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