src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecCSTab_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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "gromacs/legacyheaders/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: CubicSplineTable
  54  * VdW interaction:            None
  55  * Geometry:                   Water4-Water4
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecCSTab_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset1;
  82     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwioffset3;
  86     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  87     int              vdwjidx1A,vdwjidx1B;
  88     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     int              vdwjidx3A,vdwjidx3B;
  92     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  93     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  94     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  95     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  96     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  97     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  98     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  99     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 100     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 101     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 102     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 103     real             *charge;
 104     _fjsp_v2r8       rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 105     real             *vftab;
 106     _fjsp_v2r8       itab_tmp;
 107     _fjsp_v2r8       dummy_mask,cutoff_mask;
 108     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 109     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 110     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 111
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 124     charge           = mdatoms->chargeA;
 125
 126     vftab            = kernel_data->table_elec->data;
 127     vftabscale       = gmx_fjsp_set1_v2r8(kernel_data->table_elec->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     /* Avoid stupid compiler warnings */
 149     jnrA = jnrB = 0;
 150     j_coord_offsetA = 0;
 151     j_coord_offsetB = 0;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 172                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 173
 174         fix1             = _fjsp_setzero_v2r8();
 175         fiy1             = _fjsp_setzero_v2r8();
 176         fiz1             = _fjsp_setzero_v2r8();
 177         fix2             = _fjsp_setzero_v2r8();
 178         fiy2             = _fjsp_setzero_v2r8();
 179         fiz2             = _fjsp_setzero_v2r8();
 180         fix3             = _fjsp_setzero_v2r8();
 181         fiy3             = _fjsp_setzero_v2r8();
 182         fiz3             = _fjsp_setzero_v2r8();
 183
 184         /* Reset potential sums */
 185         velecsum         = _fjsp_setzero_v2r8();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196
 197             /* load j atom coordinates */
 198             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 199                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 200
 201             /* Calculate displacement vector */
 202             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 203             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 204             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 205             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 206             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 207             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 208             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 209             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 210             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 211             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 212             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 213             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 214             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 215             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 216             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 217             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 218             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 219             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 220             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 221             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 222             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 223             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 224             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 225             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 226             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 227             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 228             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 229
 230             /* Calculate squared distance and things based on it */
 231             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 232             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 233             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 234             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 235             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 236             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 237             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 238             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 239             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 240
 241             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 242             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 243             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 244             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 245             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 246             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 247             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 248             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 249             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 250
 251             fjx1             = _fjsp_setzero_v2r8();
 252             fjy1             = _fjsp_setzero_v2r8();
 253             fjz1             = _fjsp_setzero_v2r8();
 254             fjx2             = _fjsp_setzero_v2r8();
 255             fjy2             = _fjsp_setzero_v2r8();
 256             fjz2             = _fjsp_setzero_v2r8();
 257             fjx3             = _fjsp_setzero_v2r8();
 258             fjy3             = _fjsp_setzero_v2r8();
 259             fjz3             = _fjsp_setzero_v2r8();
 260
 261             /**************************
 262              * CALCULATE INTERACTIONS *
 263              **************************/
 264
 265             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 266
 267             /* Calculate table index by multiplying r with table scale and truncate to integer */
 268             rt               = _fjsp_mul_v2r8(r11,vftabscale);
 269             itab_tmp         = _fjsp_dtox_v2r8(rt);
 270             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 271             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 272             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 273
 274             vfconv.i[0]     *= 4;
 275             vfconv.i[1]     *= 4;
 276
 277             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 278             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 279             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 280             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 281             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 282             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 283             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 284             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 285             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 286             velec            = _fjsp_mul_v2r8(qq11,VV);
 287             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 288             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
 289
 290             /* Update potential sum for this i atom from the interaction with this j atom. */
 291             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 292
 293             fscal            = felec;
 294
 295             /* Update vectorial force */
 296             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 297             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 298             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 299
 300             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 301             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 302             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 309
 310             /* Calculate table index by multiplying r with table scale and truncate to integer */
 311             rt               = _fjsp_mul_v2r8(r12,vftabscale);
 312             itab_tmp         = _fjsp_dtox_v2r8(rt);
 313             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 314             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 315             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 316
 317             vfconv.i[0]     *= 4;
 318             vfconv.i[1]     *= 4;
 319
 320             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 321             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 322             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 323             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 324             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 325             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 326             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 327             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 328             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 329             velec            = _fjsp_mul_v2r8(qq12,VV);
 330             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 331             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 335
 336             fscal            = felec;
 337
 338             /* Update vectorial force */
 339             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 340             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 341             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 342
 343             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 344             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 345             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 346
 347             /**************************
 348              * CALCULATE INTERACTIONS *
 349              **************************/
 350
 351             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 352
 353             /* Calculate table index by multiplying r with table scale and truncate to integer */
 354             rt               = _fjsp_mul_v2r8(r13,vftabscale);
 355             itab_tmp         = _fjsp_dtox_v2r8(rt);
 356             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 357             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 358             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 359
 360             vfconv.i[0]     *= 4;
 361             vfconv.i[1]     *= 4;
 362
 363             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 364             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 365             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 366             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 367             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 368             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 369             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 370             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 371             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 372             velec            = _fjsp_mul_v2r8(qq13,VV);
 373             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 374             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 378
 379             fscal            = felec;
 380
 381             /* Update vectorial force */
 382             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 383             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 384             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 385
 386             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 387             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 388             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 389
 390             /**************************
 391              * CALCULATE INTERACTIONS *
 392              **************************/
 393
 394             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 395
 396             /* Calculate table index by multiplying r with table scale and truncate to integer */
 397             rt               = _fjsp_mul_v2r8(r21,vftabscale);
 398             itab_tmp         = _fjsp_dtox_v2r8(rt);
 399             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 400             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 401             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 402
 403             vfconv.i[0]     *= 4;
 404             vfconv.i[1]     *= 4;
 405
 406             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 407             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 408             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 409             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 410             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 411             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 412             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 413             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 414             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 415             velec            = _fjsp_mul_v2r8(qq21,VV);
 416             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 417             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 421
 422             fscal            = felec;
 423
 424             /* Update vectorial force */
 425             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 426             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 427             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 428
 429             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 430             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 431             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 438
 439             /* Calculate table index by multiplying r with table scale and truncate to integer */
 440             rt               = _fjsp_mul_v2r8(r22,vftabscale);
 441             itab_tmp         = _fjsp_dtox_v2r8(rt);
 442             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 443             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 444             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 445
 446             vfconv.i[0]     *= 4;
 447             vfconv.i[1]     *= 4;
 448
 449             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 450             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 451             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 452             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 453             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 454             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 455             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 456             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 457             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 458             velec            = _fjsp_mul_v2r8(qq22,VV);
 459             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 460             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
 461
 462             /* Update potential sum for this i atom from the interaction with this j atom. */
 463             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467             /* Update vectorial force */
 468             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 469             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 470             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 471
 472             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 473             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 474             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 475
 476             /**************************
 477              * CALCULATE INTERACTIONS *
 478              **************************/
 479
 480             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
 481
 482             /* Calculate table index by multiplying r with table scale and truncate to integer */
 483             rt               = _fjsp_mul_v2r8(r23,vftabscale);
 484             itab_tmp         = _fjsp_dtox_v2r8(rt);
 485             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 486             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 487             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 488
 489             vfconv.i[0]     *= 4;
 490             vfconv.i[1]     *= 4;
 491
 492             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 493             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 494             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 495             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 496             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 497             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 498             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 499             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 500             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 501             velec            = _fjsp_mul_v2r8(qq23,VV);
 502             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 503             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
 504
 505             /* Update potential sum for this i atom from the interaction with this j atom. */
 506             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 507
 508             fscal            = felec;
 509
 510             /* Update vectorial force */
 511             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 512             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 513             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 514
 515             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 516             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 517             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 518
 519             /**************************
 520              * CALCULATE INTERACTIONS *
 521              **************************/
 522
 523             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
 524
 525             /* Calculate table index by multiplying r with table scale and truncate to integer */
 526             rt               = _fjsp_mul_v2r8(r31,vftabscale);
 527             itab_tmp         = _fjsp_dtox_v2r8(rt);
 528             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 529             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 530             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 531
 532             vfconv.i[0]     *= 4;
 533             vfconv.i[1]     *= 4;
 534
 535             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 536             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 537             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 538             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 539             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 540             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 541             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 542             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 543             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 544             velec            = _fjsp_mul_v2r8(qq31,VV);
 545             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 546             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
 547
 548             /* Update potential sum for this i atom from the interaction with this j atom. */
 549             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 550
 551             fscal            = felec;
 552
 553             /* Update vectorial force */
 554             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 555             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 556             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 557
 558             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 559             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 560             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 561
 562             /**************************
 563              * CALCULATE INTERACTIONS *
 564              **************************/
 565
 566             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
 567
 568             /* Calculate table index by multiplying r with table scale and truncate to integer */
 569             rt               = _fjsp_mul_v2r8(r32,vftabscale);
 570             itab_tmp         = _fjsp_dtox_v2r8(rt);
 571             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 572             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 573             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 574
 575             vfconv.i[0]     *= 4;
 576             vfconv.i[1]     *= 4;
 577
 578             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 579             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 580             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 581             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 582             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 583             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 584             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 585             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 586             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 587             velec            = _fjsp_mul_v2r8(qq32,VV);
 588             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 589             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
 590
 591             /* Update potential sum for this i atom from the interaction with this j atom. */
 592             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 593
 594             fscal            = felec;
 595
 596             /* Update vectorial force */
 597             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 598             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 599             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 600
 601             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 602             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 603             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 604
 605             /**************************
 606              * CALCULATE INTERACTIONS *
 607              **************************/
 608
 609             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
 610
 611             /* Calculate table index by multiplying r with table scale and truncate to integer */
 612             rt               = _fjsp_mul_v2r8(r33,vftabscale);
 613             itab_tmp         = _fjsp_dtox_v2r8(rt);
 614             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 615             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 616             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 617
 618             vfconv.i[0]     *= 4;
 619             vfconv.i[1]     *= 4;
 620
 621             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 622             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 623             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 624             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 625             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 626             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 627             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 628             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 629             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 630             velec            = _fjsp_mul_v2r8(qq33,VV);
 631             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 632             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
 633
 634             /* Update potential sum for this i atom from the interaction with this j atom. */
 635             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 636
 637             fscal            = felec;
 638
 639             /* Update vectorial force */
 640             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 641             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 642             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 643
 644             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 645             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 646             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 647
 648             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);
 649
 650             /* Inner loop uses 414 flops */
 651         }
 652
 653         if(jidx<j_index_end)
 654         {
 655
 656             jnrA             = jjnr[jidx];
 657             j_coord_offsetA  = DIM*jnrA;
 658
 659             /* load j atom coordinates */
 660             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
 661                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 662
 663             /* Calculate displacement vector */
 664             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 665             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 666             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 667             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 668             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 669             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 670             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 671             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 672             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 673             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 674             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 675             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 676             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 677             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 678             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 679             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 680             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 681             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 682             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 683             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 684             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 685             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 686             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 687             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 688             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 689             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 690             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 691
 692             /* Calculate squared distance and things based on it */
 693             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 694             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 695             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 696             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 697             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 698             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 699             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 700             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 701             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 702
 703             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 704             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 705             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 706             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 707             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 708             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 709             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 710             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 711             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 712
 713             fjx1             = _fjsp_setzero_v2r8();
 714             fjy1             = _fjsp_setzero_v2r8();
 715             fjz1             = _fjsp_setzero_v2r8();
 716             fjx2             = _fjsp_setzero_v2r8();
 717             fjy2             = _fjsp_setzero_v2r8();
 718             fjz2             = _fjsp_setzero_v2r8();
 719             fjx3             = _fjsp_setzero_v2r8();
 720             fjy3             = _fjsp_setzero_v2r8();
 721             fjz3             = _fjsp_setzero_v2r8();
 722
 723             /**************************
 724              * CALCULATE INTERACTIONS *
 725              **************************/
 726
 727             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 728
 729             /* Calculate table index by multiplying r with table scale and truncate to integer */
 730             rt               = _fjsp_mul_v2r8(r11,vftabscale);
 731             itab_tmp         = _fjsp_dtox_v2r8(rt);
 732             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 733             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 734             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 735
 736             vfconv.i[0]     *= 4;
 737             vfconv.i[1]     *= 4;
 738
 739             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 740             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 741             F                = _fjsp_setzero_v2r8();
 742             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 743             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 744             H                = _fjsp_setzero_v2r8();
 745             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 746             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 747             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 748             velec            = _fjsp_mul_v2r8(qq11,VV);
 749             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 750             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
 751
 752             /* Update potential sum for this i atom from the interaction with this j atom. */
 753             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 754             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 755
 756             fscal            = felec;
 757
 758             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 759
 760             /* Update vectorial force */
 761             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 762             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 763             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 764
 765             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 766             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 767             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 774
 775             /* Calculate table index by multiplying r with table scale and truncate to integer */
 776             rt               = _fjsp_mul_v2r8(r12,vftabscale);
 777             itab_tmp         = _fjsp_dtox_v2r8(rt);
 778             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 779             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 780             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 781
 782             vfconv.i[0]     *= 4;
 783             vfconv.i[1]     *= 4;
 784
 785             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 786             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 787             F                = _fjsp_setzero_v2r8();
 788             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 789             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 790             H                = _fjsp_setzero_v2r8();
 791             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 792             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 793             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 794             velec            = _fjsp_mul_v2r8(qq12,VV);
 795             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 796             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
 797
 798             /* Update potential sum for this i atom from the interaction with this j atom. */
 799             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 800             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 801
 802             fscal            = felec;
 803
 804             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 805
 806             /* Update vectorial force */
 807             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 808             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 809             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 810
 811             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 812             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 813             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 814
 815             /**************************
 816              * CALCULATE INTERACTIONS *
 817              **************************/
 818
 819             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 820
 821             /* Calculate table index by multiplying r with table scale and truncate to integer */
 822             rt               = _fjsp_mul_v2r8(r13,vftabscale);
 823             itab_tmp         = _fjsp_dtox_v2r8(rt);
 824             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 825             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 826             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 827
 828             vfconv.i[0]     *= 4;
 829             vfconv.i[1]     *= 4;
 830
 831             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 832             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 833             F                = _fjsp_setzero_v2r8();
 834             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 835             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 836             H                = _fjsp_setzero_v2r8();
 837             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 838             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 839             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 840             velec            = _fjsp_mul_v2r8(qq13,VV);
 841             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 842             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
 843
 844             /* Update potential sum for this i atom from the interaction with this j atom. */
 845             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 846             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 847
 848             fscal            = felec;
 849
 850             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 851
 852             /* Update vectorial force */
 853             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 854             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 855             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 856
 857             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 858             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 859             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 860
 861             /**************************
 862              * CALCULATE INTERACTIONS *
 863              **************************/
 864
 865             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 866
 867             /* Calculate table index by multiplying r with table scale and truncate to integer */
 868             rt               = _fjsp_mul_v2r8(r21,vftabscale);
 869             itab_tmp         = _fjsp_dtox_v2r8(rt);
 870             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 871             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 872             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 873
 874             vfconv.i[0]     *= 4;
 875             vfconv.i[1]     *= 4;
 876
 877             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 878             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 879             F                = _fjsp_setzero_v2r8();
 880             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 881             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 882             H                = _fjsp_setzero_v2r8();
 883             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 884             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 885             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 886             velec            = _fjsp_mul_v2r8(qq21,VV);
 887             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 888             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
 889
 890             /* Update potential sum for this i atom from the interaction with this j atom. */
 891             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 892             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 893
 894             fscal            = felec;
 895
 896             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 897
 898             /* Update vectorial force */
 899             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 900             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 901             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 902
 903             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 904             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 905             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 906
 907             /**************************
 908              * CALCULATE INTERACTIONS *
 909              **************************/
 910
 911             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 912
 913             /* Calculate table index by multiplying r with table scale and truncate to integer */
 914             rt               = _fjsp_mul_v2r8(r22,vftabscale);
 915             itab_tmp         = _fjsp_dtox_v2r8(rt);
 916             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 917             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 918             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 919
 920             vfconv.i[0]     *= 4;
 921             vfconv.i[1]     *= 4;
 922
 923             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 924             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 925             F                = _fjsp_setzero_v2r8();
 926             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 927             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 928             H                = _fjsp_setzero_v2r8();
 929             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 930             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 931             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 932             velec            = _fjsp_mul_v2r8(qq22,VV);
 933             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 934             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
 935
 936             /* Update potential sum for this i atom from the interaction with this j atom. */
 937             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 938             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 939
 940             fscal            = felec;
 941
 942             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 943
 944             /* Update vectorial force */
 945             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 946             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 947             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 948
 949             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 950             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 951             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 952
 953             /**************************
 954              * CALCULATE INTERACTIONS *
 955              **************************/
 956
 957             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
 958
 959             /* Calculate table index by multiplying r with table scale and truncate to integer */
 960             rt               = _fjsp_mul_v2r8(r23,vftabscale);
 961             itab_tmp         = _fjsp_dtox_v2r8(rt);
 962             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 963             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 964             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 965
 966             vfconv.i[0]     *= 4;
 967             vfconv.i[1]     *= 4;
 968
 969             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 970             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 971             F                = _fjsp_setzero_v2r8();
 972             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 973             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 974             H                = _fjsp_setzero_v2r8();
 975             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 976             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 977             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 978             velec            = _fjsp_mul_v2r8(qq23,VV);
 979             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 980             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
 981
 982             /* Update potential sum for this i atom from the interaction with this j atom. */
 983             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 984             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 985
 986             fscal            = felec;
 987
 988             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 989
 990             /* Update vectorial force */
 991             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 992             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 993             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 994
 995             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 996             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 997             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 998
 999             /**************************
1000              * CALCULATE INTERACTIONS *
1001              **************************/
1002
1003             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1004
1005             /* Calculate table index by multiplying r with table scale and truncate to integer */
1006             rt               = _fjsp_mul_v2r8(r31,vftabscale);
1007             itab_tmp         = _fjsp_dtox_v2r8(rt);
1008             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1009             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1010             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1011
1012             vfconv.i[0]     *= 4;
1013             vfconv.i[1]     *= 4;
1014
1015             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1016             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1017             F                = _fjsp_setzero_v2r8();
1018             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1019             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1020             H                = _fjsp_setzero_v2r8();
1021             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1022             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1023             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
1024             velec            = _fjsp_mul_v2r8(qq31,VV);
1025             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1026             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
1027
1028             /* Update potential sum for this i atom from the interaction with this j atom. */
1029             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1030             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1031
1032             fscal            = felec;
1033
1034             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1035
1036             /* Update vectorial force */
1037             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1038             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1039             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1040
1041             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1042             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1043             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1044
1045             /**************************
1046              * CALCULATE INTERACTIONS *
1047              **************************/
1048
1049             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1050
1051             /* Calculate table index by multiplying r with table scale and truncate to integer */
1052             rt               = _fjsp_mul_v2r8(r32,vftabscale);
1053             itab_tmp         = _fjsp_dtox_v2r8(rt);
1054             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1055             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1056             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1057
1058             vfconv.i[0]     *= 4;
1059             vfconv.i[1]     *= 4;
1060
1061             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1062             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1063             F                = _fjsp_setzero_v2r8();
1064             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1065             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1066             H                = _fjsp_setzero_v2r8();
1067             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1068             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1069             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
1070             velec            = _fjsp_mul_v2r8(qq32,VV);
1071             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1072             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
1073
1074             /* Update potential sum for this i atom from the interaction with this j atom. */
1075             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1076             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1077
1078             fscal            = felec;
1079
1080             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1081
1082             /* Update vectorial force */
1083             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1084             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1085             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1086
1087             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1088             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1089             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1090
1091             /**************************
1092              * CALCULATE INTERACTIONS *
1093              **************************/
1094
1095             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1096
1097             /* Calculate table index by multiplying r with table scale and truncate to integer */
1098             rt               = _fjsp_mul_v2r8(r33,vftabscale);
1099             itab_tmp         = _fjsp_dtox_v2r8(rt);
1100             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1101             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1102             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1103
1104             vfconv.i[0]     *= 4;
1105             vfconv.i[1]     *= 4;
1106
1107             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1108             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1109             F                = _fjsp_setzero_v2r8();
1110             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1111             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1112             H                = _fjsp_setzero_v2r8();
1113             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1114             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1115             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
1116             velec            = _fjsp_mul_v2r8(qq33,VV);
1117             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1118             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
1119
1120             /* Update potential sum for this i atom from the interaction with this j atom. */
1121             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1122             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1123
1124             fscal            = felec;
1125
1126             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1127
1128             /* Update vectorial force */
1129             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1130             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1131             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1132
1133             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1134             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1135             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1136
1137             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1138
1139             /* Inner loop uses 414 flops */
1140         }
1141
1142         /* End of innermost loop */
1143
1144         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1145                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1146
1147         ggid                        = gid[iidx];
1148         /* Update potential energies */
1149         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1150
1151         /* Increment number of inner iterations */
1152         inneriter                  += j_index_end - j_index_start;
1153
1154         /* Outer loop uses 19 flops */
1155     }
1156
1157     /* Increment number of outer iterations */
1158     outeriter        += nri;
1159
1160     /* Update outer/inner flops */
1161
1162     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*414);
1163 }
1164 /*
1165  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1166  * Electrostatics interaction: CubicSplineTable
1167  * VdW interaction:            None
1168  * Geometry:                   Water4-Water4
1169  * Calculate force/pot:        Force
1170  */
1171 void
1172 nb_kernel_ElecCSTab_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1173                     (t_nblist                    * gmx_restrict       nlist,
1174                      rvec                        * gmx_restrict          xx,
1175                      rvec                        * gmx_restrict          ff,
1176                      t_forcerec                  * gmx_restrict          fr,
1177                      t_mdatoms                   * gmx_restrict     mdatoms,
1178                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1179                      t_nrnb                      * gmx_restrict        nrnb)
1180 {
1181     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1182      * just 0 for non-waters.
1183      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1184      * jnr indices corresponding to data put in the four positions in the SIMD register.
1185      */
1186     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1187     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1188     int              jnrA,jnrB;
1189     int              j_coord_offsetA,j_coord_offsetB;
1190     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1191     real             rcutoff_scalar;
1192     real             *shiftvec,*fshift,*x,*f;
1193     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1194     int              vdwioffset1;
1195     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1196     int              vdwioffset2;
1197     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1198     int              vdwioffset3;
1199     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1200     int              vdwjidx1A,vdwjidx1B;
1201     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1202     int              vdwjidx2A,vdwjidx2B;
1203     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1204     int              vdwjidx3A,vdwjidx3B;
1205     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1206     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1207     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1208     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1209     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1210     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1211     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1212     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1213     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1214     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1215     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1216     real             *charge;
1217     _fjsp_v2r8       rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1218     real             *vftab;
1219     _fjsp_v2r8       itab_tmp;
1220     _fjsp_v2r8       dummy_mask,cutoff_mask;
1221     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1222     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1223     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1224
1225     x                = xx[0];
1226     f                = ff[0];
1227
1228     nri              = nlist->nri;
1229     iinr             = nlist->iinr;
1230     jindex           = nlist->jindex;
1231     jjnr             = nlist->jjnr;
1232     shiftidx         = nlist->shift;
1233     gid              = nlist->gid;
1234     shiftvec         = fr->shift_vec[0];
1235     fshift           = fr->fshift[0];
1236     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1237     charge           = mdatoms->chargeA;
1238
1239     vftab            = kernel_data->table_elec->data;
1240     vftabscale       = gmx_fjsp_set1_v2r8(kernel_data->table_elec->scale);
1241
1242     /* Setup water-specific parameters */
1243     inr              = nlist->iinr[0];
1244     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1245     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1246     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
1247
1248     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1249     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1250     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
1251     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1252     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1253     qq13             = _fjsp_mul_v2r8(iq1,jq3);
1254     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1255     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1256     qq23             = _fjsp_mul_v2r8(iq2,jq3);
1257     qq31             = _fjsp_mul_v2r8(iq3,jq1);
1258     qq32             = _fjsp_mul_v2r8(iq3,jq2);
1259     qq33             = _fjsp_mul_v2r8(iq3,jq3);
1260
1261     /* Avoid stupid compiler warnings */
1262     jnrA = jnrB = 0;
1263     j_coord_offsetA = 0;
1264     j_coord_offsetB = 0;
1265
1266     outeriter        = 0;
1267     inneriter        = 0;
1268
1269     /* Start outer loop over neighborlists */
1270     for(iidx=0; iidx<nri; iidx++)
1271     {
1272         /* Load shift vector for this list */
1273         i_shift_offset   = DIM*shiftidx[iidx];
1274
1275         /* Load limits for loop over neighbors */
1276         j_index_start    = jindex[iidx];
1277         j_index_end      = jindex[iidx+1];
1278
1279         /* Get outer coordinate index */
1280         inr              = iinr[iidx];
1281         i_coord_offset   = DIM*inr;
1282
1283         /* Load i particle coords and add shift vector */
1284         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1285                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1286
1287         fix1             = _fjsp_setzero_v2r8();
1288         fiy1             = _fjsp_setzero_v2r8();
1289         fiz1             = _fjsp_setzero_v2r8();
1290         fix2             = _fjsp_setzero_v2r8();
1291         fiy2             = _fjsp_setzero_v2r8();
1292         fiz2             = _fjsp_setzero_v2r8();
1293         fix3             = _fjsp_setzero_v2r8();
1294         fiy3             = _fjsp_setzero_v2r8();
1295         fiz3             = _fjsp_setzero_v2r8();
1296
1297         /* Start inner kernel loop */
1298         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1299         {
1300
1301             /* Get j neighbor index, and coordinate index */
1302             jnrA             = jjnr[jidx];
1303             jnrB             = jjnr[jidx+1];
1304             j_coord_offsetA  = DIM*jnrA;
1305             j_coord_offsetB  = DIM*jnrB;
1306
1307             /* load j atom coordinates */
1308             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1309                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1310
1311             /* Calculate displacement vector */
1312             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1313             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1314             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1315             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1316             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1317             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1318             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1319             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1320             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1321             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1322             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1323             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1324             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1325             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1326             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1327             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1328             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1329             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1330             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1331             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1332             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1333             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1334             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1335             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1336             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1337             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1338             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1339
1340             /* Calculate squared distance and things based on it */
1341             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1342             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1343             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1344             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1345             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1346             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1347             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1348             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1349             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1350
1351             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1352             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1353             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1354             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1355             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1356             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1357             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1358             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1359             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1360
1361             fjx1             = _fjsp_setzero_v2r8();
1362             fjy1             = _fjsp_setzero_v2r8();
1363             fjz1             = _fjsp_setzero_v2r8();
1364             fjx2             = _fjsp_setzero_v2r8();
1365             fjy2             = _fjsp_setzero_v2r8();
1366             fjz2             = _fjsp_setzero_v2r8();
1367             fjx3             = _fjsp_setzero_v2r8();
1368             fjy3             = _fjsp_setzero_v2r8();
1369             fjz3             = _fjsp_setzero_v2r8();
1370
1371             /**************************
1372              * CALCULATE INTERACTIONS *
1373              **************************/
1374
1375             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1376
1377             /* Calculate table index by multiplying r with table scale and truncate to integer */
1378             rt               = _fjsp_mul_v2r8(r11,vftabscale);
1379             itab_tmp         = _fjsp_dtox_v2r8(rt);
1380             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1381             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1382             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1383
1384             vfconv.i[0]     *= 4;
1385             vfconv.i[1]     *= 4;
1386
1387             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1388             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1389             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1390             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1391             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1392             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1393             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1394             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1395             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1396             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
1397
1398             fscal            = felec;
1399
1400             /* Update vectorial force */
1401             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1402             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1403             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1404
1405             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1406             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1407             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1408
1409             /**************************
1410              * CALCULATE INTERACTIONS *
1411              **************************/
1412
1413             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1414
1415             /* Calculate table index by multiplying r with table scale and truncate to integer */
1416             rt               = _fjsp_mul_v2r8(r12,vftabscale);
1417             itab_tmp         = _fjsp_dtox_v2r8(rt);
1418             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1419             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1420             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1421
1422             vfconv.i[0]     *= 4;
1423             vfconv.i[1]     *= 4;
1424
1425             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1426             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1427             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1428             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1429             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1430             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1431             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1432             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1433             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1434             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
1435
1436             fscal            = felec;
1437
1438             /* Update vectorial force */
1439             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1440             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1441             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1442
1443             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1444             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1445             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1446
1447             /**************************
1448              * CALCULATE INTERACTIONS *
1449              **************************/
1450
1451             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
1452
1453             /* Calculate table index by multiplying r with table scale and truncate to integer */
1454             rt               = _fjsp_mul_v2r8(r13,vftabscale);
1455             itab_tmp         = _fjsp_dtox_v2r8(rt);
1456             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1457             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1458             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1459
1460             vfconv.i[0]     *= 4;
1461             vfconv.i[1]     *= 4;
1462
1463             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1464             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1465             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1466             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1467             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1468             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1469             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1470             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1471             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1472             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
1473
1474             fscal            = felec;
1475
1476             /* Update vectorial force */
1477             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1478             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1479             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1480
1481             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1482             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1483             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1484
1485             /**************************
1486              * CALCULATE INTERACTIONS *
1487              **************************/
1488
1489             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1490
1491             /* Calculate table index by multiplying r with table scale and truncate to integer */
1492             rt               = _fjsp_mul_v2r8(r21,vftabscale);
1493             itab_tmp         = _fjsp_dtox_v2r8(rt);
1494             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1495             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1496             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1497
1498             vfconv.i[0]     *= 4;
1499             vfconv.i[1]     *= 4;
1500
1501             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1502             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1503             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1504             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1505             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1506             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1507             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1508             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1509             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1510             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
1511
1512             fscal            = felec;
1513
1514             /* Update vectorial force */
1515             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1516             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1517             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1518
1519             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1520             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1521             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1522
1523             /**************************
1524              * CALCULATE INTERACTIONS *
1525              **************************/
1526
1527             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1528
1529             /* Calculate table index by multiplying r with table scale and truncate to integer */
1530             rt               = _fjsp_mul_v2r8(r22,vftabscale);
1531             itab_tmp         = _fjsp_dtox_v2r8(rt);
1532             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1533             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1534             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1535
1536             vfconv.i[0]     *= 4;
1537             vfconv.i[1]     *= 4;
1538
1539             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1540             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1541             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1542             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1543             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1544             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1545             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1546             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1547             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1548             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
1549
1550             fscal            = felec;
1551
1552             /* Update vectorial force */
1553             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1554             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1555             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1556
1557             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1558             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1559             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1560
1561             /**************************
1562              * CALCULATE INTERACTIONS *
1563              **************************/
1564
1565             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1566
1567             /* Calculate table index by multiplying r with table scale and truncate to integer */
1568             rt               = _fjsp_mul_v2r8(r23,vftabscale);
1569             itab_tmp         = _fjsp_dtox_v2r8(rt);
1570             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1571             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1572             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1573
1574             vfconv.i[0]     *= 4;
1575             vfconv.i[1]     *= 4;
1576
1577             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1578             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1579             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1580             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1581             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1582             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1583             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1584             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1585             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1586             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
1587
1588             fscal            = felec;
1589
1590             /* Update vectorial force */
1591             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1592             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1593             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1594
1595             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1596             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1597             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1598
1599             /**************************
1600              * CALCULATE INTERACTIONS *
1601              **************************/
1602
1603             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1604
1605             /* Calculate table index by multiplying r with table scale and truncate to integer */
1606             rt               = _fjsp_mul_v2r8(r31,vftabscale);
1607             itab_tmp         = _fjsp_dtox_v2r8(rt);
1608             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1609             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1610             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1611
1612             vfconv.i[0]     *= 4;
1613             vfconv.i[1]     *= 4;
1614
1615             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1616             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1617             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1618             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1619             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1620             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1621             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1622             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1623             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1624             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
1625
1626             fscal            = felec;
1627
1628             /* Update vectorial force */
1629             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1630             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1631             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1632
1633             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1634             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1635             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1636
1637             /**************************
1638              * CALCULATE INTERACTIONS *
1639              **************************/
1640
1641             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1642
1643             /* Calculate table index by multiplying r with table scale and truncate to integer */
1644             rt               = _fjsp_mul_v2r8(r32,vftabscale);
1645             itab_tmp         = _fjsp_dtox_v2r8(rt);
1646             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1647             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1648             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1649
1650             vfconv.i[0]     *= 4;
1651             vfconv.i[1]     *= 4;
1652
1653             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1654             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1655             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1656             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1657             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1658             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1659             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1660             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1661             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1662             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
1663
1664             fscal            = felec;
1665
1666             /* Update vectorial force */
1667             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1668             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1669             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1670
1671             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1672             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1673             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1674
1675             /**************************
1676              * CALCULATE INTERACTIONS *
1677              **************************/
1678
1679             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1680
1681             /* Calculate table index by multiplying r with table scale and truncate to integer */
1682             rt               = _fjsp_mul_v2r8(r33,vftabscale);
1683             itab_tmp         = _fjsp_dtox_v2r8(rt);
1684             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1685             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1686             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1687
1688             vfconv.i[0]     *= 4;
1689             vfconv.i[1]     *= 4;
1690
1691             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1692             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1693             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1694             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1695             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1696             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1697             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1698             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1699             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1700             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
1701
1702             fscal            = felec;
1703
1704             /* Update vectorial force */
1705             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1706             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1707             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1708
1709             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1710             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1711             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1712
1713             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);
1714
1715             /* Inner loop uses 378 flops */
1716         }
1717
1718         if(jidx<j_index_end)
1719         {
1720
1721             jnrA             = jjnr[jidx];
1722             j_coord_offsetA  = DIM*jnrA;
1723
1724             /* load j atom coordinates */
1725             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1726                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1727
1728             /* Calculate displacement vector */
1729             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1730             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1731             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1732             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1733             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1734             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1735             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1736             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1737             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1738             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1739             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1740             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1741             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1742             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1743             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1744             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1745             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1746             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1747             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1748             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1749             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1750             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1751             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1752             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1753             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1754             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1755             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1756
1757             /* Calculate squared distance and things based on it */
1758             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1759             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1760             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1761             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1762             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1763             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1764             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1765             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1766             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1767
1768             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1769             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1770             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1771             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1772             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1773             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1774             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1775             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1776             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1777
1778             fjx1             = _fjsp_setzero_v2r8();
1779             fjy1             = _fjsp_setzero_v2r8();
1780             fjz1             = _fjsp_setzero_v2r8();
1781             fjx2             = _fjsp_setzero_v2r8();
1782             fjy2             = _fjsp_setzero_v2r8();
1783             fjz2             = _fjsp_setzero_v2r8();
1784             fjx3             = _fjsp_setzero_v2r8();
1785             fjy3             = _fjsp_setzero_v2r8();
1786             fjz3             = _fjsp_setzero_v2r8();
1787
1788             /**************************
1789              * CALCULATE INTERACTIONS *
1790              **************************/
1791
1792             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1793
1794             /* Calculate table index by multiplying r with table scale and truncate to integer */
1795             rt               = _fjsp_mul_v2r8(r11,vftabscale);
1796             itab_tmp         = _fjsp_dtox_v2r8(rt);
1797             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1798             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1799             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1800
1801             vfconv.i[0]     *= 4;
1802             vfconv.i[1]     *= 4;
1803
1804             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1805             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1806             F                = _fjsp_setzero_v2r8();
1807             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1808             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1809             H                = _fjsp_setzero_v2r8();
1810             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1811             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1812             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1813             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
1814
1815             fscal            = felec;
1816
1817             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1818
1819             /* Update vectorial force */
1820             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1821             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1822             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1823
1824             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1825             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1826             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1827
1828             /**************************
1829              * CALCULATE INTERACTIONS *
1830              **************************/
1831
1832             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1833
1834             /* Calculate table index by multiplying r with table scale and truncate to integer */
1835             rt               = _fjsp_mul_v2r8(r12,vftabscale);
1836             itab_tmp         = _fjsp_dtox_v2r8(rt);
1837             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1838             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1839             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1840
1841             vfconv.i[0]     *= 4;
1842             vfconv.i[1]     *= 4;
1843
1844             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1845             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1846             F                = _fjsp_setzero_v2r8();
1847             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1848             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1849             H                = _fjsp_setzero_v2r8();
1850             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1851             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1852             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1853             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
1854
1855             fscal            = felec;
1856
1857             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1858
1859             /* Update vectorial force */
1860             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1861             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1862             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1863
1864             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1865             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1866             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1867
1868             /**************************
1869              * CALCULATE INTERACTIONS *
1870              **************************/
1871
1872             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
1873
1874             /* Calculate table index by multiplying r with table scale and truncate to integer */
1875             rt               = _fjsp_mul_v2r8(r13,vftabscale);
1876             itab_tmp         = _fjsp_dtox_v2r8(rt);
1877             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1878             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1879             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1880
1881             vfconv.i[0]     *= 4;
1882             vfconv.i[1]     *= 4;
1883
1884             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1885             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1886             F                = _fjsp_setzero_v2r8();
1887             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1888             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1889             H                = _fjsp_setzero_v2r8();
1890             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1891             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1892             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1893             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
1894
1895             fscal            = felec;
1896
1897             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1898
1899             /* Update vectorial force */
1900             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1901             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1902             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1903
1904             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1905             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1906             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1907
1908             /**************************
1909              * CALCULATE INTERACTIONS *
1910              **************************/
1911
1912             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1913
1914             /* Calculate table index by multiplying r with table scale and truncate to integer */
1915             rt               = _fjsp_mul_v2r8(r21,vftabscale);
1916             itab_tmp         = _fjsp_dtox_v2r8(rt);
1917             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1918             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1919             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1920
1921             vfconv.i[0]     *= 4;
1922             vfconv.i[1]     *= 4;
1923
1924             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1925             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1926             F                = _fjsp_setzero_v2r8();
1927             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1928             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1929             H                = _fjsp_setzero_v2r8();
1930             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1931             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1932             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1933             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
1934
1935             fscal            = felec;
1936
1937             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1938
1939             /* Update vectorial force */
1940             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1941             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1942             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1943
1944             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1945             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1946             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1947
1948             /**************************
1949              * CALCULATE INTERACTIONS *
1950              **************************/
1951
1952             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1953
1954             /* Calculate table index by multiplying r with table scale and truncate to integer */
1955             rt               = _fjsp_mul_v2r8(r22,vftabscale);
1956             itab_tmp         = _fjsp_dtox_v2r8(rt);
1957             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1958             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1959             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1960
1961             vfconv.i[0]     *= 4;
1962             vfconv.i[1]     *= 4;
1963
1964             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1965             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1966             F                = _fjsp_setzero_v2r8();
1967             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1968             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1969             H                = _fjsp_setzero_v2r8();
1970             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1971             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1972             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1973             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
1974
1975             fscal            = felec;
1976
1977             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1978
1979             /* Update vectorial force */
1980             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1981             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1982             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1983
1984             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1985             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1986             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1987
1988             /**************************
1989              * CALCULATE INTERACTIONS *
1990              **************************/
1991
1992             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1993
1994             /* Calculate table index by multiplying r with table scale and truncate to integer */
1995             rt               = _fjsp_mul_v2r8(r23,vftabscale);
1996             itab_tmp         = _fjsp_dtox_v2r8(rt);
1997             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1998             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1999             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
2000
2001             vfconv.i[0]     *= 4;
2002             vfconv.i[1]     *= 4;
2003
2004             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2005             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2006             F                = _fjsp_setzero_v2r8();
2007             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2008             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2009             H                = _fjsp_setzero_v2r8();
2010             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2011             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2012             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2013             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
2014
2015             fscal            = felec;
2016
2017             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2018
2019             /* Update vectorial force */
2020             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
2021             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
2022             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
2023
2024             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
2025             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
2026             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
2027
2028             /**************************
2029              * CALCULATE INTERACTIONS *
2030              **************************/
2031
2032             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
2033
2034             /* Calculate table index by multiplying r with table scale and truncate to integer */
2035             rt               = _fjsp_mul_v2r8(r31,vftabscale);
2036             itab_tmp         = _fjsp_dtox_v2r8(rt);
2037             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
2038             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
2039             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
2040
2041             vfconv.i[0]     *= 4;
2042             vfconv.i[1]     *= 4;
2043
2044             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2045             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2046             F                = _fjsp_setzero_v2r8();
2047             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2048             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2049             H                = _fjsp_setzero_v2r8();
2050             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2051             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2052             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2053             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
2054
2055             fscal            = felec;
2056
2057             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2058
2059             /* Update vectorial force */
2060             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
2061             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
2062             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
2063
2064             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
2065             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
2066             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
2067
2068             /**************************
2069              * CALCULATE INTERACTIONS *
2070              **************************/
2071
2072             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
2073
2074             /* Calculate table index by multiplying r with table scale and truncate to integer */
2075             rt               = _fjsp_mul_v2r8(r32,vftabscale);
2076             itab_tmp         = _fjsp_dtox_v2r8(rt);
2077             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
2078             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
2079             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
2080
2081             vfconv.i[0]     *= 4;
2082             vfconv.i[1]     *= 4;
2083
2084             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2085             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2086             F                = _fjsp_setzero_v2r8();
2087             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2088             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2089             H                = _fjsp_setzero_v2r8();
2090             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2091             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2092             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2093             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
2094
2095             fscal            = felec;
2096
2097             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2098
2099             /* Update vectorial force */
2100             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
2101             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
2102             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
2103
2104             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
2105             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
2106             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
2107
2108             /**************************
2109              * CALCULATE INTERACTIONS *
2110              **************************/
2111
2112             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
2113
2114             /* Calculate table index by multiplying r with table scale and truncate to integer */
2115             rt               = _fjsp_mul_v2r8(r33,vftabscale);
2116             itab_tmp         = _fjsp_dtox_v2r8(rt);
2117             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
2118             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
2119             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
2120
2121             vfconv.i[0]     *= 4;
2122             vfconv.i[1]     *= 4;
2123
2124             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2125             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2126             F                = _fjsp_setzero_v2r8();
2127             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2128             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2129             H                = _fjsp_setzero_v2r8();
2130             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2131             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2132             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2133             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
2134
2135             fscal            = felec;
2136
2137             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2138
2139             /* Update vectorial force */
2140             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
2141             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
2142             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
2143
2144             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
2145             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
2146             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
2147
2148             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2149
2150             /* Inner loop uses 378 flops */
2151         }
2152
2153         /* End of innermost loop */
2154
2155         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2156                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2157
2158         /* Increment number of inner iterations */
2159         inneriter                  += j_index_end - j_index_start;
2160
2161         /* Outer loop uses 18 flops */
2162     }
2163
2164     /* Increment number of outer iterations */
2165     outeriter        += nri;
2166
2167     /* Update outer/inner flops */
2168
2169     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*378);
2170 }