src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRF_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 "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "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_ElecRF_VdwCSTab_GeomW4W4_VF_sse2_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_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     krf              = _mm_set1_pd(fr->ic->k_rf);
 136     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 137     crf              = _mm_set1_pd(fr->ic->c_rf);
 138     nvdwtype         = fr->ntype;
 139     vdwparam         = fr->nbfp;
 140     vdwtype          = mdatoms->typeA;
 141
 142     vftab            = kernel_data->table_vdw->data;
 143     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 144
 145     /* Setup water-specific parameters */
 146     inr              = nlist->iinr[0];
 147     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 148     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 149     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 150     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 151
 152     jq1              = _mm_set1_pd(charge[inr+1]);
 153     jq2              = _mm_set1_pd(charge[inr+2]);
 154     jq3              = _mm_set1_pd(charge[inr+3]);
 155     vdwjidx0A        = 2*vdwtype[inr+0];
 156     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 157     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 158     qq11             = _mm_mul_pd(iq1,jq1);
 159     qq12             = _mm_mul_pd(iq1,jq2);
 160     qq13             = _mm_mul_pd(iq1,jq3);
 161     qq21             = _mm_mul_pd(iq2,jq1);
 162     qq22             = _mm_mul_pd(iq2,jq2);
 163     qq23             = _mm_mul_pd(iq2,jq3);
 164     qq31             = _mm_mul_pd(iq3,jq1);
 165     qq32             = _mm_mul_pd(iq3,jq2);
 166     qq33             = _mm_mul_pd(iq3,jq3);
 167
 168     /* Avoid stupid compiler warnings */
 169     jnrA = jnrB = 0;
 170     j_coord_offsetA = 0;
 171     j_coord_offsetB = 0;
 172
 173     outeriter        = 0;
 174     inneriter        = 0;
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 193
 194         fix0             = _mm_setzero_pd();
 195         fiy0             = _mm_setzero_pd();
 196         fiz0             = _mm_setzero_pd();
 197         fix1             = _mm_setzero_pd();
 198         fiy1             = _mm_setzero_pd();
 199         fiz1             = _mm_setzero_pd();
 200         fix2             = _mm_setzero_pd();
 201         fiy2             = _mm_setzero_pd();
 202         fiz2             = _mm_setzero_pd();
 203         fix3             = _mm_setzero_pd();
 204         fiy3             = _mm_setzero_pd();
 205         fiz3             = _mm_setzero_pd();
 206
 207         /* Reset potential sums */
 208         velecsum         = _mm_setzero_pd();
 209         vvdwsum          = _mm_setzero_pd();
 210
 211         /* Start inner kernel loop */
 212         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 213         {
 214
 215             /* Get j neighbor index, and coordinate index */
 216             jnrA             = jjnr[jidx];
 217             jnrB             = jjnr[jidx+1];
 218             j_coord_offsetA  = DIM*jnrA;
 219             j_coord_offsetB  = DIM*jnrB;
 220
 221             /* load j atom coordinates */
 222             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 223                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 224                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 225
 226             /* Calculate displacement vector */
 227             dx00             = _mm_sub_pd(ix0,jx0);
 228             dy00             = _mm_sub_pd(iy0,jy0);
 229             dz00             = _mm_sub_pd(iz0,jz0);
 230             dx11             = _mm_sub_pd(ix1,jx1);
 231             dy11             = _mm_sub_pd(iy1,jy1);
 232             dz11             = _mm_sub_pd(iz1,jz1);
 233             dx12             = _mm_sub_pd(ix1,jx2);
 234             dy12             = _mm_sub_pd(iy1,jy2);
 235             dz12             = _mm_sub_pd(iz1,jz2);
 236             dx13             = _mm_sub_pd(ix1,jx3);
 237             dy13             = _mm_sub_pd(iy1,jy3);
 238             dz13             = _mm_sub_pd(iz1,jz3);
 239             dx21             = _mm_sub_pd(ix2,jx1);
 240             dy21             = _mm_sub_pd(iy2,jy1);
 241             dz21             = _mm_sub_pd(iz2,jz1);
 242             dx22             = _mm_sub_pd(ix2,jx2);
 243             dy22             = _mm_sub_pd(iy2,jy2);
 244             dz22             = _mm_sub_pd(iz2,jz2);
 245             dx23             = _mm_sub_pd(ix2,jx3);
 246             dy23             = _mm_sub_pd(iy2,jy3);
 247             dz23             = _mm_sub_pd(iz2,jz3);
 248             dx31             = _mm_sub_pd(ix3,jx1);
 249             dy31             = _mm_sub_pd(iy3,jy1);
 250             dz31             = _mm_sub_pd(iz3,jz1);
 251             dx32             = _mm_sub_pd(ix3,jx2);
 252             dy32             = _mm_sub_pd(iy3,jy2);
 253             dz32             = _mm_sub_pd(iz3,jz2);
 254             dx33             = _mm_sub_pd(ix3,jx3);
 255             dy33             = _mm_sub_pd(iy3,jy3);
 256             dz33             = _mm_sub_pd(iz3,jz3);
 257
 258             /* Calculate squared distance and things based on it */
 259             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 260             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 261             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 262             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 263             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 264             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 265             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 266             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 267             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 268             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 269
 270             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 271             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 272             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 273             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 274             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 275             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 276             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 277             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 278             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 279             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 280
 281             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 282             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 283             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 284             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 285             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 286             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 287             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 288             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 289             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 290
 291             fjx0             = _mm_setzero_pd();
 292             fjy0             = _mm_setzero_pd();
 293             fjz0             = _mm_setzero_pd();
 294             fjx1             = _mm_setzero_pd();
 295             fjy1             = _mm_setzero_pd();
 296             fjz1             = _mm_setzero_pd();
 297             fjx2             = _mm_setzero_pd();
 298             fjy2             = _mm_setzero_pd();
 299             fjz2             = _mm_setzero_pd();
 300             fjx3             = _mm_setzero_pd();
 301             fjy3             = _mm_setzero_pd();
 302             fjz3             = _mm_setzero_pd();
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             r00              = _mm_mul_pd(rsq00,rinv00);
 309
 310             /* Calculate table index by multiplying r with table scale and truncate to integer */
 311             rt               = _mm_mul_pd(r00,vftabscale);
 312             vfitab           = _mm_cvttpd_epi32(rt);
 313             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 314             vfitab           = _mm_slli_epi32(vfitab,3);
 315
 316             /* CUBIC SPLINE TABLE DISPERSION */
 317             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 318             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 319             GMX_MM_TRANSPOSE2_PD(Y,F);
 320             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 321             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 322             GMX_MM_TRANSPOSE2_PD(G,H);
 323             Heps             = _mm_mul_pd(vfeps,H);
 324             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 325             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 326             vvdw6            = _mm_mul_pd(c6_00,VV);
 327             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 328             fvdw6            = _mm_mul_pd(c6_00,FF);
 329
 330             /* CUBIC SPLINE TABLE REPULSION */
 331             vfitab           = _mm_add_epi32(vfitab,ifour);
 332             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 333             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 334             GMX_MM_TRANSPOSE2_PD(Y,F);
 335             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 336             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 337             GMX_MM_TRANSPOSE2_PD(G,H);
 338             Heps             = _mm_mul_pd(vfeps,H);
 339             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 340             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 341             vvdw12           = _mm_mul_pd(c12_00,VV);
 342             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 343             fvdw12           = _mm_mul_pd(c12_00,FF);
 344             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 345             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 349
 350             fscal            = fvdw;
 351
 352             /* Calculate temporary vectorial force */
 353             tx               = _mm_mul_pd(fscal,dx00);
 354             ty               = _mm_mul_pd(fscal,dy00);
 355             tz               = _mm_mul_pd(fscal,dz00);
 356
 357             /* Update vectorial force */
 358             fix0             = _mm_add_pd(fix0,tx);
 359             fiy0             = _mm_add_pd(fiy0,ty);
 360             fiz0             = _mm_add_pd(fiz0,tz);
 361
 362             fjx0             = _mm_add_pd(fjx0,tx);
 363             fjy0             = _mm_add_pd(fjy0,ty);
 364             fjz0             = _mm_add_pd(fjz0,tz);
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             /* REACTION-FIELD ELECTROSTATICS */
 371             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 372             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _mm_add_pd(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379             /* Calculate temporary vectorial force */
 380             tx               = _mm_mul_pd(fscal,dx11);
 381             ty               = _mm_mul_pd(fscal,dy11);
 382             tz               = _mm_mul_pd(fscal,dz11);
 383
 384             /* Update vectorial force */
 385             fix1             = _mm_add_pd(fix1,tx);
 386             fiy1             = _mm_add_pd(fiy1,ty);
 387             fiz1             = _mm_add_pd(fiz1,tz);
 388
 389             fjx1             = _mm_add_pd(fjx1,tx);
 390             fjy1             = _mm_add_pd(fjy1,ty);
 391             fjz1             = _mm_add_pd(fjz1,tz);
 392
 393             /**************************
 394              * CALCULATE INTERACTIONS *
 395              **************************/
 396
 397             /* REACTION-FIELD ELECTROSTATICS */
 398             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 399             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 400
 401             /* Update potential sum for this i atom from the interaction with this j atom. */
 402             velecsum         = _mm_add_pd(velecsum,velec);
 403
 404             fscal            = felec;
 405
 406             /* Calculate temporary vectorial force */
 407             tx               = _mm_mul_pd(fscal,dx12);
 408             ty               = _mm_mul_pd(fscal,dy12);
 409             tz               = _mm_mul_pd(fscal,dz12);
 410
 411             /* Update vectorial force */
 412             fix1             = _mm_add_pd(fix1,tx);
 413             fiy1             = _mm_add_pd(fiy1,ty);
 414             fiz1             = _mm_add_pd(fiz1,tz);
 415
 416             fjx2             = _mm_add_pd(fjx2,tx);
 417             fjy2             = _mm_add_pd(fjy2,ty);
 418             fjz2             = _mm_add_pd(fjz2,tz);
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             /* REACTION-FIELD ELECTROSTATICS */
 425             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 426             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 427
 428             /* Update potential sum for this i atom from the interaction with this j atom. */
 429             velecsum         = _mm_add_pd(velecsum,velec);
 430
 431             fscal            = felec;
 432
 433             /* Calculate temporary vectorial force */
 434             tx               = _mm_mul_pd(fscal,dx13);
 435             ty               = _mm_mul_pd(fscal,dy13);
 436             tz               = _mm_mul_pd(fscal,dz13);
 437
 438             /* Update vectorial force */
 439             fix1             = _mm_add_pd(fix1,tx);
 440             fiy1             = _mm_add_pd(fiy1,ty);
 441             fiz1             = _mm_add_pd(fiz1,tz);
 442
 443             fjx3             = _mm_add_pd(fjx3,tx);
 444             fjy3             = _mm_add_pd(fjy3,ty);
 445             fjz3             = _mm_add_pd(fjz3,tz);
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             /* REACTION-FIELD ELECTROSTATICS */
 452             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 453             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velecsum         = _mm_add_pd(velecsum,velec);
 457
 458             fscal            = felec;
 459
 460             /* Calculate temporary vectorial force */
 461             tx               = _mm_mul_pd(fscal,dx21);
 462             ty               = _mm_mul_pd(fscal,dy21);
 463             tz               = _mm_mul_pd(fscal,dz21);
 464
 465             /* Update vectorial force */
 466             fix2             = _mm_add_pd(fix2,tx);
 467             fiy2             = _mm_add_pd(fiy2,ty);
 468             fiz2             = _mm_add_pd(fiz2,tz);
 469
 470             fjx1             = _mm_add_pd(fjx1,tx);
 471             fjy1             = _mm_add_pd(fjy1,ty);
 472             fjz1             = _mm_add_pd(fjz1,tz);
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             /* REACTION-FIELD ELECTROSTATICS */
 479             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 480             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 481
 482             /* Update potential sum for this i atom from the interaction with this j atom. */
 483             velecsum         = _mm_add_pd(velecsum,velec);
 484
 485             fscal            = felec;
 486
 487             /* Calculate temporary vectorial force */
 488             tx               = _mm_mul_pd(fscal,dx22);
 489             ty               = _mm_mul_pd(fscal,dy22);
 490             tz               = _mm_mul_pd(fscal,dz22);
 491
 492             /* Update vectorial force */
 493             fix2             = _mm_add_pd(fix2,tx);
 494             fiy2             = _mm_add_pd(fiy2,ty);
 495             fiz2             = _mm_add_pd(fiz2,tz);
 496
 497             fjx2             = _mm_add_pd(fjx2,tx);
 498             fjy2             = _mm_add_pd(fjy2,ty);
 499             fjz2             = _mm_add_pd(fjz2,tz);
 500
 501             /**************************
 502              * CALCULATE INTERACTIONS *
 503              **************************/
 504
 505             /* REACTION-FIELD ELECTROSTATICS */
 506             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 507             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velecsum         = _mm_add_pd(velecsum,velec);
 511
 512             fscal            = felec;
 513
 514             /* Calculate temporary vectorial force */
 515             tx               = _mm_mul_pd(fscal,dx23);
 516             ty               = _mm_mul_pd(fscal,dy23);
 517             tz               = _mm_mul_pd(fscal,dz23);
 518
 519             /* Update vectorial force */
 520             fix2             = _mm_add_pd(fix2,tx);
 521             fiy2             = _mm_add_pd(fiy2,ty);
 522             fiz2             = _mm_add_pd(fiz2,tz);
 523
 524             fjx3             = _mm_add_pd(fjx3,tx);
 525             fjy3             = _mm_add_pd(fjy3,ty);
 526             fjz3             = _mm_add_pd(fjz3,tz);
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             /* REACTION-FIELD ELECTROSTATICS */
 533             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 534             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 535
 536             /* Update potential sum for this i atom from the interaction with this j atom. */
 537             velecsum         = _mm_add_pd(velecsum,velec);
 538
 539             fscal            = felec;
 540
 541             /* Calculate temporary vectorial force */
 542             tx               = _mm_mul_pd(fscal,dx31);
 543             ty               = _mm_mul_pd(fscal,dy31);
 544             tz               = _mm_mul_pd(fscal,dz31);
 545
 546             /* Update vectorial force */
 547             fix3             = _mm_add_pd(fix3,tx);
 548             fiy3             = _mm_add_pd(fiy3,ty);
 549             fiz3             = _mm_add_pd(fiz3,tz);
 550
 551             fjx1             = _mm_add_pd(fjx1,tx);
 552             fjy1             = _mm_add_pd(fjy1,ty);
 553             fjz1             = _mm_add_pd(fjz1,tz);
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             /* REACTION-FIELD ELECTROSTATICS */
 560             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 561             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 562
 563             /* Update potential sum for this i atom from the interaction with this j atom. */
 564             velecsum         = _mm_add_pd(velecsum,velec);
 565
 566             fscal            = felec;
 567
 568             /* Calculate temporary vectorial force */
 569             tx               = _mm_mul_pd(fscal,dx32);
 570             ty               = _mm_mul_pd(fscal,dy32);
 571             tz               = _mm_mul_pd(fscal,dz32);
 572
 573             /* Update vectorial force */
 574             fix3             = _mm_add_pd(fix3,tx);
 575             fiy3             = _mm_add_pd(fiy3,ty);
 576             fiz3             = _mm_add_pd(fiz3,tz);
 577
 578             fjx2             = _mm_add_pd(fjx2,tx);
 579             fjy2             = _mm_add_pd(fjy2,ty);
 580             fjz2             = _mm_add_pd(fjz2,tz);
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             /* REACTION-FIELD ELECTROSTATICS */
 587             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 588             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 589
 590             /* Update potential sum for this i atom from the interaction with this j atom. */
 591             velecsum         = _mm_add_pd(velecsum,velec);
 592
 593             fscal            = felec;
 594
 595             /* Calculate temporary vectorial force */
 596             tx               = _mm_mul_pd(fscal,dx33);
 597             ty               = _mm_mul_pd(fscal,dy33);
 598             tz               = _mm_mul_pd(fscal,dz33);
 599
 600             /* Update vectorial force */
 601             fix3             = _mm_add_pd(fix3,tx);
 602             fiy3             = _mm_add_pd(fiy3,ty);
 603             fiz3             = _mm_add_pd(fiz3,tz);
 604
 605             fjx3             = _mm_add_pd(fjx3,tx);
 606             fjy3             = _mm_add_pd(fjy3,ty);
 607             fjz3             = _mm_add_pd(fjz3,tz);
 608
 609             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);
 610
 611             /* Inner loop uses 347 flops */
 612         }
 613
 614         if(jidx<j_index_end)
 615         {
 616
 617             jnrA             = jjnr[jidx];
 618             j_coord_offsetA  = DIM*jnrA;
 619
 620             /* load j atom coordinates */
 621             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 622                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 623                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 624
 625             /* Calculate displacement vector */
 626             dx00             = _mm_sub_pd(ix0,jx0);
 627             dy00             = _mm_sub_pd(iy0,jy0);
 628             dz00             = _mm_sub_pd(iz0,jz0);
 629             dx11             = _mm_sub_pd(ix1,jx1);
 630             dy11             = _mm_sub_pd(iy1,jy1);
 631             dz11             = _mm_sub_pd(iz1,jz1);
 632             dx12             = _mm_sub_pd(ix1,jx2);
 633             dy12             = _mm_sub_pd(iy1,jy2);
 634             dz12             = _mm_sub_pd(iz1,jz2);
 635             dx13             = _mm_sub_pd(ix1,jx3);
 636             dy13             = _mm_sub_pd(iy1,jy3);
 637             dz13             = _mm_sub_pd(iz1,jz3);
 638             dx21             = _mm_sub_pd(ix2,jx1);
 639             dy21             = _mm_sub_pd(iy2,jy1);
 640             dz21             = _mm_sub_pd(iz2,jz1);
 641             dx22             = _mm_sub_pd(ix2,jx2);
 642             dy22             = _mm_sub_pd(iy2,jy2);
 643             dz22             = _mm_sub_pd(iz2,jz2);
 644             dx23             = _mm_sub_pd(ix2,jx3);
 645             dy23             = _mm_sub_pd(iy2,jy3);
 646             dz23             = _mm_sub_pd(iz2,jz3);
 647             dx31             = _mm_sub_pd(ix3,jx1);
 648             dy31             = _mm_sub_pd(iy3,jy1);
 649             dz31             = _mm_sub_pd(iz3,jz1);
 650             dx32             = _mm_sub_pd(ix3,jx2);
 651             dy32             = _mm_sub_pd(iy3,jy2);
 652             dz32             = _mm_sub_pd(iz3,jz2);
 653             dx33             = _mm_sub_pd(ix3,jx3);
 654             dy33             = _mm_sub_pd(iy3,jy3);
 655             dz33             = _mm_sub_pd(iz3,jz3);
 656
 657             /* Calculate squared distance and things based on it */
 658             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 659             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 660             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 661             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 662             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 663             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 664             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 665             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 666             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 667             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 668
 669             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 670             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 671             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 672             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 673             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 674             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 675             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 676             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 677             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 678             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 679
 680             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 681             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 682             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 683             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 684             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 685             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 686             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 687             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 688             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 689
 690             fjx0             = _mm_setzero_pd();
 691             fjy0             = _mm_setzero_pd();
 692             fjz0             = _mm_setzero_pd();
 693             fjx1             = _mm_setzero_pd();
 694             fjy1             = _mm_setzero_pd();
 695             fjz1             = _mm_setzero_pd();
 696             fjx2             = _mm_setzero_pd();
 697             fjy2             = _mm_setzero_pd();
 698             fjz2             = _mm_setzero_pd();
 699             fjx3             = _mm_setzero_pd();
 700             fjy3             = _mm_setzero_pd();
 701             fjz3             = _mm_setzero_pd();
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             r00              = _mm_mul_pd(rsq00,rinv00);
 708
 709             /* Calculate table index by multiplying r with table scale and truncate to integer */
 710             rt               = _mm_mul_pd(r00,vftabscale);
 711             vfitab           = _mm_cvttpd_epi32(rt);
 712             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 713             vfitab           = _mm_slli_epi32(vfitab,3);
 714
 715             /* CUBIC SPLINE TABLE DISPERSION */
 716             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 717             F                = _mm_setzero_pd();
 718             GMX_MM_TRANSPOSE2_PD(Y,F);
 719             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 720             H                = _mm_setzero_pd();
 721             GMX_MM_TRANSPOSE2_PD(G,H);
 722             Heps             = _mm_mul_pd(vfeps,H);
 723             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 724             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 725             vvdw6            = _mm_mul_pd(c6_00,VV);
 726             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 727             fvdw6            = _mm_mul_pd(c6_00,FF);
 728
 729             /* CUBIC SPLINE TABLE REPULSION */
 730             vfitab           = _mm_add_epi32(vfitab,ifour);
 731             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 732             F                = _mm_setzero_pd();
 733             GMX_MM_TRANSPOSE2_PD(Y,F);
 734             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 735             H                = _mm_setzero_pd();
 736             GMX_MM_TRANSPOSE2_PD(G,H);
 737             Heps             = _mm_mul_pd(vfeps,H);
 738             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 739             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 740             vvdw12           = _mm_mul_pd(c12_00,VV);
 741             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 742             fvdw12           = _mm_mul_pd(c12_00,FF);
 743             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 744             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 745
 746             /* Update potential sum for this i atom from the interaction with this j atom. */
 747             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 748             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 749
 750             fscal            = fvdw;
 751
 752             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 753
 754             /* Calculate temporary vectorial force */
 755             tx               = _mm_mul_pd(fscal,dx00);
 756             ty               = _mm_mul_pd(fscal,dy00);
 757             tz               = _mm_mul_pd(fscal,dz00);
 758
 759             /* Update vectorial force */
 760             fix0             = _mm_add_pd(fix0,tx);
 761             fiy0             = _mm_add_pd(fiy0,ty);
 762             fiz0             = _mm_add_pd(fiz0,tz);
 763
 764             fjx0             = _mm_add_pd(fjx0,tx);
 765             fjy0             = _mm_add_pd(fjy0,ty);
 766             fjz0             = _mm_add_pd(fjz0,tz);
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             /* REACTION-FIELD ELECTROSTATICS */
 773             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 774             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 775
 776             /* Update potential sum for this i atom from the interaction with this j atom. */
 777             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 778             velecsum         = _mm_add_pd(velecsum,velec);
 779
 780             fscal            = felec;
 781
 782             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 783
 784             /* Calculate temporary vectorial force */
 785             tx               = _mm_mul_pd(fscal,dx11);
 786             ty               = _mm_mul_pd(fscal,dy11);
 787             tz               = _mm_mul_pd(fscal,dz11);
 788
 789             /* Update vectorial force */
 790             fix1             = _mm_add_pd(fix1,tx);
 791             fiy1             = _mm_add_pd(fiy1,ty);
 792             fiz1             = _mm_add_pd(fiz1,tz);
 793
 794             fjx1             = _mm_add_pd(fjx1,tx);
 795             fjy1             = _mm_add_pd(fjy1,ty);
 796             fjz1             = _mm_add_pd(fjz1,tz);
 797
 798             /**************************
 799              * CALCULATE INTERACTIONS *
 800              **************************/
 801
 802             /* REACTION-FIELD ELECTROSTATICS */
 803             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 804             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 805
 806             /* Update potential sum for this i atom from the interaction with this j atom. */
 807             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 808             velecsum         = _mm_add_pd(velecsum,velec);
 809
 810             fscal            = felec;
 811
 812             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 813
 814             /* Calculate temporary vectorial force */
 815             tx               = _mm_mul_pd(fscal,dx12);
 816             ty               = _mm_mul_pd(fscal,dy12);
 817             tz               = _mm_mul_pd(fscal,dz12);
 818
 819             /* Update vectorial force */
 820             fix1             = _mm_add_pd(fix1,tx);
 821             fiy1             = _mm_add_pd(fiy1,ty);
 822             fiz1             = _mm_add_pd(fiz1,tz);
 823
 824             fjx2             = _mm_add_pd(fjx2,tx);
 825             fjy2             = _mm_add_pd(fjy2,ty);
 826             fjz2             = _mm_add_pd(fjz2,tz);
 827
 828             /**************************
 829              * CALCULATE INTERACTIONS *
 830              **************************/
 831
 832             /* REACTION-FIELD ELECTROSTATICS */
 833             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 834             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 835
 836             /* Update potential sum for this i atom from the interaction with this j atom. */
 837             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 838             velecsum         = _mm_add_pd(velecsum,velec);
 839
 840             fscal            = felec;
 841
 842             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 843
 844             /* Calculate temporary vectorial force */
 845             tx               = _mm_mul_pd(fscal,dx13);
 846             ty               = _mm_mul_pd(fscal,dy13);
 847             tz               = _mm_mul_pd(fscal,dz13);
 848
 849             /* Update vectorial force */
 850             fix1             = _mm_add_pd(fix1,tx);
 851             fiy1             = _mm_add_pd(fiy1,ty);
 852             fiz1             = _mm_add_pd(fiz1,tz);
 853
 854             fjx3             = _mm_add_pd(fjx3,tx);
 855             fjy3             = _mm_add_pd(fjy3,ty);
 856             fjz3             = _mm_add_pd(fjz3,tz);
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             /* REACTION-FIELD ELECTROSTATICS */
 863             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 864             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 865
 866             /* Update potential sum for this i atom from the interaction with this j atom. */
 867             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 868             velecsum         = _mm_add_pd(velecsum,velec);
 869
 870             fscal            = felec;
 871
 872             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 873
 874             /* Calculate temporary vectorial force */
 875             tx               = _mm_mul_pd(fscal,dx21);
 876             ty               = _mm_mul_pd(fscal,dy21);
 877             tz               = _mm_mul_pd(fscal,dz21);
 878
 879             /* Update vectorial force */
 880             fix2             = _mm_add_pd(fix2,tx);
 881             fiy2             = _mm_add_pd(fiy2,ty);
 882             fiz2             = _mm_add_pd(fiz2,tz);
 883
 884             fjx1             = _mm_add_pd(fjx1,tx);
 885             fjy1             = _mm_add_pd(fjy1,ty);
 886             fjz1             = _mm_add_pd(fjz1,tz);
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             /* REACTION-FIELD ELECTROSTATICS */
 893             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 894             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 895
 896             /* Update potential sum for this i atom from the interaction with this j atom. */
 897             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 898             velecsum         = _mm_add_pd(velecsum,velec);
 899
 900             fscal            = felec;
 901
 902             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 903
 904             /* Calculate temporary vectorial force */
 905             tx               = _mm_mul_pd(fscal,dx22);
 906             ty               = _mm_mul_pd(fscal,dy22);
 907             tz               = _mm_mul_pd(fscal,dz22);
 908
 909             /* Update vectorial force */
 910             fix2             = _mm_add_pd(fix2,tx);
 911             fiy2             = _mm_add_pd(fiy2,ty);
 912             fiz2             = _mm_add_pd(fiz2,tz);
 913
 914             fjx2             = _mm_add_pd(fjx2,tx);
 915             fjy2             = _mm_add_pd(fjy2,ty);
 916             fjz2             = _mm_add_pd(fjz2,tz);
 917
 918             /**************************
 919              * CALCULATE INTERACTIONS *
 920              **************************/
 921
 922             /* REACTION-FIELD ELECTROSTATICS */
 923             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 924             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 925
 926             /* Update potential sum for this i atom from the interaction with this j atom. */
 927             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 928             velecsum         = _mm_add_pd(velecsum,velec);
 929
 930             fscal            = felec;
 931
 932             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 933
 934             /* Calculate temporary vectorial force */
 935             tx               = _mm_mul_pd(fscal,dx23);
 936             ty               = _mm_mul_pd(fscal,dy23);
 937             tz               = _mm_mul_pd(fscal,dz23);
 938
 939             /* Update vectorial force */
 940             fix2             = _mm_add_pd(fix2,tx);
 941             fiy2             = _mm_add_pd(fiy2,ty);
 942             fiz2             = _mm_add_pd(fiz2,tz);
 943
 944             fjx3             = _mm_add_pd(fjx3,tx);
 945             fjy3             = _mm_add_pd(fjy3,ty);
 946             fjz3             = _mm_add_pd(fjz3,tz);
 947
 948             /**************************
 949              * CALCULATE INTERACTIONS *
 950              **************************/
 951
 952             /* REACTION-FIELD ELECTROSTATICS */
 953             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 954             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 955
 956             /* Update potential sum for this i atom from the interaction with this j atom. */
 957             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 958             velecsum         = _mm_add_pd(velecsum,velec);
 959
 960             fscal            = felec;
 961
 962             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 963
 964             /* Calculate temporary vectorial force */
 965             tx               = _mm_mul_pd(fscal,dx31);
 966             ty               = _mm_mul_pd(fscal,dy31);
 967             tz               = _mm_mul_pd(fscal,dz31);
 968
 969             /* Update vectorial force */
 970             fix3             = _mm_add_pd(fix3,tx);
 971             fiy3             = _mm_add_pd(fiy3,ty);
 972             fiz3             = _mm_add_pd(fiz3,tz);
 973
 974             fjx1             = _mm_add_pd(fjx1,tx);
 975             fjy1             = _mm_add_pd(fjy1,ty);
 976             fjz1             = _mm_add_pd(fjz1,tz);
 977
 978             /**************************
 979              * CALCULATE INTERACTIONS *
 980              **************************/
 981
 982             /* REACTION-FIELD ELECTROSTATICS */
 983             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 984             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 985
 986             /* Update potential sum for this i atom from the interaction with this j atom. */
 987             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 988             velecsum         = _mm_add_pd(velecsum,velec);
 989
 990             fscal            = felec;
 991
 992             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 993
 994             /* Calculate temporary vectorial force */
 995             tx               = _mm_mul_pd(fscal,dx32);
 996             ty               = _mm_mul_pd(fscal,dy32);
 997             tz               = _mm_mul_pd(fscal,dz32);
 998
 999             /* Update vectorial force */
1000             fix3             = _mm_add_pd(fix3,tx);
1001             fiy3             = _mm_add_pd(fiy3,ty);
1002             fiz3             = _mm_add_pd(fiz3,tz);
1003
1004             fjx2             = _mm_add_pd(fjx2,tx);
1005             fjy2             = _mm_add_pd(fjy2,ty);
1006             fjz2             = _mm_add_pd(fjz2,tz);
1007
1008             /**************************
1009              * CALCULATE INTERACTIONS *
1010              **************************/
1011
1012             /* REACTION-FIELD ELECTROSTATICS */
1013             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1014             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1015
1016             /* Update potential sum for this i atom from the interaction with this j atom. */
1017             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1018             velecsum         = _mm_add_pd(velecsum,velec);
1019
1020             fscal            = felec;
1021
1022             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1023
1024             /* Calculate temporary vectorial force */
1025             tx               = _mm_mul_pd(fscal,dx33);
1026             ty               = _mm_mul_pd(fscal,dy33);
1027             tz               = _mm_mul_pd(fscal,dz33);
1028
1029             /* Update vectorial force */
1030             fix3             = _mm_add_pd(fix3,tx);
1031             fiy3             = _mm_add_pd(fiy3,ty);
1032             fiz3             = _mm_add_pd(fiz3,tz);
1033
1034             fjx3             = _mm_add_pd(fjx3,tx);
1035             fjy3             = _mm_add_pd(fjy3,ty);
1036             fjz3             = _mm_add_pd(fjz3,tz);
1037
1038             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1039
1040             /* Inner loop uses 347 flops */
1041         }
1042
1043         /* End of innermost loop */
1044
1045         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1046                                               f+i_coord_offset,fshift+i_shift_offset);
1047
1048         ggid                        = gid[iidx];
1049         /* Update potential energies */
1050         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1051         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1052
1053         /* Increment number of inner iterations */
1054         inneriter                  += j_index_end - j_index_start;
1055
1056         /* Outer loop uses 26 flops */
1057     }
1058
1059     /* Increment number of outer iterations */
1060     outeriter        += nri;
1061
1062     /* Update outer/inner flops */
1063
1064     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*347);
1065 }
1066 /*
1067  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse2_double
1068  * Electrostatics interaction: ReactionField
1069  * VdW interaction:            CubicSplineTable
1070  * Geometry:                   Water4-Water4
1071  * Calculate force/pot:        Force
1072  */
1073 void
1074 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse2_double
1075                     (t_nblist                    * gmx_restrict       nlist,
1076                      rvec                        * gmx_restrict          xx,
1077                      rvec                        * gmx_restrict          ff,
1078                      t_forcerec                  * gmx_restrict          fr,
1079                      t_mdatoms                   * gmx_restrict     mdatoms,
1080                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1081                      t_nrnb                      * gmx_restrict        nrnb)
1082 {
1083     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1084      * just 0 for non-waters.
1085      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1086      * jnr indices corresponding to data put in the four positions in the SIMD register.
1087      */
1088     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1089     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1090     int              jnrA,jnrB;
1091     int              j_coord_offsetA,j_coord_offsetB;
1092     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1093     real             rcutoff_scalar;
1094     real             *shiftvec,*fshift,*x,*f;
1095     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1096     int              vdwioffset0;
1097     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1098     int              vdwioffset1;
1099     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1100     int              vdwioffset2;
1101     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1102     int              vdwioffset3;
1103     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1104     int              vdwjidx0A,vdwjidx0B;
1105     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1106     int              vdwjidx1A,vdwjidx1B;
1107     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1108     int              vdwjidx2A,vdwjidx2B;
1109     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1110     int              vdwjidx3A,vdwjidx3B;
1111     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1112     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1113     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1114     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1115     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1116     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1117     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1118     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1119     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1120     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1121     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1122     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1123     real             *charge;
1124     int              nvdwtype;
1125     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1126     int              *vdwtype;
1127     real             *vdwparam;
1128     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1129     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1130     __m128i          vfitab;
1131     __m128i          ifour       = _mm_set1_epi32(4);
1132     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1133     real             *vftab;
1134     __m128d          dummy_mask,cutoff_mask;
1135     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1136     __m128d          one     = _mm_set1_pd(1.0);
1137     __m128d          two     = _mm_set1_pd(2.0);
1138     x                = xx[0];
1139     f                = ff[0];
1140
1141     nri              = nlist->nri;
1142     iinr             = nlist->iinr;
1143     jindex           = nlist->jindex;
1144     jjnr             = nlist->jjnr;
1145     shiftidx         = nlist->shift;
1146     gid              = nlist->gid;
1147     shiftvec         = fr->shift_vec[0];
1148     fshift           = fr->fshift[0];
1149     facel            = _mm_set1_pd(fr->epsfac);
1150     charge           = mdatoms->chargeA;
1151     krf              = _mm_set1_pd(fr->ic->k_rf);
1152     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1153     crf              = _mm_set1_pd(fr->ic->c_rf);
1154     nvdwtype         = fr->ntype;
1155     vdwparam         = fr->nbfp;
1156     vdwtype          = mdatoms->typeA;
1157
1158     vftab            = kernel_data->table_vdw->data;
1159     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1160
1161     /* Setup water-specific parameters */
1162     inr              = nlist->iinr[0];
1163     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1164     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1165     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1166     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1167
1168     jq1              = _mm_set1_pd(charge[inr+1]);
1169     jq2              = _mm_set1_pd(charge[inr+2]);
1170     jq3              = _mm_set1_pd(charge[inr+3]);
1171     vdwjidx0A        = 2*vdwtype[inr+0];
1172     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1173     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1174     qq11             = _mm_mul_pd(iq1,jq1);
1175     qq12             = _mm_mul_pd(iq1,jq2);
1176     qq13             = _mm_mul_pd(iq1,jq3);
1177     qq21             = _mm_mul_pd(iq2,jq1);
1178     qq22             = _mm_mul_pd(iq2,jq2);
1179     qq23             = _mm_mul_pd(iq2,jq3);
1180     qq31             = _mm_mul_pd(iq3,jq1);
1181     qq32             = _mm_mul_pd(iq3,jq2);
1182     qq33             = _mm_mul_pd(iq3,jq3);
1183
1184     /* Avoid stupid compiler warnings */
1185     jnrA = jnrB = 0;
1186     j_coord_offsetA = 0;
1187     j_coord_offsetB = 0;
1188
1189     outeriter        = 0;
1190     inneriter        = 0;
1191
1192     /* Start outer loop over neighborlists */
1193     for(iidx=0; iidx<nri; iidx++)
1194     {
1195         /* Load shift vector for this list */
1196         i_shift_offset   = DIM*shiftidx[iidx];
1197
1198         /* Load limits for loop over neighbors */
1199         j_index_start    = jindex[iidx];
1200         j_index_end      = jindex[iidx+1];
1201
1202         /* Get outer coordinate index */
1203         inr              = iinr[iidx];
1204         i_coord_offset   = DIM*inr;
1205
1206         /* Load i particle coords and add shift vector */
1207         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1208                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1209
1210         fix0             = _mm_setzero_pd();
1211         fiy0             = _mm_setzero_pd();
1212         fiz0             = _mm_setzero_pd();
1213         fix1             = _mm_setzero_pd();
1214         fiy1             = _mm_setzero_pd();
1215         fiz1             = _mm_setzero_pd();
1216         fix2             = _mm_setzero_pd();
1217         fiy2             = _mm_setzero_pd();
1218         fiz2             = _mm_setzero_pd();
1219         fix3             = _mm_setzero_pd();
1220         fiy3             = _mm_setzero_pd();
1221         fiz3             = _mm_setzero_pd();
1222
1223         /* Start inner kernel loop */
1224         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1225         {
1226
1227             /* Get j neighbor index, and coordinate index */
1228             jnrA             = jjnr[jidx];
1229             jnrB             = jjnr[jidx+1];
1230             j_coord_offsetA  = DIM*jnrA;
1231             j_coord_offsetB  = DIM*jnrB;
1232
1233             /* load j atom coordinates */
1234             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1235                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1236                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1237
1238             /* Calculate displacement vector */
1239             dx00             = _mm_sub_pd(ix0,jx0);
1240             dy00             = _mm_sub_pd(iy0,jy0);
1241             dz00             = _mm_sub_pd(iz0,jz0);
1242             dx11             = _mm_sub_pd(ix1,jx1);
1243             dy11             = _mm_sub_pd(iy1,jy1);
1244             dz11             = _mm_sub_pd(iz1,jz1);
1245             dx12             = _mm_sub_pd(ix1,jx2);
1246             dy12             = _mm_sub_pd(iy1,jy2);
1247             dz12             = _mm_sub_pd(iz1,jz2);
1248             dx13             = _mm_sub_pd(ix1,jx3);
1249             dy13             = _mm_sub_pd(iy1,jy3);
1250             dz13             = _mm_sub_pd(iz1,jz3);
1251             dx21             = _mm_sub_pd(ix2,jx1);
1252             dy21             = _mm_sub_pd(iy2,jy1);
1253             dz21             = _mm_sub_pd(iz2,jz1);
1254             dx22             = _mm_sub_pd(ix2,jx2);
1255             dy22             = _mm_sub_pd(iy2,jy2);
1256             dz22             = _mm_sub_pd(iz2,jz2);
1257             dx23             = _mm_sub_pd(ix2,jx3);
1258             dy23             = _mm_sub_pd(iy2,jy3);
1259             dz23             = _mm_sub_pd(iz2,jz3);
1260             dx31             = _mm_sub_pd(ix3,jx1);
1261             dy31             = _mm_sub_pd(iy3,jy1);
1262             dz31             = _mm_sub_pd(iz3,jz1);
1263             dx32             = _mm_sub_pd(ix3,jx2);
1264             dy32             = _mm_sub_pd(iy3,jy2);
1265             dz32             = _mm_sub_pd(iz3,jz2);
1266             dx33             = _mm_sub_pd(ix3,jx3);
1267             dy33             = _mm_sub_pd(iy3,jy3);
1268             dz33             = _mm_sub_pd(iz3,jz3);
1269
1270             /* Calculate squared distance and things based on it */
1271             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1272             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1273             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1274             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1275             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1276             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1277             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1278             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1279             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1280             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1281
1282             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1283             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1284             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1285             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1286             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1287             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1288             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1289             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1290             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1291             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1292
1293             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1294             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1295             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1296             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1297             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1298             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1299             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1300             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1301             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1302
1303             fjx0             = _mm_setzero_pd();
1304             fjy0             = _mm_setzero_pd();
1305             fjz0             = _mm_setzero_pd();
1306             fjx1             = _mm_setzero_pd();
1307             fjy1             = _mm_setzero_pd();
1308             fjz1             = _mm_setzero_pd();
1309             fjx2             = _mm_setzero_pd();
1310             fjy2             = _mm_setzero_pd();
1311             fjz2             = _mm_setzero_pd();
1312             fjx3             = _mm_setzero_pd();
1313             fjy3             = _mm_setzero_pd();
1314             fjz3             = _mm_setzero_pd();
1315
1316             /**************************
1317              * CALCULATE INTERACTIONS *
1318              **************************/
1319
1320             r00              = _mm_mul_pd(rsq00,rinv00);
1321
1322             /* Calculate table index by multiplying r with table scale and truncate to integer */
1323             rt               = _mm_mul_pd(r00,vftabscale);
1324             vfitab           = _mm_cvttpd_epi32(rt);
1325             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1326             vfitab           = _mm_slli_epi32(vfitab,3);
1327
1328             /* CUBIC SPLINE TABLE DISPERSION */
1329             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1330             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1331             GMX_MM_TRANSPOSE2_PD(Y,F);
1332             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1333             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1334             GMX_MM_TRANSPOSE2_PD(G,H);
1335             Heps             = _mm_mul_pd(vfeps,H);
1336             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1337             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1338             fvdw6            = _mm_mul_pd(c6_00,FF);
1339
1340             /* CUBIC SPLINE TABLE REPULSION */
1341             vfitab           = _mm_add_epi32(vfitab,ifour);
1342             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1343             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1344             GMX_MM_TRANSPOSE2_PD(Y,F);
1345             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1346             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1347             GMX_MM_TRANSPOSE2_PD(G,H);
1348             Heps             = _mm_mul_pd(vfeps,H);
1349             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1350             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1351             fvdw12           = _mm_mul_pd(c12_00,FF);
1352             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1353
1354             fscal            = fvdw;
1355
1356             /* Calculate temporary vectorial force */
1357             tx               = _mm_mul_pd(fscal,dx00);
1358             ty               = _mm_mul_pd(fscal,dy00);
1359             tz               = _mm_mul_pd(fscal,dz00);
1360
1361             /* Update vectorial force */
1362             fix0             = _mm_add_pd(fix0,tx);
1363             fiy0             = _mm_add_pd(fiy0,ty);
1364             fiz0             = _mm_add_pd(fiz0,tz);
1365
1366             fjx0             = _mm_add_pd(fjx0,tx);
1367             fjy0             = _mm_add_pd(fjy0,ty);
1368             fjz0             = _mm_add_pd(fjz0,tz);
1369
1370             /**************************
1371              * CALCULATE INTERACTIONS *
1372              **************************/
1373
1374             /* REACTION-FIELD ELECTROSTATICS */
1375             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1376
1377             fscal            = felec;
1378
1379             /* Calculate temporary vectorial force */
1380             tx               = _mm_mul_pd(fscal,dx11);
1381             ty               = _mm_mul_pd(fscal,dy11);
1382             tz               = _mm_mul_pd(fscal,dz11);
1383
1384             /* Update vectorial force */
1385             fix1             = _mm_add_pd(fix1,tx);
1386             fiy1             = _mm_add_pd(fiy1,ty);
1387             fiz1             = _mm_add_pd(fiz1,tz);
1388
1389             fjx1             = _mm_add_pd(fjx1,tx);
1390             fjy1             = _mm_add_pd(fjy1,ty);
1391             fjz1             = _mm_add_pd(fjz1,tz);
1392
1393             /**************************
1394              * CALCULATE INTERACTIONS *
1395              **************************/
1396
1397             /* REACTION-FIELD ELECTROSTATICS */
1398             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1399
1400             fscal            = felec;
1401
1402             /* Calculate temporary vectorial force */
1403             tx               = _mm_mul_pd(fscal,dx12);
1404             ty               = _mm_mul_pd(fscal,dy12);
1405             tz               = _mm_mul_pd(fscal,dz12);
1406
1407             /* Update vectorial force */
1408             fix1             = _mm_add_pd(fix1,tx);
1409             fiy1             = _mm_add_pd(fiy1,ty);
1410             fiz1             = _mm_add_pd(fiz1,tz);
1411
1412             fjx2             = _mm_add_pd(fjx2,tx);
1413             fjy2             = _mm_add_pd(fjy2,ty);
1414             fjz2             = _mm_add_pd(fjz2,tz);
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             /* REACTION-FIELD ELECTROSTATICS */
1421             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1422
1423             fscal            = felec;
1424
1425             /* Calculate temporary vectorial force */
1426             tx               = _mm_mul_pd(fscal,dx13);
1427             ty               = _mm_mul_pd(fscal,dy13);
1428             tz               = _mm_mul_pd(fscal,dz13);
1429
1430             /* Update vectorial force */
1431             fix1             = _mm_add_pd(fix1,tx);
1432             fiy1             = _mm_add_pd(fiy1,ty);
1433             fiz1             = _mm_add_pd(fiz1,tz);
1434
1435             fjx3             = _mm_add_pd(fjx3,tx);
1436             fjy3             = _mm_add_pd(fjy3,ty);
1437             fjz3             = _mm_add_pd(fjz3,tz);
1438
1439             /**************************
1440              * CALCULATE INTERACTIONS *
1441              **************************/
1442
1443             /* REACTION-FIELD ELECTROSTATICS */
1444             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1445
1446             fscal            = felec;
1447
1448             /* Calculate temporary vectorial force */
1449             tx               = _mm_mul_pd(fscal,dx21);
1450             ty               = _mm_mul_pd(fscal,dy21);
1451             tz               = _mm_mul_pd(fscal,dz21);
1452
1453             /* Update vectorial force */
1454             fix2             = _mm_add_pd(fix2,tx);
1455             fiy2             = _mm_add_pd(fiy2,ty);
1456             fiz2             = _mm_add_pd(fiz2,tz);
1457
1458             fjx1             = _mm_add_pd(fjx1,tx);
1459             fjy1             = _mm_add_pd(fjy1,ty);
1460             fjz1             = _mm_add_pd(fjz1,tz);
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             /* REACTION-FIELD ELECTROSTATICS */
1467             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1468
1469             fscal            = felec;
1470
1471             /* Calculate temporary vectorial force */
1472             tx               = _mm_mul_pd(fscal,dx22);
1473             ty               = _mm_mul_pd(fscal,dy22);
1474             tz               = _mm_mul_pd(fscal,dz22);
1475
1476             /* Update vectorial force */
1477             fix2             = _mm_add_pd(fix2,tx);
1478             fiy2             = _mm_add_pd(fiy2,ty);
1479             fiz2             = _mm_add_pd(fiz2,tz);
1480
1481             fjx2             = _mm_add_pd(fjx2,tx);
1482             fjy2             = _mm_add_pd(fjy2,ty);
1483             fjz2             = _mm_add_pd(fjz2,tz);
1484
1485             /**************************
1486              * CALCULATE INTERACTIONS *
1487              **************************/
1488
1489             /* REACTION-FIELD ELECTROSTATICS */
1490             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
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             /* REACTION-FIELD ELECTROSTATICS */
1513             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1514
1515             fscal            = felec;
1516
1517             /* Calculate temporary vectorial force */
1518             tx               = _mm_mul_pd(fscal,dx31);
1519             ty               = _mm_mul_pd(fscal,dy31);
1520             tz               = _mm_mul_pd(fscal,dz31);
1521
1522             /* Update vectorial force */
1523             fix3             = _mm_add_pd(fix3,tx);
1524             fiy3             = _mm_add_pd(fiy3,ty);
1525             fiz3             = _mm_add_pd(fiz3,tz);
1526
1527             fjx1             = _mm_add_pd(fjx1,tx);
1528             fjy1             = _mm_add_pd(fjy1,ty);
1529             fjz1             = _mm_add_pd(fjz1,tz);
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             /* REACTION-FIELD ELECTROSTATICS */
1536             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1537
1538             fscal            = felec;
1539
1540             /* Calculate temporary vectorial force */
1541             tx               = _mm_mul_pd(fscal,dx32);
1542             ty               = _mm_mul_pd(fscal,dy32);
1543             tz               = _mm_mul_pd(fscal,dz32);
1544
1545             /* Update vectorial force */
1546             fix3             = _mm_add_pd(fix3,tx);
1547             fiy3             = _mm_add_pd(fiy3,ty);
1548             fiz3             = _mm_add_pd(fiz3,tz);
1549
1550             fjx2             = _mm_add_pd(fjx2,tx);
1551             fjy2             = _mm_add_pd(fjy2,ty);
1552             fjz2             = _mm_add_pd(fjz2,tz);
1553
1554             /**************************
1555              * CALCULATE INTERACTIONS *
1556              **************************/
1557
1558             /* REACTION-FIELD ELECTROSTATICS */
1559             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1560
1561             fscal            = felec;
1562
1563             /* Calculate temporary vectorial force */
1564             tx               = _mm_mul_pd(fscal,dx33);
1565             ty               = _mm_mul_pd(fscal,dy33);
1566             tz               = _mm_mul_pd(fscal,dz33);
1567
1568             /* Update vectorial force */
1569             fix3             = _mm_add_pd(fix3,tx);
1570             fiy3             = _mm_add_pd(fiy3,ty);
1571             fiz3             = _mm_add_pd(fiz3,tz);
1572
1573             fjx3             = _mm_add_pd(fjx3,tx);
1574             fjy3             = _mm_add_pd(fjy3,ty);
1575             fjz3             = _mm_add_pd(fjz3,tz);
1576
1577             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);
1578
1579             /* Inner loop uses 294 flops */
1580         }
1581
1582         if(jidx<j_index_end)
1583         {
1584
1585             jnrA             = jjnr[jidx];
1586             j_coord_offsetA  = DIM*jnrA;
1587
1588             /* load j atom coordinates */
1589             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1590                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1591                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1592
1593             /* Calculate displacement vector */
1594             dx00             = _mm_sub_pd(ix0,jx0);
1595             dy00             = _mm_sub_pd(iy0,jy0);
1596             dz00             = _mm_sub_pd(iz0,jz0);
1597             dx11             = _mm_sub_pd(ix1,jx1);
1598             dy11             = _mm_sub_pd(iy1,jy1);
1599             dz11             = _mm_sub_pd(iz1,jz1);
1600             dx12             = _mm_sub_pd(ix1,jx2);
1601             dy12             = _mm_sub_pd(iy1,jy2);
1602             dz12             = _mm_sub_pd(iz1,jz2);
1603             dx13             = _mm_sub_pd(ix1,jx3);
1604             dy13             = _mm_sub_pd(iy1,jy3);
1605             dz13             = _mm_sub_pd(iz1,jz3);
1606             dx21             = _mm_sub_pd(ix2,jx1);
1607             dy21             = _mm_sub_pd(iy2,jy1);
1608             dz21             = _mm_sub_pd(iz2,jz1);
1609             dx22             = _mm_sub_pd(ix2,jx2);
1610             dy22             = _mm_sub_pd(iy2,jy2);
1611             dz22             = _mm_sub_pd(iz2,jz2);
1612             dx23             = _mm_sub_pd(ix2,jx3);
1613             dy23             = _mm_sub_pd(iy2,jy3);
1614             dz23             = _mm_sub_pd(iz2,jz3);
1615             dx31             = _mm_sub_pd(ix3,jx1);
1616             dy31             = _mm_sub_pd(iy3,jy1);
1617             dz31             = _mm_sub_pd(iz3,jz1);
1618             dx32             = _mm_sub_pd(ix3,jx2);
1619             dy32             = _mm_sub_pd(iy3,jy2);
1620             dz32             = _mm_sub_pd(iz3,jz2);
1621             dx33             = _mm_sub_pd(ix3,jx3);
1622             dy33             = _mm_sub_pd(iy3,jy3);
1623             dz33             = _mm_sub_pd(iz3,jz3);
1624
1625             /* Calculate squared distance and things based on it */
1626             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1627             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1628             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1629             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1630             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1631             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1632             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1633             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1634             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1635             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1636
1637             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1638             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1639             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1640             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1641             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1642             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1643             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1644             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1645             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1646             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1647
1648             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1649             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1650             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1651             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1652             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1653             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1654             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1655             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1656             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1657
1658             fjx0             = _mm_setzero_pd();
1659             fjy0             = _mm_setzero_pd();
1660             fjz0             = _mm_setzero_pd();
1661             fjx1             = _mm_setzero_pd();
1662             fjy1             = _mm_setzero_pd();
1663             fjz1             = _mm_setzero_pd();
1664             fjx2             = _mm_setzero_pd();
1665             fjy2             = _mm_setzero_pd();
1666             fjz2             = _mm_setzero_pd();
1667             fjx3             = _mm_setzero_pd();
1668             fjy3             = _mm_setzero_pd();
1669             fjz3             = _mm_setzero_pd();
1670
1671             /**************************
1672              * CALCULATE INTERACTIONS *
1673              **************************/
1674
1675             r00              = _mm_mul_pd(rsq00,rinv00);
1676
1677             /* Calculate table index by multiplying r with table scale and truncate to integer */
1678             rt               = _mm_mul_pd(r00,vftabscale);
1679             vfitab           = _mm_cvttpd_epi32(rt);
1680             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1681             vfitab           = _mm_slli_epi32(vfitab,3);
1682
1683             /* CUBIC SPLINE TABLE DISPERSION */
1684             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1685             F                = _mm_setzero_pd();
1686             GMX_MM_TRANSPOSE2_PD(Y,F);
1687             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1688             H                = _mm_setzero_pd();
1689             GMX_MM_TRANSPOSE2_PD(G,H);
1690             Heps             = _mm_mul_pd(vfeps,H);
1691             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1692             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1693             fvdw6            = _mm_mul_pd(c6_00,FF);
1694
1695             /* CUBIC SPLINE TABLE REPULSION */
1696             vfitab           = _mm_add_epi32(vfitab,ifour);
1697             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1698             F                = _mm_setzero_pd();
1699             GMX_MM_TRANSPOSE2_PD(Y,F);
1700             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1701             H                = _mm_setzero_pd();
1702             GMX_MM_TRANSPOSE2_PD(G,H);
1703             Heps             = _mm_mul_pd(vfeps,H);
1704             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1705             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1706             fvdw12           = _mm_mul_pd(c12_00,FF);
1707             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1708
1709             fscal            = fvdw;
1710
1711             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1712
1713             /* Calculate temporary vectorial force */
1714             tx               = _mm_mul_pd(fscal,dx00);
1715             ty               = _mm_mul_pd(fscal,dy00);
1716             tz               = _mm_mul_pd(fscal,dz00);
1717
1718             /* Update vectorial force */
1719             fix0             = _mm_add_pd(fix0,tx);
1720             fiy0             = _mm_add_pd(fiy0,ty);
1721             fiz0             = _mm_add_pd(fiz0,tz);
1722
1723             fjx0             = _mm_add_pd(fjx0,tx);
1724             fjy0             = _mm_add_pd(fjy0,ty);
1725             fjz0             = _mm_add_pd(fjz0,tz);
1726
1727             /**************************
1728              * CALCULATE INTERACTIONS *
1729              **************************/
1730
1731             /* REACTION-FIELD ELECTROSTATICS */
1732             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1733
1734             fscal            = felec;
1735
1736             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1737
1738             /* Calculate temporary vectorial force */
1739             tx               = _mm_mul_pd(fscal,dx11);
1740             ty               = _mm_mul_pd(fscal,dy11);
1741             tz               = _mm_mul_pd(fscal,dz11);
1742
1743             /* Update vectorial force */
1744             fix1             = _mm_add_pd(fix1,tx);
1745             fiy1             = _mm_add_pd(fiy1,ty);
1746             fiz1             = _mm_add_pd(fiz1,tz);
1747
1748             fjx1             = _mm_add_pd(fjx1,tx);
1749             fjy1             = _mm_add_pd(fjy1,ty);
1750             fjz1             = _mm_add_pd(fjz1,tz);
1751
1752             /**************************
1753              * CALCULATE INTERACTIONS *
1754              **************************/
1755
1756             /* REACTION-FIELD ELECTROSTATICS */
1757             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1758
1759             fscal            = felec;
1760
1761             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1762
1763             /* Calculate temporary vectorial force */
1764             tx               = _mm_mul_pd(fscal,dx12);
1765             ty               = _mm_mul_pd(fscal,dy12);
1766             tz               = _mm_mul_pd(fscal,dz12);
1767
1768             /* Update vectorial force */
1769             fix1             = _mm_add_pd(fix1,tx);
1770             fiy1             = _mm_add_pd(fiy1,ty);
1771             fiz1             = _mm_add_pd(fiz1,tz);
1772
1773             fjx2             = _mm_add_pd(fjx2,tx);
1774             fjy2             = _mm_add_pd(fjy2,ty);
1775             fjz2             = _mm_add_pd(fjz2,tz);
1776
1777             /**************************
1778              * CALCULATE INTERACTIONS *
1779              **************************/
1780
1781             /* REACTION-FIELD ELECTROSTATICS */
1782             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1783
1784             fscal            = felec;
1785
1786             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1787
1788             /* Calculate temporary vectorial force */
1789             tx               = _mm_mul_pd(fscal,dx13);
1790             ty               = _mm_mul_pd(fscal,dy13);
1791             tz               = _mm_mul_pd(fscal,dz13);
1792
1793             /* Update vectorial force */
1794             fix1             = _mm_add_pd(fix1,tx);
1795             fiy1             = _mm_add_pd(fiy1,ty);
1796             fiz1             = _mm_add_pd(fiz1,tz);
1797
1798             fjx3             = _mm_add_pd(fjx3,tx);
1799             fjy3             = _mm_add_pd(fjy3,ty);
1800             fjz3             = _mm_add_pd(fjz3,tz);
1801
1802             /**************************
1803              * CALCULATE INTERACTIONS *
1804              **************************/
1805
1806             /* REACTION-FIELD ELECTROSTATICS */
1807             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1808
1809             fscal            = felec;
1810
1811             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1812
1813             /* Calculate temporary vectorial force */
1814             tx               = _mm_mul_pd(fscal,dx21);
1815             ty               = _mm_mul_pd(fscal,dy21);
1816             tz               = _mm_mul_pd(fscal,dz21);
1817
1818             /* Update vectorial force */
1819             fix2             = _mm_add_pd(fix2,tx);
1820             fiy2             = _mm_add_pd(fiy2,ty);
1821             fiz2             = _mm_add_pd(fiz2,tz);
1822
1823             fjx1             = _mm_add_pd(fjx1,tx);
1824             fjy1             = _mm_add_pd(fjy1,ty);
1825             fjz1             = _mm_add_pd(fjz1,tz);
1826
1827             /**************************
1828              * CALCULATE INTERACTIONS *
1829              **************************/
1830
1831             /* REACTION-FIELD ELECTROSTATICS */
1832             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1833
1834             fscal            = felec;
1835
1836             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1837
1838             /* Calculate temporary vectorial force */
1839             tx               = _mm_mul_pd(fscal,dx22);
1840             ty               = _mm_mul_pd(fscal,dy22);
1841             tz               = _mm_mul_pd(fscal,dz22);
1842
1843             /* Update vectorial force */
1844             fix2             = _mm_add_pd(fix2,tx);
1845             fiy2             = _mm_add_pd(fiy2,ty);
1846             fiz2             = _mm_add_pd(fiz2,tz);
1847
1848             fjx2             = _mm_add_pd(fjx2,tx);
1849             fjy2             = _mm_add_pd(fjy2,ty);
1850             fjz2             = _mm_add_pd(fjz2,tz);
1851
1852             /**************************
1853              * CALCULATE INTERACTIONS *
1854              **************************/
1855
1856             /* REACTION-FIELD ELECTROSTATICS */
1857             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1858
1859             fscal            = felec;
1860
1861             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1862
1863             /* Calculate temporary vectorial force */
1864             tx               = _mm_mul_pd(fscal,dx23);
1865             ty               = _mm_mul_pd(fscal,dy23);
1866             tz               = _mm_mul_pd(fscal,dz23);
1867
1868             /* Update vectorial force */
1869             fix2             = _mm_add_pd(fix2,tx);
1870             fiy2             = _mm_add_pd(fiy2,ty);
1871             fiz2             = _mm_add_pd(fiz2,tz);
1872
1873             fjx3             = _mm_add_pd(fjx3,tx);
1874             fjy3             = _mm_add_pd(fjy3,ty);
1875             fjz3             = _mm_add_pd(fjz3,tz);
1876
1877             /**************************
1878              * CALCULATE INTERACTIONS *
1879              **************************/
1880
1881             /* REACTION-FIELD ELECTROSTATICS */
1882             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1883
1884             fscal            = felec;
1885
1886             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1887
1888             /* Calculate temporary vectorial force */
1889             tx               = _mm_mul_pd(fscal,dx31);
1890             ty               = _mm_mul_pd(fscal,dy31);
1891             tz               = _mm_mul_pd(fscal,dz31);
1892
1893             /* Update vectorial force */
1894             fix3             = _mm_add_pd(fix3,tx);
1895             fiy3             = _mm_add_pd(fiy3,ty);
1896             fiz3             = _mm_add_pd(fiz3,tz);
1897
1898             fjx1             = _mm_add_pd(fjx1,tx);
1899             fjy1             = _mm_add_pd(fjy1,ty);
1900             fjz1             = _mm_add_pd(fjz1,tz);
1901
1902             /**************************
1903              * CALCULATE INTERACTIONS *
1904              **************************/
1905
1906             /* REACTION-FIELD ELECTROSTATICS */
1907             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1908
1909             fscal            = felec;
1910
1911             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1912
1913             /* Calculate temporary vectorial force */
1914             tx               = _mm_mul_pd(fscal,dx32);
1915             ty               = _mm_mul_pd(fscal,dy32);
1916             tz               = _mm_mul_pd(fscal,dz32);
1917
1918             /* Update vectorial force */
1919             fix3             = _mm_add_pd(fix3,tx);
1920             fiy3             = _mm_add_pd(fiy3,ty);
1921             fiz3             = _mm_add_pd(fiz3,tz);
1922
1923             fjx2             = _mm_add_pd(fjx2,tx);
1924             fjy2             = _mm_add_pd(fjy2,ty);
1925             fjz2             = _mm_add_pd(fjz2,tz);
1926
1927             /**************************
1928              * CALCULATE INTERACTIONS *
1929              **************************/
1930
1931             /* REACTION-FIELD ELECTROSTATICS */
1932             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1933
1934             fscal            = felec;
1935
1936             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1937
1938             /* Calculate temporary vectorial force */
1939             tx               = _mm_mul_pd(fscal,dx33);
1940             ty               = _mm_mul_pd(fscal,dy33);
1941             tz               = _mm_mul_pd(fscal,dz33);
1942
1943             /* Update vectorial force */
1944             fix3             = _mm_add_pd(fix3,tx);
1945             fiy3             = _mm_add_pd(fiy3,ty);
1946             fiz3             = _mm_add_pd(fiz3,tz);
1947
1948             fjx3             = _mm_add_pd(fjx3,tx);
1949             fjy3             = _mm_add_pd(fjy3,ty);
1950             fjz3             = _mm_add_pd(fjz3,tz);
1951
1952             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1953
1954             /* Inner loop uses 294 flops */
1955         }
1956
1957         /* End of innermost loop */
1958
1959         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1960                                               f+i_coord_offset,fshift+i_shift_offset);
1961
1962         /* Increment number of inner iterations */
1963         inneriter                  += j_index_end - j_index_start;
1964
1965         /* Outer loop uses 24 flops */
1966     }
1967
1968     /* Increment number of outer iterations */
1969     outeriter        += nri;
1970
1971     /* Update outer/inner flops */
1972
1973     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*294);
1974 }