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 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: CubicSplineTable
  52  * VdW interaction:            None
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCSTab_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset1;
  80     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  81     int              vdwioffset2;
  82     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  83     int              vdwioffset3;
  84     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  85     int              vdwjidx1A,vdwjidx1B;
  86     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  87     int              vdwjidx2A,vdwjidx2B;
  88     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  89     int              vdwjidx3A,vdwjidx3B;
  90     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  91     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  92     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  93     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  94     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  95     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  96     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  97     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  98     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  99     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 100     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     _fjsp_v2r8       rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 103     real             *vftab;
 104     _fjsp_v2r8       itab_tmp;
 105     _fjsp_v2r8       dummy_mask,cutoff_mask;
 106     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 107     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 108     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 109
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123
 124     vftab            = kernel_data->table_elec->data;
 125     vftabscale       = gmx_fjsp_set1_v2r8(kernel_data->table_elec->scale);
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 130     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 131     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 132
 133     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 134     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 135     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 136     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 137     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 138     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 139     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 140     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 141     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 142     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 143     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 144     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 145
 146     /* Avoid stupid compiler warnings */
 147     jnrA = jnrB = 0;
 148     j_coord_offsetA = 0;
 149     j_coord_offsetB = 0;
 150
 151     outeriter        = 0;
 152     inneriter        = 0;
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 170                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 171
 172         fix1             = _fjsp_setzero_v2r8();
 173         fiy1             = _fjsp_setzero_v2r8();
 174         fiz1             = _fjsp_setzero_v2r8();
 175         fix2             = _fjsp_setzero_v2r8();
 176         fiy2             = _fjsp_setzero_v2r8();
 177         fiz2             = _fjsp_setzero_v2r8();
 178         fix3             = _fjsp_setzero_v2r8();
 179         fiy3             = _fjsp_setzero_v2r8();
 180         fiz3             = _fjsp_setzero_v2r8();
 181
 182         /* Reset potential sums */
 183         velecsum         = _fjsp_setzero_v2r8();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194
 195             /* load j atom coordinates */
 196             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 197                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 198
 199             /* Calculate displacement vector */
 200             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 201             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 202             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 203             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 204             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 205             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 206             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 207             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 208             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 209             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 210             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 211             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 212             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 213             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 214             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 215             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 216             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 217             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 218             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 219             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 220             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 221             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 222             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 223             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 224             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 225             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 226             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 227
 228             /* Calculate squared distance and things based on it */
 229             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 230             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 231             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 232             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 233             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 234             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 235             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 236             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 237             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 238
 239             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 240             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 241             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 242             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 243             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 244             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 245             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 246             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 247             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 248
 249             fjx1             = _fjsp_setzero_v2r8();
 250             fjy1             = _fjsp_setzero_v2r8();
 251             fjz1             = _fjsp_setzero_v2r8();
 252             fjx2             = _fjsp_setzero_v2r8();
 253             fjy2             = _fjsp_setzero_v2r8();
 254             fjz2             = _fjsp_setzero_v2r8();
 255             fjx3             = _fjsp_setzero_v2r8();
 256             fjy3             = _fjsp_setzero_v2r8();
 257             fjz3             = _fjsp_setzero_v2r8();
 258
 259             /**************************
 260              * CALCULATE INTERACTIONS *
 261              **************************/
 262
 263             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 264
 265             /* Calculate table index by multiplying r with table scale and truncate to integer */
 266             rt               = _fjsp_mul_v2r8(r11,vftabscale);
 267             itab_tmp         = _fjsp_dtox_v2r8(rt);
 268             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 269             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 270             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 271
 272             vfconv.i[0]     *= 4;
 273             vfconv.i[1]     *= 4;
 274
 275             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 276             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 277             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 278             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 279             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 280             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 281             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 282             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 283             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 284             velec            = _fjsp_mul_v2r8(qq11,VV);
 285             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 286             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
 287
 288             /* Update potential sum for this i atom from the interaction with this j atom. */
 289             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 290
 291             fscal            = felec;
 292
 293             /* Update vectorial force */
 294             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 295             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 296             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 297
 298             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 299             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 300             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 307
 308             /* Calculate table index by multiplying r with table scale and truncate to integer */
 309             rt               = _fjsp_mul_v2r8(r12,vftabscale);
 310             itab_tmp         = _fjsp_dtox_v2r8(rt);
 311             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 312             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 313             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 314
 315             vfconv.i[0]     *= 4;
 316             vfconv.i[1]     *= 4;
 317
 318             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 319             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 320             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 321             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 322             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 323             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 324             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 325             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 326             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 327             velec            = _fjsp_mul_v2r8(qq12,VV);
 328             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 329             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
 330
 331             /* Update potential sum for this i atom from the interaction with this j atom. */
 332             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 333
 334             fscal            = felec;
 335
 336             /* Update vectorial force */
 337             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 338             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 339             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 340
 341             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 342             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 343             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 350
 351             /* Calculate table index by multiplying r with table scale and truncate to integer */
 352             rt               = _fjsp_mul_v2r8(r13,vftabscale);
 353             itab_tmp         = _fjsp_dtox_v2r8(rt);
 354             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 355             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 356             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 357
 358             vfconv.i[0]     *= 4;
 359             vfconv.i[1]     *= 4;
 360
 361             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 362             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 363             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 364             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 365             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 366             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 367             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 368             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 369             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 370             velec            = _fjsp_mul_v2r8(qq13,VV);
 371             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 372             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379             /* Update vectorial force */
 380             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 381             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 382             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 383
 384             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 385             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 386             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 393
 394             /* Calculate table index by multiplying r with table scale and truncate to integer */
 395             rt               = _fjsp_mul_v2r8(r21,vftabscale);
 396             itab_tmp         = _fjsp_dtox_v2r8(rt);
 397             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 398             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 399             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 400
 401             vfconv.i[0]     *= 4;
 402             vfconv.i[1]     *= 4;
 403
 404             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 405             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 406             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 407             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 408             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 409             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 410             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 411             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 412             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 413             velec            = _fjsp_mul_v2r8(qq21,VV);
 414             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 415             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 419
 420             fscal            = felec;
 421
 422             /* Update vectorial force */
 423             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 424             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 425             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 426
 427             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 428             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 429             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 436
 437             /* Calculate table index by multiplying r with table scale and truncate to integer */
 438             rt               = _fjsp_mul_v2r8(r22,vftabscale);
 439             itab_tmp         = _fjsp_dtox_v2r8(rt);
 440             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 441             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 442             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 443
 444             vfconv.i[0]     *= 4;
 445             vfconv.i[1]     *= 4;
 446
 447             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 448             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 449             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 450             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 451             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 452             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 453             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 454             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 455             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 456             velec            = _fjsp_mul_v2r8(qq22,VV);
 457             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 458             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             /* Update vectorial force */
 466             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 467             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 468             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 469
 470             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 471             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 472             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
 479
 480             /* Calculate table index by multiplying r with table scale and truncate to integer */
 481             rt               = _fjsp_mul_v2r8(r23,vftabscale);
 482             itab_tmp         = _fjsp_dtox_v2r8(rt);
 483             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 484             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 485             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 486
 487             vfconv.i[0]     *= 4;
 488             vfconv.i[1]     *= 4;
 489
 490             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 491             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 492             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 493             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 494             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 495             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 496             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 497             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 498             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 499             velec            = _fjsp_mul_v2r8(qq23,VV);
 500             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 501             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
 502
 503             /* Update potential sum for this i atom from the interaction with this j atom. */
 504             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 505
 506             fscal            = felec;
 507
 508             /* Update vectorial force */
 509             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 510             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 511             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 512
 513             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 514             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 515             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 516
 517             /**************************
 518              * CALCULATE INTERACTIONS *
 519              **************************/
 520
 521             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
 522
 523             /* Calculate table index by multiplying r with table scale and truncate to integer */
 524             rt               = _fjsp_mul_v2r8(r31,vftabscale);
 525             itab_tmp         = _fjsp_dtox_v2r8(rt);
 526             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 527             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 528             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 529
 530             vfconv.i[0]     *= 4;
 531             vfconv.i[1]     *= 4;
 532
 533             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 534             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 535             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 536             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 537             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 538             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 539             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 540             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 541             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 542             velec            = _fjsp_mul_v2r8(qq31,VV);
 543             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 544             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
 545
 546             /* Update potential sum for this i atom from the interaction with this j atom. */
 547             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 548
 549             fscal            = felec;
 550
 551             /* Update vectorial force */
 552             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 553             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 554             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 555
 556             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 557             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 558             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 559
 560             /**************************
 561              * CALCULATE INTERACTIONS *
 562              **************************/
 563
 564             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
 565
 566             /* Calculate table index by multiplying r with table scale and truncate to integer */
 567             rt               = _fjsp_mul_v2r8(r32,vftabscale);
 568             itab_tmp         = _fjsp_dtox_v2r8(rt);
 569             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 570             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 571             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 572
 573             vfconv.i[0]     *= 4;
 574             vfconv.i[1]     *= 4;
 575
 576             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 577             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 578             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 579             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 580             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 581             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 582             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 583             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 584             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 585             velec            = _fjsp_mul_v2r8(qq32,VV);
 586             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 587             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
 588
 589             /* Update potential sum for this i atom from the interaction with this j atom. */
 590             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 591
 592             fscal            = felec;
 593
 594             /* Update vectorial force */
 595             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 596             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 597             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 598
 599             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 600             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 601             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 602
 603             /**************************
 604              * CALCULATE INTERACTIONS *
 605              **************************/
 606
 607             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
 608
 609             /* Calculate table index by multiplying r with table scale and truncate to integer */
 610             rt               = _fjsp_mul_v2r8(r33,vftabscale);
 611             itab_tmp         = _fjsp_dtox_v2r8(rt);
 612             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 613             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 614             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 615
 616             vfconv.i[0]     *= 4;
 617             vfconv.i[1]     *= 4;
 618
 619             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 620             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 621             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 622             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 623             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 624             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
 625             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 626             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 627             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 628             velec            = _fjsp_mul_v2r8(qq33,VV);
 629             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 630             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
 631
 632             /* Update potential sum for this i atom from the interaction with this j atom. */
 633             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 634
 635             fscal            = felec;
 636
 637             /* Update vectorial force */
 638             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 639             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 640             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 641
 642             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 643             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 644             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 645
 646             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);
 647
 648             /* Inner loop uses 414 flops */
 649         }
 650
 651         if(jidx<j_index_end)
 652         {
 653
 654             jnrA             = jjnr[jidx];
 655             j_coord_offsetA  = DIM*jnrA;
 656
 657             /* load j atom coordinates */
 658             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
 659                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 660
 661             /* Calculate displacement vector */
 662             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 663             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 664             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 665             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 666             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 667             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 668             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 669             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 670             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 671             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 672             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 673             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 674             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 675             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 676             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 677             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 678             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 679             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 680             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 681             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 682             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 683             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 684             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 685             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 686             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 687             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 688             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 689
 690             /* Calculate squared distance and things based on it */
 691             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 692             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 693             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 694             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 695             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 696             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 697             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 698             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 699             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 700
 701             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 702             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 703             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 704             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 705             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 706             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 707             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 708             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 709             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 710
 711             fjx1             = _fjsp_setzero_v2r8();
 712             fjy1             = _fjsp_setzero_v2r8();
 713             fjz1             = _fjsp_setzero_v2r8();
 714             fjx2             = _fjsp_setzero_v2r8();
 715             fjy2             = _fjsp_setzero_v2r8();
 716             fjz2             = _fjsp_setzero_v2r8();
 717             fjx3             = _fjsp_setzero_v2r8();
 718             fjy3             = _fjsp_setzero_v2r8();
 719             fjz3             = _fjsp_setzero_v2r8();
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
 726
 727             /* Calculate table index by multiplying r with table scale and truncate to integer */
 728             rt               = _fjsp_mul_v2r8(r11,vftabscale);
 729             itab_tmp         = _fjsp_dtox_v2r8(rt);
 730             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 731             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 732             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 733
 734             vfconv.i[0]     *= 4;
 735             vfconv.i[1]     *= 4;
 736
 737             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 738             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 739             F                = _fjsp_setzero_v2r8();
 740             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 741             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 742             H                = _fjsp_setzero_v2r8();
 743             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 744             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 745             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 746             velec            = _fjsp_mul_v2r8(qq11,VV);
 747             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 748             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
 749
 750             /* Update potential sum for this i atom from the interaction with this j atom. */
 751             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 752             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 753
 754             fscal            = felec;
 755
 756             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 757
 758             /* Update vectorial force */
 759             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 760             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 761             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 762
 763             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 764             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 765             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 766
 767             /**************************
 768              * CALCULATE INTERACTIONS *
 769              **************************/
 770
 771             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
 772
 773             /* Calculate table index by multiplying r with table scale and truncate to integer */
 774             rt               = _fjsp_mul_v2r8(r12,vftabscale);
 775             itab_tmp         = _fjsp_dtox_v2r8(rt);
 776             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 777             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 778             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 779
 780             vfconv.i[0]     *= 4;
 781             vfconv.i[1]     *= 4;
 782
 783             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 784             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 785             F                = _fjsp_setzero_v2r8();
 786             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 787             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 788             H                = _fjsp_setzero_v2r8();
 789             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 790             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 791             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 792             velec            = _fjsp_mul_v2r8(qq12,VV);
 793             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 794             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
 795
 796             /* Update potential sum for this i atom from the interaction with this j atom. */
 797             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 798             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 799
 800             fscal            = felec;
 801
 802             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 803
 804             /* Update vectorial force */
 805             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 806             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 807             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 808
 809             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 810             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 811             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 812
 813             /**************************
 814              * CALCULATE INTERACTIONS *
 815              **************************/
 816
 817             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
 818
 819             /* Calculate table index by multiplying r with table scale and truncate to integer */
 820             rt               = _fjsp_mul_v2r8(r13,vftabscale);
 821             itab_tmp         = _fjsp_dtox_v2r8(rt);
 822             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 823             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 824             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 825
 826             vfconv.i[0]     *= 4;
 827             vfconv.i[1]     *= 4;
 828
 829             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 830             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 831             F                = _fjsp_setzero_v2r8();
 832             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 833             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 834             H                = _fjsp_setzero_v2r8();
 835             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 836             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 837             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 838             velec            = _fjsp_mul_v2r8(qq13,VV);
 839             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 840             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
 841
 842             /* Update potential sum for this i atom from the interaction with this j atom. */
 843             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 844             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 845
 846             fscal            = felec;
 847
 848             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 849
 850             /* Update vectorial force */
 851             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 852             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 853             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 854
 855             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 856             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 857             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
 864
 865             /* Calculate table index by multiplying r with table scale and truncate to integer */
 866             rt               = _fjsp_mul_v2r8(r21,vftabscale);
 867             itab_tmp         = _fjsp_dtox_v2r8(rt);
 868             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 869             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 870             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 871
 872             vfconv.i[0]     *= 4;
 873             vfconv.i[1]     *= 4;
 874
 875             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 876             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 877             F                = _fjsp_setzero_v2r8();
 878             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 879             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 880             H                = _fjsp_setzero_v2r8();
 881             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 882             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 883             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 884             velec            = _fjsp_mul_v2r8(qq21,VV);
 885             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 886             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
 887
 888             /* Update potential sum for this i atom from the interaction with this j atom. */
 889             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 890             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 891
 892             fscal            = felec;
 893
 894             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 895
 896             /* Update vectorial force */
 897             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 898             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 899             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 900
 901             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 902             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 903             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
 910
 911             /* Calculate table index by multiplying r with table scale and truncate to integer */
 912             rt               = _fjsp_mul_v2r8(r22,vftabscale);
 913             itab_tmp         = _fjsp_dtox_v2r8(rt);
 914             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 915             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 916             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 917
 918             vfconv.i[0]     *= 4;
 919             vfconv.i[1]     *= 4;
 920
 921             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 922             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 923             F                = _fjsp_setzero_v2r8();
 924             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 925             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 926             H                = _fjsp_setzero_v2r8();
 927             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 928             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 929             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 930             velec            = _fjsp_mul_v2r8(qq22,VV);
 931             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 932             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
 933
 934             /* Update potential sum for this i atom from the interaction with this j atom. */
 935             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 936             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 937
 938             fscal            = felec;
 939
 940             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 941
 942             /* Update vectorial force */
 943             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 944             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 945             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 946
 947             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 948             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 949             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 950
 951             /**************************
 952              * CALCULATE INTERACTIONS *
 953              **************************/
 954
 955             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
 956
 957             /* Calculate table index by multiplying r with table scale and truncate to integer */
 958             rt               = _fjsp_mul_v2r8(r23,vftabscale);
 959             itab_tmp         = _fjsp_dtox_v2r8(rt);
 960             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 961             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 962             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 963
 964             vfconv.i[0]     *= 4;
 965             vfconv.i[1]     *= 4;
 966
 967             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 968             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 969             F                = _fjsp_setzero_v2r8();
 970             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 971             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
 972             H                = _fjsp_setzero_v2r8();
 973             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 974             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
 975             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 976             velec            = _fjsp_mul_v2r8(qq23,VV);
 977             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
 978             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
 979
 980             /* Update potential sum for this i atom from the interaction with this j atom. */
 981             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 982             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 983
 984             fscal            = felec;
 985
 986             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 987
 988             /* Update vectorial force */
 989             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 990             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 991             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 992
 993             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 994             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 995             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 996
 997             /**************************
 998              * CALCULATE INTERACTIONS *
 999              **************************/
1000
1001             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1002
1003             /* Calculate table index by multiplying r with table scale and truncate to integer */
1004             rt               = _fjsp_mul_v2r8(r31,vftabscale);
1005             itab_tmp         = _fjsp_dtox_v2r8(rt);
1006             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1007             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1008             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1009
1010             vfconv.i[0]     *= 4;
1011             vfconv.i[1]     *= 4;
1012
1013             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1014             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1015             F                = _fjsp_setzero_v2r8();
1016             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1017             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1018             H                = _fjsp_setzero_v2r8();
1019             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1020             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1021             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
1022             velec            = _fjsp_mul_v2r8(qq31,VV);
1023             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1024             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
1025
1026             /* Update potential sum for this i atom from the interaction with this j atom. */
1027             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1028             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1029
1030             fscal            = felec;
1031
1032             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1033
1034             /* Update vectorial force */
1035             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1036             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1037             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1038
1039             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1040             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1041             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1042
1043             /**************************
1044              * CALCULATE INTERACTIONS *
1045              **************************/
1046
1047             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1048
1049             /* Calculate table index by multiplying r with table scale and truncate to integer */
1050             rt               = _fjsp_mul_v2r8(r32,vftabscale);
1051             itab_tmp         = _fjsp_dtox_v2r8(rt);
1052             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1053             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1054             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1055
1056             vfconv.i[0]     *= 4;
1057             vfconv.i[1]     *= 4;
1058
1059             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1060             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1061             F                = _fjsp_setzero_v2r8();
1062             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1063             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1064             H                = _fjsp_setzero_v2r8();
1065             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1066             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1067             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
1068             velec            = _fjsp_mul_v2r8(qq32,VV);
1069             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1070             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
1071
1072             /* Update potential sum for this i atom from the interaction with this j atom. */
1073             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1074             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1075
1076             fscal            = felec;
1077
1078             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1079
1080             /* Update vectorial force */
1081             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1082             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1083             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1084
1085             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1086             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1087             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1088
1089             /**************************
1090              * CALCULATE INTERACTIONS *
1091              **************************/
1092
1093             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1094
1095             /* Calculate table index by multiplying r with table scale and truncate to integer */
1096             rt               = _fjsp_mul_v2r8(r33,vftabscale);
1097             itab_tmp         = _fjsp_dtox_v2r8(rt);
1098             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1099             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1100             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1101
1102             vfconv.i[0]     *= 4;
1103             vfconv.i[1]     *= 4;
1104
1105             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1106             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1107             F                = _fjsp_setzero_v2r8();
1108             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1109             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1110             H                = _fjsp_setzero_v2r8();
1111             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1112             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1113             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
1114             velec            = _fjsp_mul_v2r8(qq33,VV);
1115             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1116             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
1117
1118             /* Update potential sum for this i atom from the interaction with this j atom. */
1119             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1120             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1121
1122             fscal            = felec;
1123
1124             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1125
1126             /* Update vectorial force */
1127             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1128             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1129             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1130
1131             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1132             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1133             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1134
1135             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1136
1137             /* Inner loop uses 414 flops */
1138         }
1139
1140         /* End of innermost loop */
1141
1142         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1143                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1144
1145         ggid                        = gid[iidx];
1146         /* Update potential energies */
1147         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1148
1149         /* Increment number of inner iterations */
1150         inneriter                  += j_index_end - j_index_start;
1151
1152         /* Outer loop uses 19 flops */
1153     }
1154
1155     /* Increment number of outer iterations */
1156     outeriter        += nri;
1157
1158     /* Update outer/inner flops */
1159
1160     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*414);
1161 }
1162 /*
1163  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1164  * Electrostatics interaction: CubicSplineTable
1165  * VdW interaction:            None
1166  * Geometry:                   Water4-Water4
1167  * Calculate force/pot:        Force
1168  */
1169 void
1170 nb_kernel_ElecCSTab_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
1171                     (t_nblist                    * gmx_restrict       nlist,
1172                      rvec                        * gmx_restrict          xx,
1173                      rvec                        * gmx_restrict          ff,
1174                      t_forcerec                  * gmx_restrict          fr,
1175                      t_mdatoms                   * gmx_restrict     mdatoms,
1176                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1177                      t_nrnb                      * gmx_restrict        nrnb)
1178 {
1179     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1180      * just 0 for non-waters.
1181      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1182      * jnr indices corresponding to data put in the four positions in the SIMD register.
1183      */
1184     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1185     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1186     int              jnrA,jnrB;
1187     int              j_coord_offsetA,j_coord_offsetB;
1188     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1189     real             rcutoff_scalar;
1190     real             *shiftvec,*fshift,*x,*f;
1191     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1192     int              vdwioffset1;
1193     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1194     int              vdwioffset2;
1195     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1196     int              vdwioffset3;
1197     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1198     int              vdwjidx1A,vdwjidx1B;
1199     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1200     int              vdwjidx2A,vdwjidx2B;
1201     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1202     int              vdwjidx3A,vdwjidx3B;
1203     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1204     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1205     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1206     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1207     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1208     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1209     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1210     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1211     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1212     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1213     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1214     real             *charge;
1215     _fjsp_v2r8       rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1216     real             *vftab;
1217     _fjsp_v2r8       itab_tmp;
1218     _fjsp_v2r8       dummy_mask,cutoff_mask;
1219     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1220     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1221     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1222
1223     x                = xx[0];
1224     f                = ff[0];
1225
1226     nri              = nlist->nri;
1227     iinr             = nlist->iinr;
1228     jindex           = nlist->jindex;
1229     jjnr             = nlist->jjnr;
1230     shiftidx         = nlist->shift;
1231     gid              = nlist->gid;
1232     shiftvec         = fr->shift_vec[0];
1233     fshift           = fr->fshift[0];
1234     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1235     charge           = mdatoms->chargeA;
1236
1237     vftab            = kernel_data->table_elec->data;
1238     vftabscale       = gmx_fjsp_set1_v2r8(kernel_data->table_elec->scale);
1239
1240     /* Setup water-specific parameters */
1241     inr              = nlist->iinr[0];
1242     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1243     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1244     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
1245
1246     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1247     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1248     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
1249     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1250     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1251     qq13             = _fjsp_mul_v2r8(iq1,jq3);
1252     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1253     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1254     qq23             = _fjsp_mul_v2r8(iq2,jq3);
1255     qq31             = _fjsp_mul_v2r8(iq3,jq1);
1256     qq32             = _fjsp_mul_v2r8(iq3,jq2);
1257     qq33             = _fjsp_mul_v2r8(iq3,jq3);
1258
1259     /* Avoid stupid compiler warnings */
1260     jnrA = jnrB = 0;
1261     j_coord_offsetA = 0;
1262     j_coord_offsetB = 0;
1263
1264     outeriter        = 0;
1265     inneriter        = 0;
1266
1267     /* Start outer loop over neighborlists */
1268     for(iidx=0; iidx<nri; iidx++)
1269     {
1270         /* Load shift vector for this list */
1271         i_shift_offset   = DIM*shiftidx[iidx];
1272
1273         /* Load limits for loop over neighbors */
1274         j_index_start    = jindex[iidx];
1275         j_index_end      = jindex[iidx+1];
1276
1277         /* Get outer coordinate index */
1278         inr              = iinr[iidx];
1279         i_coord_offset   = DIM*inr;
1280
1281         /* Load i particle coords and add shift vector */
1282         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1283                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1284
1285         fix1             = _fjsp_setzero_v2r8();
1286         fiy1             = _fjsp_setzero_v2r8();
1287         fiz1             = _fjsp_setzero_v2r8();
1288         fix2             = _fjsp_setzero_v2r8();
1289         fiy2             = _fjsp_setzero_v2r8();
1290         fiz2             = _fjsp_setzero_v2r8();
1291         fix3             = _fjsp_setzero_v2r8();
1292         fiy3             = _fjsp_setzero_v2r8();
1293         fiz3             = _fjsp_setzero_v2r8();
1294
1295         /* Start inner kernel loop */
1296         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1297         {
1298
1299             /* Get j neighbor index, and coordinate index */
1300             jnrA             = jjnr[jidx];
1301             jnrB             = jjnr[jidx+1];
1302             j_coord_offsetA  = DIM*jnrA;
1303             j_coord_offsetB  = DIM*jnrB;
1304
1305             /* load j atom coordinates */
1306             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1307                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1308
1309             /* Calculate displacement vector */
1310             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1311             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1312             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1313             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1314             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1315             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1316             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1317             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1318             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1319             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1320             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1321             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1322             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1323             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1324             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1325             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1326             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1327             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1328             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1329             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1330             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1331             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1332             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1333             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1334             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1335             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1336             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1337
1338             /* Calculate squared distance and things based on it */
1339             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1340             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1341             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1342             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1343             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1344             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1345             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1346             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1347             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1348
1349             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1350             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1351             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1352             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1353             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1354             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1355             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1356             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1357             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1358
1359             fjx1             = _fjsp_setzero_v2r8();
1360             fjy1             = _fjsp_setzero_v2r8();
1361             fjz1             = _fjsp_setzero_v2r8();
1362             fjx2             = _fjsp_setzero_v2r8();
1363             fjy2             = _fjsp_setzero_v2r8();
1364             fjz2             = _fjsp_setzero_v2r8();
1365             fjx3             = _fjsp_setzero_v2r8();
1366             fjy3             = _fjsp_setzero_v2r8();
1367             fjz3             = _fjsp_setzero_v2r8();
1368
1369             /**************************
1370              * CALCULATE INTERACTIONS *
1371              **************************/
1372
1373             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1374
1375             /* Calculate table index by multiplying r with table scale and truncate to integer */
1376             rt               = _fjsp_mul_v2r8(r11,vftabscale);
1377             itab_tmp         = _fjsp_dtox_v2r8(rt);
1378             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1379             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1380             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1381
1382             vfconv.i[0]     *= 4;
1383             vfconv.i[1]     *= 4;
1384
1385             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1386             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1387             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1388             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1389             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1390             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1391             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1392             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1393             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1394             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
1395
1396             fscal            = felec;
1397
1398             /* Update vectorial force */
1399             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1400             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1401             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1402
1403             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1404             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1405             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1406
1407             /**************************
1408              * CALCULATE INTERACTIONS *
1409              **************************/
1410
1411             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1412
1413             /* Calculate table index by multiplying r with table scale and truncate to integer */
1414             rt               = _fjsp_mul_v2r8(r12,vftabscale);
1415             itab_tmp         = _fjsp_dtox_v2r8(rt);
1416             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1417             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1418             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1419
1420             vfconv.i[0]     *= 4;
1421             vfconv.i[1]     *= 4;
1422
1423             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1424             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1425             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1426             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1427             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1428             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1429             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1430             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1431             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1432             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
1433
1434             fscal            = felec;
1435
1436             /* Update vectorial force */
1437             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1438             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1439             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1440
1441             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1442             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1443             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1444
1445             /**************************
1446              * CALCULATE INTERACTIONS *
1447              **************************/
1448
1449             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
1450
1451             /* Calculate table index by multiplying r with table scale and truncate to integer */
1452             rt               = _fjsp_mul_v2r8(r13,vftabscale);
1453             itab_tmp         = _fjsp_dtox_v2r8(rt);
1454             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1455             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1456             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1457
1458             vfconv.i[0]     *= 4;
1459             vfconv.i[1]     *= 4;
1460
1461             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1462             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1463             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1464             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1465             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1466             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1467             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1468             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1469             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1470             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
1471
1472             fscal            = felec;
1473
1474             /* Update vectorial force */
1475             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1476             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1477             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1478
1479             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1480             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1481             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1482
1483             /**************************
1484              * CALCULATE INTERACTIONS *
1485              **************************/
1486
1487             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1488
1489             /* Calculate table index by multiplying r with table scale and truncate to integer */
1490             rt               = _fjsp_mul_v2r8(r21,vftabscale);
1491             itab_tmp         = _fjsp_dtox_v2r8(rt);
1492             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1493             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1494             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1495
1496             vfconv.i[0]     *= 4;
1497             vfconv.i[1]     *= 4;
1498
1499             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1500             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1501             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1502             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1503             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1504             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1505             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1506             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1507             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1508             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
1509
1510             fscal            = felec;
1511
1512             /* Update vectorial force */
1513             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1514             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1515             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1516
1517             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1518             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1519             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1520
1521             /**************************
1522              * CALCULATE INTERACTIONS *
1523              **************************/
1524
1525             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1526
1527             /* Calculate table index by multiplying r with table scale and truncate to integer */
1528             rt               = _fjsp_mul_v2r8(r22,vftabscale);
1529             itab_tmp         = _fjsp_dtox_v2r8(rt);
1530             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1531             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1532             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1533
1534             vfconv.i[0]     *= 4;
1535             vfconv.i[1]     *= 4;
1536
1537             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1538             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1539             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1540             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1541             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1542             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1543             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1544             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1545             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1546             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
1547
1548             fscal            = felec;
1549
1550             /* Update vectorial force */
1551             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1552             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1553             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1554
1555             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1556             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1557             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1558
1559             /**************************
1560              * CALCULATE INTERACTIONS *
1561              **************************/
1562
1563             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1564
1565             /* Calculate table index by multiplying r with table scale and truncate to integer */
1566             rt               = _fjsp_mul_v2r8(r23,vftabscale);
1567             itab_tmp         = _fjsp_dtox_v2r8(rt);
1568             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1569             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1570             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1571
1572             vfconv.i[0]     *= 4;
1573             vfconv.i[1]     *= 4;
1574
1575             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1576             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1577             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1578             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1579             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1580             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1581             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1582             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1583             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1584             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
1585
1586             fscal            = felec;
1587
1588             /* Update vectorial force */
1589             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1590             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1591             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1592
1593             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1594             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1595             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1596
1597             /**************************
1598              * CALCULATE INTERACTIONS *
1599              **************************/
1600
1601             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
1602
1603             /* Calculate table index by multiplying r with table scale and truncate to integer */
1604             rt               = _fjsp_mul_v2r8(r31,vftabscale);
1605             itab_tmp         = _fjsp_dtox_v2r8(rt);
1606             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1607             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1608             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1609
1610             vfconv.i[0]     *= 4;
1611             vfconv.i[1]     *= 4;
1612
1613             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1614             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1615             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1616             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1617             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1618             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1619             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1620             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1621             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1622             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
1623
1624             fscal            = felec;
1625
1626             /* Update vectorial force */
1627             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1628             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1629             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1630
1631             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1632             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1633             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1634
1635             /**************************
1636              * CALCULATE INTERACTIONS *
1637              **************************/
1638
1639             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
1640
1641             /* Calculate table index by multiplying r with table scale and truncate to integer */
1642             rt               = _fjsp_mul_v2r8(r32,vftabscale);
1643             itab_tmp         = _fjsp_dtox_v2r8(rt);
1644             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1645             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1646             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1647
1648             vfconv.i[0]     *= 4;
1649             vfconv.i[1]     *= 4;
1650
1651             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1652             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1653             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1654             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1655             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1656             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1657             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1658             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1659             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1660             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
1661
1662             fscal            = felec;
1663
1664             /* Update vectorial force */
1665             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1666             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1667             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1668
1669             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1670             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1671             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1672
1673             /**************************
1674              * CALCULATE INTERACTIONS *
1675              **************************/
1676
1677             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
1678
1679             /* Calculate table index by multiplying r with table scale and truncate to integer */
1680             rt               = _fjsp_mul_v2r8(r33,vftabscale);
1681             itab_tmp         = _fjsp_dtox_v2r8(rt);
1682             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1683             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1684             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1685
1686             vfconv.i[0]     *= 4;
1687             vfconv.i[1]     *= 4;
1688
1689             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1690             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1691             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1692             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1693             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1694             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] +2);
1695             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1696             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1697             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1698             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
1699
1700             fscal            = felec;
1701
1702             /* Update vectorial force */
1703             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1704             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1705             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1706
1707             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1708             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1709             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1710
1711             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);
1712
1713             /* Inner loop uses 378 flops */
1714         }
1715
1716         if(jidx<j_index_end)
1717         {
1718
1719             jnrA             = jjnr[jidx];
1720             j_coord_offsetA  = DIM*jnrA;
1721
1722             /* load j atom coordinates */
1723             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1724                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1725
1726             /* Calculate displacement vector */
1727             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1728             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1729             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1730             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1731             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1732             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1733             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1734             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1735             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1736             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1737             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1738             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1739             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1740             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1741             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1742             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1743             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1744             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1745             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1746             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1747             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1748             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1749             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1750             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1751             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1752             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1753             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1754
1755             /* Calculate squared distance and things based on it */
1756             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1757             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1758             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1759             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1760             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1761             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1762             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1763             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1764             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1765
1766             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1767             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1768             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1769             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1770             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1771             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1772             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1773             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1774             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1775
1776             fjx1             = _fjsp_setzero_v2r8();
1777             fjy1             = _fjsp_setzero_v2r8();
1778             fjz1             = _fjsp_setzero_v2r8();
1779             fjx2             = _fjsp_setzero_v2r8();
1780             fjy2             = _fjsp_setzero_v2r8();
1781             fjz2             = _fjsp_setzero_v2r8();
1782             fjx3             = _fjsp_setzero_v2r8();
1783             fjy3             = _fjsp_setzero_v2r8();
1784             fjz3             = _fjsp_setzero_v2r8();
1785
1786             /**************************
1787              * CALCULATE INTERACTIONS *
1788              **************************/
1789
1790             r11              = _fjsp_mul_v2r8(rsq11,rinv11);
1791
1792             /* Calculate table index by multiplying r with table scale and truncate to integer */
1793             rt               = _fjsp_mul_v2r8(r11,vftabscale);
1794             itab_tmp         = _fjsp_dtox_v2r8(rt);
1795             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1796             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1797             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1798
1799             vfconv.i[0]     *= 4;
1800             vfconv.i[1]     *= 4;
1801
1802             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1803             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1804             F                = _fjsp_setzero_v2r8();
1805             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1806             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1807             H                = _fjsp_setzero_v2r8();
1808             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1809             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1810             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1811             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq11,FF),_fjsp_mul_v2r8(vftabscale,rinv11)));
1812
1813             fscal            = felec;
1814
1815             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1816
1817             /* Update vectorial force */
1818             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1819             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1820             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1821
1822             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1823             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1824             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1825
1826             /**************************
1827              * CALCULATE INTERACTIONS *
1828              **************************/
1829
1830             r12              = _fjsp_mul_v2r8(rsq12,rinv12);
1831
1832             /* Calculate table index by multiplying r with table scale and truncate to integer */
1833             rt               = _fjsp_mul_v2r8(r12,vftabscale);
1834             itab_tmp         = _fjsp_dtox_v2r8(rt);
1835             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1836             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1837             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1838
1839             vfconv.i[0]     *= 4;
1840             vfconv.i[1]     *= 4;
1841
1842             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1843             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1844             F                = _fjsp_setzero_v2r8();
1845             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1846             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1847             H                = _fjsp_setzero_v2r8();
1848             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1849             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1850             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1851             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq12,FF),_fjsp_mul_v2r8(vftabscale,rinv12)));
1852
1853             fscal            = felec;
1854
1855             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1856
1857             /* Update vectorial force */
1858             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1859             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1860             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1861
1862             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1863             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1864             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1865
1866             /**************************
1867              * CALCULATE INTERACTIONS *
1868              **************************/
1869
1870             r13              = _fjsp_mul_v2r8(rsq13,rinv13);
1871
1872             /* Calculate table index by multiplying r with table scale and truncate to integer */
1873             rt               = _fjsp_mul_v2r8(r13,vftabscale);
1874             itab_tmp         = _fjsp_dtox_v2r8(rt);
1875             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1876             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1877             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1878
1879             vfconv.i[0]     *= 4;
1880             vfconv.i[1]     *= 4;
1881
1882             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1883             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1884             F                = _fjsp_setzero_v2r8();
1885             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1886             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1887             H                = _fjsp_setzero_v2r8();
1888             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1889             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1890             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1891             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq13,FF),_fjsp_mul_v2r8(vftabscale,rinv13)));
1892
1893             fscal            = felec;
1894
1895             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1896
1897             /* Update vectorial force */
1898             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1899             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1900             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1901
1902             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1903             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1904             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1905
1906             /**************************
1907              * CALCULATE INTERACTIONS *
1908              **************************/
1909
1910             r21              = _fjsp_mul_v2r8(rsq21,rinv21);
1911
1912             /* Calculate table index by multiplying r with table scale and truncate to integer */
1913             rt               = _fjsp_mul_v2r8(r21,vftabscale);
1914             itab_tmp         = _fjsp_dtox_v2r8(rt);
1915             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1916             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1917             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1918
1919             vfconv.i[0]     *= 4;
1920             vfconv.i[1]     *= 4;
1921
1922             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1923             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1924             F                = _fjsp_setzero_v2r8();
1925             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1926             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1927             H                = _fjsp_setzero_v2r8();
1928             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1929             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1930             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1931             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq21,FF),_fjsp_mul_v2r8(vftabscale,rinv21)));
1932
1933             fscal            = felec;
1934
1935             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1936
1937             /* Update vectorial force */
1938             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1939             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1940             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1941
1942             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1943             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1944             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1945
1946             /**************************
1947              * CALCULATE INTERACTIONS *
1948              **************************/
1949
1950             r22              = _fjsp_mul_v2r8(rsq22,rinv22);
1951
1952             /* Calculate table index by multiplying r with table scale and truncate to integer */
1953             rt               = _fjsp_mul_v2r8(r22,vftabscale);
1954             itab_tmp         = _fjsp_dtox_v2r8(rt);
1955             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1956             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1957             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1958
1959             vfconv.i[0]     *= 4;
1960             vfconv.i[1]     *= 4;
1961
1962             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1963             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1964             F                = _fjsp_setzero_v2r8();
1965             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1966             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
1967             H                = _fjsp_setzero_v2r8();
1968             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1969             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
1970             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
1971             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq22,FF),_fjsp_mul_v2r8(vftabscale,rinv22)));
1972
1973             fscal            = felec;
1974
1975             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1976
1977             /* Update vectorial force */
1978             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1979             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1980             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1981
1982             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1983             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1984             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1985
1986             /**************************
1987              * CALCULATE INTERACTIONS *
1988              **************************/
1989
1990             r23              = _fjsp_mul_v2r8(rsq23,rinv23);
1991
1992             /* Calculate table index by multiplying r with table scale and truncate to integer */
1993             rt               = _fjsp_mul_v2r8(r23,vftabscale);
1994             itab_tmp         = _fjsp_dtox_v2r8(rt);
1995             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1996             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1997             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1998
1999             vfconv.i[0]     *= 4;
2000             vfconv.i[1]     *= 4;
2001
2002             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2003             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2004             F                = _fjsp_setzero_v2r8();
2005             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2006             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2007             H                = _fjsp_setzero_v2r8();
2008             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2009             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2010             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2011             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq23,FF),_fjsp_mul_v2r8(vftabscale,rinv23)));
2012
2013             fscal            = felec;
2014
2015             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2016
2017             /* Update vectorial force */
2018             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
2019             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
2020             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
2021
2022             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
2023             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
2024             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
2025
2026             /**************************
2027              * CALCULATE INTERACTIONS *
2028              **************************/
2029
2030             r31              = _fjsp_mul_v2r8(rsq31,rinv31);
2031
2032             /* Calculate table index by multiplying r with table scale and truncate to integer */
2033             rt               = _fjsp_mul_v2r8(r31,vftabscale);
2034             itab_tmp         = _fjsp_dtox_v2r8(rt);
2035             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
2036             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
2037             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
2038
2039             vfconv.i[0]     *= 4;
2040             vfconv.i[1]     *= 4;
2041
2042             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2043             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2044             F                = _fjsp_setzero_v2r8();
2045             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2046             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2047             H                = _fjsp_setzero_v2r8();
2048             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2049             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2050             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2051             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq31,FF),_fjsp_mul_v2r8(vftabscale,rinv31)));
2052
2053             fscal            = felec;
2054
2055             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2056
2057             /* Update vectorial force */
2058             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
2059             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
2060             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
2061
2062             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
2063             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
2064             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
2065
2066             /**************************
2067              * CALCULATE INTERACTIONS *
2068              **************************/
2069
2070             r32              = _fjsp_mul_v2r8(rsq32,rinv32);
2071
2072             /* Calculate table index by multiplying r with table scale and truncate to integer */
2073             rt               = _fjsp_mul_v2r8(r32,vftabscale);
2074             itab_tmp         = _fjsp_dtox_v2r8(rt);
2075             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
2076             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
2077             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
2078
2079             vfconv.i[0]     *= 4;
2080             vfconv.i[1]     *= 4;
2081
2082             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2083             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2084             F                = _fjsp_setzero_v2r8();
2085             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2086             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2087             H                = _fjsp_setzero_v2r8();
2088             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2089             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2090             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2091             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq32,FF),_fjsp_mul_v2r8(vftabscale,rinv32)));
2092
2093             fscal            = felec;
2094
2095             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2096
2097             /* Update vectorial force */
2098             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
2099             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
2100             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
2101
2102             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
2103             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
2104             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
2105
2106             /**************************
2107              * CALCULATE INTERACTIONS *
2108              **************************/
2109
2110             r33              = _fjsp_mul_v2r8(rsq33,rinv33);
2111
2112             /* Calculate table index by multiplying r with table scale and truncate to integer */
2113             rt               = _fjsp_mul_v2r8(r33,vftabscale);
2114             itab_tmp         = _fjsp_dtox_v2r8(rt);
2115             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
2116             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
2117             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
2118
2119             vfconv.i[0]     *= 4;
2120             vfconv.i[1]     *= 4;
2121
2122             /* CUBIC SPLINE TABLE ELECTROSTATICS */
2123             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
2124             F                = _fjsp_setzero_v2r8();
2125             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
2126             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] +2);
2127             H                = _fjsp_setzero_v2r8();
2128             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
2129             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(vfeps,H,G),F);
2130             FF               = _fjsp_madd_v2r8(_fjsp_madd_v2r8(twovfeps,H,G),vfeps,Fp);
2131             felec            = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_mul_v2r8(qq33,FF),_fjsp_mul_v2r8(vftabscale,rinv33)));
2132
2133             fscal            = felec;
2134
2135             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2136
2137             /* Update vectorial force */
2138             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
2139             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
2140             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
2141
2142             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
2143             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
2144             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
2145
2146             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2147
2148             /* Inner loop uses 378 flops */
2149         }
2150
2151         /* End of innermost loop */
2152
2153         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2154                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2155
2156         /* Increment number of inner iterations */
2157         inneriter                  += j_index_end - j_index_start;
2158
2159         /* Outer loop uses 18 flops */
2160     }
2161
2162     /* Increment number of outer iterations */
2163     outeriter        += nri;
2164
2165     /* Update outer/inner flops */
2166
2167     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*378);
2168 }