src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRF_VdwCSTab_GeomW4W4_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_single 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_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_sse2_single
  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_sse2_single
  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,C,D refer to j loop unrolling done with SSE, e.g. for the four 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,jnrC,jnrD;
  61     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             shX,shY,shZ,rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwioffset3;
  73     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  75     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  77     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  79     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  81     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  99     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 100     __m128i          vfitab;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 103     real             *vftab;
 104     __m128           dummy_mask,cutoff_mask;
 105     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 106     __m128           one     = _mm_set1_ps(1.0);
 107     __m128           two     = _mm_set1_ps(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_ps(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     krf              = _mm_set1_ps(fr->ic->k_rf);
 122     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 123     crf              = _mm_set1_ps(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_ps(kernel_data->table_vdw->scale);
 130
 131     /* Setup water-specific parameters */
 132     inr              = nlist->iinr[0];
 133     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 134     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 135     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 136     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 137
 138     jq1              = _mm_set1_ps(charge[inr+1]);
 139     jq2              = _mm_set1_ps(charge[inr+2]);
 140     jq3              = _mm_set1_ps(charge[inr+3]);
 141     vdwjidx0A        = 2*vdwtype[inr+0];
 142     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 143     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 144     qq11             = _mm_mul_ps(iq1,jq1);
 145     qq12             = _mm_mul_ps(iq1,jq2);
 146     qq13             = _mm_mul_ps(iq1,jq3);
 147     qq21             = _mm_mul_ps(iq2,jq1);
 148     qq22             = _mm_mul_ps(iq2,jq2);
 149     qq23             = _mm_mul_ps(iq2,jq3);
 150     qq31             = _mm_mul_ps(iq3,jq1);
 151     qq32             = _mm_mul_ps(iq3,jq2);
 152     qq33             = _mm_mul_ps(iq3,jq3);
 153
 154     /* Avoid stupid compiler warnings */
 155     jnrA = jnrB = jnrC = jnrD = 0;
 156     j_coord_offsetA = 0;
 157     j_coord_offsetB = 0;
 158     j_coord_offsetC = 0;
 159     j_coord_offsetD = 0;
 160
 161     outeriter        = 0;
 162     inneriter        = 0;
 163
 164     /* Start outer loop over neighborlists */
 165     for(iidx=0; iidx<nri; iidx++)
 166     {
 167         /* Load shift vector for this list */
 168         i_shift_offset   = DIM*shiftidx[iidx];
 169         shX              = shiftvec[i_shift_offset+XX];
 170         shY              = shiftvec[i_shift_offset+YY];
 171         shZ              = shiftvec[i_shift_offset+ZZ];
 172
 173         /* Load limits for loop over neighbors */
 174         j_index_start    = jindex[iidx];
 175         j_index_end      = jindex[iidx+1];
 176
 177         /* Get outer coordinate index */
 178         inr              = iinr[iidx];
 179         i_coord_offset   = DIM*inr;
 180
 181         /* Load i particle coords and add shift vector */
 182         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 183         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 184         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 185         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 186         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 187         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 188         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 189         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 190         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 191         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 192         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 193         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 194
 195         fix0             = _mm_setzero_ps();
 196         fiy0             = _mm_setzero_ps();
 197         fiz0             = _mm_setzero_ps();
 198         fix1             = _mm_setzero_ps();
 199         fiy1             = _mm_setzero_ps();
 200         fiz1             = _mm_setzero_ps();
 201         fix2             = _mm_setzero_ps();
 202         fiy2             = _mm_setzero_ps();
 203         fiz2             = _mm_setzero_ps();
 204         fix3             = _mm_setzero_ps();
 205         fiy3             = _mm_setzero_ps();
 206         fiz3             = _mm_setzero_ps();
 207
 208         /* Reset potential sums */
 209         velecsum         = _mm_setzero_ps();
 210         vvdwsum          = _mm_setzero_ps();
 211
 212         /* Start inner kernel loop */
 213         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 214         {
 215
 216             /* Get j neighbor index, and coordinate index */
 217             jnrA             = jjnr[jidx];
 218             jnrB             = jjnr[jidx+1];
 219             jnrC             = jjnr[jidx+2];
 220             jnrD             = jjnr[jidx+3];
 221
 222             j_coord_offsetA  = DIM*jnrA;
 223             j_coord_offsetB  = DIM*jnrB;
 224             j_coord_offsetC  = DIM*jnrC;
 225             j_coord_offsetD  = DIM*jnrD;
 226
 227             /* load j atom coordinates */
 228             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 229                                               x+j_coord_offsetC,x+j_coord_offsetD,
 230                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 231                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 232
 233             /* Calculate displacement vector */
 234             dx00             = _mm_sub_ps(ix0,jx0);
 235             dy00             = _mm_sub_ps(iy0,jy0);
 236             dz00             = _mm_sub_ps(iz0,jz0);
 237             dx11             = _mm_sub_ps(ix1,jx1);
 238             dy11             = _mm_sub_ps(iy1,jy1);
 239             dz11             = _mm_sub_ps(iz1,jz1);
 240             dx12             = _mm_sub_ps(ix1,jx2);
 241             dy12             = _mm_sub_ps(iy1,jy2);
 242             dz12             = _mm_sub_ps(iz1,jz2);
 243             dx13             = _mm_sub_ps(ix1,jx3);
 244             dy13             = _mm_sub_ps(iy1,jy3);
 245             dz13             = _mm_sub_ps(iz1,jz3);
 246             dx21             = _mm_sub_ps(ix2,jx1);
 247             dy21             = _mm_sub_ps(iy2,jy1);
 248             dz21             = _mm_sub_ps(iz2,jz1);
 249             dx22             = _mm_sub_ps(ix2,jx2);
 250             dy22             = _mm_sub_ps(iy2,jy2);
 251             dz22             = _mm_sub_ps(iz2,jz2);
 252             dx23             = _mm_sub_ps(ix2,jx3);
 253             dy23             = _mm_sub_ps(iy2,jy3);
 254             dz23             = _mm_sub_ps(iz2,jz3);
 255             dx31             = _mm_sub_ps(ix3,jx1);
 256             dy31             = _mm_sub_ps(iy3,jy1);
 257             dz31             = _mm_sub_ps(iz3,jz1);
 258             dx32             = _mm_sub_ps(ix3,jx2);
 259             dy32             = _mm_sub_ps(iy3,jy2);
 260             dz32             = _mm_sub_ps(iz3,jz2);
 261             dx33             = _mm_sub_ps(ix3,jx3);
 262             dy33             = _mm_sub_ps(iy3,jy3);
 263             dz33             = _mm_sub_ps(iz3,jz3);
 264
 265             /* Calculate squared distance and things based on it */
 266             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 267             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 268             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 269             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 270             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 271             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 272             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 273             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 274             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 275             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 276
 277             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 278             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 279             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 280             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 281             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 282             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 283             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 284             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 285             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 286             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 287
 288             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 289             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 290             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 291             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 292             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 293             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 294             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 295             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 296             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 297
 298             fjx0             = _mm_setzero_ps();
 299             fjy0             = _mm_setzero_ps();
 300             fjz0             = _mm_setzero_ps();
 301             fjx1             = _mm_setzero_ps();
 302             fjy1             = _mm_setzero_ps();
 303             fjz1             = _mm_setzero_ps();
 304             fjx2             = _mm_setzero_ps();
 305             fjy2             = _mm_setzero_ps();
 306             fjz2             = _mm_setzero_ps();
 307             fjx3             = _mm_setzero_ps();
 308             fjy3             = _mm_setzero_ps();
 309             fjz3             = _mm_setzero_ps();
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             r00              = _mm_mul_ps(rsq00,rinv00);
 316
 317             /* Calculate table index by multiplying r with table scale and truncate to integer */
 318             rt               = _mm_mul_ps(r00,vftabscale);
 319             vfitab           = _mm_cvttps_epi32(rt);
 320             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 321             vfitab           = _mm_slli_epi32(vfitab,3);
 322
 323             /* CUBIC SPLINE TABLE DISPERSION */
 324             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 325             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 326             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 327             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 328             _MM_TRANSPOSE4_PS(Y,F,G,H);
 329             Heps             = _mm_mul_ps(vfeps,H);
 330             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 331             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 332             vvdw6            = _mm_mul_ps(c6_00,VV);
 333             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 334             fvdw6            = _mm_mul_ps(c6_00,FF);
 335
 336             /* CUBIC SPLINE TABLE REPULSION */
 337             vfitab           = _mm_add_epi32(vfitab,ifour);
 338             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 339             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 340             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 341             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 342             _MM_TRANSPOSE4_PS(Y,F,G,H);
 343             Heps             = _mm_mul_ps(vfeps,H);
 344             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 345             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 346             vvdw12           = _mm_mul_ps(c12_00,VV);
 347             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 348             fvdw12           = _mm_mul_ps(c12_00,FF);
 349             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 350             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 354
 355             fscal            = fvdw;
 356
 357             /* Calculate temporary vectorial force */
 358             tx               = _mm_mul_ps(fscal,dx00);
 359             ty               = _mm_mul_ps(fscal,dy00);
 360             tz               = _mm_mul_ps(fscal,dz00);
 361
 362             /* Update vectorial force */
 363             fix0             = _mm_add_ps(fix0,tx);
 364             fiy0             = _mm_add_ps(fiy0,ty);
 365             fiz0             = _mm_add_ps(fiz0,tz);
 366
 367             fjx0             = _mm_add_ps(fjx0,tx);
 368             fjy0             = _mm_add_ps(fjy0,ty);
 369             fjz0             = _mm_add_ps(fjz0,tz);
 370
 371             /**************************
 372              * CALCULATE INTERACTIONS *
 373              **************************/
 374
 375             /* REACTION-FIELD ELECTROSTATICS */
 376             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 377             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velecsum         = _mm_add_ps(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             /* Calculate temporary vectorial force */
 385             tx               = _mm_mul_ps(fscal,dx11);
 386             ty               = _mm_mul_ps(fscal,dy11);
 387             tz               = _mm_mul_ps(fscal,dz11);
 388
 389             /* Update vectorial force */
 390             fix1             = _mm_add_ps(fix1,tx);
 391             fiy1             = _mm_add_ps(fiy1,ty);
 392             fiz1             = _mm_add_ps(fiz1,tz);
 393
 394             fjx1             = _mm_add_ps(fjx1,tx);
 395             fjy1             = _mm_add_ps(fjy1,ty);
 396             fjz1             = _mm_add_ps(fjz1,tz);
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             /* REACTION-FIELD ELECTROSTATICS */
 403             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 404             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velecsum         = _mm_add_ps(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411             /* Calculate temporary vectorial force */
 412             tx               = _mm_mul_ps(fscal,dx12);
 413             ty               = _mm_mul_ps(fscal,dy12);
 414             tz               = _mm_mul_ps(fscal,dz12);
 415
 416             /* Update vectorial force */
 417             fix1             = _mm_add_ps(fix1,tx);
 418             fiy1             = _mm_add_ps(fiy1,ty);
 419             fiz1             = _mm_add_ps(fiz1,tz);
 420
 421             fjx2             = _mm_add_ps(fjx2,tx);
 422             fjy2             = _mm_add_ps(fjy2,ty);
 423             fjz2             = _mm_add_ps(fjz2,tz);
 424
 425             /**************************
 426              * CALCULATE INTERACTIONS *
 427              **************************/
 428
 429             /* REACTION-FIELD ELECTROSTATICS */
 430             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 431             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velecsum         = _mm_add_ps(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             /* Calculate temporary vectorial force */
 439             tx               = _mm_mul_ps(fscal,dx13);
 440             ty               = _mm_mul_ps(fscal,dy13);
 441             tz               = _mm_mul_ps(fscal,dz13);
 442
 443             /* Update vectorial force */
 444             fix1             = _mm_add_ps(fix1,tx);
 445             fiy1             = _mm_add_ps(fiy1,ty);
 446             fiz1             = _mm_add_ps(fiz1,tz);
 447
 448             fjx3             = _mm_add_ps(fjx3,tx);
 449             fjy3             = _mm_add_ps(fjy3,ty);
 450             fjz3             = _mm_add_ps(fjz3,tz);
 451
 452             /**************************
 453              * CALCULATE INTERACTIONS *
 454              **************************/
 455
 456             /* REACTION-FIELD ELECTROSTATICS */
 457             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 458             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velecsum         = _mm_add_ps(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             /* Calculate temporary vectorial force */
 466             tx               = _mm_mul_ps(fscal,dx21);
 467             ty               = _mm_mul_ps(fscal,dy21);
 468             tz               = _mm_mul_ps(fscal,dz21);
 469
 470             /* Update vectorial force */
 471             fix2             = _mm_add_ps(fix2,tx);
 472             fiy2             = _mm_add_ps(fiy2,ty);
 473             fiz2             = _mm_add_ps(fiz2,tz);
 474
 475             fjx1             = _mm_add_ps(fjx1,tx);
 476             fjy1             = _mm_add_ps(fjy1,ty);
 477             fjz1             = _mm_add_ps(fjz1,tz);
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             /* REACTION-FIELD ELECTROSTATICS */
 484             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 485             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 486
 487             /* Update potential sum for this i atom from the interaction with this j atom. */
 488             velecsum         = _mm_add_ps(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492             /* Calculate temporary vectorial force */
 493             tx               = _mm_mul_ps(fscal,dx22);
 494             ty               = _mm_mul_ps(fscal,dy22);
 495             tz               = _mm_mul_ps(fscal,dz22);
 496
 497             /* Update vectorial force */
 498             fix2             = _mm_add_ps(fix2,tx);
 499             fiy2             = _mm_add_ps(fiy2,ty);
 500             fiz2             = _mm_add_ps(fiz2,tz);
 501
 502             fjx2             = _mm_add_ps(fjx2,tx);
 503             fjy2             = _mm_add_ps(fjy2,ty);
 504             fjz2             = _mm_add_ps(fjz2,tz);
 505
 506             /**************************
 507              * CALCULATE INTERACTIONS *
 508              **************************/
 509
 510             /* REACTION-FIELD ELECTROSTATICS */
 511             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 512             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 513
 514             /* Update potential sum for this i atom from the interaction with this j atom. */
 515             velecsum         = _mm_add_ps(velecsum,velec);
 516
 517             fscal            = felec;
 518
 519             /* Calculate temporary vectorial force */
 520             tx               = _mm_mul_ps(fscal,dx23);
 521             ty               = _mm_mul_ps(fscal,dy23);
 522             tz               = _mm_mul_ps(fscal,dz23);
 523
 524             /* Update vectorial force */
 525             fix2             = _mm_add_ps(fix2,tx);
 526             fiy2             = _mm_add_ps(fiy2,ty);
 527             fiz2             = _mm_add_ps(fiz2,tz);
 528
 529             fjx3             = _mm_add_ps(fjx3,tx);
 530             fjy3             = _mm_add_ps(fjy3,ty);
 531             fjz3             = _mm_add_ps(fjz3,tz);
 532
 533             /**************************
 534              * CALCULATE INTERACTIONS *
 535              **************************/
 536
 537             /* REACTION-FIELD ELECTROSTATICS */
 538             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 539             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 540
 541             /* Update potential sum for this i atom from the interaction with this j atom. */
 542             velecsum         = _mm_add_ps(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546             /* Calculate temporary vectorial force */
 547             tx               = _mm_mul_ps(fscal,dx31);
 548             ty               = _mm_mul_ps(fscal,dy31);
 549             tz               = _mm_mul_ps(fscal,dz31);
 550
 551             /* Update vectorial force */
 552             fix3             = _mm_add_ps(fix3,tx);
 553             fiy3             = _mm_add_ps(fiy3,ty);
 554             fiz3             = _mm_add_ps(fiz3,tz);
 555
 556             fjx1             = _mm_add_ps(fjx1,tx);
 557             fjy1             = _mm_add_ps(fjy1,ty);
 558             fjz1             = _mm_add_ps(fjz1,tz);
 559
 560             /**************************
 561              * CALCULATE INTERACTIONS *
 562              **************************/
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 566             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 567
 568             /* Update potential sum for this i atom from the interaction with this j atom. */
 569             velecsum         = _mm_add_ps(velecsum,velec);
 570
 571             fscal            = felec;
 572
 573             /* Calculate temporary vectorial force */
 574             tx               = _mm_mul_ps(fscal,dx32);
 575             ty               = _mm_mul_ps(fscal,dy32);
 576             tz               = _mm_mul_ps(fscal,dz32);
 577
 578             /* Update vectorial force */
 579             fix3             = _mm_add_ps(fix3,tx);
 580             fiy3             = _mm_add_ps(fiy3,ty);
 581             fiz3             = _mm_add_ps(fiz3,tz);
 582
 583             fjx2             = _mm_add_ps(fjx2,tx);
 584             fjy2             = _mm_add_ps(fjy2,ty);
 585             fjz2             = _mm_add_ps(fjz2,tz);
 586
 587             /**************************
 588              * CALCULATE INTERACTIONS *
 589              **************************/
 590
 591             /* REACTION-FIELD ELECTROSTATICS */
 592             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 593             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 594
 595             /* Update potential sum for this i atom from the interaction with this j atom. */
 596             velecsum         = _mm_add_ps(velecsum,velec);
 597
 598             fscal            = felec;
 599
 600             /* Calculate temporary vectorial force */
 601             tx               = _mm_mul_ps(fscal,dx33);
 602             ty               = _mm_mul_ps(fscal,dy33);
 603             tz               = _mm_mul_ps(fscal,dz33);
 604
 605             /* Update vectorial force */
 606             fix3             = _mm_add_ps(fix3,tx);
 607             fiy3             = _mm_add_ps(fiy3,ty);
 608             fiz3             = _mm_add_ps(fiz3,tz);
 609
 610             fjx3             = _mm_add_ps(fjx3,tx);
 611             fjy3             = _mm_add_ps(fjy3,ty);
 612             fjz3             = _mm_add_ps(fjz3,tz);
 613
 614             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 615                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 616                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 617                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 618
 619             /* Inner loop uses 347 flops */
 620         }
 621
 622         if(jidx<j_index_end)
 623         {
 624
 625             /* Get j neighbor index, and coordinate index */
 626             jnrA             = jjnr[jidx];
 627             jnrB             = jjnr[jidx+1];
 628             jnrC             = jjnr[jidx+2];
 629             jnrD             = jjnr[jidx+3];
 630
 631             /* Sign of each element will be negative for non-real atoms.
 632              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 633              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 634              */
 635             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 636             jnrA       = (jnrA>=0) ? jnrA : 0;
 637             jnrB       = (jnrB>=0) ? jnrB : 0;
 638             jnrC       = (jnrC>=0) ? jnrC : 0;
 639             jnrD       = (jnrD>=0) ? jnrD : 0;
 640
 641             j_coord_offsetA  = DIM*jnrA;
 642             j_coord_offsetB  = DIM*jnrB;
 643             j_coord_offsetC  = DIM*jnrC;
 644             j_coord_offsetD  = DIM*jnrD;
 645
 646             /* load j atom coordinates */
 647             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 648                                               x+j_coord_offsetC,x+j_coord_offsetD,
 649                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 650                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 651
 652             /* Calculate displacement vector */
 653             dx00             = _mm_sub_ps(ix0,jx0);
 654             dy00             = _mm_sub_ps(iy0,jy0);
 655             dz00             = _mm_sub_ps(iz0,jz0);
 656             dx11             = _mm_sub_ps(ix1,jx1);
 657             dy11             = _mm_sub_ps(iy1,jy1);
 658             dz11             = _mm_sub_ps(iz1,jz1);
 659             dx12             = _mm_sub_ps(ix1,jx2);
 660             dy12             = _mm_sub_ps(iy1,jy2);
 661             dz12             = _mm_sub_ps(iz1,jz2);
 662             dx13             = _mm_sub_ps(ix1,jx3);
 663             dy13             = _mm_sub_ps(iy1,jy3);
 664             dz13             = _mm_sub_ps(iz1,jz3);
 665             dx21             = _mm_sub_ps(ix2,jx1);
 666             dy21             = _mm_sub_ps(iy2,jy1);
 667             dz21             = _mm_sub_ps(iz2,jz1);
 668             dx22             = _mm_sub_ps(ix2,jx2);
 669             dy22             = _mm_sub_ps(iy2,jy2);
 670             dz22             = _mm_sub_ps(iz2,jz2);
 671             dx23             = _mm_sub_ps(ix2,jx3);
 672             dy23             = _mm_sub_ps(iy2,jy3);
 673             dz23             = _mm_sub_ps(iz2,jz3);
 674             dx31             = _mm_sub_ps(ix3,jx1);
 675             dy31             = _mm_sub_ps(iy3,jy1);
 676             dz31             = _mm_sub_ps(iz3,jz1);
 677             dx32             = _mm_sub_ps(ix3,jx2);
 678             dy32             = _mm_sub_ps(iy3,jy2);
 679             dz32             = _mm_sub_ps(iz3,jz2);
 680             dx33             = _mm_sub_ps(ix3,jx3);
 681             dy33             = _mm_sub_ps(iy3,jy3);
 682             dz33             = _mm_sub_ps(iz3,jz3);
 683
 684             /* Calculate squared distance and things based on it */
 685             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 686             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 687             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 688             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 689             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 690             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 691             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 692             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 693             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 694             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 695
 696             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 697             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 698             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 699             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 700             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 701             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 702             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 703             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 704             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 705             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 706
 707             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 708             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 709             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 710             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 711             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 712             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 713             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 714             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 715             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 716
 717             fjx0             = _mm_setzero_ps();
 718             fjy0             = _mm_setzero_ps();
 719             fjz0             = _mm_setzero_ps();
 720             fjx1             = _mm_setzero_ps();
 721             fjy1             = _mm_setzero_ps();
 722             fjz1             = _mm_setzero_ps();
 723             fjx2             = _mm_setzero_ps();
 724             fjy2             = _mm_setzero_ps();
 725             fjz2             = _mm_setzero_ps();
 726             fjx3             = _mm_setzero_ps();
 727             fjy3             = _mm_setzero_ps();
 728             fjz3             = _mm_setzero_ps();
 729
 730             /**************************
 731              * CALCULATE INTERACTIONS *
 732              **************************/
 733
 734             r00              = _mm_mul_ps(rsq00,rinv00);
 735             r00              = _mm_andnot_ps(dummy_mask,r00);
 736
 737             /* Calculate table index by multiplying r with table scale and truncate to integer */
 738             rt               = _mm_mul_ps(r00,vftabscale);
 739             vfitab           = _mm_cvttps_epi32(rt);
 740             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 741             vfitab           = _mm_slli_epi32(vfitab,3);
 742
 743             /* CUBIC SPLINE TABLE DISPERSION */
 744             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 745             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 746             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 747             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 748             _MM_TRANSPOSE4_PS(Y,F,G,H);
 749             Heps             = _mm_mul_ps(vfeps,H);
 750             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 751             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 752             vvdw6            = _mm_mul_ps(c6_00,VV);
 753             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 754             fvdw6            = _mm_mul_ps(c6_00,FF);
 755
 756             /* CUBIC SPLINE TABLE REPULSION */
 757             vfitab           = _mm_add_epi32(vfitab,ifour);
 758             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 759             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 760             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 761             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 762             _MM_TRANSPOSE4_PS(Y,F,G,H);
 763             Heps             = _mm_mul_ps(vfeps,H);
 764             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 765             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 766             vvdw12           = _mm_mul_ps(c12_00,VV);
 767             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 768             fvdw12           = _mm_mul_ps(c12_00,FF);
 769             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 770             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 771
 772             /* Update potential sum for this i atom from the interaction with this j atom. */
 773             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 774             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 775
 776             fscal            = fvdw;
 777
 778             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 779
 780             /* Calculate temporary vectorial force */
 781             tx               = _mm_mul_ps(fscal,dx00);
 782             ty               = _mm_mul_ps(fscal,dy00);
 783             tz               = _mm_mul_ps(fscal,dz00);
 784
 785             /* Update vectorial force */
 786             fix0             = _mm_add_ps(fix0,tx);
 787             fiy0             = _mm_add_ps(fiy0,ty);
 788             fiz0             = _mm_add_ps(fiz0,tz);
 789
 790             fjx0             = _mm_add_ps(fjx0,tx);
 791             fjy0             = _mm_add_ps(fjy0,ty);
 792             fjz0             = _mm_add_ps(fjz0,tz);
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 800             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_andnot_ps(dummy_mask,velec);
 804             velecsum         = _mm_add_ps(velecsum,velec);
 805
 806             fscal            = felec;
 807
 808             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 809
 810             /* Calculate temporary vectorial force */
 811             tx               = _mm_mul_ps(fscal,dx11);
 812             ty               = _mm_mul_ps(fscal,dy11);
 813             tz               = _mm_mul_ps(fscal,dz11);
 814
 815             /* Update vectorial force */
 816             fix1             = _mm_add_ps(fix1,tx);
 817             fiy1             = _mm_add_ps(fiy1,ty);
 818             fiz1             = _mm_add_ps(fiz1,tz);
 819
 820             fjx1             = _mm_add_ps(fjx1,tx);
 821             fjy1             = _mm_add_ps(fjy1,ty);
 822             fjz1             = _mm_add_ps(fjz1,tz);
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             /* REACTION-FIELD ELECTROSTATICS */
 829             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 830             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 831
 832             /* Update potential sum for this i atom from the interaction with this j atom. */
 833             velec            = _mm_andnot_ps(dummy_mask,velec);
 834             velecsum         = _mm_add_ps(velecsum,velec);
 835
 836             fscal            = felec;
 837
 838             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 839
 840             /* Calculate temporary vectorial force */
 841             tx               = _mm_mul_ps(fscal,dx12);
 842             ty               = _mm_mul_ps(fscal,dy12);
 843             tz               = _mm_mul_ps(fscal,dz12);
 844
 845             /* Update vectorial force */
 846             fix1             = _mm_add_ps(fix1,tx);
 847             fiy1             = _mm_add_ps(fiy1,ty);
 848             fiz1             = _mm_add_ps(fiz1,tz);
 849
 850             fjx2             = _mm_add_ps(fjx2,tx);
 851             fjy2             = _mm_add_ps(fjy2,ty);
 852             fjz2             = _mm_add_ps(fjz2,tz);
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             /* REACTION-FIELD ELECTROSTATICS */
 859             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 860             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 861
 862             /* Update potential sum for this i atom from the interaction with this j atom. */
 863             velec            = _mm_andnot_ps(dummy_mask,velec);
 864             velecsum         = _mm_add_ps(velecsum,velec);
 865
 866             fscal            = felec;
 867
 868             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 869
 870             /* Calculate temporary vectorial force */
 871             tx               = _mm_mul_ps(fscal,dx13);
 872             ty               = _mm_mul_ps(fscal,dy13);
 873             tz               = _mm_mul_ps(fscal,dz13);
 874
 875             /* Update vectorial force */
 876             fix1             = _mm_add_ps(fix1,tx);
 877             fiy1             = _mm_add_ps(fiy1,ty);
 878             fiz1             = _mm_add_ps(fiz1,tz);
 879
 880             fjx3             = _mm_add_ps(fjx3,tx);
 881             fjy3             = _mm_add_ps(fjy3,ty);
 882             fjz3             = _mm_add_ps(fjz3,tz);
 883
 884             /**************************
 885              * CALCULATE INTERACTIONS *
 886              **************************/
 887
 888             /* REACTION-FIELD ELECTROSTATICS */
 889             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 890             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 891
 892             /* Update potential sum for this i atom from the interaction with this j atom. */
 893             velec            = _mm_andnot_ps(dummy_mask,velec);
 894             velecsum         = _mm_add_ps(velecsum,velec);
 895
 896             fscal            = felec;
 897
 898             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 899
 900             /* Calculate temporary vectorial force */
 901             tx               = _mm_mul_ps(fscal,dx21);
 902             ty               = _mm_mul_ps(fscal,dy21);
 903             tz               = _mm_mul_ps(fscal,dz21);
 904
 905             /* Update vectorial force */
 906             fix2             = _mm_add_ps(fix2,tx);
 907             fiy2             = _mm_add_ps(fiy2,ty);
 908             fiz2             = _mm_add_ps(fiz2,tz);
 909
 910             fjx1             = _mm_add_ps(fjx1,tx);
 911             fjy1             = _mm_add_ps(fjy1,ty);
 912             fjz1             = _mm_add_ps(fjz1,tz);
 913
 914             /**************************
 915              * CALCULATE INTERACTIONS *
 916              **************************/
 917
 918             /* REACTION-FIELD ELECTROSTATICS */
 919             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 920             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 921
 922             /* Update potential sum for this i atom from the interaction with this j atom. */
 923             velec            = _mm_andnot_ps(dummy_mask,velec);
 924             velecsum         = _mm_add_ps(velecsum,velec);
 925
 926             fscal            = felec;
 927
 928             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 929
 930             /* Calculate temporary vectorial force */
 931             tx               = _mm_mul_ps(fscal,dx22);
 932             ty               = _mm_mul_ps(fscal,dy22);
 933             tz               = _mm_mul_ps(fscal,dz22);
 934
 935             /* Update vectorial force */
 936             fix2             = _mm_add_ps(fix2,tx);
 937             fiy2             = _mm_add_ps(fiy2,ty);
 938             fiz2             = _mm_add_ps(fiz2,tz);
 939
 940             fjx2             = _mm_add_ps(fjx2,tx);
 941             fjy2             = _mm_add_ps(fjy2,ty);
 942             fjz2             = _mm_add_ps(fjz2,tz);
 943
 944             /**************************
 945              * CALCULATE INTERACTIONS *
 946              **************************/
 947
 948             /* REACTION-FIELD ELECTROSTATICS */
 949             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 950             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 951
 952             /* Update potential sum for this i atom from the interaction with this j atom. */
 953             velec            = _mm_andnot_ps(dummy_mask,velec);
 954             velecsum         = _mm_add_ps(velecsum,velec);
 955
 956             fscal            = felec;
 957
 958             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 959
 960             /* Calculate temporary vectorial force */
 961             tx               = _mm_mul_ps(fscal,dx23);
 962             ty               = _mm_mul_ps(fscal,dy23);
 963             tz               = _mm_mul_ps(fscal,dz23);
 964
 965             /* Update vectorial force */
 966             fix2             = _mm_add_ps(fix2,tx);
 967             fiy2             = _mm_add_ps(fiy2,ty);
 968             fiz2             = _mm_add_ps(fiz2,tz);
 969
 970             fjx3             = _mm_add_ps(fjx3,tx);
 971             fjy3             = _mm_add_ps(fjy3,ty);
 972             fjz3             = _mm_add_ps(fjz3,tz);
 973
 974             /**************************
 975              * CALCULATE INTERACTIONS *
 976              **************************/
 977
 978             /* REACTION-FIELD ELECTROSTATICS */
 979             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 980             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 981
 982             /* Update potential sum for this i atom from the interaction with this j atom. */
 983             velec            = _mm_andnot_ps(dummy_mask,velec);
 984             velecsum         = _mm_add_ps(velecsum,velec);
 985
 986             fscal            = felec;
 987
 988             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 989
 990             /* Calculate temporary vectorial force */
 991             tx               = _mm_mul_ps(fscal,dx31);
 992             ty               = _mm_mul_ps(fscal,dy31);
 993             tz               = _mm_mul_ps(fscal,dz31);
 994
 995             /* Update vectorial force */
 996             fix3             = _mm_add_ps(fix3,tx);
 997             fiy3             = _mm_add_ps(fiy3,ty);
 998             fiz3             = _mm_add_ps(fiz3,tz);
 999
1000             fjx1             = _mm_add_ps(fjx1,tx);
1001             fjy1             = _mm_add_ps(fjy1,ty);
1002             fjz1             = _mm_add_ps(fjz1,tz);
1003
1004             /**************************
1005              * CALCULATE INTERACTIONS *
1006              **************************/
1007
1008             /* REACTION-FIELD ELECTROSTATICS */
1009             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1010             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1011
1012             /* Update potential sum for this i atom from the interaction with this j atom. */
1013             velec            = _mm_andnot_ps(dummy_mask,velec);
1014             velecsum         = _mm_add_ps(velecsum,velec);
1015
1016             fscal            = felec;
1017
1018             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1019
1020             /* Calculate temporary vectorial force */
1021             tx               = _mm_mul_ps(fscal,dx32);
1022             ty               = _mm_mul_ps(fscal,dy32);
1023             tz               = _mm_mul_ps(fscal,dz32);
1024
1025             /* Update vectorial force */
1026             fix3             = _mm_add_ps(fix3,tx);
1027             fiy3             = _mm_add_ps(fiy3,ty);
1028             fiz3             = _mm_add_ps(fiz3,tz);
1029
1030             fjx2             = _mm_add_ps(fjx2,tx);
1031             fjy2             = _mm_add_ps(fjy2,ty);
1032             fjz2             = _mm_add_ps(fjz2,tz);
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             /* REACTION-FIELD ELECTROSTATICS */
1039             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1040             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1041
1042             /* Update potential sum for this i atom from the interaction with this j atom. */
1043             velec            = _mm_andnot_ps(dummy_mask,velec);
1044             velecsum         = _mm_add_ps(velecsum,velec);
1045
1046             fscal            = felec;
1047
1048             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1049
1050             /* Calculate temporary vectorial force */
1051             tx               = _mm_mul_ps(fscal,dx33);
1052             ty               = _mm_mul_ps(fscal,dy33);
1053             tz               = _mm_mul_ps(fscal,dz33);
1054
1055             /* Update vectorial force */
1056             fix3             = _mm_add_ps(fix3,tx);
1057             fiy3             = _mm_add_ps(fiy3,ty);
1058             fiz3             = _mm_add_ps(fiz3,tz);
1059
1060             fjx3             = _mm_add_ps(fjx3,tx);
1061             fjy3             = _mm_add_ps(fjy3,ty);
1062             fjz3             = _mm_add_ps(fjz3,tz);
1063
1064             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1065                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1066                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1067                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1068
1069             /* Inner loop uses 348 flops */
1070         }
1071
1072         /* End of innermost loop */
1073
1074         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1075                                               f+i_coord_offset,fshift+i_shift_offset);
1076
1077         ggid                        = gid[iidx];
1078         /* Update potential energies */
1079         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1080         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1081
1082         /* Increment number of inner iterations */
1083         inneriter                  += j_index_end - j_index_start;
1084
1085         /* Outer loop uses 38 flops */
1086     }
1087
1088     /* Increment number of outer iterations */
1089     outeriter        += nri;
1090
1091     /* Update outer/inner flops */
1092
1093     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*38 + inneriter*348);
1094 }
1095 /*
1096  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse2_single
1097  * Electrostatics interaction: ReactionField
1098  * VdW interaction:            CubicSplineTable
1099  * Geometry:                   Water4-Water4
1100  * Calculate force/pot:        Force
1101  */
1102 void
1103 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse2_single
1104                     (t_nblist * gmx_restrict                nlist,
1105                      rvec * gmx_restrict                    xx,
1106                      rvec * gmx_restrict                    ff,
1107                      t_forcerec * gmx_restrict              fr,
1108                      t_mdatoms * gmx_restrict               mdatoms,
1109                      nb_kernel_data_t * gmx_restrict        kernel_data,
1110                      t_nrnb * gmx_restrict                  nrnb)
1111 {
1112     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1113      * just 0 for non-waters.
1114      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1115      * jnr indices corresponding to data put in the four positions in the SIMD register.
1116      */
1117     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1118     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1119     int              jnrA,jnrB,jnrC,jnrD;
1120     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1121     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1122     real             shX,shY,shZ,rcutoff_scalar;
1123     real             *shiftvec,*fshift,*x,*f;
1124     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1125     int              vdwioffset0;
1126     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1127     int              vdwioffset1;
1128     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1129     int              vdwioffset2;
1130     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1131     int              vdwioffset3;
1132     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1133     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1134     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1135     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1136     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1137     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1138     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1139     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1140     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1141     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1142     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1143     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1144     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1145     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1146     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1147     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1148     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1149     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1150     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1151     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1152     real             *charge;
1153     int              nvdwtype;
1154     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1155     int              *vdwtype;
1156     real             *vdwparam;
1157     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1158     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1159     __m128i          vfitab;
1160     __m128i          ifour       = _mm_set1_epi32(4);
1161     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1162     real             *vftab;
1163     __m128           dummy_mask,cutoff_mask;
1164     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1165     __m128           one     = _mm_set1_ps(1.0);
1166     __m128           two     = _mm_set1_ps(2.0);
1167     x                = xx[0];
1168     f                = ff[0];
1169
1170     nri              = nlist->nri;
1171     iinr             = nlist->iinr;
1172     jindex           = nlist->jindex;
1173     jjnr             = nlist->jjnr;
1174     shiftidx         = nlist->shift;
1175     gid              = nlist->gid;
1176     shiftvec         = fr->shift_vec[0];
1177     fshift           = fr->fshift[0];
1178     facel            = _mm_set1_ps(fr->epsfac);
1179     charge           = mdatoms->chargeA;
1180     krf              = _mm_set1_ps(fr->ic->k_rf);
1181     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1182     crf              = _mm_set1_ps(fr->ic->c_rf);
1183     nvdwtype         = fr->ntype;
1184     vdwparam         = fr->nbfp;
1185     vdwtype          = mdatoms->typeA;
1186
1187     vftab            = kernel_data->table_vdw->data;
1188     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1189
1190     /* Setup water-specific parameters */
1191     inr              = nlist->iinr[0];
1192     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1193     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1194     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1195     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1196
1197     jq1              = _mm_set1_ps(charge[inr+1]);
1198     jq2              = _mm_set1_ps(charge[inr+2]);
1199     jq3              = _mm_set1_ps(charge[inr+3]);
1200     vdwjidx0A        = 2*vdwtype[inr+0];
1201     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1202     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1203     qq11             = _mm_mul_ps(iq1,jq1);
1204     qq12             = _mm_mul_ps(iq1,jq2);
1205     qq13             = _mm_mul_ps(iq1,jq3);
1206     qq21             = _mm_mul_ps(iq2,jq1);
1207     qq22             = _mm_mul_ps(iq2,jq2);
1208     qq23             = _mm_mul_ps(iq2,jq3);
1209     qq31             = _mm_mul_ps(iq3,jq1);
1210     qq32             = _mm_mul_ps(iq3,jq2);
1211     qq33             = _mm_mul_ps(iq3,jq3);
1212
1213     /* Avoid stupid compiler warnings */
1214     jnrA = jnrB = jnrC = jnrD = 0;
1215     j_coord_offsetA = 0;
1216     j_coord_offsetB = 0;
1217     j_coord_offsetC = 0;
1218     j_coord_offsetD = 0;
1219
1220     outeriter        = 0;
1221     inneriter        = 0;
1222
1223     /* Start outer loop over neighborlists */
1224     for(iidx=0; iidx<nri; iidx++)
1225     {
1226         /* Load shift vector for this list */
1227         i_shift_offset   = DIM*shiftidx[iidx];
1228         shX              = shiftvec[i_shift_offset+XX];
1229         shY              = shiftvec[i_shift_offset+YY];
1230         shZ              = shiftvec[i_shift_offset+ZZ];
1231
1232         /* Load limits for loop over neighbors */
1233         j_index_start    = jindex[iidx];
1234         j_index_end      = jindex[iidx+1];
1235
1236         /* Get outer coordinate index */
1237         inr              = iinr[iidx];
1238         i_coord_offset   = DIM*inr;
1239
1240         /* Load i particle coords and add shift vector */
1241         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1242         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1243         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1244         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1245         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1246         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1247         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1248         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1249         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1250         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1251         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1252         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1253
1254         fix0             = _mm_setzero_ps();
1255         fiy0             = _mm_setzero_ps();
1256         fiz0             = _mm_setzero_ps();
1257         fix1             = _mm_setzero_ps();
1258         fiy1             = _mm_setzero_ps();
1259         fiz1             = _mm_setzero_ps();
1260         fix2             = _mm_setzero_ps();
1261         fiy2             = _mm_setzero_ps();
1262         fiz2             = _mm_setzero_ps();
1263         fix3             = _mm_setzero_ps();
1264         fiy3             = _mm_setzero_ps();
1265         fiz3             = _mm_setzero_ps();
1266
1267         /* Start inner kernel loop */
1268         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1269         {
1270
1271             /* Get j neighbor index, and coordinate index */
1272             jnrA             = jjnr[jidx];
1273             jnrB             = jjnr[jidx+1];
1274             jnrC             = jjnr[jidx+2];
1275             jnrD             = jjnr[jidx+3];
1276
1277             j_coord_offsetA  = DIM*jnrA;
1278             j_coord_offsetB  = DIM*jnrB;
1279             j_coord_offsetC  = DIM*jnrC;
1280             j_coord_offsetD  = DIM*jnrD;
1281
1282             /* load j atom coordinates */
1283             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1284                                               x+j_coord_offsetC,x+j_coord_offsetD,
1285                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1286                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1287
1288             /* Calculate displacement vector */
1289             dx00             = _mm_sub_ps(ix0,jx0);
1290             dy00             = _mm_sub_ps(iy0,jy0);
1291             dz00             = _mm_sub_ps(iz0,jz0);
1292             dx11             = _mm_sub_ps(ix1,jx1);
1293             dy11             = _mm_sub_ps(iy1,jy1);
1294             dz11             = _mm_sub_ps(iz1,jz1);
1295             dx12             = _mm_sub_ps(ix1,jx2);
1296             dy12             = _mm_sub_ps(iy1,jy2);
1297             dz12             = _mm_sub_ps(iz1,jz2);
1298             dx13             = _mm_sub_ps(ix1,jx3);
1299             dy13             = _mm_sub_ps(iy1,jy3);
1300             dz13             = _mm_sub_ps(iz1,jz3);
1301             dx21             = _mm_sub_ps(ix2,jx1);
1302             dy21             = _mm_sub_ps(iy2,jy1);
1303             dz21             = _mm_sub_ps(iz2,jz1);
1304             dx22             = _mm_sub_ps(ix2,jx2);
1305             dy22             = _mm_sub_ps(iy2,jy2);
1306             dz22             = _mm_sub_ps(iz2,jz2);
1307             dx23             = _mm_sub_ps(ix2,jx3);
1308             dy23             = _mm_sub_ps(iy2,jy3);
1309             dz23             = _mm_sub_ps(iz2,jz3);
1310             dx31             = _mm_sub_ps(ix3,jx1);
1311             dy31             = _mm_sub_ps(iy3,jy1);
1312             dz31             = _mm_sub_ps(iz3,jz1);
1313             dx32             = _mm_sub_ps(ix3,jx2);
1314             dy32             = _mm_sub_ps(iy3,jy2);
1315             dz32             = _mm_sub_ps(iz3,jz2);
1316             dx33             = _mm_sub_ps(ix3,jx3);
1317             dy33             = _mm_sub_ps(iy3,jy3);
1318             dz33             = _mm_sub_ps(iz3,jz3);
1319
1320             /* Calculate squared distance and things based on it */
1321             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1322             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1323             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1324             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1325             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1326             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1327             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1328             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1329             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1330             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1331
1332             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1333             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1334             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1335             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1336             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1337             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1338             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1339             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1340             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1341             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1342
1343             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1344             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1345             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1346             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1347             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1348             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1349             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1350             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1351             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1352
1353             fjx0             = _mm_setzero_ps();
1354             fjy0             = _mm_setzero_ps();
1355             fjz0             = _mm_setzero_ps();
1356             fjx1             = _mm_setzero_ps();
1357             fjy1             = _mm_setzero_ps();
1358             fjz1             = _mm_setzero_ps();
1359             fjx2             = _mm_setzero_ps();
1360             fjy2             = _mm_setzero_ps();
1361             fjz2             = _mm_setzero_ps();
1362             fjx3             = _mm_setzero_ps();
1363             fjy3             = _mm_setzero_ps();
1364             fjz3             = _mm_setzero_ps();
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             r00              = _mm_mul_ps(rsq00,rinv00);
1371
1372             /* Calculate table index by multiplying r with table scale and truncate to integer */
1373             rt               = _mm_mul_ps(r00,vftabscale);
1374             vfitab           = _mm_cvttps_epi32(rt);
1375             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1376             vfitab           = _mm_slli_epi32(vfitab,3);
1377
1378             /* CUBIC SPLINE TABLE DISPERSION */
1379             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1380             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1381             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1382             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1383             _MM_TRANSPOSE4_PS(Y,F,G,H);
1384             Heps             = _mm_mul_ps(vfeps,H);
1385             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1386             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1387             fvdw6            = _mm_mul_ps(c6_00,FF);
1388
1389             /* CUBIC SPLINE TABLE REPULSION */
1390             vfitab           = _mm_add_epi32(vfitab,ifour);
1391             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1392             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1393             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1394             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1395             _MM_TRANSPOSE4_PS(Y,F,G,H);
1396             Heps             = _mm_mul_ps(vfeps,H);
1397             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1398             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1399             fvdw12           = _mm_mul_ps(c12_00,FF);
1400             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1401
1402             fscal            = fvdw;
1403
1404             /* Calculate temporary vectorial force */
1405             tx               = _mm_mul_ps(fscal,dx00);
1406             ty               = _mm_mul_ps(fscal,dy00);
1407             tz               = _mm_mul_ps(fscal,dz00);
1408
1409             /* Update vectorial force */
1410             fix0             = _mm_add_ps(fix0,tx);
1411             fiy0             = _mm_add_ps(fiy0,ty);
1412             fiz0             = _mm_add_ps(fiz0,tz);
1413
1414             fjx0             = _mm_add_ps(fjx0,tx);
1415             fjy0             = _mm_add_ps(fjy0,ty);
1416             fjz0             = _mm_add_ps(fjz0,tz);
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             /* REACTION-FIELD ELECTROSTATICS */
1423             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1424
1425             fscal            = felec;
1426
1427             /* Calculate temporary vectorial force */
1428             tx               = _mm_mul_ps(fscal,dx11);
1429             ty               = _mm_mul_ps(fscal,dy11);
1430             tz               = _mm_mul_ps(fscal,dz11);
1431
1432             /* Update vectorial force */
1433             fix1             = _mm_add_ps(fix1,tx);
1434             fiy1             = _mm_add_ps(fiy1,ty);
1435             fiz1             = _mm_add_ps(fiz1,tz);
1436
1437             fjx1             = _mm_add_ps(fjx1,tx);
1438             fjy1             = _mm_add_ps(fjy1,ty);
1439             fjz1             = _mm_add_ps(fjz1,tz);
1440
1441             /**************************
1442              * CALCULATE INTERACTIONS *
1443              **************************/
1444
1445             /* REACTION-FIELD ELECTROSTATICS */
1446             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1447
1448             fscal            = felec;
1449
1450             /* Calculate temporary vectorial force */
1451             tx               = _mm_mul_ps(fscal,dx12);
1452             ty               = _mm_mul_ps(fscal,dy12);
1453             tz               = _mm_mul_ps(fscal,dz12);
1454
1455             /* Update vectorial force */
1456             fix1             = _mm_add_ps(fix1,tx);
1457             fiy1             = _mm_add_ps(fiy1,ty);
1458             fiz1             = _mm_add_ps(fiz1,tz);
1459
1460             fjx2             = _mm_add_ps(fjx2,tx);
1461             fjy2             = _mm_add_ps(fjy2,ty);
1462             fjz2             = _mm_add_ps(fjz2,tz);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             /* REACTION-FIELD ELECTROSTATICS */
1469             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1470
1471             fscal            = felec;
1472
1473             /* Calculate temporary vectorial force */
1474             tx               = _mm_mul_ps(fscal,dx13);
1475             ty               = _mm_mul_ps(fscal,dy13);
1476             tz               = _mm_mul_ps(fscal,dz13);
1477
1478             /* Update vectorial force */
1479             fix1             = _mm_add_ps(fix1,tx);
1480             fiy1             = _mm_add_ps(fiy1,ty);
1481             fiz1             = _mm_add_ps(fiz1,tz);
1482
1483             fjx3             = _mm_add_ps(fjx3,tx);
1484             fjy3             = _mm_add_ps(fjy3,ty);
1485             fjz3             = _mm_add_ps(fjz3,tz);
1486
1487             /**************************
1488              * CALCULATE INTERACTIONS *
1489              **************************/
1490
1491             /* REACTION-FIELD ELECTROSTATICS */
1492             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1493
1494             fscal            = felec;
1495
1496             /* Calculate temporary vectorial force */
1497             tx               = _mm_mul_ps(fscal,dx21);
1498             ty               = _mm_mul_ps(fscal,dy21);
1499             tz               = _mm_mul_ps(fscal,dz21);
1500
1501             /* Update vectorial force */
1502             fix2             = _mm_add_ps(fix2,tx);
1503             fiy2             = _mm_add_ps(fiy2,ty);
1504             fiz2             = _mm_add_ps(fiz2,tz);
1505
1506             fjx1             = _mm_add_ps(fjx1,tx);
1507             fjy1             = _mm_add_ps(fjy1,ty);
1508             fjz1             = _mm_add_ps(fjz1,tz);
1509
1510             /**************************
1511              * CALCULATE INTERACTIONS *
1512              **************************/
1513
1514             /* REACTION-FIELD ELECTROSTATICS */
1515             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1516
1517             fscal            = felec;
1518
1519             /* Calculate temporary vectorial force */
1520             tx               = _mm_mul_ps(fscal,dx22);
1521             ty               = _mm_mul_ps(fscal,dy22);
1522             tz               = _mm_mul_ps(fscal,dz22);
1523
1524             /* Update vectorial force */
1525             fix2             = _mm_add_ps(fix2,tx);
1526             fiy2             = _mm_add_ps(fiy2,ty);
1527             fiz2             = _mm_add_ps(fiz2,tz);
1528
1529             fjx2             = _mm_add_ps(fjx2,tx);
1530             fjy2             = _mm_add_ps(fjy2,ty);
1531             fjz2             = _mm_add_ps(fjz2,tz);
1532
1533             /**************************
1534              * CALCULATE INTERACTIONS *
1535              **************************/
1536
1537             /* REACTION-FIELD ELECTROSTATICS */
1538             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1539
1540             fscal            = felec;
1541
1542             /* Calculate temporary vectorial force */
1543             tx               = _mm_mul_ps(fscal,dx23);
1544             ty               = _mm_mul_ps(fscal,dy23);
1545             tz               = _mm_mul_ps(fscal,dz23);
1546
1547             /* Update vectorial force */
1548             fix2             = _mm_add_ps(fix2,tx);
1549             fiy2             = _mm_add_ps(fiy2,ty);
1550             fiz2             = _mm_add_ps(fiz2,tz);
1551
1552             fjx3             = _mm_add_ps(fjx3,tx);
1553             fjy3             = _mm_add_ps(fjy3,ty);
1554             fjz3             = _mm_add_ps(fjz3,tz);
1555
1556             /**************************
1557              * CALCULATE INTERACTIONS *
1558              **************************/
1559
1560             /* REACTION-FIELD ELECTROSTATICS */
1561             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1562
1563             fscal            = felec;
1564
1565             /* Calculate temporary vectorial force */
1566             tx               = _mm_mul_ps(fscal,dx31);
1567             ty               = _mm_mul_ps(fscal,dy31);
1568             tz               = _mm_mul_ps(fscal,dz31);
1569
1570             /* Update vectorial force */
1571             fix3             = _mm_add_ps(fix3,tx);
1572             fiy3             = _mm_add_ps(fiy3,ty);
1573             fiz3             = _mm_add_ps(fiz3,tz);
1574
1575             fjx1             = _mm_add_ps(fjx1,tx);
1576             fjy1             = _mm_add_ps(fjy1,ty);
1577             fjz1             = _mm_add_ps(fjz1,tz);
1578
1579             /**************************
1580              * CALCULATE INTERACTIONS *
1581              **************************/
1582
1583             /* REACTION-FIELD ELECTROSTATICS */
1584             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1585
1586             fscal            = felec;
1587
1588             /* Calculate temporary vectorial force */
1589             tx               = _mm_mul_ps(fscal,dx32);
1590             ty               = _mm_mul_ps(fscal,dy32);
1591             tz               = _mm_mul_ps(fscal,dz32);
1592
1593             /* Update vectorial force */
1594             fix3             = _mm_add_ps(fix3,tx);
1595             fiy3             = _mm_add_ps(fiy3,ty);
1596             fiz3             = _mm_add_ps(fiz3,tz);
1597
1598             fjx2             = _mm_add_ps(fjx2,tx);
1599             fjy2             = _mm_add_ps(fjy2,ty);
1600             fjz2             = _mm_add_ps(fjz2,tz);
1601
1602             /**************************
1603              * CALCULATE INTERACTIONS *
1604              **************************/
1605
1606             /* REACTION-FIELD ELECTROSTATICS */
1607             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1608
1609             fscal            = felec;
1610
1611             /* Calculate temporary vectorial force */
1612             tx               = _mm_mul_ps(fscal,dx33);
1613             ty               = _mm_mul_ps(fscal,dy33);
1614             tz               = _mm_mul_ps(fscal,dz33);
1615
1616             /* Update vectorial force */
1617             fix3             = _mm_add_ps(fix3,tx);
1618             fiy3             = _mm_add_ps(fiy3,ty);
1619             fiz3             = _mm_add_ps(fiz3,tz);
1620
1621             fjx3             = _mm_add_ps(fjx3,tx);
1622             fjy3             = _mm_add_ps(fjy3,ty);
1623             fjz3             = _mm_add_ps(fjz3,tz);
1624
1625             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1626                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1627                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1628                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1629
1630             /* Inner loop uses 294 flops */
1631         }
1632
1633         if(jidx<j_index_end)
1634         {
1635
1636             /* Get j neighbor index, and coordinate index */
1637             jnrA             = jjnr[jidx];
1638             jnrB             = jjnr[jidx+1];
1639             jnrC             = jjnr[jidx+2];
1640             jnrD             = jjnr[jidx+3];
1641
1642             /* Sign of each element will be negative for non-real atoms.
1643              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1644              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1645              */
1646             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1647             jnrA       = (jnrA>=0) ? jnrA : 0;
1648             jnrB       = (jnrB>=0) ? jnrB : 0;
1649             jnrC       = (jnrC>=0) ? jnrC : 0;
1650             jnrD       = (jnrD>=0) ? jnrD : 0;
1651
1652             j_coord_offsetA  = DIM*jnrA;
1653             j_coord_offsetB  = DIM*jnrB;
1654             j_coord_offsetC  = DIM*jnrC;
1655             j_coord_offsetD  = DIM*jnrD;
1656
1657             /* load j atom coordinates */
1658             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1659                                               x+j_coord_offsetC,x+j_coord_offsetD,
1660                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1661                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1662
1663             /* Calculate displacement vector */
1664             dx00             = _mm_sub_ps(ix0,jx0);
1665             dy00             = _mm_sub_ps(iy0,jy0);
1666             dz00             = _mm_sub_ps(iz0,jz0);
1667             dx11             = _mm_sub_ps(ix1,jx1);
1668             dy11             = _mm_sub_ps(iy1,jy1);
1669             dz11             = _mm_sub_ps(iz1,jz1);
1670             dx12             = _mm_sub_ps(ix1,jx2);
1671             dy12             = _mm_sub_ps(iy1,jy2);
1672             dz12             = _mm_sub_ps(iz1,jz2);
1673             dx13             = _mm_sub_ps(ix1,jx3);
1674             dy13             = _mm_sub_ps(iy1,jy3);
1675             dz13             = _mm_sub_ps(iz1,jz3);
1676             dx21             = _mm_sub_ps(ix2,jx1);
1677             dy21             = _mm_sub_ps(iy2,jy1);
1678             dz21             = _mm_sub_ps(iz2,jz1);
1679             dx22             = _mm_sub_ps(ix2,jx2);
1680             dy22             = _mm_sub_ps(iy2,jy2);
1681             dz22             = _mm_sub_ps(iz2,jz2);
1682             dx23             = _mm_sub_ps(ix2,jx3);
1683             dy23             = _mm_sub_ps(iy2,jy3);
1684             dz23             = _mm_sub_ps(iz2,jz3);
1685             dx31             = _mm_sub_ps(ix3,jx1);
1686             dy31             = _mm_sub_ps(iy3,jy1);
1687             dz31             = _mm_sub_ps(iz3,jz1);
1688             dx32             = _mm_sub_ps(ix3,jx2);
1689             dy32             = _mm_sub_ps(iy3,jy2);
1690             dz32             = _mm_sub_ps(iz3,jz2);
1691             dx33             = _mm_sub_ps(ix3,jx3);
1692             dy33             = _mm_sub_ps(iy3,jy3);
1693             dz33             = _mm_sub_ps(iz3,jz3);
1694
1695             /* Calculate squared distance and things based on it */
1696             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1697             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1698             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1699             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1700             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1701             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1702             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1703             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1704             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1705             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1706
1707             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1708             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1709             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1710             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1711             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1712             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1713             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1714             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1715             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1716             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1717
1718             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1719             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1720             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1721             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1722             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1723             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1724             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1725             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1726             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1727
1728             fjx0             = _mm_setzero_ps();
1729             fjy0             = _mm_setzero_ps();
1730             fjz0             = _mm_setzero_ps();
1731             fjx1             = _mm_setzero_ps();
1732             fjy1             = _mm_setzero_ps();
1733             fjz1             = _mm_setzero_ps();
1734             fjx2             = _mm_setzero_ps();
1735             fjy2             = _mm_setzero_ps();
1736             fjz2             = _mm_setzero_ps();
1737             fjx3             = _mm_setzero_ps();
1738             fjy3             = _mm_setzero_ps();
1739             fjz3             = _mm_setzero_ps();
1740
1741             /**************************
1742              * CALCULATE INTERACTIONS *
1743              **************************/
1744
1745             r00              = _mm_mul_ps(rsq00,rinv00);
1746             r00              = _mm_andnot_ps(dummy_mask,r00);
1747
1748             /* Calculate table index by multiplying r with table scale and truncate to integer */
1749             rt               = _mm_mul_ps(r00,vftabscale);
1750             vfitab           = _mm_cvttps_epi32(rt);
1751             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1752             vfitab           = _mm_slli_epi32(vfitab,3);
1753
1754             /* CUBIC SPLINE TABLE DISPERSION */
1755             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1756             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1757             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1758             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1759             _MM_TRANSPOSE4_PS(Y,F,G,H);
1760             Heps             = _mm_mul_ps(vfeps,H);
1761             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1762             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1763             fvdw6            = _mm_mul_ps(c6_00,FF);
1764
1765             /* CUBIC SPLINE TABLE REPULSION */
1766             vfitab           = _mm_add_epi32(vfitab,ifour);
1767             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1768             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1769             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1770             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1771             _MM_TRANSPOSE4_PS(Y,F,G,H);
1772             Heps             = _mm_mul_ps(vfeps,H);
1773             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1774             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1775             fvdw12           = _mm_mul_ps(c12_00,FF);
1776             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1777
1778             fscal            = fvdw;
1779
1780             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1781
1782             /* Calculate temporary vectorial force */
1783             tx               = _mm_mul_ps(fscal,dx00);
1784             ty               = _mm_mul_ps(fscal,dy00);
1785             tz               = _mm_mul_ps(fscal,dz00);
1786
1787             /* Update vectorial force */
1788             fix0             = _mm_add_ps(fix0,tx);
1789             fiy0             = _mm_add_ps(fiy0,ty);
1790             fiz0             = _mm_add_ps(fiz0,tz);
1791
1792             fjx0             = _mm_add_ps(fjx0,tx);
1793             fjy0             = _mm_add_ps(fjy0,ty);
1794             fjz0             = _mm_add_ps(fjz0,tz);
1795
1796             /**************************
1797              * CALCULATE INTERACTIONS *
1798              **************************/
1799
1800             /* REACTION-FIELD ELECTROSTATICS */
1801             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1802
1803             fscal            = felec;
1804
1805             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1806
1807             /* Calculate temporary vectorial force */
1808             tx               = _mm_mul_ps(fscal,dx11);
1809             ty               = _mm_mul_ps(fscal,dy11);
1810             tz               = _mm_mul_ps(fscal,dz11);
1811
1812             /* Update vectorial force */
1813             fix1             = _mm_add_ps(fix1,tx);
1814             fiy1             = _mm_add_ps(fiy1,ty);
1815             fiz1             = _mm_add_ps(fiz1,tz);
1816
1817             fjx1             = _mm_add_ps(fjx1,tx);
1818             fjy1             = _mm_add_ps(fjy1,ty);
1819             fjz1             = _mm_add_ps(fjz1,tz);
1820
1821             /**************************
1822              * CALCULATE INTERACTIONS *
1823              **************************/
1824
1825             /* REACTION-FIELD ELECTROSTATICS */
1826             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1827
1828             fscal            = felec;
1829
1830             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1831
1832             /* Calculate temporary vectorial force */
1833             tx               = _mm_mul_ps(fscal,dx12);
1834             ty               = _mm_mul_ps(fscal,dy12);
1835             tz               = _mm_mul_ps(fscal,dz12);
1836
1837             /* Update vectorial force */
1838             fix1             = _mm_add_ps(fix1,tx);
1839             fiy1             = _mm_add_ps(fiy1,ty);
1840             fiz1             = _mm_add_ps(fiz1,tz);
1841
1842             fjx2             = _mm_add_ps(fjx2,tx);
1843             fjy2             = _mm_add_ps(fjy2,ty);
1844             fjz2             = _mm_add_ps(fjz2,tz);
1845
1846             /**************************
1847              * CALCULATE INTERACTIONS *
1848              **************************/
1849
1850             /* REACTION-FIELD ELECTROSTATICS */
1851             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1852
1853             fscal            = felec;
1854
1855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1856
1857             /* Calculate temporary vectorial force */
1858             tx               = _mm_mul_ps(fscal,dx13);
1859             ty               = _mm_mul_ps(fscal,dy13);
1860             tz               = _mm_mul_ps(fscal,dz13);
1861
1862             /* Update vectorial force */
1863             fix1             = _mm_add_ps(fix1,tx);
1864             fiy1             = _mm_add_ps(fiy1,ty);
1865             fiz1             = _mm_add_ps(fiz1,tz);
1866
1867             fjx3             = _mm_add_ps(fjx3,tx);
1868             fjy3             = _mm_add_ps(fjy3,ty);
1869             fjz3             = _mm_add_ps(fjz3,tz);
1870
1871             /**************************
1872              * CALCULATE INTERACTIONS *
1873              **************************/
1874
1875             /* REACTION-FIELD ELECTROSTATICS */
1876             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1877
1878             fscal            = felec;
1879
1880             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1881
1882             /* Calculate temporary vectorial force */
1883             tx               = _mm_mul_ps(fscal,dx21);
1884             ty               = _mm_mul_ps(fscal,dy21);
1885             tz               = _mm_mul_ps(fscal,dz21);
1886
1887             /* Update vectorial force */
1888             fix2             = _mm_add_ps(fix2,tx);
1889             fiy2             = _mm_add_ps(fiy2,ty);
1890             fiz2             = _mm_add_ps(fiz2,tz);
1891
1892             fjx1             = _mm_add_ps(fjx1,tx);
1893             fjy1             = _mm_add_ps(fjy1,ty);
1894             fjz1             = _mm_add_ps(fjz1,tz);
1895
1896             /**************************
1897              * CALCULATE INTERACTIONS *
1898              **************************/
1899
1900             /* REACTION-FIELD ELECTROSTATICS */
1901             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1902
1903             fscal            = felec;
1904
1905             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1906
1907             /* Calculate temporary vectorial force */
1908             tx               = _mm_mul_ps(fscal,dx22);
1909             ty               = _mm_mul_ps(fscal,dy22);
1910             tz               = _mm_mul_ps(fscal,dz22);
1911
1912             /* Update vectorial force */
1913             fix2             = _mm_add_ps(fix2,tx);
1914             fiy2             = _mm_add_ps(fiy2,ty);
1915             fiz2             = _mm_add_ps(fiz2,tz);
1916
1917             fjx2             = _mm_add_ps(fjx2,tx);
1918             fjy2             = _mm_add_ps(fjy2,ty);
1919             fjz2             = _mm_add_ps(fjz2,tz);
1920
1921             /**************************
1922              * CALCULATE INTERACTIONS *
1923              **************************/
1924
1925             /* REACTION-FIELD ELECTROSTATICS */
1926             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1927
1928             fscal            = felec;
1929
1930             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1931
1932             /* Calculate temporary vectorial force */
1933             tx               = _mm_mul_ps(fscal,dx23);
1934             ty               = _mm_mul_ps(fscal,dy23);
1935             tz               = _mm_mul_ps(fscal,dz23);
1936
1937             /* Update vectorial force */
1938             fix2             = _mm_add_ps(fix2,tx);
1939             fiy2             = _mm_add_ps(fiy2,ty);
1940             fiz2             = _mm_add_ps(fiz2,tz);
1941
1942             fjx3             = _mm_add_ps(fjx3,tx);
1943             fjy3             = _mm_add_ps(fjy3,ty);
1944             fjz3             = _mm_add_ps(fjz3,tz);
1945
1946             /**************************
1947              * CALCULATE INTERACTIONS *
1948              **************************/
1949
1950             /* REACTION-FIELD ELECTROSTATICS */
1951             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1952
1953             fscal            = felec;
1954
1955             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1956
1957             /* Calculate temporary vectorial force */
1958             tx               = _mm_mul_ps(fscal,dx31);
1959             ty               = _mm_mul_ps(fscal,dy31);
1960             tz               = _mm_mul_ps(fscal,dz31);
1961
1962             /* Update vectorial force */
1963             fix3             = _mm_add_ps(fix3,tx);
1964             fiy3             = _mm_add_ps(fiy3,ty);
1965             fiz3             = _mm_add_ps(fiz3,tz);
1966
1967             fjx1             = _mm_add_ps(fjx1,tx);
1968             fjy1             = _mm_add_ps(fjy1,ty);
1969             fjz1             = _mm_add_ps(fjz1,tz);
1970
1971             /**************************
1972              * CALCULATE INTERACTIONS *
1973              **************************/
1974
1975             /* REACTION-FIELD ELECTROSTATICS */
1976             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1977
1978             fscal            = felec;
1979
1980             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1981
1982             /* Calculate temporary vectorial force */
1983             tx               = _mm_mul_ps(fscal,dx32);
1984             ty               = _mm_mul_ps(fscal,dy32);
1985             tz               = _mm_mul_ps(fscal,dz32);
1986
1987             /* Update vectorial force */
1988             fix3             = _mm_add_ps(fix3,tx);
1989             fiy3             = _mm_add_ps(fiy3,ty);
1990             fiz3             = _mm_add_ps(fiz3,tz);
1991
1992             fjx2             = _mm_add_ps(fjx2,tx);
1993             fjy2             = _mm_add_ps(fjy2,ty);
1994             fjz2             = _mm_add_ps(fjz2,tz);
1995
1996             /**************************
1997              * CALCULATE INTERACTIONS *
1998              **************************/
1999
2000             /* REACTION-FIELD ELECTROSTATICS */
2001             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2002
2003             fscal            = felec;
2004
2005             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2006
2007             /* Calculate temporary vectorial force */
2008             tx               = _mm_mul_ps(fscal,dx33);
2009             ty               = _mm_mul_ps(fscal,dy33);
2010             tz               = _mm_mul_ps(fscal,dz33);
2011
2012             /* Update vectorial force */
2013             fix3             = _mm_add_ps(fix3,tx);
2014             fiy3             = _mm_add_ps(fiy3,ty);
2015             fiz3             = _mm_add_ps(fiz3,tz);
2016
2017             fjx3             = _mm_add_ps(fjx3,tx);
2018             fjy3             = _mm_add_ps(fjy3,ty);
2019             fjz3             = _mm_add_ps(fjz3,tz);
2020
2021             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
2022                                                    f+j_coord_offsetC,f+j_coord_offsetD,
2023                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2024                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2025
2026             /* Inner loop uses 295 flops */
2027         }
2028
2029         /* End of innermost loop */
2030
2031         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2032                                               f+i_coord_offset,fshift+i_shift_offset);
2033
2034         /* Increment number of inner iterations */
2035         inneriter                  += j_index_end - j_index_start;
2036
2037         /* Outer loop uses 36 flops */
2038     }
2039
2040     /* Increment number of outer iterations */
2041     outeriter        += nri;
2042
2043     /* Update outer/inner flops */
2044
2045     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*36 + inneriter*295);
2046 }