src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_sse2_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 sse2_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_double.h"
  50 #include "kernelutil_x86_sse2_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_VF_sse2_double
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_VF_sse2_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B;
  91     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B;
  93     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B;
  95     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B;
  97     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 116     __m128i          vfitab;
 117     __m128i          ifour       = _mm_set1_epi32(4);
 118     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 119     real             *vftab;
 120     __m128d          dummy_mask,cutoff_mask;
 121     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 122     __m128d          one     = _mm_set1_pd(1.0);
 123     __m128d          two     = _mm_set1_pd(2.0);
 124     x                = xx[0];
 125     f                = ff[0];
 126
 127     nri              = nlist->nri;
 128     iinr             = nlist->iinr;
 129     jindex           = nlist->jindex;
 130     jjnr             = nlist->jjnr;
 131     shiftidx         = nlist->shift;
 132     gid              = nlist->gid;
 133     shiftvec         = fr->shift_vec[0];
 134     fshift           = fr->fshift[0];
 135     facel            = _mm_set1_pd(fr->epsfac);
 136     charge           = mdatoms->chargeA;
 137     nvdwtype         = fr->ntype;
 138     vdwparam         = fr->nbfp;
 139     vdwtype          = mdatoms->typeA;
 140
 141     vftab            = kernel_data->table_vdw->data;
 142     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 143
 144     /* Setup water-specific parameters */
 145     inr              = nlist->iinr[0];
 146     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 147     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 148     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 149     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 150
 151     jq1              = _mm_set1_pd(charge[inr+1]);
 152     jq2              = _mm_set1_pd(charge[inr+2]);
 153     jq3              = _mm_set1_pd(charge[inr+3]);
 154     vdwjidx0A        = 2*vdwtype[inr+0];
 155     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 156     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 157     qq11             = _mm_mul_pd(iq1,jq1);
 158     qq12             = _mm_mul_pd(iq1,jq2);
 159     qq13             = _mm_mul_pd(iq1,jq3);
 160     qq21             = _mm_mul_pd(iq2,jq1);
 161     qq22             = _mm_mul_pd(iq2,jq2);
 162     qq23             = _mm_mul_pd(iq2,jq3);
 163     qq31             = _mm_mul_pd(iq3,jq1);
 164     qq32             = _mm_mul_pd(iq3,jq2);
 165     qq33             = _mm_mul_pd(iq3,jq3);
 166
 167     /* Avoid stupid compiler warnings */
 168     jnrA = jnrB = 0;
 169     j_coord_offsetA = 0;
 170     j_coord_offsetB = 0;
 171
 172     outeriter        = 0;
 173     inneriter        = 0;
 174
 175     /* Start outer loop over neighborlists */
 176     for(iidx=0; iidx<nri; iidx++)
 177     {
 178         /* Load shift vector for this list */
 179         i_shift_offset   = DIM*shiftidx[iidx];
 180
 181         /* Load limits for loop over neighbors */
 182         j_index_start    = jindex[iidx];
 183         j_index_end      = jindex[iidx+1];
 184
 185         /* Get outer coordinate index */
 186         inr              = iinr[iidx];
 187         i_coord_offset   = DIM*inr;
 188
 189         /* Load i particle coords and add shift vector */
 190         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 191                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 192
 193         fix0             = _mm_setzero_pd();
 194         fiy0             = _mm_setzero_pd();
 195         fiz0             = _mm_setzero_pd();
 196         fix1             = _mm_setzero_pd();
 197         fiy1             = _mm_setzero_pd();
 198         fiz1             = _mm_setzero_pd();
 199         fix2             = _mm_setzero_pd();
 200         fiy2             = _mm_setzero_pd();
 201         fiz2             = _mm_setzero_pd();
 202         fix3             = _mm_setzero_pd();
 203         fiy3             = _mm_setzero_pd();
 204         fiz3             = _mm_setzero_pd();
 205
 206         /* Reset potential sums */
 207         velecsum         = _mm_setzero_pd();
 208         vvdwsum          = _mm_setzero_pd();
 209
 210         /* Start inner kernel loop */
 211         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 212         {
 213
 214             /* Get j neighbor index, and coordinate index */
 215             jnrA             = jjnr[jidx];
 216             jnrB             = jjnr[jidx+1];
 217             j_coord_offsetA  = DIM*jnrA;
 218             j_coord_offsetB  = DIM*jnrB;
 219
 220             /* load j atom coordinates */
 221             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 222                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 223                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 224
 225             /* Calculate displacement vector */
 226             dx00             = _mm_sub_pd(ix0,jx0);
 227             dy00             = _mm_sub_pd(iy0,jy0);
 228             dz00             = _mm_sub_pd(iz0,jz0);
 229             dx11             = _mm_sub_pd(ix1,jx1);
 230             dy11             = _mm_sub_pd(iy1,jy1);
 231             dz11             = _mm_sub_pd(iz1,jz1);
 232             dx12             = _mm_sub_pd(ix1,jx2);
 233             dy12             = _mm_sub_pd(iy1,jy2);
 234             dz12             = _mm_sub_pd(iz1,jz2);
 235             dx13             = _mm_sub_pd(ix1,jx3);
 236             dy13             = _mm_sub_pd(iy1,jy3);
 237             dz13             = _mm_sub_pd(iz1,jz3);
 238             dx21             = _mm_sub_pd(ix2,jx1);
 239             dy21             = _mm_sub_pd(iy2,jy1);
 240             dz21             = _mm_sub_pd(iz2,jz1);
 241             dx22             = _mm_sub_pd(ix2,jx2);
 242             dy22             = _mm_sub_pd(iy2,jy2);
 243             dz22             = _mm_sub_pd(iz2,jz2);
 244             dx23             = _mm_sub_pd(ix2,jx3);
 245             dy23             = _mm_sub_pd(iy2,jy3);
 246             dz23             = _mm_sub_pd(iz2,jz3);
 247             dx31             = _mm_sub_pd(ix3,jx1);
 248             dy31             = _mm_sub_pd(iy3,jy1);
 249             dz31             = _mm_sub_pd(iz3,jz1);
 250             dx32             = _mm_sub_pd(ix3,jx2);
 251             dy32             = _mm_sub_pd(iy3,jy2);
 252             dz32             = _mm_sub_pd(iz3,jz2);
 253             dx33             = _mm_sub_pd(ix3,jx3);
 254             dy33             = _mm_sub_pd(iy3,jy3);
 255             dz33             = _mm_sub_pd(iz3,jz3);
 256
 257             /* Calculate squared distance and things based on it */
 258             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 259             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 260             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 261             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 262             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 263             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 264             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 265             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 266             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 267             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 268
 269             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 270             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 271             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 272             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 273             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 274             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 275             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 276             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 277             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 278             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 279
 280             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 281             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 282             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 283             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 284             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 285             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 286             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 287             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 288             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 289
 290             fjx0             = _mm_setzero_pd();
 291             fjy0             = _mm_setzero_pd();
 292             fjz0             = _mm_setzero_pd();
 293             fjx1             = _mm_setzero_pd();
 294             fjy1             = _mm_setzero_pd();
 295             fjz1             = _mm_setzero_pd();
 296             fjx2             = _mm_setzero_pd();
 297             fjy2             = _mm_setzero_pd();
 298             fjz2             = _mm_setzero_pd();
 299             fjx3             = _mm_setzero_pd();
 300             fjy3             = _mm_setzero_pd();
 301             fjz3             = _mm_setzero_pd();
 302
 303             /**************************
 304              * CALCULATE INTERACTIONS *
 305              **************************/
 306
 307             r00              = _mm_mul_pd(rsq00,rinv00);
 308
 309             /* Calculate table index by multiplying r with table scale and truncate to integer */
 310             rt               = _mm_mul_pd(r00,vftabscale);
 311             vfitab           = _mm_cvttpd_epi32(rt);
 312             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 313             vfitab           = _mm_slli_epi32(vfitab,3);
 314
 315             /* CUBIC SPLINE TABLE DISPERSION */
 316             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 317             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 318             GMX_MM_TRANSPOSE2_PD(Y,F);
 319             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 320             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 321             GMX_MM_TRANSPOSE2_PD(G,H);
 322             Heps             = _mm_mul_pd(vfeps,H);
 323             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 324             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 325             vvdw6            = _mm_mul_pd(c6_00,VV);
 326             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 327             fvdw6            = _mm_mul_pd(c6_00,FF);
 328
 329             /* CUBIC SPLINE TABLE REPULSION */
 330             vfitab           = _mm_add_epi32(vfitab,ifour);
 331             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 332             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 333             GMX_MM_TRANSPOSE2_PD(Y,F);
 334             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 335             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 336             GMX_MM_TRANSPOSE2_PD(G,H);
 337             Heps             = _mm_mul_pd(vfeps,H);
 338             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 339             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 340             vvdw12           = _mm_mul_pd(c12_00,VV);
 341             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 342             fvdw12           = _mm_mul_pd(c12_00,FF);
 343             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 344             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 348
 349             fscal            = fvdw;
 350
 351             /* Calculate temporary vectorial force */
 352             tx               = _mm_mul_pd(fscal,dx00);
 353             ty               = _mm_mul_pd(fscal,dy00);
 354             tz               = _mm_mul_pd(fscal,dz00);
 355
 356             /* Update vectorial force */
 357             fix0             = _mm_add_pd(fix0,tx);
 358             fiy0             = _mm_add_pd(fiy0,ty);
 359             fiz0             = _mm_add_pd(fiz0,tz);
 360
 361             fjx0             = _mm_add_pd(fjx0,tx);
 362             fjy0             = _mm_add_pd(fjy0,ty);
 363             fjz0             = _mm_add_pd(fjz0,tz);
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             /* COULOMB ELECTROSTATICS */
 370             velec            = _mm_mul_pd(qq11,rinv11);
 371             felec            = _mm_mul_pd(velec,rinvsq11);
 372
 373             /* Update potential sum for this i atom from the interaction with this j atom. */
 374             velecsum         = _mm_add_pd(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             /* Calculate temporary vectorial force */
 379             tx               = _mm_mul_pd(fscal,dx11);
 380             ty               = _mm_mul_pd(fscal,dy11);
 381             tz               = _mm_mul_pd(fscal,dz11);
 382
 383             /* Update vectorial force */
 384             fix1             = _mm_add_pd(fix1,tx);
 385             fiy1             = _mm_add_pd(fiy1,ty);
 386             fiz1             = _mm_add_pd(fiz1,tz);
 387
 388             fjx1             = _mm_add_pd(fjx1,tx);
 389             fjy1             = _mm_add_pd(fjy1,ty);
 390             fjz1             = _mm_add_pd(fjz1,tz);
 391
 392             /**************************
 393              * CALCULATE INTERACTIONS *
 394              **************************/
 395
 396             /* COULOMB ELECTROSTATICS */
 397             velec            = _mm_mul_pd(qq12,rinv12);
 398             felec            = _mm_mul_pd(velec,rinvsq12);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             velecsum         = _mm_add_pd(velecsum,velec);
 402
 403             fscal            = felec;
 404
 405             /* Calculate temporary vectorial force */
 406             tx               = _mm_mul_pd(fscal,dx12);
 407             ty               = _mm_mul_pd(fscal,dy12);
 408             tz               = _mm_mul_pd(fscal,dz12);
 409
 410             /* Update vectorial force */
 411             fix1             = _mm_add_pd(fix1,tx);
 412             fiy1             = _mm_add_pd(fiy1,ty);
 413             fiz1             = _mm_add_pd(fiz1,tz);
 414
 415             fjx2             = _mm_add_pd(fjx2,tx);
 416             fjy2             = _mm_add_pd(fjy2,ty);
 417             fjz2             = _mm_add_pd(fjz2,tz);
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             /* COULOMB ELECTROSTATICS */
 424             velec            = _mm_mul_pd(qq13,rinv13);
 425             felec            = _mm_mul_pd(velec,rinvsq13);
 426
 427             /* Update potential sum for this i atom from the interaction with this j atom. */
 428             velecsum         = _mm_add_pd(velecsum,velec);
 429
 430             fscal            = felec;
 431
 432             /* Calculate temporary vectorial force */
 433             tx               = _mm_mul_pd(fscal,dx13);
 434             ty               = _mm_mul_pd(fscal,dy13);
 435             tz               = _mm_mul_pd(fscal,dz13);
 436
 437             /* Update vectorial force */
 438             fix1             = _mm_add_pd(fix1,tx);
 439             fiy1             = _mm_add_pd(fiy1,ty);
 440             fiz1             = _mm_add_pd(fiz1,tz);
 441
 442             fjx3             = _mm_add_pd(fjx3,tx);
 443             fjy3             = _mm_add_pd(fjy3,ty);
 444             fjz3             = _mm_add_pd(fjz3,tz);
 445
 446             /**************************
 447              * CALCULATE INTERACTIONS *
 448              **************************/
 449
 450             /* COULOMB ELECTROSTATICS */
 451             velec            = _mm_mul_pd(qq21,rinv21);
 452             felec            = _mm_mul_pd(velec,rinvsq21);
 453
 454             /* Update potential sum for this i atom from the interaction with this j atom. */
 455             velecsum         = _mm_add_pd(velecsum,velec);
 456
 457             fscal            = felec;
 458
 459             /* Calculate temporary vectorial force */
 460             tx               = _mm_mul_pd(fscal,dx21);
 461             ty               = _mm_mul_pd(fscal,dy21);
 462             tz               = _mm_mul_pd(fscal,dz21);
 463
 464             /* Update vectorial force */
 465             fix2             = _mm_add_pd(fix2,tx);
 466             fiy2             = _mm_add_pd(fiy2,ty);
 467             fiz2             = _mm_add_pd(fiz2,tz);
 468
 469             fjx1             = _mm_add_pd(fjx1,tx);
 470             fjy1             = _mm_add_pd(fjy1,ty);
 471             fjz1             = _mm_add_pd(fjz1,tz);
 472
 473             /**************************
 474              * CALCULATE INTERACTIONS *
 475              **************************/
 476
 477             /* COULOMB ELECTROSTATICS */
 478             velec            = _mm_mul_pd(qq22,rinv22);
 479             felec            = _mm_mul_pd(velec,rinvsq22);
 480
 481             /* Update potential sum for this i atom from the interaction with this j atom. */
 482             velecsum         = _mm_add_pd(velecsum,velec);
 483
 484             fscal            = felec;
 485
 486             /* Calculate temporary vectorial force */
 487             tx               = _mm_mul_pd(fscal,dx22);
 488             ty               = _mm_mul_pd(fscal,dy22);
 489             tz               = _mm_mul_pd(fscal,dz22);
 490
 491             /* Update vectorial force */
 492             fix2             = _mm_add_pd(fix2,tx);
 493             fiy2             = _mm_add_pd(fiy2,ty);
 494             fiz2             = _mm_add_pd(fiz2,tz);
 495
 496             fjx2             = _mm_add_pd(fjx2,tx);
 497             fjy2             = _mm_add_pd(fjy2,ty);
 498             fjz2             = _mm_add_pd(fjz2,tz);
 499
 500             /**************************
 501              * CALCULATE INTERACTIONS *
 502              **************************/
 503
 504             /* COULOMB ELECTROSTATICS */
 505             velec            = _mm_mul_pd(qq23,rinv23);
 506             felec            = _mm_mul_pd(velec,rinvsq23);
 507
 508             /* Update potential sum for this i atom from the interaction with this j atom. */
 509             velecsum         = _mm_add_pd(velecsum,velec);
 510
 511             fscal            = felec;
 512
 513             /* Calculate temporary vectorial force */
 514             tx               = _mm_mul_pd(fscal,dx23);
 515             ty               = _mm_mul_pd(fscal,dy23);
 516             tz               = _mm_mul_pd(fscal,dz23);
 517
 518             /* Update vectorial force */
 519             fix2             = _mm_add_pd(fix2,tx);
 520             fiy2             = _mm_add_pd(fiy2,ty);
 521             fiz2             = _mm_add_pd(fiz2,tz);
 522
 523             fjx3             = _mm_add_pd(fjx3,tx);
 524             fjy3             = _mm_add_pd(fjy3,ty);
 525             fjz3             = _mm_add_pd(fjz3,tz);
 526
 527             /**************************
 528              * CALCULATE INTERACTIONS *
 529              **************************/
 530
 531             /* COULOMB ELECTROSTATICS */
 532             velec            = _mm_mul_pd(qq31,rinv31);
 533             felec            = _mm_mul_pd(velec,rinvsq31);
 534
 535             /* Update potential sum for this i atom from the interaction with this j atom. */
 536             velecsum         = _mm_add_pd(velecsum,velec);
 537
 538             fscal            = felec;
 539
 540             /* Calculate temporary vectorial force */
 541             tx               = _mm_mul_pd(fscal,dx31);
 542             ty               = _mm_mul_pd(fscal,dy31);
 543             tz               = _mm_mul_pd(fscal,dz31);
 544
 545             /* Update vectorial force */
 546             fix3             = _mm_add_pd(fix3,tx);
 547             fiy3             = _mm_add_pd(fiy3,ty);
 548             fiz3             = _mm_add_pd(fiz3,tz);
 549
 550             fjx1             = _mm_add_pd(fjx1,tx);
 551             fjy1             = _mm_add_pd(fjy1,ty);
 552             fjz1             = _mm_add_pd(fjz1,tz);
 553
 554             /**************************
 555              * CALCULATE INTERACTIONS *
 556              **************************/
 557
 558             /* COULOMB ELECTROSTATICS */
 559             velec            = _mm_mul_pd(qq32,rinv32);
 560             felec            = _mm_mul_pd(velec,rinvsq32);
 561
 562             /* Update potential sum for this i atom from the interaction with this j atom. */
 563             velecsum         = _mm_add_pd(velecsum,velec);
 564
 565             fscal            = felec;
 566
 567             /* Calculate temporary vectorial force */
 568             tx               = _mm_mul_pd(fscal,dx32);
 569             ty               = _mm_mul_pd(fscal,dy32);
 570             tz               = _mm_mul_pd(fscal,dz32);
 571
 572             /* Update vectorial force */
 573             fix3             = _mm_add_pd(fix3,tx);
 574             fiy3             = _mm_add_pd(fiy3,ty);
 575             fiz3             = _mm_add_pd(fiz3,tz);
 576
 577             fjx2             = _mm_add_pd(fjx2,tx);
 578             fjy2             = _mm_add_pd(fjy2,ty);
 579             fjz2             = _mm_add_pd(fjz2,tz);
 580
 581             /**************************
 582              * CALCULATE INTERACTIONS *
 583              **************************/
 584
 585             /* COULOMB ELECTROSTATICS */
 586             velec            = _mm_mul_pd(qq33,rinv33);
 587             felec            = _mm_mul_pd(velec,rinvsq33);
 588
 589             /* Update potential sum for this i atom from the interaction with this j atom. */
 590             velecsum         = _mm_add_pd(velecsum,velec);
 591
 592             fscal            = felec;
 593
 594             /* Calculate temporary vectorial force */
 595             tx               = _mm_mul_pd(fscal,dx33);
 596             ty               = _mm_mul_pd(fscal,dy33);
 597             tz               = _mm_mul_pd(fscal,dz33);
 598
 599             /* Update vectorial force */
 600             fix3             = _mm_add_pd(fix3,tx);
 601             fiy3             = _mm_add_pd(fiy3,ty);
 602             fiz3             = _mm_add_pd(fiz3,tz);
 603
 604             fjx3             = _mm_add_pd(fjx3,tx);
 605             fjy3             = _mm_add_pd(fjy3,ty);
 606             fjz3             = _mm_add_pd(fjz3,tz);
 607
 608             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 609
 610             /* Inner loop uses 311 flops */
 611         }
 612
 613         if(jidx<j_index_end)
 614         {
 615
 616             jnrA             = jjnr[jidx];
 617             j_coord_offsetA  = DIM*jnrA;
 618
 619             /* load j atom coordinates */
 620             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 621                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 622                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 623
 624             /* Calculate displacement vector */
 625             dx00             = _mm_sub_pd(ix0,jx0);
 626             dy00             = _mm_sub_pd(iy0,jy0);
 627             dz00             = _mm_sub_pd(iz0,jz0);
 628             dx11             = _mm_sub_pd(ix1,jx1);
 629             dy11             = _mm_sub_pd(iy1,jy1);
 630             dz11             = _mm_sub_pd(iz1,jz1);
 631             dx12             = _mm_sub_pd(ix1,jx2);
 632             dy12             = _mm_sub_pd(iy1,jy2);
 633             dz12             = _mm_sub_pd(iz1,jz2);
 634             dx13             = _mm_sub_pd(ix1,jx3);
 635             dy13             = _mm_sub_pd(iy1,jy3);
 636             dz13             = _mm_sub_pd(iz1,jz3);
 637             dx21             = _mm_sub_pd(ix2,jx1);
 638             dy21             = _mm_sub_pd(iy2,jy1);
 639             dz21             = _mm_sub_pd(iz2,jz1);
 640             dx22             = _mm_sub_pd(ix2,jx2);
 641             dy22             = _mm_sub_pd(iy2,jy2);
 642             dz22             = _mm_sub_pd(iz2,jz2);
 643             dx23             = _mm_sub_pd(ix2,jx3);
 644             dy23             = _mm_sub_pd(iy2,jy3);
 645             dz23             = _mm_sub_pd(iz2,jz3);
 646             dx31             = _mm_sub_pd(ix3,jx1);
 647             dy31             = _mm_sub_pd(iy3,jy1);
 648             dz31             = _mm_sub_pd(iz3,jz1);
 649             dx32             = _mm_sub_pd(ix3,jx2);
 650             dy32             = _mm_sub_pd(iy3,jy2);
 651             dz32             = _mm_sub_pd(iz3,jz2);
 652             dx33             = _mm_sub_pd(ix3,jx3);
 653             dy33             = _mm_sub_pd(iy3,jy3);
 654             dz33             = _mm_sub_pd(iz3,jz3);
 655
 656             /* Calculate squared distance and things based on it */
 657             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 658             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 659             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 660             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 661             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 662             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 663             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 664             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 665             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 666             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 667
 668             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 669             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 670             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 671             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 672             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 673             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 674             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 675             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 676             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 677             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 678
 679             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 680             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 681             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 682             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 683             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 684             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 685             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 686             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 687             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 688
 689             fjx0             = _mm_setzero_pd();
 690             fjy0             = _mm_setzero_pd();
 691             fjz0             = _mm_setzero_pd();
 692             fjx1             = _mm_setzero_pd();
 693             fjy1             = _mm_setzero_pd();
 694             fjz1             = _mm_setzero_pd();
 695             fjx2             = _mm_setzero_pd();
 696             fjy2             = _mm_setzero_pd();
 697             fjz2             = _mm_setzero_pd();
 698             fjx3             = _mm_setzero_pd();
 699             fjy3             = _mm_setzero_pd();
 700             fjz3             = _mm_setzero_pd();
 701
 702             /**************************
 703              * CALCULATE INTERACTIONS *
 704              **************************/
 705
 706             r00              = _mm_mul_pd(rsq00,rinv00);
 707
 708             /* Calculate table index by multiplying r with table scale and truncate to integer */
 709             rt               = _mm_mul_pd(r00,vftabscale);
 710             vfitab           = _mm_cvttpd_epi32(rt);
 711             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 712             vfitab           = _mm_slli_epi32(vfitab,3);
 713
 714             /* CUBIC SPLINE TABLE DISPERSION */
 715             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 716             F                = _mm_setzero_pd();
 717             GMX_MM_TRANSPOSE2_PD(Y,F);
 718             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 719             H                = _mm_setzero_pd();
 720             GMX_MM_TRANSPOSE2_PD(G,H);
 721             Heps             = _mm_mul_pd(vfeps,H);
 722             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 723             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 724             vvdw6            = _mm_mul_pd(c6_00,VV);
 725             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 726             fvdw6            = _mm_mul_pd(c6_00,FF);
 727
 728             /* CUBIC SPLINE TABLE REPULSION */
 729             vfitab           = _mm_add_epi32(vfitab,ifour);
 730             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 731             F                = _mm_setzero_pd();
 732             GMX_MM_TRANSPOSE2_PD(Y,F);
 733             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 734             H                = _mm_setzero_pd();
 735             GMX_MM_TRANSPOSE2_PD(G,H);
 736             Heps             = _mm_mul_pd(vfeps,H);
 737             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 738             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 739             vvdw12           = _mm_mul_pd(c12_00,VV);
 740             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 741             fvdw12           = _mm_mul_pd(c12_00,FF);
 742             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 743             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 744
 745             /* Update potential sum for this i atom from the interaction with this j atom. */
 746             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 747             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 748
 749             fscal            = fvdw;
 750
 751             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 752
 753             /* Calculate temporary vectorial force */
 754             tx               = _mm_mul_pd(fscal,dx00);
 755             ty               = _mm_mul_pd(fscal,dy00);
 756             tz               = _mm_mul_pd(fscal,dz00);
 757
 758             /* Update vectorial force */
 759             fix0             = _mm_add_pd(fix0,tx);
 760             fiy0             = _mm_add_pd(fiy0,ty);
 761             fiz0             = _mm_add_pd(fiz0,tz);
 762
 763             fjx0             = _mm_add_pd(fjx0,tx);
 764             fjy0             = _mm_add_pd(fjy0,ty);
 765             fjz0             = _mm_add_pd(fjz0,tz);
 766
 767             /**************************
 768              * CALCULATE INTERACTIONS *
 769              **************************/
 770
 771             /* COULOMB ELECTROSTATICS */
 772             velec            = _mm_mul_pd(qq11,rinv11);
 773             felec            = _mm_mul_pd(velec,rinvsq11);
 774
 775             /* Update potential sum for this i atom from the interaction with this j atom. */
 776             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 777             velecsum         = _mm_add_pd(velecsum,velec);
 778
 779             fscal            = felec;
 780
 781             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 782
 783             /* Calculate temporary vectorial force */
 784             tx               = _mm_mul_pd(fscal,dx11);
 785             ty               = _mm_mul_pd(fscal,dy11);
 786             tz               = _mm_mul_pd(fscal,dz11);
 787
 788             /* Update vectorial force */
 789             fix1             = _mm_add_pd(fix1,tx);
 790             fiy1             = _mm_add_pd(fiy1,ty);
 791             fiz1             = _mm_add_pd(fiz1,tz);
 792
 793             fjx1             = _mm_add_pd(fjx1,tx);
 794             fjy1             = _mm_add_pd(fjy1,ty);
 795             fjz1             = _mm_add_pd(fjz1,tz);
 796
 797             /**************************
 798              * CALCULATE INTERACTIONS *
 799              **************************/
 800
 801             /* COULOMB ELECTROSTATICS */
 802             velec            = _mm_mul_pd(qq12,rinv12);
 803             felec            = _mm_mul_pd(velec,rinvsq12);
 804
 805             /* Update potential sum for this i atom from the interaction with this j atom. */
 806             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 807             velecsum         = _mm_add_pd(velecsum,velec);
 808
 809             fscal            = felec;
 810
 811             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 812
 813             /* Calculate temporary vectorial force */
 814             tx               = _mm_mul_pd(fscal,dx12);
 815             ty               = _mm_mul_pd(fscal,dy12);
 816             tz               = _mm_mul_pd(fscal,dz12);
 817
 818             /* Update vectorial force */
 819             fix1             = _mm_add_pd(fix1,tx);
 820             fiy1             = _mm_add_pd(fiy1,ty);
 821             fiz1             = _mm_add_pd(fiz1,tz);
 822
 823             fjx2             = _mm_add_pd(fjx2,tx);
 824             fjy2             = _mm_add_pd(fjy2,ty);
 825             fjz2             = _mm_add_pd(fjz2,tz);
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             /* COULOMB ELECTROSTATICS */
 832             velec            = _mm_mul_pd(qq13,rinv13);
 833             felec            = _mm_mul_pd(velec,rinvsq13);
 834
 835             /* Update potential sum for this i atom from the interaction with this j atom. */
 836             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 837             velecsum         = _mm_add_pd(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 842
 843             /* Calculate temporary vectorial force */
 844             tx               = _mm_mul_pd(fscal,dx13);
 845             ty               = _mm_mul_pd(fscal,dy13);
 846             tz               = _mm_mul_pd(fscal,dz13);
 847
 848             /* Update vectorial force */
 849             fix1             = _mm_add_pd(fix1,tx);
 850             fiy1             = _mm_add_pd(fiy1,ty);
 851             fiz1             = _mm_add_pd(fiz1,tz);
 852
 853             fjx3             = _mm_add_pd(fjx3,tx);
 854             fjy3             = _mm_add_pd(fjy3,ty);
 855             fjz3             = _mm_add_pd(fjz3,tz);
 856
 857             /**************************
 858              * CALCULATE INTERACTIONS *
 859              **************************/
 860
 861             /* COULOMB ELECTROSTATICS */
 862             velec            = _mm_mul_pd(qq21,rinv21);
 863             felec            = _mm_mul_pd(velec,rinvsq21);
 864
 865             /* Update potential sum for this i atom from the interaction with this j atom. */
 866             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 867             velecsum         = _mm_add_pd(velecsum,velec);
 868
 869             fscal            = felec;
 870
 871             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 872
 873             /* Calculate temporary vectorial force */
 874             tx               = _mm_mul_pd(fscal,dx21);
 875             ty               = _mm_mul_pd(fscal,dy21);
 876             tz               = _mm_mul_pd(fscal,dz21);
 877
 878             /* Update vectorial force */
 879             fix2             = _mm_add_pd(fix2,tx);
 880             fiy2             = _mm_add_pd(fiy2,ty);
 881             fiz2             = _mm_add_pd(fiz2,tz);
 882
 883             fjx1             = _mm_add_pd(fjx1,tx);
 884             fjy1             = _mm_add_pd(fjy1,ty);
 885             fjz1             = _mm_add_pd(fjz1,tz);
 886
 887             /**************************
 888              * CALCULATE INTERACTIONS *
 889              **************************/
 890
 891             /* COULOMB ELECTROSTATICS */
 892             velec            = _mm_mul_pd(qq22,rinv22);
 893             felec            = _mm_mul_pd(velec,rinvsq22);
 894
 895             /* Update potential sum for this i atom from the interaction with this j atom. */
 896             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 897             velecsum         = _mm_add_pd(velecsum,velec);
 898
 899             fscal            = felec;
 900
 901             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 902
 903             /* Calculate temporary vectorial force */
 904             tx               = _mm_mul_pd(fscal,dx22);
 905             ty               = _mm_mul_pd(fscal,dy22);
 906             tz               = _mm_mul_pd(fscal,dz22);
 907
 908             /* Update vectorial force */
 909             fix2             = _mm_add_pd(fix2,tx);
 910             fiy2             = _mm_add_pd(fiy2,ty);
 911             fiz2             = _mm_add_pd(fiz2,tz);
 912
 913             fjx2             = _mm_add_pd(fjx2,tx);
 914             fjy2             = _mm_add_pd(fjy2,ty);
 915             fjz2             = _mm_add_pd(fjz2,tz);
 916
 917             /**************************
 918              * CALCULATE INTERACTIONS *
 919              **************************/
 920
 921             /* COULOMB ELECTROSTATICS */
 922             velec            = _mm_mul_pd(qq23,rinv23);
 923             felec            = _mm_mul_pd(velec,rinvsq23);
 924
 925             /* Update potential sum for this i atom from the interaction with this j atom. */
 926             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 927             velecsum         = _mm_add_pd(velecsum,velec);
 928
 929             fscal            = felec;
 930
 931             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 932
 933             /* Calculate temporary vectorial force */
 934             tx               = _mm_mul_pd(fscal,dx23);
 935             ty               = _mm_mul_pd(fscal,dy23);
 936             tz               = _mm_mul_pd(fscal,dz23);
 937
 938             /* Update vectorial force */
 939             fix2             = _mm_add_pd(fix2,tx);
 940             fiy2             = _mm_add_pd(fiy2,ty);
 941             fiz2             = _mm_add_pd(fiz2,tz);
 942
 943             fjx3             = _mm_add_pd(fjx3,tx);
 944             fjy3             = _mm_add_pd(fjy3,ty);
 945             fjz3             = _mm_add_pd(fjz3,tz);
 946
 947             /**************************
 948              * CALCULATE INTERACTIONS *
 949              **************************/
 950
 951             /* COULOMB ELECTROSTATICS */
 952             velec            = _mm_mul_pd(qq31,rinv31);
 953             felec            = _mm_mul_pd(velec,rinvsq31);
 954
 955             /* Update potential sum for this i atom from the interaction with this j atom. */
 956             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 957             velecsum         = _mm_add_pd(velecsum,velec);
 958
 959             fscal            = felec;
 960
 961             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 962
 963             /* Calculate temporary vectorial force */
 964             tx               = _mm_mul_pd(fscal,dx31);
 965             ty               = _mm_mul_pd(fscal,dy31);
 966             tz               = _mm_mul_pd(fscal,dz31);
 967
 968             /* Update vectorial force */
 969             fix3             = _mm_add_pd(fix3,tx);
 970             fiy3             = _mm_add_pd(fiy3,ty);
 971             fiz3             = _mm_add_pd(fiz3,tz);
 972
 973             fjx1             = _mm_add_pd(fjx1,tx);
 974             fjy1             = _mm_add_pd(fjy1,ty);
 975             fjz1             = _mm_add_pd(fjz1,tz);
 976
 977             /**************************
 978              * CALCULATE INTERACTIONS *
 979              **************************/
 980
 981             /* COULOMB ELECTROSTATICS */
 982             velec            = _mm_mul_pd(qq32,rinv32);
 983             felec            = _mm_mul_pd(velec,rinvsq32);
 984
 985             /* Update potential sum for this i atom from the interaction with this j atom. */
 986             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 987             velecsum         = _mm_add_pd(velecsum,velec);
 988
 989             fscal            = felec;
 990
 991             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 992
 993             /* Calculate temporary vectorial force */
 994             tx               = _mm_mul_pd(fscal,dx32);
 995             ty               = _mm_mul_pd(fscal,dy32);
 996             tz               = _mm_mul_pd(fscal,dz32);
 997
 998             /* Update vectorial force */
 999             fix3             = _mm_add_pd(fix3,tx);
1000             fiy3             = _mm_add_pd(fiy3,ty);
1001             fiz3             = _mm_add_pd(fiz3,tz);
1002
1003             fjx2             = _mm_add_pd(fjx2,tx);
1004             fjy2             = _mm_add_pd(fjy2,ty);
1005             fjz2             = _mm_add_pd(fjz2,tz);
1006
1007             /**************************
1008              * CALCULATE INTERACTIONS *
1009              **************************/
1010
1011             /* COULOMB ELECTROSTATICS */
1012             velec            = _mm_mul_pd(qq33,rinv33);
1013             felec            = _mm_mul_pd(velec,rinvsq33);
1014
1015             /* Update potential sum for this i atom from the interaction with this j atom. */
1016             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1017             velecsum         = _mm_add_pd(velecsum,velec);
1018
1019             fscal            = felec;
1020
1021             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1022
1023             /* Calculate temporary vectorial force */
1024             tx               = _mm_mul_pd(fscal,dx33);
1025             ty               = _mm_mul_pd(fscal,dy33);
1026             tz               = _mm_mul_pd(fscal,dz33);
1027
1028             /* Update vectorial force */
1029             fix3             = _mm_add_pd(fix3,tx);
1030             fiy3             = _mm_add_pd(fiy3,ty);
1031             fiz3             = _mm_add_pd(fiz3,tz);
1032
1033             fjx3             = _mm_add_pd(fjx3,tx);
1034             fjy3             = _mm_add_pd(fjy3,ty);
1035             fjz3             = _mm_add_pd(fjz3,tz);
1036
1037             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1038
1039             /* Inner loop uses 311 flops */
1040         }
1041
1042         /* End of innermost loop */
1043
1044         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1045                                               f+i_coord_offset,fshift+i_shift_offset);
1046
1047         ggid                        = gid[iidx];
1048         /* Update potential energies */
1049         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1050         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1051
1052         /* Increment number of inner iterations */
1053         inneriter                  += j_index_end - j_index_start;
1054
1055         /* Outer loop uses 26 flops */
1056     }
1057
1058     /* Increment number of outer iterations */
1059     outeriter        += nri;
1060
1061     /* Update outer/inner flops */
1062
1063     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*311);
1064 }
1065 /*
1066  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_sse2_double
1067  * Electrostatics interaction: Coulomb
1068  * VdW interaction:            CubicSplineTable
1069  * Geometry:                   Water4-Water4
1070  * Calculate force/pot:        Force
1071  */
1072 void
1073 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_sse2_double
1074                     (t_nblist                    * gmx_restrict       nlist,
1075                      rvec                        * gmx_restrict          xx,
1076                      rvec                        * gmx_restrict          ff,
1077                      t_forcerec                  * gmx_restrict          fr,
1078                      t_mdatoms                   * gmx_restrict     mdatoms,
1079                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1080                      t_nrnb                      * gmx_restrict        nrnb)
1081 {
1082     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1083      * just 0 for non-waters.
1084      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1085      * jnr indices corresponding to data put in the four positions in the SIMD register.
1086      */
1087     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1088     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1089     int              jnrA,jnrB;
1090     int              j_coord_offsetA,j_coord_offsetB;
1091     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1092     real             rcutoff_scalar;
1093     real             *shiftvec,*fshift,*x,*f;
1094     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1095     int              vdwioffset0;
1096     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1097     int              vdwioffset1;
1098     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1099     int              vdwioffset2;
1100     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1101     int              vdwioffset3;
1102     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1103     int              vdwjidx0A,vdwjidx0B;
1104     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1105     int              vdwjidx1A,vdwjidx1B;
1106     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1107     int              vdwjidx2A,vdwjidx2B;
1108     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1109     int              vdwjidx3A,vdwjidx3B;
1110     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1111     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1112     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1113     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1114     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1115     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1116     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1117     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1118     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1119     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1120     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1121     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1122     real             *charge;
1123     int              nvdwtype;
1124     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1125     int              *vdwtype;
1126     real             *vdwparam;
1127     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1128     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1129     __m128i          vfitab;
1130     __m128i          ifour       = _mm_set1_epi32(4);
1131     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1132     real             *vftab;
1133     __m128d          dummy_mask,cutoff_mask;
1134     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1135     __m128d          one     = _mm_set1_pd(1.0);
1136     __m128d          two     = _mm_set1_pd(2.0);
1137     x                = xx[0];
1138     f                = ff[0];
1139
1140     nri              = nlist->nri;
1141     iinr             = nlist->iinr;
1142     jindex           = nlist->jindex;
1143     jjnr             = nlist->jjnr;
1144     shiftidx         = nlist->shift;
1145     gid              = nlist->gid;
1146     shiftvec         = fr->shift_vec[0];
1147     fshift           = fr->fshift[0];
1148     facel            = _mm_set1_pd(fr->epsfac);
1149     charge           = mdatoms->chargeA;
1150     nvdwtype         = fr->ntype;
1151     vdwparam         = fr->nbfp;
1152     vdwtype          = mdatoms->typeA;
1153
1154     vftab            = kernel_data->table_vdw->data;
1155     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1156
1157     /* Setup water-specific parameters */
1158     inr              = nlist->iinr[0];
1159     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1160     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1161     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1162     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1163
1164     jq1              = _mm_set1_pd(charge[inr+1]);
1165     jq2              = _mm_set1_pd(charge[inr+2]);
1166     jq3              = _mm_set1_pd(charge[inr+3]);
1167     vdwjidx0A        = 2*vdwtype[inr+0];
1168     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1169     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1170     qq11             = _mm_mul_pd(iq1,jq1);
1171     qq12             = _mm_mul_pd(iq1,jq2);
1172     qq13             = _mm_mul_pd(iq1,jq3);
1173     qq21             = _mm_mul_pd(iq2,jq1);
1174     qq22             = _mm_mul_pd(iq2,jq2);
1175     qq23             = _mm_mul_pd(iq2,jq3);
1176     qq31             = _mm_mul_pd(iq3,jq1);
1177     qq32             = _mm_mul_pd(iq3,jq2);
1178     qq33             = _mm_mul_pd(iq3,jq3);
1179
1180     /* Avoid stupid compiler warnings */
1181     jnrA = jnrB = 0;
1182     j_coord_offsetA = 0;
1183     j_coord_offsetB = 0;
1184
1185     outeriter        = 0;
1186     inneriter        = 0;
1187
1188     /* Start outer loop over neighborlists */
1189     for(iidx=0; iidx<nri; iidx++)
1190     {
1191         /* Load shift vector for this list */
1192         i_shift_offset   = DIM*shiftidx[iidx];
1193
1194         /* Load limits for loop over neighbors */
1195         j_index_start    = jindex[iidx];
1196         j_index_end      = jindex[iidx+1];
1197
1198         /* Get outer coordinate index */
1199         inr              = iinr[iidx];
1200         i_coord_offset   = DIM*inr;
1201
1202         /* Load i particle coords and add shift vector */
1203         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1204                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1205
1206         fix0             = _mm_setzero_pd();
1207         fiy0             = _mm_setzero_pd();
1208         fiz0             = _mm_setzero_pd();
1209         fix1             = _mm_setzero_pd();
1210         fiy1             = _mm_setzero_pd();
1211         fiz1             = _mm_setzero_pd();
1212         fix2             = _mm_setzero_pd();
1213         fiy2             = _mm_setzero_pd();
1214         fiz2             = _mm_setzero_pd();
1215         fix3             = _mm_setzero_pd();
1216         fiy3             = _mm_setzero_pd();
1217         fiz3             = _mm_setzero_pd();
1218
1219         /* Start inner kernel loop */
1220         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1221         {
1222
1223             /* Get j neighbor index, and coordinate index */
1224             jnrA             = jjnr[jidx];
1225             jnrB             = jjnr[jidx+1];
1226             j_coord_offsetA  = DIM*jnrA;
1227             j_coord_offsetB  = DIM*jnrB;
1228
1229             /* load j atom coordinates */
1230             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1231                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1232                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1233
1234             /* Calculate displacement vector */
1235             dx00             = _mm_sub_pd(ix0,jx0);
1236             dy00             = _mm_sub_pd(iy0,jy0);
1237             dz00             = _mm_sub_pd(iz0,jz0);
1238             dx11             = _mm_sub_pd(ix1,jx1);
1239             dy11             = _mm_sub_pd(iy1,jy1);
1240             dz11             = _mm_sub_pd(iz1,jz1);
1241             dx12             = _mm_sub_pd(ix1,jx2);
1242             dy12             = _mm_sub_pd(iy1,jy2);
1243             dz12             = _mm_sub_pd(iz1,jz2);
1244             dx13             = _mm_sub_pd(ix1,jx3);
1245             dy13             = _mm_sub_pd(iy1,jy3);
1246             dz13             = _mm_sub_pd(iz1,jz3);
1247             dx21             = _mm_sub_pd(ix2,jx1);
1248             dy21             = _mm_sub_pd(iy2,jy1);
1249             dz21             = _mm_sub_pd(iz2,jz1);
1250             dx22             = _mm_sub_pd(ix2,jx2);
1251             dy22             = _mm_sub_pd(iy2,jy2);
1252             dz22             = _mm_sub_pd(iz2,jz2);
1253             dx23             = _mm_sub_pd(ix2,jx3);
1254             dy23             = _mm_sub_pd(iy2,jy3);
1255             dz23             = _mm_sub_pd(iz2,jz3);
1256             dx31             = _mm_sub_pd(ix3,jx1);
1257             dy31             = _mm_sub_pd(iy3,jy1);
1258             dz31             = _mm_sub_pd(iz3,jz1);
1259             dx32             = _mm_sub_pd(ix3,jx2);
1260             dy32             = _mm_sub_pd(iy3,jy2);
1261             dz32             = _mm_sub_pd(iz3,jz2);
1262             dx33             = _mm_sub_pd(ix3,jx3);
1263             dy33             = _mm_sub_pd(iy3,jy3);
1264             dz33             = _mm_sub_pd(iz3,jz3);
1265
1266             /* Calculate squared distance and things based on it */
1267             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1268             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1269             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1270             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1271             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1272             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1273             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1274             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1275             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1276             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1277
1278             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1279             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1280             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1281             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1282             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1283             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1284             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1285             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1286             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1287             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1288
1289             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1290             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1291             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1292             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1293             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1294             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1295             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1296             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1297             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1298
1299             fjx0             = _mm_setzero_pd();
1300             fjy0             = _mm_setzero_pd();
1301             fjz0             = _mm_setzero_pd();
1302             fjx1             = _mm_setzero_pd();
1303             fjy1             = _mm_setzero_pd();
1304             fjz1             = _mm_setzero_pd();
1305             fjx2             = _mm_setzero_pd();
1306             fjy2             = _mm_setzero_pd();
1307             fjz2             = _mm_setzero_pd();
1308             fjx3             = _mm_setzero_pd();
1309             fjy3             = _mm_setzero_pd();
1310             fjz3             = _mm_setzero_pd();
1311
1312             /**************************
1313              * CALCULATE INTERACTIONS *
1314              **************************/
1315
1316             r00              = _mm_mul_pd(rsq00,rinv00);
1317
1318             /* Calculate table index by multiplying r with table scale and truncate to integer */
1319             rt               = _mm_mul_pd(r00,vftabscale);
1320             vfitab           = _mm_cvttpd_epi32(rt);
1321             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1322             vfitab           = _mm_slli_epi32(vfitab,3);
1323
1324             /* CUBIC SPLINE TABLE DISPERSION */
1325             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1326             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1327             GMX_MM_TRANSPOSE2_PD(Y,F);
1328             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1329             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1330             GMX_MM_TRANSPOSE2_PD(G,H);
1331             Heps             = _mm_mul_pd(vfeps,H);
1332             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1333             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1334             fvdw6            = _mm_mul_pd(c6_00,FF);
1335
1336             /* CUBIC SPLINE TABLE REPULSION */
1337             vfitab           = _mm_add_epi32(vfitab,ifour);
1338             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1339             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1340             GMX_MM_TRANSPOSE2_PD(Y,F);
1341             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1342             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1343             GMX_MM_TRANSPOSE2_PD(G,H);
1344             Heps             = _mm_mul_pd(vfeps,H);
1345             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1346             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1347             fvdw12           = _mm_mul_pd(c12_00,FF);
1348             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1349
1350             fscal            = fvdw;
1351
1352             /* Calculate temporary vectorial force */
1353             tx               = _mm_mul_pd(fscal,dx00);
1354             ty               = _mm_mul_pd(fscal,dy00);
1355             tz               = _mm_mul_pd(fscal,dz00);
1356
1357             /* Update vectorial force */
1358             fix0             = _mm_add_pd(fix0,tx);
1359             fiy0             = _mm_add_pd(fiy0,ty);
1360             fiz0             = _mm_add_pd(fiz0,tz);
1361
1362             fjx0             = _mm_add_pd(fjx0,tx);
1363             fjy0             = _mm_add_pd(fjy0,ty);
1364             fjz0             = _mm_add_pd(fjz0,tz);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* COULOMB ELECTROSTATICS */
1371             velec            = _mm_mul_pd(qq11,rinv11);
1372             felec            = _mm_mul_pd(velec,rinvsq11);
1373
1374             fscal            = felec;
1375
1376             /* Calculate temporary vectorial force */
1377             tx               = _mm_mul_pd(fscal,dx11);
1378             ty               = _mm_mul_pd(fscal,dy11);
1379             tz               = _mm_mul_pd(fscal,dz11);
1380
1381             /* Update vectorial force */
1382             fix1             = _mm_add_pd(fix1,tx);
1383             fiy1             = _mm_add_pd(fiy1,ty);
1384             fiz1             = _mm_add_pd(fiz1,tz);
1385
1386             fjx1             = _mm_add_pd(fjx1,tx);
1387             fjy1             = _mm_add_pd(fjy1,ty);
1388             fjz1             = _mm_add_pd(fjz1,tz);
1389
1390             /**************************
1391              * CALCULATE INTERACTIONS *
1392              **************************/
1393
1394             /* COULOMB ELECTROSTATICS */
1395             velec            = _mm_mul_pd(qq12,rinv12);
1396             felec            = _mm_mul_pd(velec,rinvsq12);
1397
1398             fscal            = felec;
1399
1400             /* Calculate temporary vectorial force */
1401             tx               = _mm_mul_pd(fscal,dx12);
1402             ty               = _mm_mul_pd(fscal,dy12);
1403             tz               = _mm_mul_pd(fscal,dz12);
1404
1405             /* Update vectorial force */
1406             fix1             = _mm_add_pd(fix1,tx);
1407             fiy1             = _mm_add_pd(fiy1,ty);
1408             fiz1             = _mm_add_pd(fiz1,tz);
1409
1410             fjx2             = _mm_add_pd(fjx2,tx);
1411             fjy2             = _mm_add_pd(fjy2,ty);
1412             fjz2             = _mm_add_pd(fjz2,tz);
1413
1414             /**************************
1415              * CALCULATE INTERACTIONS *
1416              **************************/
1417
1418             /* COULOMB ELECTROSTATICS */
1419             velec            = _mm_mul_pd(qq13,rinv13);
1420             felec            = _mm_mul_pd(velec,rinvsq13);
1421
1422             fscal            = felec;
1423
1424             /* Calculate temporary vectorial force */
1425             tx               = _mm_mul_pd(fscal,dx13);
1426             ty               = _mm_mul_pd(fscal,dy13);
1427             tz               = _mm_mul_pd(fscal,dz13);
1428
1429             /* Update vectorial force */
1430             fix1             = _mm_add_pd(fix1,tx);
1431             fiy1             = _mm_add_pd(fiy1,ty);
1432             fiz1             = _mm_add_pd(fiz1,tz);
1433
1434             fjx3             = _mm_add_pd(fjx3,tx);
1435             fjy3             = _mm_add_pd(fjy3,ty);
1436             fjz3             = _mm_add_pd(fjz3,tz);
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             /* COULOMB ELECTROSTATICS */
1443             velec            = _mm_mul_pd(qq21,rinv21);
1444             felec            = _mm_mul_pd(velec,rinvsq21);
1445
1446             fscal            = felec;
1447
1448             /* Calculate temporary vectorial force */
1449             tx               = _mm_mul_pd(fscal,dx21);
1450             ty               = _mm_mul_pd(fscal,dy21);
1451             tz               = _mm_mul_pd(fscal,dz21);
1452
1453             /* Update vectorial force */
1454             fix2             = _mm_add_pd(fix2,tx);
1455             fiy2             = _mm_add_pd(fiy2,ty);
1456             fiz2             = _mm_add_pd(fiz2,tz);
1457
1458             fjx1             = _mm_add_pd(fjx1,tx);
1459             fjy1             = _mm_add_pd(fjy1,ty);
1460             fjz1             = _mm_add_pd(fjz1,tz);
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             /* COULOMB ELECTROSTATICS */
1467             velec            = _mm_mul_pd(qq22,rinv22);
1468             felec            = _mm_mul_pd(velec,rinvsq22);
1469
1470             fscal            = felec;
1471
1472             /* Calculate temporary vectorial force */
1473             tx               = _mm_mul_pd(fscal,dx22);
1474             ty               = _mm_mul_pd(fscal,dy22);
1475             tz               = _mm_mul_pd(fscal,dz22);
1476
1477             /* Update vectorial force */
1478             fix2             = _mm_add_pd(fix2,tx);
1479             fiy2             = _mm_add_pd(fiy2,ty);
1480             fiz2             = _mm_add_pd(fiz2,tz);
1481
1482             fjx2             = _mm_add_pd(fjx2,tx);
1483             fjy2             = _mm_add_pd(fjy2,ty);
1484             fjz2             = _mm_add_pd(fjz2,tz);
1485
1486             /**************************
1487              * CALCULATE INTERACTIONS *
1488              **************************/
1489
1490             /* COULOMB ELECTROSTATICS */
1491             velec            = _mm_mul_pd(qq23,rinv23);
1492             felec            = _mm_mul_pd(velec,rinvsq23);
1493
1494             fscal            = felec;
1495
1496             /* Calculate temporary vectorial force */
1497             tx               = _mm_mul_pd(fscal,dx23);
1498             ty               = _mm_mul_pd(fscal,dy23);
1499             tz               = _mm_mul_pd(fscal,dz23);
1500
1501             /* Update vectorial force */
1502             fix2             = _mm_add_pd(fix2,tx);
1503             fiy2             = _mm_add_pd(fiy2,ty);
1504             fiz2             = _mm_add_pd(fiz2,tz);
1505
1506             fjx3             = _mm_add_pd(fjx3,tx);
1507             fjy3             = _mm_add_pd(fjy3,ty);
1508             fjz3             = _mm_add_pd(fjz3,tz);
1509
1510             /**************************
1511              * CALCULATE INTERACTIONS *
1512              **************************/
1513
1514             /* COULOMB ELECTROSTATICS */
1515             velec            = _mm_mul_pd(qq31,rinv31);
1516             felec            = _mm_mul_pd(velec,rinvsq31);
1517
1518             fscal            = felec;
1519
1520             /* Calculate temporary vectorial force */
1521             tx               = _mm_mul_pd(fscal,dx31);
1522             ty               = _mm_mul_pd(fscal,dy31);
1523             tz               = _mm_mul_pd(fscal,dz31);
1524
1525             /* Update vectorial force */
1526             fix3             = _mm_add_pd(fix3,tx);
1527             fiy3             = _mm_add_pd(fiy3,ty);
1528             fiz3             = _mm_add_pd(fiz3,tz);
1529
1530             fjx1             = _mm_add_pd(fjx1,tx);
1531             fjy1             = _mm_add_pd(fjy1,ty);
1532             fjz1             = _mm_add_pd(fjz1,tz);
1533
1534             /**************************
1535              * CALCULATE INTERACTIONS *
1536              **************************/
1537
1538             /* COULOMB ELECTROSTATICS */
1539             velec            = _mm_mul_pd(qq32,rinv32);
1540             felec            = _mm_mul_pd(velec,rinvsq32);
1541
1542             fscal            = felec;
1543
1544             /* Calculate temporary vectorial force */
1545             tx               = _mm_mul_pd(fscal,dx32);
1546             ty               = _mm_mul_pd(fscal,dy32);
1547             tz               = _mm_mul_pd(fscal,dz32);
1548
1549             /* Update vectorial force */
1550             fix3             = _mm_add_pd(fix3,tx);
1551             fiy3             = _mm_add_pd(fiy3,ty);
1552             fiz3             = _mm_add_pd(fiz3,tz);
1553
1554             fjx2             = _mm_add_pd(fjx2,tx);
1555             fjy2             = _mm_add_pd(fjy2,ty);
1556             fjz2             = _mm_add_pd(fjz2,tz);
1557
1558             /**************************
1559              * CALCULATE INTERACTIONS *
1560              **************************/
1561
1562             /* COULOMB ELECTROSTATICS */
1563             velec            = _mm_mul_pd(qq33,rinv33);
1564             felec            = _mm_mul_pd(velec,rinvsq33);
1565
1566             fscal            = felec;
1567
1568             /* Calculate temporary vectorial force */
1569             tx               = _mm_mul_pd(fscal,dx33);
1570             ty               = _mm_mul_pd(fscal,dy33);
1571             tz               = _mm_mul_pd(fscal,dz33);
1572
1573             /* Update vectorial force */
1574             fix3             = _mm_add_pd(fix3,tx);
1575             fiy3             = _mm_add_pd(fiy3,ty);
1576             fiz3             = _mm_add_pd(fiz3,tz);
1577
1578             fjx3             = _mm_add_pd(fjx3,tx);
1579             fjy3             = _mm_add_pd(fjy3,ty);
1580             fjz3             = _mm_add_pd(fjz3,tz);
1581
1582             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1583
1584             /* Inner loop uses 294 flops */
1585         }
1586
1587         if(jidx<j_index_end)
1588         {
1589
1590             jnrA             = jjnr[jidx];
1591             j_coord_offsetA  = DIM*jnrA;
1592
1593             /* load j atom coordinates */
1594             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1595                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1596                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1597
1598             /* Calculate displacement vector */
1599             dx00             = _mm_sub_pd(ix0,jx0);
1600             dy00             = _mm_sub_pd(iy0,jy0);
1601             dz00             = _mm_sub_pd(iz0,jz0);
1602             dx11             = _mm_sub_pd(ix1,jx1);
1603             dy11             = _mm_sub_pd(iy1,jy1);
1604             dz11             = _mm_sub_pd(iz1,jz1);
1605             dx12             = _mm_sub_pd(ix1,jx2);
1606             dy12             = _mm_sub_pd(iy1,jy2);
1607             dz12             = _mm_sub_pd(iz1,jz2);
1608             dx13             = _mm_sub_pd(ix1,jx3);
1609             dy13             = _mm_sub_pd(iy1,jy3);
1610             dz13             = _mm_sub_pd(iz1,jz3);
1611             dx21             = _mm_sub_pd(ix2,jx1);
1612             dy21             = _mm_sub_pd(iy2,jy1);
1613             dz21             = _mm_sub_pd(iz2,jz1);
1614             dx22             = _mm_sub_pd(ix2,jx2);
1615             dy22             = _mm_sub_pd(iy2,jy2);
1616             dz22             = _mm_sub_pd(iz2,jz2);
1617             dx23             = _mm_sub_pd(ix2,jx3);
1618             dy23             = _mm_sub_pd(iy2,jy3);
1619             dz23             = _mm_sub_pd(iz2,jz3);
1620             dx31             = _mm_sub_pd(ix3,jx1);
1621             dy31             = _mm_sub_pd(iy3,jy1);
1622             dz31             = _mm_sub_pd(iz3,jz1);
1623             dx32             = _mm_sub_pd(ix3,jx2);
1624             dy32             = _mm_sub_pd(iy3,jy2);
1625             dz32             = _mm_sub_pd(iz3,jz2);
1626             dx33             = _mm_sub_pd(ix3,jx3);
1627             dy33             = _mm_sub_pd(iy3,jy3);
1628             dz33             = _mm_sub_pd(iz3,jz3);
1629
1630             /* Calculate squared distance and things based on it */
1631             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1632             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1633             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1634             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1635             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1636             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1637             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1638             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1639             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1640             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1641
1642             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1643             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1644             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1645             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1646             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1647             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1648             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1649             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1650             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1651             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1652
1653             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1654             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1655             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1656             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1657             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1658             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1659             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1660             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1661             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1662
1663             fjx0             = _mm_setzero_pd();
1664             fjy0             = _mm_setzero_pd();
1665             fjz0             = _mm_setzero_pd();
1666             fjx1             = _mm_setzero_pd();
1667             fjy1             = _mm_setzero_pd();
1668             fjz1             = _mm_setzero_pd();
1669             fjx2             = _mm_setzero_pd();
1670             fjy2             = _mm_setzero_pd();
1671             fjz2             = _mm_setzero_pd();
1672             fjx3             = _mm_setzero_pd();
1673             fjy3             = _mm_setzero_pd();
1674             fjz3             = _mm_setzero_pd();
1675
1676             /**************************
1677              * CALCULATE INTERACTIONS *
1678              **************************/
1679
1680             r00              = _mm_mul_pd(rsq00,rinv00);
1681
1682             /* Calculate table index by multiplying r with table scale and truncate to integer */
1683             rt               = _mm_mul_pd(r00,vftabscale);
1684             vfitab           = _mm_cvttpd_epi32(rt);
1685             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1686             vfitab           = _mm_slli_epi32(vfitab,3);
1687
1688             /* CUBIC SPLINE TABLE DISPERSION */
1689             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1690             F                = _mm_setzero_pd();
1691             GMX_MM_TRANSPOSE2_PD(Y,F);
1692             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1693             H                = _mm_setzero_pd();
1694             GMX_MM_TRANSPOSE2_PD(G,H);
1695             Heps             = _mm_mul_pd(vfeps,H);
1696             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1697             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1698             fvdw6            = _mm_mul_pd(c6_00,FF);
1699
1700             /* CUBIC SPLINE TABLE REPULSION */
1701             vfitab           = _mm_add_epi32(vfitab,ifour);
1702             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1703             F                = _mm_setzero_pd();
1704             GMX_MM_TRANSPOSE2_PD(Y,F);
1705             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1706             H                = _mm_setzero_pd();
1707             GMX_MM_TRANSPOSE2_PD(G,H);
1708             Heps             = _mm_mul_pd(vfeps,H);
1709             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1710             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1711             fvdw12           = _mm_mul_pd(c12_00,FF);
1712             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1713
1714             fscal            = fvdw;
1715
1716             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1717
1718             /* Calculate temporary vectorial force */
1719             tx               = _mm_mul_pd(fscal,dx00);
1720             ty               = _mm_mul_pd(fscal,dy00);
1721             tz               = _mm_mul_pd(fscal,dz00);
1722
1723             /* Update vectorial force */
1724             fix0             = _mm_add_pd(fix0,tx);
1725             fiy0             = _mm_add_pd(fiy0,ty);
1726             fiz0             = _mm_add_pd(fiz0,tz);
1727
1728             fjx0             = _mm_add_pd(fjx0,tx);
1729             fjy0             = _mm_add_pd(fjy0,ty);
1730             fjz0             = _mm_add_pd(fjz0,tz);
1731
1732             /**************************
1733              * CALCULATE INTERACTIONS *
1734              **************************/
1735
1736             /* COULOMB ELECTROSTATICS */
1737             velec            = _mm_mul_pd(qq11,rinv11);
1738             felec            = _mm_mul_pd(velec,rinvsq11);
1739
1740             fscal            = felec;
1741
1742             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1743
1744             /* Calculate temporary vectorial force */
1745             tx               = _mm_mul_pd(fscal,dx11);
1746             ty               = _mm_mul_pd(fscal,dy11);
1747             tz               = _mm_mul_pd(fscal,dz11);
1748
1749             /* Update vectorial force */
1750             fix1             = _mm_add_pd(fix1,tx);
1751             fiy1             = _mm_add_pd(fiy1,ty);
1752             fiz1             = _mm_add_pd(fiz1,tz);
1753
1754             fjx1             = _mm_add_pd(fjx1,tx);
1755             fjy1             = _mm_add_pd(fjy1,ty);
1756             fjz1             = _mm_add_pd(fjz1,tz);
1757
1758             /**************************
1759              * CALCULATE INTERACTIONS *
1760              **************************/
1761
1762             /* COULOMB ELECTROSTATICS */
1763             velec            = _mm_mul_pd(qq12,rinv12);
1764             felec            = _mm_mul_pd(velec,rinvsq12);
1765
1766             fscal            = felec;
1767
1768             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1769
1770             /* Calculate temporary vectorial force */
1771             tx               = _mm_mul_pd(fscal,dx12);
1772             ty               = _mm_mul_pd(fscal,dy12);
1773             tz               = _mm_mul_pd(fscal,dz12);
1774
1775             /* Update vectorial force */
1776             fix1             = _mm_add_pd(fix1,tx);
1777             fiy1             = _mm_add_pd(fiy1,ty);
1778             fiz1             = _mm_add_pd(fiz1,tz);
1779
1780             fjx2             = _mm_add_pd(fjx2,tx);
1781             fjy2             = _mm_add_pd(fjy2,ty);
1782             fjz2             = _mm_add_pd(fjz2,tz);
1783
1784             /**************************
1785              * CALCULATE INTERACTIONS *
1786              **************************/
1787
1788             /* COULOMB ELECTROSTATICS */
1789             velec            = _mm_mul_pd(qq13,rinv13);
1790             felec            = _mm_mul_pd(velec,rinvsq13);
1791
1792             fscal            = felec;
1793
1794             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1795
1796             /* Calculate temporary vectorial force */
1797             tx               = _mm_mul_pd(fscal,dx13);
1798             ty               = _mm_mul_pd(fscal,dy13);
1799             tz               = _mm_mul_pd(fscal,dz13);
1800
1801             /* Update vectorial force */
1802             fix1             = _mm_add_pd(fix1,tx);
1803             fiy1             = _mm_add_pd(fiy1,ty);
1804             fiz1             = _mm_add_pd(fiz1,tz);
1805
1806             fjx3             = _mm_add_pd(fjx3,tx);
1807             fjy3             = _mm_add_pd(fjy3,ty);
1808             fjz3             = _mm_add_pd(fjz3,tz);
1809
1810             /**************************
1811              * CALCULATE INTERACTIONS *
1812              **************************/
1813
1814             /* COULOMB ELECTROSTATICS */
1815             velec            = _mm_mul_pd(qq21,rinv21);
1816             felec            = _mm_mul_pd(velec,rinvsq21);
1817
1818             fscal            = felec;
1819
1820             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1821
1822             /* Calculate temporary vectorial force */
1823             tx               = _mm_mul_pd(fscal,dx21);
1824             ty               = _mm_mul_pd(fscal,dy21);
1825             tz               = _mm_mul_pd(fscal,dz21);
1826
1827             /* Update vectorial force */
1828             fix2             = _mm_add_pd(fix2,tx);
1829             fiy2             = _mm_add_pd(fiy2,ty);
1830             fiz2             = _mm_add_pd(fiz2,tz);
1831
1832             fjx1             = _mm_add_pd(fjx1,tx);
1833             fjy1             = _mm_add_pd(fjy1,ty);
1834             fjz1             = _mm_add_pd(fjz1,tz);
1835
1836             /**************************
1837              * CALCULATE INTERACTIONS *
1838              **************************/
1839
1840             /* COULOMB ELECTROSTATICS */
1841             velec            = _mm_mul_pd(qq22,rinv22);
1842             felec            = _mm_mul_pd(velec,rinvsq22);
1843
1844             fscal            = felec;
1845
1846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1847
1848             /* Calculate temporary vectorial force */
1849             tx               = _mm_mul_pd(fscal,dx22);
1850             ty               = _mm_mul_pd(fscal,dy22);
1851             tz               = _mm_mul_pd(fscal,dz22);
1852
1853             /* Update vectorial force */
1854             fix2             = _mm_add_pd(fix2,tx);
1855             fiy2             = _mm_add_pd(fiy2,ty);
1856             fiz2             = _mm_add_pd(fiz2,tz);
1857
1858             fjx2             = _mm_add_pd(fjx2,tx);
1859             fjy2             = _mm_add_pd(fjy2,ty);
1860             fjz2             = _mm_add_pd(fjz2,tz);
1861
1862             /**************************
1863              * CALCULATE INTERACTIONS *
1864              **************************/
1865
1866             /* COULOMB ELECTROSTATICS */
1867             velec            = _mm_mul_pd(qq23,rinv23);
1868             felec            = _mm_mul_pd(velec,rinvsq23);
1869
1870             fscal            = felec;
1871
1872             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1873
1874             /* Calculate temporary vectorial force */
1875             tx               = _mm_mul_pd(fscal,dx23);
1876             ty               = _mm_mul_pd(fscal,dy23);
1877             tz               = _mm_mul_pd(fscal,dz23);
1878
1879             /* Update vectorial force */
1880             fix2             = _mm_add_pd(fix2,tx);
1881             fiy2             = _mm_add_pd(fiy2,ty);
1882             fiz2             = _mm_add_pd(fiz2,tz);
1883
1884             fjx3             = _mm_add_pd(fjx3,tx);
1885             fjy3             = _mm_add_pd(fjy3,ty);
1886             fjz3             = _mm_add_pd(fjz3,tz);
1887
1888             /**************************
1889              * CALCULATE INTERACTIONS *
1890              **************************/
1891
1892             /* COULOMB ELECTROSTATICS */
1893             velec            = _mm_mul_pd(qq31,rinv31);
1894             felec            = _mm_mul_pd(velec,rinvsq31);
1895
1896             fscal            = felec;
1897
1898             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1899
1900             /* Calculate temporary vectorial force */
1901             tx               = _mm_mul_pd(fscal,dx31);
1902             ty               = _mm_mul_pd(fscal,dy31);
1903             tz               = _mm_mul_pd(fscal,dz31);
1904
1905             /* Update vectorial force */
1906             fix3             = _mm_add_pd(fix3,tx);
1907             fiy3             = _mm_add_pd(fiy3,ty);
1908             fiz3             = _mm_add_pd(fiz3,tz);
1909
1910             fjx1             = _mm_add_pd(fjx1,tx);
1911             fjy1             = _mm_add_pd(fjy1,ty);
1912             fjz1             = _mm_add_pd(fjz1,tz);
1913
1914             /**************************
1915              * CALCULATE INTERACTIONS *
1916              **************************/
1917
1918             /* COULOMB ELECTROSTATICS */
1919             velec            = _mm_mul_pd(qq32,rinv32);
1920             felec            = _mm_mul_pd(velec,rinvsq32);
1921
1922             fscal            = felec;
1923
1924             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1925
1926             /* Calculate temporary vectorial force */
1927             tx               = _mm_mul_pd(fscal,dx32);
1928             ty               = _mm_mul_pd(fscal,dy32);
1929             tz               = _mm_mul_pd(fscal,dz32);
1930
1931             /* Update vectorial force */
1932             fix3             = _mm_add_pd(fix3,tx);
1933             fiy3             = _mm_add_pd(fiy3,ty);
1934             fiz3             = _mm_add_pd(fiz3,tz);
1935
1936             fjx2             = _mm_add_pd(fjx2,tx);
1937             fjy2             = _mm_add_pd(fjy2,ty);
1938             fjz2             = _mm_add_pd(fjz2,tz);
1939
1940             /**************************
1941              * CALCULATE INTERACTIONS *
1942              **************************/
1943
1944             /* COULOMB ELECTROSTATICS */
1945             velec            = _mm_mul_pd(qq33,rinv33);
1946             felec            = _mm_mul_pd(velec,rinvsq33);
1947
1948             fscal            = felec;
1949
1950             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1951
1952             /* Calculate temporary vectorial force */
1953             tx               = _mm_mul_pd(fscal,dx33);
1954             ty               = _mm_mul_pd(fscal,dy33);
1955             tz               = _mm_mul_pd(fscal,dz33);
1956
1957             /* Update vectorial force */
1958             fix3             = _mm_add_pd(fix3,tx);
1959             fiy3             = _mm_add_pd(fiy3,ty);
1960             fiz3             = _mm_add_pd(fiz3,tz);
1961
1962             fjx3             = _mm_add_pd(fjx3,tx);
1963             fjy3             = _mm_add_pd(fjy3,ty);
1964             fjz3             = _mm_add_pd(fjz3,tz);
1965
1966             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1967
1968             /* Inner loop uses 294 flops */
1969         }
1970
1971         /* End of innermost loop */
1972
1973         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1974                                               f+i_coord_offset,fshift+i_shift_offset);
1975
1976         /* Increment number of inner iterations */
1977         inneriter                  += j_index_end - j_index_start;
1978
1979         /* Outer loop uses 24 flops */
1980     }
1981
1982     /* Increment number of outer iterations */
1983     outeriter        += nri;
1984
1985     /* Update outer/inner flops */
1986
1987     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*294);
1988 }