src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRF_VdwCSTab_GeomW4W4_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_sse4_1_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_sse4_1_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwioffset3;
  73     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B;
  75     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B;
  77     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B;
  79     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B;
  81     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  99     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 100     __m128i          vfitab;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 103     real             *vftab;
 104     __m128d          dummy_mask,cutoff_mask;
 105     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 106     __m128d          one     = _mm_set1_pd(1.0);
 107     __m128d          two     = _mm_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     krf              = _mm_set1_pd(fr->ic->k_rf);
 122     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 123     crf              = _mm_set1_pd(fr->ic->c_rf);
 124     nvdwtype         = fr->ntype;
 125     vdwparam         = fr->nbfp;
 126     vdwtype          = mdatoms->typeA;
 127
 128     vftab            = kernel_data->table_vdw->data;
 129     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 130
 131     /* Setup water-specific parameters */
 132     inr              = nlist->iinr[0];
 133     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 134     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 135     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 136     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 137
 138     jq1              = _mm_set1_pd(charge[inr+1]);
 139     jq2              = _mm_set1_pd(charge[inr+2]);
 140     jq3              = _mm_set1_pd(charge[inr+3]);
 141     vdwjidx0A        = 2*vdwtype[inr+0];
 142     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 143     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 144     qq11             = _mm_mul_pd(iq1,jq1);
 145     qq12             = _mm_mul_pd(iq1,jq2);
 146     qq13             = _mm_mul_pd(iq1,jq3);
 147     qq21             = _mm_mul_pd(iq2,jq1);
 148     qq22             = _mm_mul_pd(iq2,jq2);
 149     qq23             = _mm_mul_pd(iq2,jq3);
 150     qq31             = _mm_mul_pd(iq3,jq1);
 151     qq32             = _mm_mul_pd(iq3,jq2);
 152     qq33             = _mm_mul_pd(iq3,jq3);
 153
 154     /* Avoid stupid compiler warnings */
 155     jnrA = jnrB = 0;
 156     j_coord_offsetA = 0;
 157     j_coord_offsetB = 0;
 158
 159     outeriter        = 0;
 160     inneriter        = 0;
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 178                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 179
 180         fix0             = _mm_setzero_pd();
 181         fiy0             = _mm_setzero_pd();
 182         fiz0             = _mm_setzero_pd();
 183         fix1             = _mm_setzero_pd();
 184         fiy1             = _mm_setzero_pd();
 185         fiz1             = _mm_setzero_pd();
 186         fix2             = _mm_setzero_pd();
 187         fiy2             = _mm_setzero_pd();
 188         fiz2             = _mm_setzero_pd();
 189         fix3             = _mm_setzero_pd();
 190         fiy3             = _mm_setzero_pd();
 191         fiz3             = _mm_setzero_pd();
 192
 193         /* Reset potential sums */
 194         velecsum         = _mm_setzero_pd();
 195         vvdwsum          = _mm_setzero_pd();
 196
 197         /* Start inner kernel loop */
 198         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 199         {
 200
 201             /* Get j neighbor index, and coordinate index */
 202             jnrA             = jjnr[jidx];
 203             jnrB             = jjnr[jidx+1];
 204             j_coord_offsetA  = DIM*jnrA;
 205             j_coord_offsetB  = DIM*jnrB;
 206
 207             /* load j atom coordinates */
 208             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 209                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 210                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 211
 212             /* Calculate displacement vector */
 213             dx00             = _mm_sub_pd(ix0,jx0);
 214             dy00             = _mm_sub_pd(iy0,jy0);
 215             dz00             = _mm_sub_pd(iz0,jz0);
 216             dx11             = _mm_sub_pd(ix1,jx1);
 217             dy11             = _mm_sub_pd(iy1,jy1);
 218             dz11             = _mm_sub_pd(iz1,jz1);
 219             dx12             = _mm_sub_pd(ix1,jx2);
 220             dy12             = _mm_sub_pd(iy1,jy2);
 221             dz12             = _mm_sub_pd(iz1,jz2);
 222             dx13             = _mm_sub_pd(ix1,jx3);
 223             dy13             = _mm_sub_pd(iy1,jy3);
 224             dz13             = _mm_sub_pd(iz1,jz3);
 225             dx21             = _mm_sub_pd(ix2,jx1);
 226             dy21             = _mm_sub_pd(iy2,jy1);
 227             dz21             = _mm_sub_pd(iz2,jz1);
 228             dx22             = _mm_sub_pd(ix2,jx2);
 229             dy22             = _mm_sub_pd(iy2,jy2);
 230             dz22             = _mm_sub_pd(iz2,jz2);
 231             dx23             = _mm_sub_pd(ix2,jx3);
 232             dy23             = _mm_sub_pd(iy2,jy3);
 233             dz23             = _mm_sub_pd(iz2,jz3);
 234             dx31             = _mm_sub_pd(ix3,jx1);
 235             dy31             = _mm_sub_pd(iy3,jy1);
 236             dz31             = _mm_sub_pd(iz3,jz1);
 237             dx32             = _mm_sub_pd(ix3,jx2);
 238             dy32             = _mm_sub_pd(iy3,jy2);
 239             dz32             = _mm_sub_pd(iz3,jz2);
 240             dx33             = _mm_sub_pd(ix3,jx3);
 241             dy33             = _mm_sub_pd(iy3,jy3);
 242             dz33             = _mm_sub_pd(iz3,jz3);
 243
 244             /* Calculate squared distance and things based on it */
 245             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 246             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 247             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 248             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 249             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 250             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 251             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 252             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 253             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 254             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 255
 256             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 257             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 258             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 259             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 260             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 261             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 262             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 263             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 264             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 265             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 266
 267             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 268             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 269             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 270             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 271             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 272             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 273             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 274             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 275             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 276
 277             fjx0             = _mm_setzero_pd();
 278             fjy0             = _mm_setzero_pd();
 279             fjz0             = _mm_setzero_pd();
 280             fjx1             = _mm_setzero_pd();
 281             fjy1             = _mm_setzero_pd();
 282             fjz1             = _mm_setzero_pd();
 283             fjx2             = _mm_setzero_pd();
 284             fjy2             = _mm_setzero_pd();
 285             fjz2             = _mm_setzero_pd();
 286             fjx3             = _mm_setzero_pd();
 287             fjy3             = _mm_setzero_pd();
 288             fjz3             = _mm_setzero_pd();
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             r00              = _mm_mul_pd(rsq00,rinv00);
 295
 296             /* Calculate table index by multiplying r with table scale and truncate to integer */
 297             rt               = _mm_mul_pd(r00,vftabscale);
 298             vfitab           = _mm_cvttpd_epi32(rt);
 299             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 300             vfitab           = _mm_slli_epi32(vfitab,3);
 301
 302             /* CUBIC SPLINE TABLE DISPERSION */
 303             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 304             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 305             GMX_MM_TRANSPOSE2_PD(Y,F);
 306             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 307             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 308             GMX_MM_TRANSPOSE2_PD(G,H);
 309             Heps             = _mm_mul_pd(vfeps,H);
 310             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 311             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 312             vvdw6            = _mm_mul_pd(c6_00,VV);
 313             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 314             fvdw6            = _mm_mul_pd(c6_00,FF);
 315
 316             /* CUBIC SPLINE TABLE REPULSION */
 317             vfitab           = _mm_add_epi32(vfitab,ifour);
 318             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 319             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 320             GMX_MM_TRANSPOSE2_PD(Y,F);
 321             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 322             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 323             GMX_MM_TRANSPOSE2_PD(G,H);
 324             Heps             = _mm_mul_pd(vfeps,H);
 325             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 326             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 327             vvdw12           = _mm_mul_pd(c12_00,VV);
 328             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 329             fvdw12           = _mm_mul_pd(c12_00,FF);
 330             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 331             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 335
 336             fscal            = fvdw;
 337
 338             /* Calculate temporary vectorial force */
 339             tx               = _mm_mul_pd(fscal,dx00);
 340             ty               = _mm_mul_pd(fscal,dy00);
 341             tz               = _mm_mul_pd(fscal,dz00);
 342
 343             /* Update vectorial force */
 344             fix0             = _mm_add_pd(fix0,tx);
 345             fiy0             = _mm_add_pd(fiy0,ty);
 346             fiz0             = _mm_add_pd(fiz0,tz);
 347
 348             fjx0             = _mm_add_pd(fjx0,tx);
 349             fjy0             = _mm_add_pd(fjy0,ty);
 350             fjz0             = _mm_add_pd(fjz0,tz);
 351
 352             /**************************
 353              * CALCULATE INTERACTIONS *
 354              **************************/
 355
 356             /* REACTION-FIELD ELECTROSTATICS */
 357             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 358             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 359
 360             /* Update potential sum for this i atom from the interaction with this j atom. */
 361             velecsum         = _mm_add_pd(velecsum,velec);
 362
 363             fscal            = felec;
 364
 365             /* Calculate temporary vectorial force */
 366             tx               = _mm_mul_pd(fscal,dx11);
 367             ty               = _mm_mul_pd(fscal,dy11);
 368             tz               = _mm_mul_pd(fscal,dz11);
 369
 370             /* Update vectorial force */
 371             fix1             = _mm_add_pd(fix1,tx);
 372             fiy1             = _mm_add_pd(fiy1,ty);
 373             fiz1             = _mm_add_pd(fiz1,tz);
 374
 375             fjx1             = _mm_add_pd(fjx1,tx);
 376             fjy1             = _mm_add_pd(fjy1,ty);
 377             fjz1             = _mm_add_pd(fjz1,tz);
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* REACTION-FIELD ELECTROSTATICS */
 384             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 385             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velecsum         = _mm_add_pd(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392             /* Calculate temporary vectorial force */
 393             tx               = _mm_mul_pd(fscal,dx12);
 394             ty               = _mm_mul_pd(fscal,dy12);
 395             tz               = _mm_mul_pd(fscal,dz12);
 396
 397             /* Update vectorial force */
 398             fix1             = _mm_add_pd(fix1,tx);
 399             fiy1             = _mm_add_pd(fiy1,ty);
 400             fiz1             = _mm_add_pd(fiz1,tz);
 401
 402             fjx2             = _mm_add_pd(fjx2,tx);
 403             fjy2             = _mm_add_pd(fjy2,ty);
 404             fjz2             = _mm_add_pd(fjz2,tz);
 405
 406             /**************************
 407              * CALCULATE INTERACTIONS *
 408              **************************/
 409
 410             /* REACTION-FIELD ELECTROSTATICS */
 411             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 412             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velecsum         = _mm_add_pd(velecsum,velec);
 416
 417             fscal            = felec;
 418
 419             /* Calculate temporary vectorial force */
 420             tx               = _mm_mul_pd(fscal,dx13);
 421             ty               = _mm_mul_pd(fscal,dy13);
 422             tz               = _mm_mul_pd(fscal,dz13);
 423
 424             /* Update vectorial force */
 425             fix1             = _mm_add_pd(fix1,tx);
 426             fiy1             = _mm_add_pd(fiy1,ty);
 427             fiz1             = _mm_add_pd(fiz1,tz);
 428
 429             fjx3             = _mm_add_pd(fjx3,tx);
 430             fjy3             = _mm_add_pd(fjy3,ty);
 431             fjz3             = _mm_add_pd(fjz3,tz);
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 439             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 440
 441             /* Update potential sum for this i atom from the interaction with this j atom. */
 442             velecsum         = _mm_add_pd(velecsum,velec);
 443
 444             fscal            = felec;
 445
 446             /* Calculate temporary vectorial force */
 447             tx               = _mm_mul_pd(fscal,dx21);
 448             ty               = _mm_mul_pd(fscal,dy21);
 449             tz               = _mm_mul_pd(fscal,dz21);
 450
 451             /* Update vectorial force */
 452             fix2             = _mm_add_pd(fix2,tx);
 453             fiy2             = _mm_add_pd(fiy2,ty);
 454             fiz2             = _mm_add_pd(fiz2,tz);
 455
 456             fjx1             = _mm_add_pd(fjx1,tx);
 457             fjy1             = _mm_add_pd(fjy1,ty);
 458             fjz1             = _mm_add_pd(fjz1,tz);
 459
 460             /**************************
 461              * CALCULATE INTERACTIONS *
 462              **************************/
 463
 464             /* REACTION-FIELD ELECTROSTATICS */
 465             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 466             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 467
 468             /* Update potential sum for this i atom from the interaction with this j atom. */
 469             velecsum         = _mm_add_pd(velecsum,velec);
 470
 471             fscal            = felec;
 472
 473             /* Calculate temporary vectorial force */
 474             tx               = _mm_mul_pd(fscal,dx22);
 475             ty               = _mm_mul_pd(fscal,dy22);
 476             tz               = _mm_mul_pd(fscal,dz22);
 477
 478             /* Update vectorial force */
 479             fix2             = _mm_add_pd(fix2,tx);
 480             fiy2             = _mm_add_pd(fiy2,ty);
 481             fiz2             = _mm_add_pd(fiz2,tz);
 482
 483             fjx2             = _mm_add_pd(fjx2,tx);
 484             fjy2             = _mm_add_pd(fjy2,ty);
 485             fjz2             = _mm_add_pd(fjz2,tz);
 486
 487             /**************************
 488              * CALCULATE INTERACTIONS *
 489              **************************/
 490
 491             /* REACTION-FIELD ELECTROSTATICS */
 492             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 493             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 494
 495             /* Update potential sum for this i atom from the interaction with this j atom. */
 496             velecsum         = _mm_add_pd(velecsum,velec);
 497
 498             fscal            = felec;
 499
 500             /* Calculate temporary vectorial force */
 501             tx               = _mm_mul_pd(fscal,dx23);
 502             ty               = _mm_mul_pd(fscal,dy23);
 503             tz               = _mm_mul_pd(fscal,dz23);
 504
 505             /* Update vectorial force */
 506             fix2             = _mm_add_pd(fix2,tx);
 507             fiy2             = _mm_add_pd(fiy2,ty);
 508             fiz2             = _mm_add_pd(fiz2,tz);
 509
 510             fjx3             = _mm_add_pd(fjx3,tx);
 511             fjy3             = _mm_add_pd(fjy3,ty);
 512             fjz3             = _mm_add_pd(fjz3,tz);
 513
 514             /**************************
 515              * CALCULATE INTERACTIONS *
 516              **************************/
 517
 518             /* REACTION-FIELD ELECTROSTATICS */
 519             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 520             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 521
 522             /* Update potential sum for this i atom from the interaction with this j atom. */
 523             velecsum         = _mm_add_pd(velecsum,velec);
 524
 525             fscal            = felec;
 526
 527             /* Calculate temporary vectorial force */
 528             tx               = _mm_mul_pd(fscal,dx31);
 529             ty               = _mm_mul_pd(fscal,dy31);
 530             tz               = _mm_mul_pd(fscal,dz31);
 531
 532             /* Update vectorial force */
 533             fix3             = _mm_add_pd(fix3,tx);
 534             fiy3             = _mm_add_pd(fiy3,ty);
 535             fiz3             = _mm_add_pd(fiz3,tz);
 536
 537             fjx1             = _mm_add_pd(fjx1,tx);
 538             fjy1             = _mm_add_pd(fjy1,ty);
 539             fjz1             = _mm_add_pd(fjz1,tz);
 540
 541             /**************************
 542              * CALCULATE INTERACTIONS *
 543              **************************/
 544
 545             /* REACTION-FIELD ELECTROSTATICS */
 546             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 547             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 548
 549             /* Update potential sum for this i atom from the interaction with this j atom. */
 550             velecsum         = _mm_add_pd(velecsum,velec);
 551
 552             fscal            = felec;
 553
 554             /* Calculate temporary vectorial force */
 555             tx               = _mm_mul_pd(fscal,dx32);
 556             ty               = _mm_mul_pd(fscal,dy32);
 557             tz               = _mm_mul_pd(fscal,dz32);
 558
 559             /* Update vectorial force */
 560             fix3             = _mm_add_pd(fix3,tx);
 561             fiy3             = _mm_add_pd(fiy3,ty);
 562             fiz3             = _mm_add_pd(fiz3,tz);
 563
 564             fjx2             = _mm_add_pd(fjx2,tx);
 565             fjy2             = _mm_add_pd(fjy2,ty);
 566             fjz2             = _mm_add_pd(fjz2,tz);
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             /* REACTION-FIELD ELECTROSTATICS */
 573             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 574             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 575
 576             /* Update potential sum for this i atom from the interaction with this j atom. */
 577             velecsum         = _mm_add_pd(velecsum,velec);
 578
 579             fscal            = felec;
 580
 581             /* Calculate temporary vectorial force */
 582             tx               = _mm_mul_pd(fscal,dx33);
 583             ty               = _mm_mul_pd(fscal,dy33);
 584             tz               = _mm_mul_pd(fscal,dz33);
 585
 586             /* Update vectorial force */
 587             fix3             = _mm_add_pd(fix3,tx);
 588             fiy3             = _mm_add_pd(fiy3,ty);
 589             fiz3             = _mm_add_pd(fiz3,tz);
 590
 591             fjx3             = _mm_add_pd(fjx3,tx);
 592             fjy3             = _mm_add_pd(fjy3,ty);
 593             fjz3             = _mm_add_pd(fjz3,tz);
 594
 595             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);
 596
 597             /* Inner loop uses 347 flops */
 598         }
 599
 600         if(jidx<j_index_end)
 601         {
 602
 603             jnrA             = jjnr[jidx];
 604             j_coord_offsetA  = DIM*jnrA;
 605
 606             /* load j atom coordinates */
 607             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 608                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 609                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 610
 611             /* Calculate displacement vector */
 612             dx00             = _mm_sub_pd(ix0,jx0);
 613             dy00             = _mm_sub_pd(iy0,jy0);
 614             dz00             = _mm_sub_pd(iz0,jz0);
 615             dx11             = _mm_sub_pd(ix1,jx1);
 616             dy11             = _mm_sub_pd(iy1,jy1);
 617             dz11             = _mm_sub_pd(iz1,jz1);
 618             dx12             = _mm_sub_pd(ix1,jx2);
 619             dy12             = _mm_sub_pd(iy1,jy2);
 620             dz12             = _mm_sub_pd(iz1,jz2);
 621             dx13             = _mm_sub_pd(ix1,jx3);
 622             dy13             = _mm_sub_pd(iy1,jy3);
 623             dz13             = _mm_sub_pd(iz1,jz3);
 624             dx21             = _mm_sub_pd(ix2,jx1);
 625             dy21             = _mm_sub_pd(iy2,jy1);
 626             dz21             = _mm_sub_pd(iz2,jz1);
 627             dx22             = _mm_sub_pd(ix2,jx2);
 628             dy22             = _mm_sub_pd(iy2,jy2);
 629             dz22             = _mm_sub_pd(iz2,jz2);
 630             dx23             = _mm_sub_pd(ix2,jx3);
 631             dy23             = _mm_sub_pd(iy2,jy3);
 632             dz23             = _mm_sub_pd(iz2,jz3);
 633             dx31             = _mm_sub_pd(ix3,jx1);
 634             dy31             = _mm_sub_pd(iy3,jy1);
 635             dz31             = _mm_sub_pd(iz3,jz1);
 636             dx32             = _mm_sub_pd(ix3,jx2);
 637             dy32             = _mm_sub_pd(iy3,jy2);
 638             dz32             = _mm_sub_pd(iz3,jz2);
 639             dx33             = _mm_sub_pd(ix3,jx3);
 640             dy33             = _mm_sub_pd(iy3,jy3);
 641             dz33             = _mm_sub_pd(iz3,jz3);
 642
 643             /* Calculate squared distance and things based on it */
 644             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 645             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 646             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 647             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 648             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 649             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 650             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 651             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 652             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 653             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 654
 655             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 656             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 657             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 658             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 659             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 660             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 661             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 662             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 663             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 664             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 665
 666             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 667             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 668             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 669             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 670             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 671             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 672             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 673             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 674             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 675
 676             fjx0             = _mm_setzero_pd();
 677             fjy0             = _mm_setzero_pd();
 678             fjz0             = _mm_setzero_pd();
 679             fjx1             = _mm_setzero_pd();
 680             fjy1             = _mm_setzero_pd();
 681             fjz1             = _mm_setzero_pd();
 682             fjx2             = _mm_setzero_pd();
 683             fjy2             = _mm_setzero_pd();
 684             fjz2             = _mm_setzero_pd();
 685             fjx3             = _mm_setzero_pd();
 686             fjy3             = _mm_setzero_pd();
 687             fjz3             = _mm_setzero_pd();
 688
 689             /**************************
 690              * CALCULATE INTERACTIONS *
 691              **************************/
 692
 693             r00              = _mm_mul_pd(rsq00,rinv00);
 694
 695             /* Calculate table index by multiplying r with table scale and truncate to integer */
 696             rt               = _mm_mul_pd(r00,vftabscale);
 697             vfitab           = _mm_cvttpd_epi32(rt);
 698             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 699             vfitab           = _mm_slli_epi32(vfitab,3);
 700
 701             /* CUBIC SPLINE TABLE DISPERSION */
 702             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 703             F                = _mm_setzero_pd();
 704             GMX_MM_TRANSPOSE2_PD(Y,F);
 705             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 706             H                = _mm_setzero_pd();
 707             GMX_MM_TRANSPOSE2_PD(G,H);
 708             Heps             = _mm_mul_pd(vfeps,H);
 709             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 710             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 711             vvdw6            = _mm_mul_pd(c6_00,VV);
 712             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 713             fvdw6            = _mm_mul_pd(c6_00,FF);
 714
 715             /* CUBIC SPLINE TABLE REPULSION */
 716             vfitab           = _mm_add_epi32(vfitab,ifour);
 717             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 718             F                = _mm_setzero_pd();
 719             GMX_MM_TRANSPOSE2_PD(Y,F);
 720             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 721             H                = _mm_setzero_pd();
 722             GMX_MM_TRANSPOSE2_PD(G,H);
 723             Heps             = _mm_mul_pd(vfeps,H);
 724             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 725             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 726             vvdw12           = _mm_mul_pd(c12_00,VV);
 727             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 728             fvdw12           = _mm_mul_pd(c12_00,FF);
 729             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 730             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 731
 732             /* Update potential sum for this i atom from the interaction with this j atom. */
 733             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 734             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 735
 736             fscal            = fvdw;
 737
 738             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 739
 740             /* Calculate temporary vectorial force */
 741             tx               = _mm_mul_pd(fscal,dx00);
 742             ty               = _mm_mul_pd(fscal,dy00);
 743             tz               = _mm_mul_pd(fscal,dz00);
 744
 745             /* Update vectorial force */
 746             fix0             = _mm_add_pd(fix0,tx);
 747             fiy0             = _mm_add_pd(fiy0,ty);
 748             fiz0             = _mm_add_pd(fiz0,tz);
 749
 750             fjx0             = _mm_add_pd(fjx0,tx);
 751             fjy0             = _mm_add_pd(fjy0,ty);
 752             fjz0             = _mm_add_pd(fjz0,tz);
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             /* REACTION-FIELD ELECTROSTATICS */
 759             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 760             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 761
 762             /* Update potential sum for this i atom from the interaction with this j atom. */
 763             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 764             velecsum         = _mm_add_pd(velecsum,velec);
 765
 766             fscal            = felec;
 767
 768             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 769
 770             /* Calculate temporary vectorial force */
 771             tx               = _mm_mul_pd(fscal,dx11);
 772             ty               = _mm_mul_pd(fscal,dy11);
 773             tz               = _mm_mul_pd(fscal,dz11);
 774
 775             /* Update vectorial force */
 776             fix1             = _mm_add_pd(fix1,tx);
 777             fiy1             = _mm_add_pd(fiy1,ty);
 778             fiz1             = _mm_add_pd(fiz1,tz);
 779
 780             fjx1             = _mm_add_pd(fjx1,tx);
 781             fjy1             = _mm_add_pd(fjy1,ty);
 782             fjz1             = _mm_add_pd(fjz1,tz);
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             /* REACTION-FIELD ELECTROSTATICS */
 789             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 790             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 791
 792             /* Update potential sum for this i atom from the interaction with this j atom. */
 793             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 794             velecsum         = _mm_add_pd(velecsum,velec);
 795
 796             fscal            = felec;
 797
 798             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 799
 800             /* Calculate temporary vectorial force */
 801             tx               = _mm_mul_pd(fscal,dx12);
 802             ty               = _mm_mul_pd(fscal,dy12);
 803             tz               = _mm_mul_pd(fscal,dz12);
 804
 805             /* Update vectorial force */
 806             fix1             = _mm_add_pd(fix1,tx);
 807             fiy1             = _mm_add_pd(fiy1,ty);
 808             fiz1             = _mm_add_pd(fiz1,tz);
 809
 810             fjx2             = _mm_add_pd(fjx2,tx);
 811             fjy2             = _mm_add_pd(fjy2,ty);
 812             fjz2             = _mm_add_pd(fjz2,tz);
 813
 814             /**************************
 815              * CALCULATE INTERACTIONS *
 816              **************************/
 817
 818             /* REACTION-FIELD ELECTROSTATICS */
 819             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 820             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 821
 822             /* Update potential sum for this i atom from the interaction with this j atom. */
 823             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 824             velecsum         = _mm_add_pd(velecsum,velec);
 825
 826             fscal            = felec;
 827
 828             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 829
 830             /* Calculate temporary vectorial force */
 831             tx               = _mm_mul_pd(fscal,dx13);
 832             ty               = _mm_mul_pd(fscal,dy13);
 833             tz               = _mm_mul_pd(fscal,dz13);
 834
 835             /* Update vectorial force */
 836             fix1             = _mm_add_pd(fix1,tx);
 837             fiy1             = _mm_add_pd(fiy1,ty);
 838             fiz1             = _mm_add_pd(fiz1,tz);
 839
 840             fjx3             = _mm_add_pd(fjx3,tx);
 841             fjy3             = _mm_add_pd(fjy3,ty);
 842             fjz3             = _mm_add_pd(fjz3,tz);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* REACTION-FIELD ELECTROSTATICS */
 849             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 850             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 851
 852             /* Update potential sum for this i atom from the interaction with this j atom. */
 853             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 854             velecsum         = _mm_add_pd(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 859
 860             /* Calculate temporary vectorial force */
 861             tx               = _mm_mul_pd(fscal,dx21);
 862             ty               = _mm_mul_pd(fscal,dy21);
 863             tz               = _mm_mul_pd(fscal,dz21);
 864
 865             /* Update vectorial force */
 866             fix2             = _mm_add_pd(fix2,tx);
 867             fiy2             = _mm_add_pd(fiy2,ty);
 868             fiz2             = _mm_add_pd(fiz2,tz);
 869
 870             fjx1             = _mm_add_pd(fjx1,tx);
 871             fjy1             = _mm_add_pd(fjy1,ty);
 872             fjz1             = _mm_add_pd(fjz1,tz);
 873
 874             /**************************
 875              * CALCULATE INTERACTIONS *
 876              **************************/
 877
 878             /* REACTION-FIELD ELECTROSTATICS */
 879             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 880             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 881
 882             /* Update potential sum for this i atom from the interaction with this j atom. */
 883             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 884             velecsum         = _mm_add_pd(velecsum,velec);
 885
 886             fscal            = felec;
 887
 888             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 889
 890             /* Calculate temporary vectorial force */
 891             tx               = _mm_mul_pd(fscal,dx22);
 892             ty               = _mm_mul_pd(fscal,dy22);
 893             tz               = _mm_mul_pd(fscal,dz22);
 894
 895             /* Update vectorial force */
 896             fix2             = _mm_add_pd(fix2,tx);
 897             fiy2             = _mm_add_pd(fiy2,ty);
 898             fiz2             = _mm_add_pd(fiz2,tz);
 899
 900             fjx2             = _mm_add_pd(fjx2,tx);
 901             fjy2             = _mm_add_pd(fjy2,ty);
 902             fjz2             = _mm_add_pd(fjz2,tz);
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             /* REACTION-FIELD ELECTROSTATICS */
 909             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 910             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 911
 912             /* Update potential sum for this i atom from the interaction with this j atom. */
 913             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 914             velecsum         = _mm_add_pd(velecsum,velec);
 915
 916             fscal            = felec;
 917
 918             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 919
 920             /* Calculate temporary vectorial force */
 921             tx               = _mm_mul_pd(fscal,dx23);
 922             ty               = _mm_mul_pd(fscal,dy23);
 923             tz               = _mm_mul_pd(fscal,dz23);
 924
 925             /* Update vectorial force */
 926             fix2             = _mm_add_pd(fix2,tx);
 927             fiy2             = _mm_add_pd(fiy2,ty);
 928             fiz2             = _mm_add_pd(fiz2,tz);
 929
 930             fjx3             = _mm_add_pd(fjx3,tx);
 931             fjy3             = _mm_add_pd(fjy3,ty);
 932             fjz3             = _mm_add_pd(fjz3,tz);
 933
 934             /**************************
 935              * CALCULATE INTERACTIONS *
 936              **************************/
 937
 938             /* REACTION-FIELD ELECTROSTATICS */
 939             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 940             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 941
 942             /* Update potential sum for this i atom from the interaction with this j atom. */
 943             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 944             velecsum         = _mm_add_pd(velecsum,velec);
 945
 946             fscal            = felec;
 947
 948             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 949
 950             /* Calculate temporary vectorial force */
 951             tx               = _mm_mul_pd(fscal,dx31);
 952             ty               = _mm_mul_pd(fscal,dy31);
 953             tz               = _mm_mul_pd(fscal,dz31);
 954
 955             /* Update vectorial force */
 956             fix3             = _mm_add_pd(fix3,tx);
 957             fiy3             = _mm_add_pd(fiy3,ty);
 958             fiz3             = _mm_add_pd(fiz3,tz);
 959
 960             fjx1             = _mm_add_pd(fjx1,tx);
 961             fjy1             = _mm_add_pd(fjy1,ty);
 962             fjz1             = _mm_add_pd(fjz1,tz);
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             /* REACTION-FIELD ELECTROSTATICS */
 969             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 970             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 971
 972             /* Update potential sum for this i atom from the interaction with this j atom. */
 973             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 974             velecsum         = _mm_add_pd(velecsum,velec);
 975
 976             fscal            = felec;
 977
 978             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 979
 980             /* Calculate temporary vectorial force */
 981             tx               = _mm_mul_pd(fscal,dx32);
 982             ty               = _mm_mul_pd(fscal,dy32);
 983             tz               = _mm_mul_pd(fscal,dz32);
 984
 985             /* Update vectorial force */
 986             fix3             = _mm_add_pd(fix3,tx);
 987             fiy3             = _mm_add_pd(fiy3,ty);
 988             fiz3             = _mm_add_pd(fiz3,tz);
 989
 990             fjx2             = _mm_add_pd(fjx2,tx);
 991             fjy2             = _mm_add_pd(fjy2,ty);
 992             fjz2             = _mm_add_pd(fjz2,tz);
 993
 994             /**************************
 995              * CALCULATE INTERACTIONS *
 996              **************************/
 997
 998             /* REACTION-FIELD ELECTROSTATICS */
 999             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1000             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1001
1002             /* Update potential sum for this i atom from the interaction with this j atom. */
1003             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1004             velecsum         = _mm_add_pd(velecsum,velec);
1005
1006             fscal            = felec;
1007
1008             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1009
1010             /* Calculate temporary vectorial force */
1011             tx               = _mm_mul_pd(fscal,dx33);
1012             ty               = _mm_mul_pd(fscal,dy33);
1013             tz               = _mm_mul_pd(fscal,dz33);
1014
1015             /* Update vectorial force */
1016             fix3             = _mm_add_pd(fix3,tx);
1017             fiy3             = _mm_add_pd(fiy3,ty);
1018             fiz3             = _mm_add_pd(fiz3,tz);
1019
1020             fjx3             = _mm_add_pd(fjx3,tx);
1021             fjy3             = _mm_add_pd(fjy3,ty);
1022             fjz3             = _mm_add_pd(fjz3,tz);
1023
1024             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1025
1026             /* Inner loop uses 347 flops */
1027         }
1028
1029         /* End of innermost loop */
1030
1031         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1032                                               f+i_coord_offset,fshift+i_shift_offset);
1033
1034         ggid                        = gid[iidx];
1035         /* Update potential energies */
1036         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1037         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1038
1039         /* Increment number of inner iterations */
1040         inneriter                  += j_index_end - j_index_start;
1041
1042         /* Outer loop uses 26 flops */
1043     }
1044
1045     /* Increment number of outer iterations */
1046     outeriter        += nri;
1047
1048     /* Update outer/inner flops */
1049
1050     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*347);
1051 }
1052 /*
1053  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse4_1_double
1054  * Electrostatics interaction: ReactionField
1055  * VdW interaction:            CubicSplineTable
1056  * Geometry:                   Water4-Water4
1057  * Calculate force/pot:        Force
1058  */
1059 void
1060 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse4_1_double
1061                     (t_nblist * gmx_restrict                nlist,
1062                      rvec * gmx_restrict                    xx,
1063                      rvec * gmx_restrict                    ff,
1064                      t_forcerec * gmx_restrict              fr,
1065                      t_mdatoms * gmx_restrict               mdatoms,
1066                      nb_kernel_data_t * gmx_restrict        kernel_data,
1067                      t_nrnb * gmx_restrict                  nrnb)
1068 {
1069     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1070      * just 0 for non-waters.
1071      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1072      * jnr indices corresponding to data put in the four positions in the SIMD register.
1073      */
1074     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1075     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1076     int              jnrA,jnrB;
1077     int              j_coord_offsetA,j_coord_offsetB;
1078     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1079     real             rcutoff_scalar;
1080     real             *shiftvec,*fshift,*x,*f;
1081     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1082     int              vdwioffset0;
1083     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1084     int              vdwioffset1;
1085     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1086     int              vdwioffset2;
1087     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1088     int              vdwioffset3;
1089     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1090     int              vdwjidx0A,vdwjidx0B;
1091     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1092     int              vdwjidx1A,vdwjidx1B;
1093     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1094     int              vdwjidx2A,vdwjidx2B;
1095     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1096     int              vdwjidx3A,vdwjidx3B;
1097     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1098     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1099     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1109     real             *charge;
1110     int              nvdwtype;
1111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1112     int              *vdwtype;
1113     real             *vdwparam;
1114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1116     __m128i          vfitab;
1117     __m128i          ifour       = _mm_set1_epi32(4);
1118     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1119     real             *vftab;
1120     __m128d          dummy_mask,cutoff_mask;
1121     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1122     __m128d          one     = _mm_set1_pd(1.0);
1123     __m128d          two     = _mm_set1_pd(2.0);
1124     x                = xx[0];
1125     f                = ff[0];
1126
1127     nri              = nlist->nri;
1128     iinr             = nlist->iinr;
1129     jindex           = nlist->jindex;
1130     jjnr             = nlist->jjnr;
1131     shiftidx         = nlist->shift;
1132     gid              = nlist->gid;
1133     shiftvec         = fr->shift_vec[0];
1134     fshift           = fr->fshift[0];
1135     facel            = _mm_set1_pd(fr->epsfac);
1136     charge           = mdatoms->chargeA;
1137     krf              = _mm_set1_pd(fr->ic->k_rf);
1138     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1139     crf              = _mm_set1_pd(fr->ic->c_rf);
1140     nvdwtype         = fr->ntype;
1141     vdwparam         = fr->nbfp;
1142     vdwtype          = mdatoms->typeA;
1143
1144     vftab            = kernel_data->table_vdw->data;
1145     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1146
1147     /* Setup water-specific parameters */
1148     inr              = nlist->iinr[0];
1149     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1150     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1151     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1152     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1153
1154     jq1              = _mm_set1_pd(charge[inr+1]);
1155     jq2              = _mm_set1_pd(charge[inr+2]);
1156     jq3              = _mm_set1_pd(charge[inr+3]);
1157     vdwjidx0A        = 2*vdwtype[inr+0];
1158     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1159     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1160     qq11             = _mm_mul_pd(iq1,jq1);
1161     qq12             = _mm_mul_pd(iq1,jq2);
1162     qq13             = _mm_mul_pd(iq1,jq3);
1163     qq21             = _mm_mul_pd(iq2,jq1);
1164     qq22             = _mm_mul_pd(iq2,jq2);
1165     qq23             = _mm_mul_pd(iq2,jq3);
1166     qq31             = _mm_mul_pd(iq3,jq1);
1167     qq32             = _mm_mul_pd(iq3,jq2);
1168     qq33             = _mm_mul_pd(iq3,jq3);
1169
1170     /* Avoid stupid compiler warnings */
1171     jnrA = jnrB = 0;
1172     j_coord_offsetA = 0;
1173     j_coord_offsetB = 0;
1174
1175     outeriter        = 0;
1176     inneriter        = 0;
1177
1178     /* Start outer loop over neighborlists */
1179     for(iidx=0; iidx<nri; iidx++)
1180     {
1181         /* Load shift vector for this list */
1182         i_shift_offset   = DIM*shiftidx[iidx];
1183
1184         /* Load limits for loop over neighbors */
1185         j_index_start    = jindex[iidx];
1186         j_index_end      = jindex[iidx+1];
1187
1188         /* Get outer coordinate index */
1189         inr              = iinr[iidx];
1190         i_coord_offset   = DIM*inr;
1191
1192         /* Load i particle coords and add shift vector */
1193         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1194                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1195
1196         fix0             = _mm_setzero_pd();
1197         fiy0             = _mm_setzero_pd();
1198         fiz0             = _mm_setzero_pd();
1199         fix1             = _mm_setzero_pd();
1200         fiy1             = _mm_setzero_pd();
1201         fiz1             = _mm_setzero_pd();
1202         fix2             = _mm_setzero_pd();
1203         fiy2             = _mm_setzero_pd();
1204         fiz2             = _mm_setzero_pd();
1205         fix3             = _mm_setzero_pd();
1206         fiy3             = _mm_setzero_pd();
1207         fiz3             = _mm_setzero_pd();
1208
1209         /* Start inner kernel loop */
1210         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1211         {
1212
1213             /* Get j neighbor index, and coordinate index */
1214             jnrA             = jjnr[jidx];
1215             jnrB             = jjnr[jidx+1];
1216             j_coord_offsetA  = DIM*jnrA;
1217             j_coord_offsetB  = DIM*jnrB;
1218
1219             /* load j atom coordinates */
1220             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1221                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1222                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1223
1224             /* Calculate displacement vector */
1225             dx00             = _mm_sub_pd(ix0,jx0);
1226             dy00             = _mm_sub_pd(iy0,jy0);
1227             dz00             = _mm_sub_pd(iz0,jz0);
1228             dx11             = _mm_sub_pd(ix1,jx1);
1229             dy11             = _mm_sub_pd(iy1,jy1);
1230             dz11             = _mm_sub_pd(iz1,jz1);
1231             dx12             = _mm_sub_pd(ix1,jx2);
1232             dy12             = _mm_sub_pd(iy1,jy2);
1233             dz12             = _mm_sub_pd(iz1,jz2);
1234             dx13             = _mm_sub_pd(ix1,jx3);
1235             dy13             = _mm_sub_pd(iy1,jy3);
1236             dz13             = _mm_sub_pd(iz1,jz3);
1237             dx21             = _mm_sub_pd(ix2,jx1);
1238             dy21             = _mm_sub_pd(iy2,jy1);
1239             dz21             = _mm_sub_pd(iz2,jz1);
1240             dx22             = _mm_sub_pd(ix2,jx2);
1241             dy22             = _mm_sub_pd(iy2,jy2);
1242             dz22             = _mm_sub_pd(iz2,jz2);
1243             dx23             = _mm_sub_pd(ix2,jx3);
1244             dy23             = _mm_sub_pd(iy2,jy3);
1245             dz23             = _mm_sub_pd(iz2,jz3);
1246             dx31             = _mm_sub_pd(ix3,jx1);
1247             dy31             = _mm_sub_pd(iy3,jy1);
1248             dz31             = _mm_sub_pd(iz3,jz1);
1249             dx32             = _mm_sub_pd(ix3,jx2);
1250             dy32             = _mm_sub_pd(iy3,jy2);
1251             dz32             = _mm_sub_pd(iz3,jz2);
1252             dx33             = _mm_sub_pd(ix3,jx3);
1253             dy33             = _mm_sub_pd(iy3,jy3);
1254             dz33             = _mm_sub_pd(iz3,jz3);
1255
1256             /* Calculate squared distance and things based on it */
1257             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1258             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1259             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1260             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1261             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1262             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1263             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1264             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1265             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1266             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1267
1268             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1269             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1270             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1271             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1272             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1273             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1274             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1275             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1276             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1277             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1278
1279             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1280             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1281             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1282             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1283             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1284             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1285             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1286             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1287             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1288
1289             fjx0             = _mm_setzero_pd();
1290             fjy0             = _mm_setzero_pd();
1291             fjz0             = _mm_setzero_pd();
1292             fjx1             = _mm_setzero_pd();
1293             fjy1             = _mm_setzero_pd();
1294             fjz1             = _mm_setzero_pd();
1295             fjx2             = _mm_setzero_pd();
1296             fjy2             = _mm_setzero_pd();
1297             fjz2             = _mm_setzero_pd();
1298             fjx3             = _mm_setzero_pd();
1299             fjy3             = _mm_setzero_pd();
1300             fjz3             = _mm_setzero_pd();
1301
1302             /**************************
1303              * CALCULATE INTERACTIONS *
1304              **************************/
1305
1306             r00              = _mm_mul_pd(rsq00,rinv00);
1307
1308             /* Calculate table index by multiplying r with table scale and truncate to integer */
1309             rt               = _mm_mul_pd(r00,vftabscale);
1310             vfitab           = _mm_cvttpd_epi32(rt);
1311             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1312             vfitab           = _mm_slli_epi32(vfitab,3);
1313
1314             /* CUBIC SPLINE TABLE DISPERSION */
1315             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1316             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1317             GMX_MM_TRANSPOSE2_PD(Y,F);
1318             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1319             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1320             GMX_MM_TRANSPOSE2_PD(G,H);
1321             Heps             = _mm_mul_pd(vfeps,H);
1322             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1323             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1324             fvdw6            = _mm_mul_pd(c6_00,FF);
1325
1326             /* CUBIC SPLINE TABLE REPULSION */
1327             vfitab           = _mm_add_epi32(vfitab,ifour);
1328             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1329             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1330             GMX_MM_TRANSPOSE2_PD(Y,F);
1331             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1332             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1333             GMX_MM_TRANSPOSE2_PD(G,H);
1334             Heps             = _mm_mul_pd(vfeps,H);
1335             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1336             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1337             fvdw12           = _mm_mul_pd(c12_00,FF);
1338             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1339
1340             fscal            = fvdw;
1341
1342             /* Calculate temporary vectorial force */
1343             tx               = _mm_mul_pd(fscal,dx00);
1344             ty               = _mm_mul_pd(fscal,dy00);
1345             tz               = _mm_mul_pd(fscal,dz00);
1346
1347             /* Update vectorial force */
1348             fix0             = _mm_add_pd(fix0,tx);
1349             fiy0             = _mm_add_pd(fiy0,ty);
1350             fiz0             = _mm_add_pd(fiz0,tz);
1351
1352             fjx0             = _mm_add_pd(fjx0,tx);
1353             fjy0             = _mm_add_pd(fjy0,ty);
1354             fjz0             = _mm_add_pd(fjz0,tz);
1355
1356             /**************************
1357              * CALCULATE INTERACTIONS *
1358              **************************/
1359
1360             /* REACTION-FIELD ELECTROSTATICS */
1361             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1362
1363             fscal            = felec;
1364
1365             /* Calculate temporary vectorial force */
1366             tx               = _mm_mul_pd(fscal,dx11);
1367             ty               = _mm_mul_pd(fscal,dy11);
1368             tz               = _mm_mul_pd(fscal,dz11);
1369
1370             /* Update vectorial force */
1371             fix1             = _mm_add_pd(fix1,tx);
1372             fiy1             = _mm_add_pd(fiy1,ty);
1373             fiz1             = _mm_add_pd(fiz1,tz);
1374
1375             fjx1             = _mm_add_pd(fjx1,tx);
1376             fjy1             = _mm_add_pd(fjy1,ty);
1377             fjz1             = _mm_add_pd(fjz1,tz);
1378
1379             /**************************
1380              * CALCULATE INTERACTIONS *
1381              **************************/
1382
1383             /* REACTION-FIELD ELECTROSTATICS */
1384             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1385
1386             fscal            = felec;
1387
1388             /* Calculate temporary vectorial force */
1389             tx               = _mm_mul_pd(fscal,dx12);
1390             ty               = _mm_mul_pd(fscal,dy12);
1391             tz               = _mm_mul_pd(fscal,dz12);
1392
1393             /* Update vectorial force */
1394             fix1             = _mm_add_pd(fix1,tx);
1395             fiy1             = _mm_add_pd(fiy1,ty);
1396             fiz1             = _mm_add_pd(fiz1,tz);
1397
1398             fjx2             = _mm_add_pd(fjx2,tx);
1399             fjy2             = _mm_add_pd(fjy2,ty);
1400             fjz2             = _mm_add_pd(fjz2,tz);
1401
1402             /**************************
1403              * CALCULATE INTERACTIONS *
1404              **************************/
1405
1406             /* REACTION-FIELD ELECTROSTATICS */
1407             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1408
1409             fscal            = felec;
1410
1411             /* Calculate temporary vectorial force */
1412             tx               = _mm_mul_pd(fscal,dx13);
1413             ty               = _mm_mul_pd(fscal,dy13);
1414             tz               = _mm_mul_pd(fscal,dz13);
1415
1416             /* Update vectorial force */
1417             fix1             = _mm_add_pd(fix1,tx);
1418             fiy1             = _mm_add_pd(fiy1,ty);
1419             fiz1             = _mm_add_pd(fiz1,tz);
1420
1421             fjx3             = _mm_add_pd(fjx3,tx);
1422             fjy3             = _mm_add_pd(fjy3,ty);
1423             fjz3             = _mm_add_pd(fjz3,tz);
1424
1425             /**************************
1426              * CALCULATE INTERACTIONS *
1427              **************************/
1428
1429             /* REACTION-FIELD ELECTROSTATICS */
1430             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1431
1432             fscal            = felec;
1433
1434             /* Calculate temporary vectorial force */
1435             tx               = _mm_mul_pd(fscal,dx21);
1436             ty               = _mm_mul_pd(fscal,dy21);
1437             tz               = _mm_mul_pd(fscal,dz21);
1438
1439             /* Update vectorial force */
1440             fix2             = _mm_add_pd(fix2,tx);
1441             fiy2             = _mm_add_pd(fiy2,ty);
1442             fiz2             = _mm_add_pd(fiz2,tz);
1443
1444             fjx1             = _mm_add_pd(fjx1,tx);
1445             fjy1             = _mm_add_pd(fjy1,ty);
1446             fjz1             = _mm_add_pd(fjz1,tz);
1447
1448             /**************************
1449              * CALCULATE INTERACTIONS *
1450              **************************/
1451
1452             /* REACTION-FIELD ELECTROSTATICS */
1453             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1454
1455             fscal            = felec;
1456
1457             /* Calculate temporary vectorial force */
1458             tx               = _mm_mul_pd(fscal,dx22);
1459             ty               = _mm_mul_pd(fscal,dy22);
1460             tz               = _mm_mul_pd(fscal,dz22);
1461
1462             /* Update vectorial force */
1463             fix2             = _mm_add_pd(fix2,tx);
1464             fiy2             = _mm_add_pd(fiy2,ty);
1465             fiz2             = _mm_add_pd(fiz2,tz);
1466
1467             fjx2             = _mm_add_pd(fjx2,tx);
1468             fjy2             = _mm_add_pd(fjy2,ty);
1469             fjz2             = _mm_add_pd(fjz2,tz);
1470
1471             /**************************
1472              * CALCULATE INTERACTIONS *
1473              **************************/
1474
1475             /* REACTION-FIELD ELECTROSTATICS */
1476             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1477
1478             fscal            = felec;
1479
1480             /* Calculate temporary vectorial force */
1481             tx               = _mm_mul_pd(fscal,dx23);
1482             ty               = _mm_mul_pd(fscal,dy23);
1483             tz               = _mm_mul_pd(fscal,dz23);
1484
1485             /* Update vectorial force */
1486             fix2             = _mm_add_pd(fix2,tx);
1487             fiy2             = _mm_add_pd(fiy2,ty);
1488             fiz2             = _mm_add_pd(fiz2,tz);
1489
1490             fjx3             = _mm_add_pd(fjx3,tx);
1491             fjy3             = _mm_add_pd(fjy3,ty);
1492             fjz3             = _mm_add_pd(fjz3,tz);
1493
1494             /**************************
1495              * CALCULATE INTERACTIONS *
1496              **************************/
1497
1498             /* REACTION-FIELD ELECTROSTATICS */
1499             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1500
1501             fscal            = felec;
1502
1503             /* Calculate temporary vectorial force */
1504             tx               = _mm_mul_pd(fscal,dx31);
1505             ty               = _mm_mul_pd(fscal,dy31);
1506             tz               = _mm_mul_pd(fscal,dz31);
1507
1508             /* Update vectorial force */
1509             fix3             = _mm_add_pd(fix3,tx);
1510             fiy3             = _mm_add_pd(fiy3,ty);
1511             fiz3             = _mm_add_pd(fiz3,tz);
1512
1513             fjx1             = _mm_add_pd(fjx1,tx);
1514             fjy1             = _mm_add_pd(fjy1,ty);
1515             fjz1             = _mm_add_pd(fjz1,tz);
1516
1517             /**************************
1518              * CALCULATE INTERACTIONS *
1519              **************************/
1520
1521             /* REACTION-FIELD ELECTROSTATICS */
1522             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1523
1524             fscal            = felec;
1525
1526             /* Calculate temporary vectorial force */
1527             tx               = _mm_mul_pd(fscal,dx32);
1528             ty               = _mm_mul_pd(fscal,dy32);
1529             tz               = _mm_mul_pd(fscal,dz32);
1530
1531             /* Update vectorial force */
1532             fix3             = _mm_add_pd(fix3,tx);
1533             fiy3             = _mm_add_pd(fiy3,ty);
1534             fiz3             = _mm_add_pd(fiz3,tz);
1535
1536             fjx2             = _mm_add_pd(fjx2,tx);
1537             fjy2             = _mm_add_pd(fjy2,ty);
1538             fjz2             = _mm_add_pd(fjz2,tz);
1539
1540             /**************************
1541              * CALCULATE INTERACTIONS *
1542              **************************/
1543
1544             /* REACTION-FIELD ELECTROSTATICS */
1545             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1546
1547             fscal            = felec;
1548
1549             /* Calculate temporary vectorial force */
1550             tx               = _mm_mul_pd(fscal,dx33);
1551             ty               = _mm_mul_pd(fscal,dy33);
1552             tz               = _mm_mul_pd(fscal,dz33);
1553
1554             /* Update vectorial force */
1555             fix3             = _mm_add_pd(fix3,tx);
1556             fiy3             = _mm_add_pd(fiy3,ty);
1557             fiz3             = _mm_add_pd(fiz3,tz);
1558
1559             fjx3             = _mm_add_pd(fjx3,tx);
1560             fjy3             = _mm_add_pd(fjy3,ty);
1561             fjz3             = _mm_add_pd(fjz3,tz);
1562
1563             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);
1564
1565             /* Inner loop uses 294 flops */
1566         }
1567
1568         if(jidx<j_index_end)
1569         {
1570
1571             jnrA             = jjnr[jidx];
1572             j_coord_offsetA  = DIM*jnrA;
1573
1574             /* load j atom coordinates */
1575             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1576                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1577                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1578
1579             /* Calculate displacement vector */
1580             dx00             = _mm_sub_pd(ix0,jx0);
1581             dy00             = _mm_sub_pd(iy0,jy0);
1582             dz00             = _mm_sub_pd(iz0,jz0);
1583             dx11             = _mm_sub_pd(ix1,jx1);
1584             dy11             = _mm_sub_pd(iy1,jy1);
1585             dz11             = _mm_sub_pd(iz1,jz1);
1586             dx12             = _mm_sub_pd(ix1,jx2);
1587             dy12             = _mm_sub_pd(iy1,jy2);
1588             dz12             = _mm_sub_pd(iz1,jz2);
1589             dx13             = _mm_sub_pd(ix1,jx3);
1590             dy13             = _mm_sub_pd(iy1,jy3);
1591             dz13             = _mm_sub_pd(iz1,jz3);
1592             dx21             = _mm_sub_pd(ix2,jx1);
1593             dy21             = _mm_sub_pd(iy2,jy1);
1594             dz21             = _mm_sub_pd(iz2,jz1);
1595             dx22             = _mm_sub_pd(ix2,jx2);
1596             dy22             = _mm_sub_pd(iy2,jy2);
1597             dz22             = _mm_sub_pd(iz2,jz2);
1598             dx23             = _mm_sub_pd(ix2,jx3);
1599             dy23             = _mm_sub_pd(iy2,jy3);
1600             dz23             = _mm_sub_pd(iz2,jz3);
1601             dx31             = _mm_sub_pd(ix3,jx1);
1602             dy31             = _mm_sub_pd(iy3,jy1);
1603             dz31             = _mm_sub_pd(iz3,jz1);
1604             dx32             = _mm_sub_pd(ix3,jx2);
1605             dy32             = _mm_sub_pd(iy3,jy2);
1606             dz32             = _mm_sub_pd(iz3,jz2);
1607             dx33             = _mm_sub_pd(ix3,jx3);
1608             dy33             = _mm_sub_pd(iy3,jy3);
1609             dz33             = _mm_sub_pd(iz3,jz3);
1610
1611             /* Calculate squared distance and things based on it */
1612             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1613             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1614             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1615             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1616             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1617             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1618             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1619             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1620             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1621             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1622
1623             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1624             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1625             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1626             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1627             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1628             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1629             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1630             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1631             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1632             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1633
1634             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1635             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1636             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1637             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1638             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1639             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1640             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1641             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1642             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1643
1644             fjx0             = _mm_setzero_pd();
1645             fjy0             = _mm_setzero_pd();
1646             fjz0             = _mm_setzero_pd();
1647             fjx1             = _mm_setzero_pd();
1648             fjy1             = _mm_setzero_pd();
1649             fjz1             = _mm_setzero_pd();
1650             fjx2             = _mm_setzero_pd();
1651             fjy2             = _mm_setzero_pd();
1652             fjz2             = _mm_setzero_pd();
1653             fjx3             = _mm_setzero_pd();
1654             fjy3             = _mm_setzero_pd();
1655             fjz3             = _mm_setzero_pd();
1656
1657             /**************************
1658              * CALCULATE INTERACTIONS *
1659              **************************/
1660
1661             r00              = _mm_mul_pd(rsq00,rinv00);
1662
1663             /* Calculate table index by multiplying r with table scale and truncate to integer */
1664             rt               = _mm_mul_pd(r00,vftabscale);
1665             vfitab           = _mm_cvttpd_epi32(rt);
1666             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1667             vfitab           = _mm_slli_epi32(vfitab,3);
1668
1669             /* CUBIC SPLINE TABLE DISPERSION */
1670             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1671             F                = _mm_setzero_pd();
1672             GMX_MM_TRANSPOSE2_PD(Y,F);
1673             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1674             H                = _mm_setzero_pd();
1675             GMX_MM_TRANSPOSE2_PD(G,H);
1676             Heps             = _mm_mul_pd(vfeps,H);
1677             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1678             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1679             fvdw6            = _mm_mul_pd(c6_00,FF);
1680
1681             /* CUBIC SPLINE TABLE REPULSION */
1682             vfitab           = _mm_add_epi32(vfitab,ifour);
1683             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1684             F                = _mm_setzero_pd();
1685             GMX_MM_TRANSPOSE2_PD(Y,F);
1686             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1687             H                = _mm_setzero_pd();
1688             GMX_MM_TRANSPOSE2_PD(G,H);
1689             Heps             = _mm_mul_pd(vfeps,H);
1690             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1691             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1692             fvdw12           = _mm_mul_pd(c12_00,FF);
1693             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1694
1695             fscal            = fvdw;
1696
1697             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1698
1699             /* Calculate temporary vectorial force */
1700             tx               = _mm_mul_pd(fscal,dx00);
1701             ty               = _mm_mul_pd(fscal,dy00);
1702             tz               = _mm_mul_pd(fscal,dz00);
1703
1704             /* Update vectorial force */
1705             fix0             = _mm_add_pd(fix0,tx);
1706             fiy0             = _mm_add_pd(fiy0,ty);
1707             fiz0             = _mm_add_pd(fiz0,tz);
1708
1709             fjx0             = _mm_add_pd(fjx0,tx);
1710             fjy0             = _mm_add_pd(fjy0,ty);
1711             fjz0             = _mm_add_pd(fjz0,tz);
1712
1713             /**************************
1714              * CALCULATE INTERACTIONS *
1715              **************************/
1716
1717             /* REACTION-FIELD ELECTROSTATICS */
1718             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1719
1720             fscal            = felec;
1721
1722             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1723
1724             /* Calculate temporary vectorial force */
1725             tx               = _mm_mul_pd(fscal,dx11);
1726             ty               = _mm_mul_pd(fscal,dy11);
1727             tz               = _mm_mul_pd(fscal,dz11);
1728
1729             /* Update vectorial force */
1730             fix1             = _mm_add_pd(fix1,tx);
1731             fiy1             = _mm_add_pd(fiy1,ty);
1732             fiz1             = _mm_add_pd(fiz1,tz);
1733
1734             fjx1             = _mm_add_pd(fjx1,tx);
1735             fjy1             = _mm_add_pd(fjy1,ty);
1736             fjz1             = _mm_add_pd(fjz1,tz);
1737
1738             /**************************
1739              * CALCULATE INTERACTIONS *
1740              **************************/
1741
1742             /* REACTION-FIELD ELECTROSTATICS */
1743             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1744
1745             fscal            = felec;
1746
1747             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1748
1749             /* Calculate temporary vectorial force */
1750             tx               = _mm_mul_pd(fscal,dx12);
1751             ty               = _mm_mul_pd(fscal,dy12);
1752             tz               = _mm_mul_pd(fscal,dz12);
1753
1754             /* Update vectorial force */
1755             fix1             = _mm_add_pd(fix1,tx);
1756             fiy1             = _mm_add_pd(fiy1,ty);
1757             fiz1             = _mm_add_pd(fiz1,tz);
1758
1759             fjx2             = _mm_add_pd(fjx2,tx);
1760             fjy2             = _mm_add_pd(fjy2,ty);
1761             fjz2             = _mm_add_pd(fjz2,tz);
1762
1763             /**************************
1764              * CALCULATE INTERACTIONS *
1765              **************************/
1766
1767             /* REACTION-FIELD ELECTROSTATICS */
1768             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1769
1770             fscal            = felec;
1771
1772             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1773
1774             /* Calculate temporary vectorial force */
1775             tx               = _mm_mul_pd(fscal,dx13);
1776             ty               = _mm_mul_pd(fscal,dy13);
1777             tz               = _mm_mul_pd(fscal,dz13);
1778
1779             /* Update vectorial force */
1780             fix1             = _mm_add_pd(fix1,tx);
1781             fiy1             = _mm_add_pd(fiy1,ty);
1782             fiz1             = _mm_add_pd(fiz1,tz);
1783
1784             fjx3             = _mm_add_pd(fjx3,tx);
1785             fjy3             = _mm_add_pd(fjy3,ty);
1786             fjz3             = _mm_add_pd(fjz3,tz);
1787
1788             /**************************
1789              * CALCULATE INTERACTIONS *
1790              **************************/
1791
1792             /* REACTION-FIELD ELECTROSTATICS */
1793             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1794
1795             fscal            = felec;
1796
1797             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1798
1799             /* Calculate temporary vectorial force */
1800             tx               = _mm_mul_pd(fscal,dx21);
1801             ty               = _mm_mul_pd(fscal,dy21);
1802             tz               = _mm_mul_pd(fscal,dz21);
1803
1804             /* Update vectorial force */
1805             fix2             = _mm_add_pd(fix2,tx);
1806             fiy2             = _mm_add_pd(fiy2,ty);
1807             fiz2             = _mm_add_pd(fiz2,tz);
1808
1809             fjx1             = _mm_add_pd(fjx1,tx);
1810             fjy1             = _mm_add_pd(fjy1,ty);
1811             fjz1             = _mm_add_pd(fjz1,tz);
1812
1813             /**************************
1814              * CALCULATE INTERACTIONS *
1815              **************************/
1816
1817             /* REACTION-FIELD ELECTROSTATICS */
1818             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1819
1820             fscal            = felec;
1821
1822             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1823
1824             /* Calculate temporary vectorial force */
1825             tx               = _mm_mul_pd(fscal,dx22);
1826             ty               = _mm_mul_pd(fscal,dy22);
1827             tz               = _mm_mul_pd(fscal,dz22);
1828
1829             /* Update vectorial force */
1830             fix2             = _mm_add_pd(fix2,tx);
1831             fiy2             = _mm_add_pd(fiy2,ty);
1832             fiz2             = _mm_add_pd(fiz2,tz);
1833
1834             fjx2             = _mm_add_pd(fjx2,tx);
1835             fjy2             = _mm_add_pd(fjy2,ty);
1836             fjz2             = _mm_add_pd(fjz2,tz);
1837
1838             /**************************
1839              * CALCULATE INTERACTIONS *
1840              **************************/
1841
1842             /* REACTION-FIELD ELECTROSTATICS */
1843             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1844
1845             fscal            = felec;
1846
1847             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1848
1849             /* Calculate temporary vectorial force */
1850             tx               = _mm_mul_pd(fscal,dx23);
1851             ty               = _mm_mul_pd(fscal,dy23);
1852             tz               = _mm_mul_pd(fscal,dz23);
1853
1854             /* Update vectorial force */
1855             fix2             = _mm_add_pd(fix2,tx);
1856             fiy2             = _mm_add_pd(fiy2,ty);
1857             fiz2             = _mm_add_pd(fiz2,tz);
1858
1859             fjx3             = _mm_add_pd(fjx3,tx);
1860             fjy3             = _mm_add_pd(fjy3,ty);
1861             fjz3             = _mm_add_pd(fjz3,tz);
1862
1863             /**************************
1864              * CALCULATE INTERACTIONS *
1865              **************************/
1866
1867             /* REACTION-FIELD ELECTROSTATICS */
1868             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1869
1870             fscal            = felec;
1871
1872             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1873
1874             /* Calculate temporary vectorial force */
1875             tx               = _mm_mul_pd(fscal,dx31);
1876             ty               = _mm_mul_pd(fscal,dy31);
1877             tz               = _mm_mul_pd(fscal,dz31);
1878
1879             /* Update vectorial force */
1880             fix3             = _mm_add_pd(fix3,tx);
1881             fiy3             = _mm_add_pd(fiy3,ty);
1882             fiz3             = _mm_add_pd(fiz3,tz);
1883
1884             fjx1             = _mm_add_pd(fjx1,tx);
1885             fjy1             = _mm_add_pd(fjy1,ty);
1886             fjz1             = _mm_add_pd(fjz1,tz);
1887
1888             /**************************
1889              * CALCULATE INTERACTIONS *
1890              **************************/
1891
1892             /* REACTION-FIELD ELECTROSTATICS */
1893             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1894
1895             fscal            = felec;
1896
1897             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1898
1899             /* Calculate temporary vectorial force */
1900             tx               = _mm_mul_pd(fscal,dx32);
1901             ty               = _mm_mul_pd(fscal,dy32);
1902             tz               = _mm_mul_pd(fscal,dz32);
1903
1904             /* Update vectorial force */
1905             fix3             = _mm_add_pd(fix3,tx);
1906             fiy3             = _mm_add_pd(fiy3,ty);
1907             fiz3             = _mm_add_pd(fiz3,tz);
1908
1909             fjx2             = _mm_add_pd(fjx2,tx);
1910             fjy2             = _mm_add_pd(fjy2,ty);
1911             fjz2             = _mm_add_pd(fjz2,tz);
1912
1913             /**************************
1914              * CALCULATE INTERACTIONS *
1915              **************************/
1916
1917             /* REACTION-FIELD ELECTROSTATICS */
1918             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1919
1920             fscal            = felec;
1921
1922             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1923
1924             /* Calculate temporary vectorial force */
1925             tx               = _mm_mul_pd(fscal,dx33);
1926             ty               = _mm_mul_pd(fscal,dy33);
1927             tz               = _mm_mul_pd(fscal,dz33);
1928
1929             /* Update vectorial force */
1930             fix3             = _mm_add_pd(fix3,tx);
1931             fiy3             = _mm_add_pd(fiy3,ty);
1932             fiz3             = _mm_add_pd(fiz3,tz);
1933
1934             fjx3             = _mm_add_pd(fjx3,tx);
1935             fjy3             = _mm_add_pd(fjy3,ty);
1936             fjz3             = _mm_add_pd(fjz3,tz);
1937
1938             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1939
1940             /* Inner loop uses 294 flops */
1941         }
1942
1943         /* End of innermost loop */
1944
1945         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1946                                               f+i_coord_offset,fshift+i_shift_offset);
1947
1948         /* Increment number of inner iterations */
1949         inneriter                  += j_index_end - j_index_start;
1950
1951         /* Outer loop uses 24 flops */
1952     }
1953
1954     /* Increment number of outer iterations */
1955     outeriter        += nri;
1956
1957     /* Update outer/inner flops */
1958
1959     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*294);
1960 }