jnrA = jnrB = 0;
1403     j_coord_offsetA = 0;
1404     j_coord_offsetB = 0;
1405
1406     outeriter        = 0;
1407     inneriter        = 0;
1408
1409     /* Start outer loop over neighborlists */
1410     for(iidx=0; iidx<nri; iidx++)
1411     {
1412         /* Load shift vector for this list */
1413         i_shift_offset   = DIM*shiftidx[iidx];
1414
1415         /* Load limits for loop over neighbors */
1416         j_index_start    = jindex[iidx];
1417         j_index_end      = jindex[iidx+1];
1418
1419         /* Get outer coordinate index */
1420         inr              = iinr[iidx];
1421         i_coord_offset   = DIM*inr;
1422
1423         /* Load i particle coords and add shift vector */
1424         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1425                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1426
1427         fix1             = _fjsp_setzero_v2r8();
1428         fiy1             = _fjsp_setzero_v2r8();
1429         fiz1             = _fjsp_setzero_v2r8();
1430         fix2             = _fjsp_setzero_v2r8();
1431         fiy2             = _fjsp_setzero_v2r8();
1432         fiz2             = _fjsp_setzero_v2r8();
1433         fix3             = _fjsp_setzero_v2r8();
1434         fiy3             = _fjsp_setzero_v2r8();
1435         fiz3             = _fjsp_setzero_v2r8();
1436
1437         /* Start inner kernel loop */
1438         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1439         {
1440
1441             /* Get j neighbor index, and coordinate index */
1442             jnrA             = jjnr[jidx];
1443             jnrB             = jjnr[jidx+1];
1444             j_coord_offsetA  = DIM*jnrA;
1445             j_coord_offsetB  = DIM*jnrB;
1446
1447             /* load j atom coordinates */
1448             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1449                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1450
1451             /* Calculate displacement vector */
1452             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1453             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1454             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1455             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1456             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1457             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1458             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1459             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1460             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1461             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1462             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1463             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1464             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1465             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1466             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1467             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1468             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1469             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1470             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1471             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1472             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1473             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1474             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1475             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1476             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1477             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1478             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1479
1480             /* Calculate squared distance and things based on it */
1481             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1482             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1483             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1484             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1485             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1486             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1487             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1488             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1489             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1490
1491             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1492             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1493             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1494             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1495             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1496             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1497             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1498             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1499             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1500
1501             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1502             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1503             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1504             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1505             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1506             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1507             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1508             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1509             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1510
1511             fjx1             = _fjsp_setzero_v2r8();
1512             fjy1             = _fjsp_setzero_v2r8();
1513             fjz1             = _fjsp_setzero_v2r8();
1514             fjx2             = _fjsp_setzero_v2r8();
1515             fjy2             = _fjsp_setzero_v2r8();
1516             fjz2             = _fjsp_setzero_v2r8();
1517             fjx3             = _fjsp_setzero_v2r8();
1518             fjy3             = _fjsp_setzero_v2r8();
1519             fjz3             = _fjsp_setzero_v2r8();
1520
1521             /**************************
1522              * CALCULATE INTERACTIONS *
1523              **************************/
1524
1525             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1526             {
1527
1528             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1529
1530             /* EWALD ELECTROSTATICS */
1531
1532             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1533             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1534             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1535             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1536             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1537
1538             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1539                                          &ewtabF,&ewtabFn);
1540             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1541             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1542
1543             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1544
1545             fscal            = felec;
1546
1547             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1548
1549             /* Update vectorial force */
1550             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1551             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1552             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1553
1554             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1555             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1556             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1557
1558             }
1559
1560             /**************************
1561              * CALCULATE INTERACTIONS *
1562              **************************/
1563
1564             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1565             {
1566
1567             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1568
1569             /* EWALD ELECTROSTATICS */
1570
1571             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1572             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
1573             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1574             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1575             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1576
1577             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1578                                          &ewtabF,&ewtabFn);
1579             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1580             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
1581
1582             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1583
1584             fscal            = felec;
1585
1586             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1587
1588             /* Update vectorial force */
1589             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1590             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1591             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1592
1593             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1594             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1595             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1596
1597             }
1598
1599             /**************************
1600              * CALCULATE INTERACTIONS *
1601              **************************/
1602
1603             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
1604             {
1605
1606             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
1607
1608             /* EWALD ELECTROSTATICS */
1609
1610             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1611             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
1612             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1613             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1614             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1615
1616             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1617                                          &ewtabF,&ewtabFn);
1618             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1619             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
1620
1621             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
1622
1623             fscal            = felec;
1624
1625             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1626
1627             /* Update vectorial force */
1628             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1629             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1630             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1631
1632             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1633             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1634             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1635
1636             }
1637
1638             /**************************
1639              * CALCULATE INTERACTIONS *
1640              **************************/
1641
1642             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1643             {
1644
1645             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1646
1647             /* EWALD ELECTROSTATICS */
1648
1649             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1650             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
1651             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1652             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1653             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1654
1655             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1656                                          &ewtabF,&ewtabFn);
1657             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1658             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
1659
1660             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1661
1662             fscal            = felec;
1663
1664             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1665
1666             /* Update vectorial force */
1667             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1668             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1669             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1670
1671             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1672             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1673             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1674
1675             }
1676
1677             /**************************
1678              * CALCULATE INTERACTIONS *
1679              **************************/
1680
1681             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1682             {
1683
1684             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1685
1686             /* EWALD ELECTROSTATICS */
1687
1688             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1689             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
1690             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1691             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1692             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1693
1694             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1695                                          &ewtabF,&ewtabFn);
1696             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1697             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
1698
1699             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1700
1701             fscal            = felec;
1702
1703             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1704
1705             /* Update vectorial force */
1706             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1707             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1708             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1709
1710             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1711             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1712             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1713
1714             }
1715
1716             /**************************
1717              * CALCULATE INTERACTIONS *
1718              **************************/
1719
1720             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
1721             {
1722
1723             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1724
1725             /* EWALD ELECTROSTATICS */
1726
1727             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1728             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
1729             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1730             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1731             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1732
1733             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1734                                          &ewtabF,&ewtabFn);
1735             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1736             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
1737
1738             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
1739
1740             fscal            = felec;
1741
1742             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1743
1744             /* Update vectorial force */
1745             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1746             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1747             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1748
1749             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1750             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1751             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1752
1753             }
1754
1755             /**************************
1756              * CALCULATE INTERACTIONS *
1757              **************************/
1758
1759             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
1760             {
1761
1762             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1763
1764             /* EWALD ELECTROSTATICS */
1765
1766             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1767             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
1768             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1769             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1770             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1771
1772             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1773                                          &ewtabF,&ewtabFn);
1774             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1775             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
1776
1777             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
1778
1779             fscal            = felec;
1780
1781             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1782
1783             /* Update vectorial force */
1784             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1785             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1786             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1787
1788             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1789             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1790             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1791
1792             }
1793
1794             /**************************
1795              * CALCULATE INTERACTIONS *
1796              **************************/
1797
1798             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
1799             {
1800
1801             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1802
1803             /* EWALD ELECTROSTATICS */
1804
1805             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1806             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
1807             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1808             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1809             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1810
1811             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1812                                          &ewtabF,&ewtabFn);
1813             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1814             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
1815
1816             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
1817
1818             fscal            = felec;
1819
1820             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1821
1822             /* Update vectorial force */
1823             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1824             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1825             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1826
1827             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1828             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1829             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1830
1831             }
1832
1833             /**************************
1834              * CALCULATE INTERACTIONS *
1835              **************************/
1836
1837             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
1838             {
1839
1840             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1841
1842             /* EWALD ELECTROSTATICS */
1843
1844             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1845             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
1846             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1847             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1848             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1849
1850             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
1851                                          &ewtabF,&ewtabFn);
1852             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1853             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
1854
1855             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
1856
1857             fscal            = felec;
1858
1859             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1860
1861             /* Update vectorial force */
1862             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1863             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1864             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1865
1866             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1867             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1868             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1869
1870             }
1871
1872             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);
1873
1874             /* Inner loop uses 378 flops */
1875         }
1876
1877         if(jidx<j_index_end)
1878         {
1879
1880             jnrA             = jjnr[jidx];
1881             j_coord_offsetA  = DIM*jnrA;
1882
1883             /* load j atom coordinates */
1884             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1885                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1886
1887             /* Calculate displacement vector */
1888             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1889             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1890             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1891             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1892             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1893             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1894             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1895             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1896             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1897             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1898             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1899             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1900             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1901             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1902             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1903             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1904             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1905             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1906             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1907             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1908             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1909             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1910             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1911             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1912             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1913             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1914             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1915
1916             /* Calculate squared distance and things based on it */
1917             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1918             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1919             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1920             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1921             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1922             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1923             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1924             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1925             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1926
1927             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1928             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1929             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1930             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1931             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1932             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1933             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1934             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1935             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1936
1937             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1938             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1939             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1940             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1941             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1942             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1943             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1944             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1945             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1946
1947             fjx1             = _fjsp_setzero_v2r8();
1948             fjy1             = _fjsp_setzero_v2r8();
1949             fjz1             = _fjsp_setzero_v2r8();
1950             fjx2             = _fjsp_setzero_v2r8();
1951             fjy2             = _fjsp_setzero_v2r8();
1952             fjz2             = _fjsp_setzero_v2r8();
1953             fjx3             = _fjsp_setzero_v2r8();
1954             fjy3             = _fjsp_setzero_v2r8();
1955             fjz3             = _fjsp_setzero_v2r8();
1956
1957             /**************************
1958              * CALCULATE INTERACTIONS *
1959              **************************/
1960
1961             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1962             {
1963
1964             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1965
1966             /* EWALD ELECTROSTATICS */
1967
1968             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1969             ewrt             = _fjsp_mul_v2r8(r11,ewtabscale);
1970             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1971             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1972             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1973
1974             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1975             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1976             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,rinv11),_fjsp_sub_v2r8(rinvsq11,felec));
1977
1978             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1979
1980             fscal            = felec;
1981
1982             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1983
1984             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1985
1986             /* Update vectorial force */
1987             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1988             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1989             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1990
1991             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1992             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1993             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1994
1995             }
1996
1997             /**************************
1998              * CALCULATE INTERACTIONS *
1999              **************************/
2000
2001             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
2002             {
2003
2004             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
2005
2006             /* EWALD ELECTROSTATICS */
2007
2008             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2009             ewrt             = _fjsp_mul_v2r8(r12,ewtabscale);
2010             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2011             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2012             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2013
2014             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2015             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2016             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,rinv12),_fjsp_sub_v2r8(rinvsq12,felec));
2017
2018             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
2019
2020             fscal            = felec;
2021
2022             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2023
2024             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2025
2026             /* Update vectorial force */
2027             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
2028             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
2029             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
2030
2031             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
2032             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
2033             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
2034
2035             }
2036
2037             /**************************
2038              * CALCULATE INTERACTIONS *
2039              **************************/
2040
2041             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
2042             {
2043
2044             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
2045
2046             /* EWALD ELECTROSTATICS */
2047
2048             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2049             ewrt             = _fjsp_mul_v2r8(r13,ewtabscale);
2050             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2051             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2052             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2053
2054             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2055             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2056             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,rinv13),_fjsp_sub_v2r8(rinvsq13,felec));
2057
2058             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
2059
2060             fscal            = felec;
2061
2062             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2063
2064             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2065
2066             /* Update vectorial force */
2067             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
2068             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
2069             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
2070
2071             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
2072             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
2073             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
2074
2075             }
2076
2077             /**************************
2078              * CALCULATE INTERACTIONS *
2079              **************************/
2080
2081             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
2082             {
2083
2084             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
2085
2086             /* EWALD ELECTROSTATICS */
2087
2088             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2089             ewrt             = _fjsp_mul_v2r8(r21,ewtabscale);
2090             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2091             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2092             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2093
2094             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2095             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2096             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,rinv21),_fjsp_sub_v2r8(rinvsq21,felec));
2097
2098             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
2099
2100             fscal            = felec;
2101
2102             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2103
2104             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2105
2106             /* Update vectorial force */
2107             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
2108             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
2109             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
2110
2111             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
2112             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
2113             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
2114
2115             }
2116
2117             /**************************
2118              * CALCULATE INTERACTIONS *
2119              **************************/
2120
2121             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
2122             {
2123
2124             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
2125
2126             /* EWALD ELECTROSTATICS */
2127
2128             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2129             ewrt             = _fjsp_mul_v2r8(r22,ewtabscale);
2130             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2131             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2132             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2133
2134             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2135             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2136             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,rinv22),_fjsp_sub_v2r8(rinvsq22,felec));
2137
2138             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
2139
2140             fscal            = felec;
2141
2142             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2143
2144             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2145
2146             /* Update vectorial force */
2147             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
2148             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
2149             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
2150
2151             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
2152             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
2153             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
2154
2155             }
2156
2157             /**************************
2158              * CALCULATE INTERACTIONS *
2159              **************************/
2160
2161             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
2162             {
2163
2164             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
2165
2166             /* EWALD ELECTROSTATICS */
2167
2168             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2169             ewrt             = _fjsp_mul_v2r8(r23,ewtabscale);
2170             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2171             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2172             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2173
2174             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2175             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2176             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,rinv23),_fjsp_sub_v2r8(rinvsq23,felec));
2177
2178             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
2179
2180             fscal            = felec;
2181
2182             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2183
2184             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2185
2186             /* Update vectorial force */
2187             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
2188             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
2189             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
2190
2191             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
2192             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
2193             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
2194
2195             }
2196
2197             /**************************
2198              * CALCULATE INTERACTIONS *
2199              **************************/
2200
2201             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
2202             {
2203
2204             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
2205
2206             /* EWALD ELECTROSTATICS */
2207
2208             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2209             ewrt             = _fjsp_mul_v2r8(r31,ewtabscale);
2210             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2211             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2212             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2213
2214             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2215             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2216             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,rinv31),_fjsp_sub_v2r8(rinvsq31,felec));
2217
2218             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
2219
2220             fscal            = felec;
2221
2222             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2223
2224             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2225
2226             /* Update vectorial force */
2227             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
2228             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
2229             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
2230
2231             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
2232             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
2233             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
2234
2235             }
2236
2237             /**************************
2238              * CALCULATE INTERACTIONS *
2239              **************************/
2240
2241             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
2242             {
2243
2244             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
2245
2246             /* EWALD ELECTROSTATICS */
2247
2248             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2249             ewrt             = _fjsp_mul_v2r8(r32,ewtabscale);
2250             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2251             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2252             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2253
2254             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2255             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2256             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,rinv32),_fjsp_sub_v2r8(rinvsq32,felec));
2257
2258             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
2259
2260             fscal            = felec;
2261
2262             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2263
2264             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2265
2266             /* Update vectorial force */
2267             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
2268             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
2269             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
2270
2271             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
2272             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
2273             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
2274
2275             }
2276
2277             /**************************
2278              * CALCULATE INTERACTIONS *
2279              **************************/
2280
2281             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
2282             {
2283
2284             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
2285
2286             /* EWALD ELECTROSTATICS */
2287
2288             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2289             ewrt             = _fjsp_mul_v2r8(r33,ewtabscale);
2290             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
2291             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
2292             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
2293
2294             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
2295             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
2296             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,rinv33),_fjsp_sub_v2r8(rinvsq33,felec));
2297
2298             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
2299
2300             fscal            = felec;
2301
2302             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2303
2304             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2305
2306             /* Update vectorial force */
2307             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
2308             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
2309             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
2310
2311             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
2312             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
2313             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
2314
2315             }
2316
2317             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2318
2319             /* Inner loop uses 378 flops */
2320         }
2321
2322         /* End of innermost loop */
2323
2324         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2325                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2326
2327         /* Increment number of inner iterations */
2328         inneriter                  += j_index_end - j_index_start;
2329
2330         /* Outer loop uses 18 flops */
2331     }
2332
2333     /* Increment number of outer iterations */
2334     outeriter        += nri;
2335
2336     /* Update outer/inner flops */
2337
2338     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*378);
2339 }