src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_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_ElecCoul_VdwCSTab_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_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     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     vftab            = kernel_data->table_vdw->data;
 126     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 127
 128     /* Setup water-specific parameters */
 129     inr              = nlist->iinr[0];
 130     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 131     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 132     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 133     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 134
 135     jq1              = _mm_set1_ps(charge[inr+1]);
 136     jq2              = _mm_set1_ps(charge[inr+2]);
 137     jq3              = _mm_set1_ps(charge[inr+3]);
 138     vdwjidx0A        = 2*vdwtype[inr+0];
 139     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 140     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 141     qq11             = _mm_mul_ps(iq1,jq1);
 142     qq12             = _mm_mul_ps(iq1,jq2);
 143     qq13             = _mm_mul_ps(iq1,jq3);
 144     qq21             = _mm_mul_ps(iq2,jq1);
 145     qq22             = _mm_mul_ps(iq2,jq2);
 146     qq23             = _mm_mul_ps(iq2,jq3);
 147     qq31             = _mm_mul_ps(iq3,jq1);
 148     qq32             = _mm_mul_ps(iq3,jq2);
 149     qq33             = _mm_mul_ps(iq3,jq3);
 150
 151     /* Avoid stupid compiler warnings */
 152     jnrA = jnrB = jnrC = jnrD = 0;
 153     j_coord_offsetA = 0;
 154     j_coord_offsetB = 0;
 155     j_coord_offsetC = 0;
 156     j_coord_offsetD = 0;
 157
 158     outeriter        = 0;
 159     inneriter        = 0;
 160
 161     /* Start outer loop over neighborlists */
 162     for(iidx=0; iidx<nri; iidx++)
 163     {
 164         /* Load shift vector for this list */
 165         i_shift_offset   = DIM*shiftidx[iidx];
 166         shX              = shiftvec[i_shift_offset+XX];
 167         shY              = shiftvec[i_shift_offset+YY];
 168         shZ              = shiftvec[i_shift_offset+ZZ];
 169
 170         /* Load limits for loop over neighbors */
 171         j_index_start    = jindex[iidx];
 172         j_index_end      = jindex[iidx+1];
 173
 174         /* Get outer coordinate index */
 175         inr              = iinr[iidx];
 176         i_coord_offset   = DIM*inr;
 177
 178         /* Load i particle coords and add shift vector */
 179         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 180         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 181         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 182         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 183         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 184         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 185         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 186         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 187         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 188         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 189         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 190         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 191
 192         fix0             = _mm_setzero_ps();
 193         fiy0             = _mm_setzero_ps();
 194         fiz0             = _mm_setzero_ps();
 195         fix1             = _mm_setzero_ps();
 196         fiy1             = _mm_setzero_ps();
 197         fiz1             = _mm_setzero_ps();
 198         fix2             = _mm_setzero_ps();
 199         fiy2             = _mm_setzero_ps();
 200         fiz2             = _mm_setzero_ps();
 201         fix3             = _mm_setzero_ps();
 202         fiy3             = _mm_setzero_ps();
 203         fiz3             = _mm_setzero_ps();
 204
 205         /* Reset potential sums */
 206         velecsum         = _mm_setzero_ps();
 207         vvdwsum          = _mm_setzero_ps();
 208
 209         /* Start inner kernel loop */
 210         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 211         {
 212
 213             /* Get j neighbor index, and coordinate index */
 214             jnrA             = jjnr[jidx];
 215             jnrB             = jjnr[jidx+1];
 216             jnrC             = jjnr[jidx+2];
 217             jnrD             = jjnr[jidx+3];
 218
 219             j_coord_offsetA  = DIM*jnrA;
 220             j_coord_offsetB  = DIM*jnrB;
 221             j_coord_offsetC  = DIM*jnrC;
 222             j_coord_offsetD  = DIM*jnrD;
 223
 224             /* load j atom coordinates */
 225             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 226                                               x+j_coord_offsetC,x+j_coord_offsetD,
 227                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 228                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 229
 230             /* Calculate displacement vector */
 231             dx00             = _mm_sub_ps(ix0,jx0);
 232             dy00             = _mm_sub_ps(iy0,jy0);
 233             dz00             = _mm_sub_ps(iz0,jz0);
 234             dx11             = _mm_sub_ps(ix1,jx1);
 235             dy11             = _mm_sub_ps(iy1,jy1);
 236             dz11             = _mm_sub_ps(iz1,jz1);
 237             dx12             = _mm_sub_ps(ix1,jx2);
 238             dy12             = _mm_sub_ps(iy1,jy2);
 239             dz12             = _mm_sub_ps(iz1,jz2);
 240             dx13             = _mm_sub_ps(ix1,jx3);
 241             dy13             = _mm_sub_ps(iy1,jy3);
 242             dz13             = _mm_sub_ps(iz1,jz3);
 243             dx21             = _mm_sub_ps(ix2,jx1);
 244             dy21             = _mm_sub_ps(iy2,jy1);
 245             dz21             = _mm_sub_ps(iz2,jz1);
 246             dx22             = _mm_sub_ps(ix2,jx2);
 247             dy22             = _mm_sub_ps(iy2,jy2);
 248             dz22             = _mm_sub_ps(iz2,jz2);
 249             dx23             = _mm_sub_ps(ix2,jx3);
 250             dy23             = _mm_sub_ps(iy2,jy3);
 251             dz23             = _mm_sub_ps(iz2,jz3);
 252             dx31             = _mm_sub_ps(ix3,jx1);
 253             dy31             = _mm_sub_ps(iy3,jy1);
 254             dz31             = _mm_sub_ps(iz3,jz1);
 255             dx32             = _mm_sub_ps(ix3,jx2);
 256             dy32             = _mm_sub_ps(iy3,jy2);
 257             dz32             = _mm_sub_ps(iz3,jz2);
 258             dx33             = _mm_sub_ps(ix3,jx3);
 259             dy33             = _mm_sub_ps(iy3,jy3);
 260             dz33             = _mm_sub_ps(iz3,jz3);
 261
 262             /* Calculate squared distance and things based on it */
 263             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 264             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 265             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 266             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 267             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 268             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 269             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 270             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 271             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 272             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 273
 274             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 275             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 276             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 277             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 278             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 279             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 280             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 281             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 282             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 283             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 284
 285             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 286             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 287             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 288             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 289             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 290             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 291             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 292             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 293             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 294
 295             fjx0             = _mm_setzero_ps();
 296             fjy0             = _mm_setzero_ps();
 297             fjz0             = _mm_setzero_ps();
 298             fjx1             = _mm_setzero_ps();
 299             fjy1             = _mm_setzero_ps();
 300             fjz1             = _mm_setzero_ps();
 301             fjx2             = _mm_setzero_ps();
 302             fjy2             = _mm_setzero_ps();
 303             fjz2             = _mm_setzero_ps();
 304             fjx3             = _mm_setzero_ps();
 305             fjy3             = _mm_setzero_ps();
 306             fjz3             = _mm_setzero_ps();
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             r00              = _mm_mul_ps(rsq00,rinv00);
 313
 314             /* Calculate table index by multiplying r with table scale and truncate to integer */
 315             rt               = _mm_mul_ps(r00,vftabscale);
 316             vfitab           = _mm_cvttps_epi32(rt);
 317             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 318             vfitab           = _mm_slli_epi32(vfitab,3);
 319
 320             /* CUBIC SPLINE TABLE DISPERSION */
 321             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 322             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 323             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 324             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 325             _MM_TRANSPOSE4_PS(Y,F,G,H);
 326             Heps             = _mm_mul_ps(vfeps,H);
 327             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 328             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 329             vvdw6            = _mm_mul_ps(c6_00,VV);
 330             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 331             fvdw6            = _mm_mul_ps(c6_00,FF);
 332
 333             /* CUBIC SPLINE TABLE REPULSION */
 334             vfitab           = _mm_add_epi32(vfitab,ifour);
 335             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 336             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 337             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 338             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 339             _MM_TRANSPOSE4_PS(Y,F,G,H);
 340             Heps             = _mm_mul_ps(vfeps,H);
 341             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 342             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 343             vvdw12           = _mm_mul_ps(c12_00,VV);
 344             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 345             fvdw12           = _mm_mul_ps(c12_00,FF);
 346             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 347             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 351
 352             fscal            = fvdw;
 353
 354             /* Calculate temporary vectorial force */
 355             tx               = _mm_mul_ps(fscal,dx00);
 356             ty               = _mm_mul_ps(fscal,dy00);
 357             tz               = _mm_mul_ps(fscal,dz00);
 358
 359             /* Update vectorial force */
 360             fix0             = _mm_add_ps(fix0,tx);
 361             fiy0             = _mm_add_ps(fiy0,ty);
 362             fiz0             = _mm_add_ps(fiz0,tz);
 363
 364             fjx0             = _mm_add_ps(fjx0,tx);
 365             fjy0             = _mm_add_ps(fjy0,ty);
 366             fjz0             = _mm_add_ps(fjz0,tz);
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             /* COULOMB ELECTROSTATICS */
 373             velec            = _mm_mul_ps(qq11,rinv11);
 374             felec            = _mm_mul_ps(velec,rinvsq11);
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velecsum         = _mm_add_ps(velecsum,velec);
 378
 379             fscal            = felec;
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = _mm_mul_ps(fscal,dx11);
 383             ty               = _mm_mul_ps(fscal,dy11);
 384             tz               = _mm_mul_ps(fscal,dz11);
 385
 386             /* Update vectorial force */
 387             fix1             = _mm_add_ps(fix1,tx);
 388             fiy1             = _mm_add_ps(fiy1,ty);
 389             fiz1             = _mm_add_ps(fiz1,tz);
 390
 391             fjx1             = _mm_add_ps(fjx1,tx);
 392             fjy1             = _mm_add_ps(fjy1,ty);
 393             fjz1             = _mm_add_ps(fjz1,tz);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             /* COULOMB ELECTROSTATICS */
 400             velec            = _mm_mul_ps(qq12,rinv12);
 401             felec            = _mm_mul_ps(velec,rinvsq12);
 402
 403             /* Update potential sum for this i atom from the interaction with this j atom. */
 404             velecsum         = _mm_add_ps(velecsum,velec);
 405
 406             fscal            = felec;
 407
 408             /* Calculate temporary vectorial force */
 409             tx               = _mm_mul_ps(fscal,dx12);
 410             ty               = _mm_mul_ps(fscal,dy12);
 411             tz               = _mm_mul_ps(fscal,dz12);
 412
 413             /* Update vectorial force */
 414             fix1             = _mm_add_ps(fix1,tx);
 415             fiy1             = _mm_add_ps(fiy1,ty);
 416             fiz1             = _mm_add_ps(fiz1,tz);
 417
 418             fjx2             = _mm_add_ps(fjx2,tx);
 419             fjy2             = _mm_add_ps(fjy2,ty);
 420             fjz2             = _mm_add_ps(fjz2,tz);
 421
 422             /**************************
 423              * CALCULATE INTERACTIONS *
 424              **************************/
 425
 426             /* COULOMB ELECTROSTATICS */
 427             velec            = _mm_mul_ps(qq13,rinv13);
 428             felec            = _mm_mul_ps(velec,rinvsq13);
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velecsum         = _mm_add_ps(velecsum,velec);
 432
 433             fscal            = felec;
 434
 435             /* Calculate temporary vectorial force */
 436             tx               = _mm_mul_ps(fscal,dx13);
 437             ty               = _mm_mul_ps(fscal,dy13);
 438             tz               = _mm_mul_ps(fscal,dz13);
 439
 440             /* Update vectorial force */
 441             fix1             = _mm_add_ps(fix1,tx);
 442             fiy1             = _mm_add_ps(fiy1,ty);
 443             fiz1             = _mm_add_ps(fiz1,tz);
 444
 445             fjx3             = _mm_add_ps(fjx3,tx);
 446             fjy3             = _mm_add_ps(fjy3,ty);
 447             fjz3             = _mm_add_ps(fjz3,tz);
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             /* COULOMB ELECTROSTATICS */
 454             velec            = _mm_mul_ps(qq21,rinv21);
 455             felec            = _mm_mul_ps(velec,rinvsq21);
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velecsum         = _mm_add_ps(velecsum,velec);
 459
 460             fscal            = felec;
 461
 462             /* Calculate temporary vectorial force */
 463             tx               = _mm_mul_ps(fscal,dx21);
 464             ty               = _mm_mul_ps(fscal,dy21);
 465             tz               = _mm_mul_ps(fscal,dz21);
 466
 467             /* Update vectorial force */
 468             fix2             = _mm_add_ps(fix2,tx);
 469             fiy2             = _mm_add_ps(fiy2,ty);
 470             fiz2             = _mm_add_ps(fiz2,tz);
 471
 472             fjx1             = _mm_add_ps(fjx1,tx);
 473             fjy1             = _mm_add_ps(fjy1,ty);
 474             fjz1             = _mm_add_ps(fjz1,tz);
 475
 476             /**************************
 477              * CALCULATE INTERACTIONS *
 478              **************************/
 479
 480             /* COULOMB ELECTROSTATICS */
 481             velec            = _mm_mul_ps(qq22,rinv22);
 482             felec            = _mm_mul_ps(velec,rinvsq22);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velecsum         = _mm_add_ps(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             /* Calculate temporary vectorial force */
 490             tx               = _mm_mul_ps(fscal,dx22);
 491             ty               = _mm_mul_ps(fscal,dy22);
 492             tz               = _mm_mul_ps(fscal,dz22);
 493
 494             /* Update vectorial force */
 495             fix2             = _mm_add_ps(fix2,tx);
 496             fiy2             = _mm_add_ps(fiy2,ty);
 497             fiz2             = _mm_add_ps(fiz2,tz);
 498
 499             fjx2             = _mm_add_ps(fjx2,tx);
 500             fjy2             = _mm_add_ps(fjy2,ty);
 501             fjz2             = _mm_add_ps(fjz2,tz);
 502
 503             /**************************
 504              * CALCULATE INTERACTIONS *
 505              **************************/
 506
 507             /* COULOMB ELECTROSTATICS */
 508             velec            = _mm_mul_ps(qq23,rinv23);
 509             felec            = _mm_mul_ps(velec,rinvsq23);
 510
 511             /* Update potential sum for this i atom from the interaction with this j atom. */
 512             velecsum         = _mm_add_ps(velecsum,velec);
 513
 514             fscal            = felec;
 515
 516             /* Calculate temporary vectorial force */
 517             tx               = _mm_mul_ps(fscal,dx23);
 518             ty               = _mm_mul_ps(fscal,dy23);
 519             tz               = _mm_mul_ps(fscal,dz23);
 520
 521             /* Update vectorial force */
 522             fix2             = _mm_add_ps(fix2,tx);
 523             fiy2             = _mm_add_ps(fiy2,ty);
 524             fiz2             = _mm_add_ps(fiz2,tz);
 525
 526             fjx3             = _mm_add_ps(fjx3,tx);
 527             fjy3             = _mm_add_ps(fjy3,ty);
 528             fjz3             = _mm_add_ps(fjz3,tz);
 529
 530             /**************************
 531              * CALCULATE INTERACTIONS *
 532              **************************/
 533
 534             /* COULOMB ELECTROSTATICS */
 535             velec            = _mm_mul_ps(qq31,rinv31);
 536             felec            = _mm_mul_ps(velec,rinvsq31);
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velecsum         = _mm_add_ps(velecsum,velec);
 540
 541             fscal            = felec;
 542
 543             /* Calculate temporary vectorial force */
 544             tx               = _mm_mul_ps(fscal,dx31);
 545             ty               = _mm_mul_ps(fscal,dy31);
 546             tz               = _mm_mul_ps(fscal,dz31);
 547
 548             /* Update vectorial force */
 549             fix3             = _mm_add_ps(fix3,tx);
 550             fiy3             = _mm_add_ps(fiy3,ty);
 551             fiz3             = _mm_add_ps(fiz3,tz);
 552
 553             fjx1             = _mm_add_ps(fjx1,tx);
 554             fjy1             = _mm_add_ps(fjy1,ty);
 555             fjz1             = _mm_add_ps(fjz1,tz);
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             /* COULOMB ELECTROSTATICS */
 562             velec            = _mm_mul_ps(qq32,rinv32);
 563             felec            = _mm_mul_ps(velec,rinvsq32);
 564
 565             /* Update potential sum for this i atom from the interaction with this j atom. */
 566             velecsum         = _mm_add_ps(velecsum,velec);
 567
 568             fscal            = felec;
 569
 570             /* Calculate temporary vectorial force */
 571             tx               = _mm_mul_ps(fscal,dx32);
 572             ty               = _mm_mul_ps(fscal,dy32);
 573             tz               = _mm_mul_ps(fscal,dz32);
 574
 575             /* Update vectorial force */
 576             fix3             = _mm_add_ps(fix3,tx);
 577             fiy3             = _mm_add_ps(fiy3,ty);
 578             fiz3             = _mm_add_ps(fiz3,tz);
 579
 580             fjx2             = _mm_add_ps(fjx2,tx);
 581             fjy2             = _mm_add_ps(fjy2,ty);
 582             fjz2             = _mm_add_ps(fjz2,tz);
 583
 584             /**************************
 585              * CALCULATE INTERACTIONS *
 586              **************************/
 587
 588             /* COULOMB ELECTROSTATICS */
 589             velec            = _mm_mul_ps(qq33,rinv33);
 590             felec            = _mm_mul_ps(velec,rinvsq33);
 591
 592             /* Update potential sum for this i atom from the interaction with this j atom. */
 593             velecsum         = _mm_add_ps(velecsum,velec);
 594
 595             fscal            = felec;
 596
 597             /* Calculate temporary vectorial force */
 598             tx               = _mm_mul_ps(fscal,dx33);
 599             ty               = _mm_mul_ps(fscal,dy33);
 600             tz               = _mm_mul_ps(fscal,dz33);
 601
 602             /* Update vectorial force */
 603             fix3             = _mm_add_ps(fix3,tx);
 604             fiy3             = _mm_add_ps(fiy3,ty);
 605             fiz3             = _mm_add_ps(fiz3,tz);
 606
 607             fjx3             = _mm_add_ps(fjx3,tx);
 608             fjy3             = _mm_add_ps(fjy3,ty);
 609             fjz3             = _mm_add_ps(fjz3,tz);
 610
 611             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 612                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 613                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 614                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 615
 616             /* Inner loop uses 311 flops */
 617         }
 618
 619         if(jidx<j_index_end)
 620         {
 621
 622             /* Get j neighbor index, and coordinate index */
 623             jnrA             = jjnr[jidx];
 624             jnrB             = jjnr[jidx+1];
 625             jnrC             = jjnr[jidx+2];
 626             jnrD             = jjnr[jidx+3];
 627
 628             /* Sign of each element will be negative for non-real atoms.
 629              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 630              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 631              */
 632             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 633             jnrA       = (jnrA>=0) ? jnrA : 0;
 634             jnrB       = (jnrB>=0) ? jnrB : 0;
 635             jnrC       = (jnrC>=0) ? jnrC : 0;
 636             jnrD       = (jnrD>=0) ? jnrD : 0;
 637
 638             j_coord_offsetA  = DIM*jnrA;
 639             j_coord_offsetB  = DIM*jnrB;
 640             j_coord_offsetC  = DIM*jnrC;
 641             j_coord_offsetD  = DIM*jnrD;
 642
 643             /* load j atom coordinates */
 644             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 645                                               x+j_coord_offsetC,x+j_coord_offsetD,
 646                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 647                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 648
 649             /* Calculate displacement vector */
 650             dx00             = _mm_sub_ps(ix0,jx0);
 651             dy00             = _mm_sub_ps(iy0,jy0);
 652             dz00             = _mm_sub_ps(iz0,jz0);
 653             dx11             = _mm_sub_ps(ix1,jx1);
 654             dy11             = _mm_sub_ps(iy1,jy1);
 655             dz11             = _mm_sub_ps(iz1,jz1);
 656             dx12             = _mm_sub_ps(ix1,jx2);
 657             dy12             = _mm_sub_ps(iy1,jy2);
 658             dz12             = _mm_sub_ps(iz1,jz2);
 659             dx13             = _mm_sub_ps(ix1,jx3);
 660             dy13             = _mm_sub_ps(iy1,jy3);
 661             dz13             = _mm_sub_ps(iz1,jz3);
 662             dx21             = _mm_sub_ps(ix2,jx1);
 663             dy21             = _mm_sub_ps(iy2,jy1);
 664             dz21             = _mm_sub_ps(iz2,jz1);
 665             dx22             = _mm_sub_ps(ix2,jx2);
 666             dy22             = _mm_sub_ps(iy2,jy2);
 667             dz22             = _mm_sub_ps(iz2,jz2);
 668             dx23             = _mm_sub_ps(ix2,jx3);
 669             dy23             = _mm_sub_ps(iy2,jy3);
 670             dz23             = _mm_sub_ps(iz2,jz3);
 671             dx31             = _mm_sub_ps(ix3,jx1);
 672             dy31             = _mm_sub_ps(iy3,jy1);
 673             dz31             = _mm_sub_ps(iz3,jz1);
 674             dx32             = _mm_sub_ps(ix3,jx2);
 675             dy32             = _mm_sub_ps(iy3,jy2);
 676             dz32             = _mm_sub_ps(iz3,jz2);
 677             dx33             = _mm_sub_ps(ix3,jx3);
 678             dy33             = _mm_sub_ps(iy3,jy3);
 679             dz33             = _mm_sub_ps(iz3,jz3);
 680
 681             /* Calculate squared distance and things based on it */
 682             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 683             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 684             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 685             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 686             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 687             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 688             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 689             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 690             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 691             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 692
 693             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 694             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 695             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 696             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 697             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 698             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 699             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 700             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 701             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 702             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 703
 704             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 705             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 706             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 707             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 708             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 709             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 710             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 711             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 712             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 713
 714             fjx0             = _mm_setzero_ps();
 715             fjy0             = _mm_setzero_ps();
 716             fjz0             = _mm_setzero_ps();
 717             fjx1             = _mm_setzero_ps();
 718             fjy1             = _mm_setzero_ps();
 719             fjz1             = _mm_setzero_ps();
 720             fjx2             = _mm_setzero_ps();
 721             fjy2             = _mm_setzero_ps();
 722             fjz2             = _mm_setzero_ps();
 723             fjx3             = _mm_setzero_ps();
 724             fjy3             = _mm_setzero_ps();
 725             fjz3             = _mm_setzero_ps();
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             r00              = _mm_mul_ps(rsq00,rinv00);
 732             r00              = _mm_andnot_ps(dummy_mask,r00);
 733
 734             /* Calculate table index by multiplying r with table scale and truncate to integer */
 735             rt               = _mm_mul_ps(r00,vftabscale);
 736             vfitab           = _mm_cvttps_epi32(rt);
 737             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 738             vfitab           = _mm_slli_epi32(vfitab,3);
 739
 740             /* CUBIC SPLINE TABLE DISPERSION */
 741             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 742             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 743             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 744             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 745             _MM_TRANSPOSE4_PS(Y,F,G,H);
 746             Heps             = _mm_mul_ps(vfeps,H);
 747             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 748             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 749             vvdw6            = _mm_mul_ps(c6_00,VV);
 750             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 751             fvdw6            = _mm_mul_ps(c6_00,FF);
 752
 753             /* CUBIC SPLINE TABLE REPULSION */
 754             vfitab           = _mm_add_epi32(vfitab,ifour);
 755             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 756             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 757             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 758             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 759             _MM_TRANSPOSE4_PS(Y,F,G,H);
 760             Heps             = _mm_mul_ps(vfeps,H);
 761             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 762             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 763             vvdw12           = _mm_mul_ps(c12_00,VV);
 764             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 765             fvdw12           = _mm_mul_ps(c12_00,FF);
 766             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 767             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 768
 769             /* Update potential sum for this i atom from the interaction with this j atom. */
 770             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 771             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 772
 773             fscal            = fvdw;
 774
 775             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 776
 777             /* Calculate temporary vectorial force */
 778             tx               = _mm_mul_ps(fscal,dx00);
 779             ty               = _mm_mul_ps(fscal,dy00);
 780             tz               = _mm_mul_ps(fscal,dz00);
 781
 782             /* Update vectorial force */
 783             fix0             = _mm_add_ps(fix0,tx);
 784             fiy0             = _mm_add_ps(fiy0,ty);
 785             fiz0             = _mm_add_ps(fiz0,tz);
 786
 787             fjx0             = _mm_add_ps(fjx0,tx);
 788             fjy0             = _mm_add_ps(fjy0,ty);
 789             fjz0             = _mm_add_ps(fjz0,tz);
 790
 791             /**************************
 792              * CALCULATE INTERACTIONS *
 793              **************************/
 794
 795             /* COULOMB ELECTROSTATICS */
 796             velec            = _mm_mul_ps(qq11,rinv11);
 797             felec            = _mm_mul_ps(velec,rinvsq11);
 798
 799             /* Update potential sum for this i atom from the interaction with this j atom. */
 800             velec            = _mm_andnot_ps(dummy_mask,velec);
 801             velecsum         = _mm_add_ps(velecsum,velec);
 802
 803             fscal            = felec;
 804
 805             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 806
 807             /* Calculate temporary vectorial force */
 808             tx               = _mm_mul_ps(fscal,dx11);
 809             ty               = _mm_mul_ps(fscal,dy11);
 810             tz               = _mm_mul_ps(fscal,dz11);
 811
 812             /* Update vectorial force */
 813             fix1             = _mm_add_ps(fix1,tx);
 814             fiy1             = _mm_add_ps(fiy1,ty);
 815             fiz1             = _mm_add_ps(fiz1,tz);
 816
 817             fjx1             = _mm_add_ps(fjx1,tx);
 818             fjy1             = _mm_add_ps(fjy1,ty);
 819             fjz1             = _mm_add_ps(fjz1,tz);
 820
 821             /**************************
 822              * CALCULATE INTERACTIONS *
 823              **************************/
 824
 825             /* COULOMB ELECTROSTATICS */
 826             velec            = _mm_mul_ps(qq12,rinv12);
 827             felec            = _mm_mul_ps(velec,rinvsq12);
 828
 829             /* Update potential sum for this i atom from the interaction with this j atom. */
 830             velec            = _mm_andnot_ps(dummy_mask,velec);
 831             velecsum         = _mm_add_ps(velecsum,velec);
 832
 833             fscal            = felec;
 834
 835             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 836
 837             /* Calculate temporary vectorial force */
 838             tx               = _mm_mul_ps(fscal,dx12);
 839             ty               = _mm_mul_ps(fscal,dy12);
 840             tz               = _mm_mul_ps(fscal,dz12);
 841
 842             /* Update vectorial force */
 843             fix1             = _mm_add_ps(fix1,tx);
 844             fiy1             = _mm_add_ps(fiy1,ty);
 845             fiz1             = _mm_add_ps(fiz1,tz);
 846
 847             fjx2             = _mm_add_ps(fjx2,tx);
 848             fjy2             = _mm_add_ps(fjy2,ty);
 849             fjz2             = _mm_add_ps(fjz2,tz);
 850
 851             /**************************
 852              * CALCULATE INTERACTIONS *
 853              **************************/
 854
 855             /* COULOMB ELECTROSTATICS */
 856             velec            = _mm_mul_ps(qq13,rinv13);
 857             felec            = _mm_mul_ps(velec,rinvsq13);
 858
 859             /* Update potential sum for this i atom from the interaction with this j atom. */
 860             velec            = _mm_andnot_ps(dummy_mask,velec);
 861             velecsum         = _mm_add_ps(velecsum,velec);
 862
 863             fscal            = felec;
 864
 865             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 866
 867             /* Calculate temporary vectorial force */
 868             tx               = _mm_mul_ps(fscal,dx13);
 869             ty               = _mm_mul_ps(fscal,dy13);
 870             tz               = _mm_mul_ps(fscal,dz13);
 871
 872             /* Update vectorial force */
 873             fix1             = _mm_add_ps(fix1,tx);
 874             fiy1             = _mm_add_ps(fiy1,ty);
 875             fiz1             = _mm_add_ps(fiz1,tz);
 876
 877             fjx3             = _mm_add_ps(fjx3,tx);
 878             fjy3             = _mm_add_ps(fjy3,ty);
 879             fjz3             = _mm_add_ps(fjz3,tz);
 880
 881             /**************************
 882              * CALCULATE INTERACTIONS *
 883              **************************/
 884
 885             /* COULOMB ELECTROSTATICS */
 886             velec            = _mm_mul_ps(qq21,rinv21);
 887             felec            = _mm_mul_ps(velec,rinvsq21);
 888
 889             /* Update potential sum for this i atom from the interaction with this j atom. */
 890             velec            = _mm_andnot_ps(dummy_mask,velec);
 891             velecsum         = _mm_add_ps(velecsum,velec);
 892
 893             fscal            = felec;
 894
 895             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 896
 897             /* Calculate temporary vectorial force */
 898             tx               = _mm_mul_ps(fscal,dx21);
 899             ty               = _mm_mul_ps(fscal,dy21);
 900             tz               = _mm_mul_ps(fscal,dz21);
 901
 902             /* Update vectorial force */
 903             fix2             = _mm_add_ps(fix2,tx);
 904             fiy2             = _mm_add_ps(fiy2,ty);
 905             fiz2             = _mm_add_ps(fiz2,tz);
 906
 907             fjx1             = _mm_add_ps(fjx1,tx);
 908             fjy1             = _mm_add_ps(fjy1,ty);
 909             fjz1             = _mm_add_ps(fjz1,tz);
 910
 911             /**************************
 912              * CALCULATE INTERACTIONS *
 913              **************************/
 914
 915             /* COULOMB ELECTROSTATICS */
 916             velec            = _mm_mul_ps(qq22,rinv22);
 917             felec            = _mm_mul_ps(velec,rinvsq22);
 918
 919             /* Update potential sum for this i atom from the interaction with this j atom. */
 920             velec            = _mm_andnot_ps(dummy_mask,velec);
 921             velecsum         = _mm_add_ps(velecsum,velec);
 922
 923             fscal            = felec;
 924
 925             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 926
 927             /* Calculate temporary vectorial force */
 928             tx               = _mm_mul_ps(fscal,dx22);
 929             ty               = _mm_mul_ps(fscal,dy22);
 930             tz               = _mm_mul_ps(fscal,dz22);
 931
 932             /* Update vectorial force */
 933             fix2             = _mm_add_ps(fix2,tx);
 934             fiy2             = _mm_add_ps(fiy2,ty);
 935             fiz2             = _mm_add_ps(fiz2,tz);
 936
 937             fjx2             = _mm_add_ps(fjx2,tx);
 938             fjy2             = _mm_add_ps(fjy2,ty);
 939             fjz2             = _mm_add_ps(fjz2,tz);
 940
 941             /**************************
 942              * CALCULATE INTERACTIONS *
 943              **************************/
 944
 945             /* COULOMB ELECTROSTATICS */
 946             velec            = _mm_mul_ps(qq23,rinv23);
 947             felec            = _mm_mul_ps(velec,rinvsq23);
 948
 949             /* Update potential sum for this i atom from the interaction with this j atom. */
 950             velec            = _mm_andnot_ps(dummy_mask,velec);
 951             velecsum         = _mm_add_ps(velecsum,velec);
 952
 953             fscal            = felec;
 954
 955             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 956
 957             /* Calculate temporary vectorial force */
 958             tx               = _mm_mul_ps(fscal,dx23);
 959             ty               = _mm_mul_ps(fscal,dy23);
 960             tz               = _mm_mul_ps(fscal,dz23);
 961
 962             /* Update vectorial force */
 963             fix2             = _mm_add_ps(fix2,tx);
 964             fiy2             = _mm_add_ps(fiy2,ty);
 965             fiz2             = _mm_add_ps(fiz2,tz);
 966
 967             fjx3             = _mm_add_ps(fjx3,tx);
 968             fjy3             = _mm_add_ps(fjy3,ty);
 969             fjz3             = _mm_add_ps(fjz3,tz);
 970
 971             /**************************
 972              * CALCULATE INTERACTIONS *
 973              **************************/
 974
 975             /* COULOMB ELECTROSTATICS */
 976             velec            = _mm_mul_ps(qq31,rinv31);
 977             felec            = _mm_mul_ps(velec,rinvsq31);
 978
 979             /* Update potential sum for this i atom from the interaction with this j atom. */
 980             velec            = _mm_andnot_ps(dummy_mask,velec);
 981             velecsum         = _mm_add_ps(velecsum,velec);
 982
 983             fscal            = felec;
 984
 985             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 986
 987             /* Calculate temporary vectorial force */
 988             tx               = _mm_mul_ps(fscal,dx31);
 989             ty               = _mm_mul_ps(fscal,dy31);
 990             tz               = _mm_mul_ps(fscal,dz31);
 991
 992             /* Update vectorial force */
 993             fix3             = _mm_add_ps(fix3,tx);
 994             fiy3             = _mm_add_ps(fiy3,ty);
 995             fiz3             = _mm_add_ps(fiz3,tz);
 996
 997             fjx1             = _mm_add_ps(fjx1,tx);
 998             fjy1             = _mm_add_ps(fjy1,ty);
 999             fjz1             = _mm_add_ps(fjz1,tz);
1000
1001             /**************************
1002              * CALCULATE INTERACTIONS *
1003              **************************/
1004
1005             /* COULOMB ELECTROSTATICS */
1006             velec            = _mm_mul_ps(qq32,rinv32);
1007             felec            = _mm_mul_ps(velec,rinvsq32);
1008
1009             /* Update potential sum for this i atom from the interaction with this j atom. */
1010             velec            = _mm_andnot_ps(dummy_mask,velec);
1011             velecsum         = _mm_add_ps(velecsum,velec);
1012
1013             fscal            = felec;
1014
1015             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1016
1017             /* Calculate temporary vectorial force */
1018             tx               = _mm_mul_ps(fscal,dx32);
1019             ty               = _mm_mul_ps(fscal,dy32);
1020             tz               = _mm_mul_ps(fscal,dz32);
1021
1022             /* Update vectorial force */
1023             fix3             = _mm_add_ps(fix3,tx);
1024             fiy3             = _mm_add_ps(fiy3,ty);
1025             fiz3             = _mm_add_ps(fiz3,tz);
1026
1027             fjx2             = _mm_add_ps(fjx2,tx);
1028             fjy2             = _mm_add_ps(fjy2,ty);
1029             fjz2             = _mm_add_ps(fjz2,tz);
1030
1031             /**************************
1032              * CALCULATE INTERACTIONS *
1033              **************************/
1034
1035             /* COULOMB ELECTROSTATICS */
1036             velec            = _mm_mul_ps(qq33,rinv33);
1037             felec            = _mm_mul_ps(velec,rinvsq33);
1038
1039             /* Update potential sum for this i atom from the interaction with this j atom. */
1040             velec            = _mm_andnot_ps(dummy_mask,velec);
1041             velecsum         = _mm_add_ps(velecsum,velec);
1042
1043             fscal            = felec;
1044
1045             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1046
1047             /* Calculate temporary vectorial force */
1048             tx               = _mm_mul_ps(fscal,dx33);
1049             ty               = _mm_mul_ps(fscal,dy33);
1050             tz               = _mm_mul_ps(fscal,dz33);
1051
1052             /* Update vectorial force */
1053             fix3             = _mm_add_ps(fix3,tx);
1054             fiy3             = _mm_add_ps(fiy3,ty);
1055             fiz3             = _mm_add_ps(fiz3,tz);
1056
1057             fjx3             = _mm_add_ps(fjx3,tx);
1058             fjy3             = _mm_add_ps(fjy3,ty);
1059             fjz3             = _mm_add_ps(fjz3,tz);
1060
1061             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1062                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1063                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1064                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1065
1066             /* Inner loop uses 312 flops */
1067         }
1068
1069         /* End of innermost loop */
1070
1071         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1072                                               f+i_coord_offset,fshift+i_shift_offset);
1073
1074         ggid                        = gid[iidx];
1075         /* Update potential energies */
1076         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1077         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1078
1079         /* Increment number of inner iterations */
1080         inneriter                  += j_index_end - j_index_start;
1081
1082         /* Outer loop uses 38 flops */
1083     }
1084
1085     /* Increment number of outer iterations */
1086     outeriter        += nri;
1087
1088     /* Update outer/inner flops */
1089
1090     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*38 + inneriter*312);
1091 }
1092 /*
1093  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_sse2_single
1094  * Electrostatics interaction: Coulomb
1095  * VdW interaction:            CubicSplineTable
1096  * Geometry:                   Water4-Water4
1097  * Calculate force/pot:        Force
1098  */
1099 void
1100 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_sse2_single
1101                     (t_nblist * gmx_restrict                nlist,
1102                      rvec * gmx_restrict                    xx,
1103                      rvec * gmx_restrict                    ff,
1104                      t_forcerec * gmx_restrict              fr,
1105                      t_mdatoms * gmx_restrict               mdatoms,
1106                      nb_kernel_data_t * gmx_restrict        kernel_data,
1107                      t_nrnb * gmx_restrict                  nrnb)
1108 {
1109     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1110      * just 0 for non-waters.
1111      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1112      * jnr indices corresponding to data put in the four positions in the SIMD register.
1113      */
1114     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1115     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1116     int              jnrA,jnrB,jnrC,jnrD;
1117     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1118     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1119     real             shX,shY,shZ,rcutoff_scalar;
1120     real             *shiftvec,*fshift,*x,*f;
1121     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1122     int              vdwioffset0;
1123     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1124     int              vdwioffset1;
1125     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1126     int              vdwioffset2;
1127     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1128     int              vdwioffset3;
1129     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1130     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1131     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1132     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1133     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1134     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1135     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1136     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1137     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1138     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1139     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1140     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1141     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1142     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1143     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1144     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1145     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1146     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1147     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1148     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1149     real             *charge;
1150     int              nvdwtype;
1151     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1152     int              *vdwtype;
1153     real             *vdwparam;
1154     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1155     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1156     __m128i          vfitab;
1157     __m128i          ifour       = _mm_set1_epi32(4);
1158     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1159     real             *vftab;
1160     __m128           dummy_mask,cutoff_mask;
1161     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1162     __m128           one     = _mm_set1_ps(1.0);
1163     __m128           two     = _mm_set1_ps(2.0);
1164     x                = xx[0];
1165     f                = ff[0];
1166
1167     nri              = nlist->nri;
1168     iinr             = nlist->iinr;
1169     jindex           = nlist->jindex;
1170     jjnr             = nlist->jjnr;
1171     shiftidx         = nlist->shift;
1172     gid              = nlist->gid;
1173     shiftvec         = fr->shift_vec[0];
1174     fshift           = fr->fshift[0];
1175     facel            = _mm_set1_ps(fr->epsfac);
1176     charge           = mdatoms->chargeA;
1177     nvdwtype         = fr->ntype;
1178     vdwparam         = fr->nbfp;
1179     vdwtype          = mdatoms->typeA;
1180
1181     vftab            = kernel_data->table_vdw->data;
1182     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1183
1184     /* Setup water-specific parameters */
1185     inr              = nlist->iinr[0];
1186     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1187     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1188     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1189     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1190
1191     jq1              = _mm_set1_ps(charge[inr+1]);
1192     jq2              = _mm_set1_ps(charge[inr+2]);
1193     jq3              = _mm_set1_ps(charge[inr+3]);
1194     vdwjidx0A        = 2*vdwtype[inr+0];
1195     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1196     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1197     qq11             = _mm_mul_ps(iq1,jq1);
1198     qq12             = _mm_mul_ps(iq1,jq2);
1199     qq13             = _mm_mul_ps(iq1,jq3);
1200     qq21             = _mm_mul_ps(iq2,jq1);
1201     qq22             = _mm_mul_ps(iq2,jq2);
1202     qq23             = _mm_mul_ps(iq2,jq3);
1203     qq31             = _mm_mul_ps(iq3,jq1);
1204     qq32             = _mm_mul_ps(iq3,jq2);
1205     qq33             = _mm_mul_ps(iq3,jq3);
1206
1207     /* Avoid stupid compiler warnings */
1208     jnrA = jnrB = jnrC = jnrD = 0;
1209     j_coord_offsetA = 0;
1210     j_coord_offsetB = 0;
1211     j_coord_offsetC = 0;
1212     j_coord_offsetD = 0;
1213
1214     outeriter        = 0;
1215     inneriter        = 0;
1216
1217     /* Start outer loop over neighborlists */
1218     for(iidx=0; iidx<nri; iidx++)
1219     {
1220         /* Load shift vector for this list */
1221         i_shift_offset   = DIM*shiftidx[iidx];
1222         shX              = shiftvec[i_shift_offset+XX];
1223         shY              = shiftvec[i_shift_offset+YY];
1224         shZ              = shiftvec[i_shift_offset+ZZ];
1225
1226         /* Load limits for loop over neighbors */
1227         j_index_start    = jindex[iidx];
1228         j_index_end      = jindex[iidx+1];
1229
1230         /* Get outer coordinate index */
1231         inr              = iinr[iidx];
1232         i_coord_offset   = DIM*inr;
1233
1234         /* Load i particle coords and add shift vector */
1235         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1236         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1237         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1238         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1239         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1240         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1241         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1242         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1243         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1244         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1245         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1246         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1247
1248         fix0             = _mm_setzero_ps();
1249         fiy0             = _mm_setzero_ps();
1250         fiz0             = _mm_setzero_ps();
1251         fix1             = _mm_setzero_ps();
1252         fiy1             = _mm_setzero_ps();
1253         fiz1             = _mm_setzero_ps();
1254         fix2             = _mm_setzero_ps();
1255         fiy2             = _mm_setzero_ps();
1256         fiz2             = _mm_setzero_ps();
1257         fix3             = _mm_setzero_ps();
1258         fiy3             = _mm_setzero_ps();
1259         fiz3             = _mm_setzero_ps();
1260
1261         /* Start inner kernel loop */
1262         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1263         {
1264
1265             /* Get j neighbor index, and coordinate index */
1266             jnrA             = jjnr[jidx];
1267             jnrB             = jjnr[jidx+1];
1268             jnrC             = jjnr[jidx+2];
1269             jnrD             = jjnr[jidx+3];
1270
1271             j_coord_offsetA  = DIM*jnrA;
1272             j_coord_offsetB  = DIM*jnrB;
1273             j_coord_offsetC  = DIM*jnrC;
1274             j_coord_offsetD  = DIM*jnrD;
1275
1276             /* load j atom coordinates */
1277             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1278                                               x+j_coord_offsetC,x+j_coord_offsetD,
1279                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1280                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1281
1282             /* Calculate displacement vector */
1283             dx00             = _mm_sub_ps(ix0,jx0);
1284             dy00             = _mm_sub_ps(iy0,jy0);
1285             dz00             = _mm_sub_ps(iz0,jz0);
1286             dx11             = _mm_sub_ps(ix1,jx1);
1287             dy11             = _mm_sub_ps(iy1,jy1);
1288             dz11             = _mm_sub_ps(iz1,jz1);
1289             dx12             = _mm_sub_ps(ix1,jx2);
1290             dy12             = _mm_sub_ps(iy1,jy2);
1291             dz12             = _mm_sub_ps(iz1,jz2);
1292             dx13             = _mm_sub_ps(ix1,jx3);
1293             dy13             = _mm_sub_ps(iy1,jy3);
1294             dz13             = _mm_sub_ps(iz1,jz3);
1295             dx21             = _mm_sub_ps(ix2,jx1);
1296             dy21             = _mm_sub_ps(iy2,jy1);
1297             dz21             = _mm_sub_ps(iz2,jz1);
1298             dx22             = _mm_sub_ps(ix2,jx2);
1299             dy22             = _mm_sub_ps(iy2,jy2);
1300             dz22             = _mm_sub_ps(iz2,jz2);
1301             dx23             = _mm_sub_ps(ix2,jx3);
1302             dy23             = _mm_sub_ps(iy2,jy3);
1303             dz23             = _mm_sub_ps(iz2,jz3);
1304             dx31             = _mm_sub_ps(ix3,jx1);
1305             dy31             = _mm_sub_ps(iy3,jy1);
1306             dz31             = _mm_sub_ps(iz3,jz1);
1307             dx32             = _mm_sub_ps(ix3,jx2);
1308             dy32             = _mm_sub_ps(iy3,jy2);
1309             dz32             = _mm_sub_ps(iz3,jz2);
1310             dx33             = _mm_sub_ps(ix3,jx3);
1311             dy33             = _mm_sub_ps(iy3,jy3);
1312             dz33             = _mm_sub_ps(iz3,jz3);
1313
1314             /* Calculate squared distance and things based on it */
1315             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1316             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1317             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1318             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1319             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1320             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1321             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1322             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1323             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1324             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1325
1326             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1327             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1328             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1329             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1330             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1331             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1332             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1333             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1334             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1335             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1336
1337             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1338             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1339             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1340             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1341             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1342             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1343             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1344             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1345             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1346
1347             fjx0             = _mm_setzero_ps();
1348             fjy0             = _mm_setzero_ps();
1349             fjz0             = _mm_setzero_ps();
1350             fjx1             = _mm_setzero_ps();
1351             fjy1             = _mm_setzero_ps();
1352             fjz1             = _mm_setzero_ps();
1353             fjx2             = _mm_setzero_ps();
1354             fjy2             = _mm_setzero_ps();
1355             fjz2             = _mm_setzero_ps();
1356             fjx3             = _mm_setzero_ps();
1357             fjy3             = _mm_setzero_ps();
1358             fjz3             = _mm_setzero_ps();
1359
1360             /**************************
1361              * CALCULATE INTERACTIONS *
1362              **************************/
1363
1364             r00              = _mm_mul_ps(rsq00,rinv00);
1365
1366             /* Calculate table index by multiplying r with table scale and truncate to integer */
1367             rt               = _mm_mul_ps(r00,vftabscale);
1368             vfitab           = _mm_cvttps_epi32(rt);
1369             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1370             vfitab           = _mm_slli_epi32(vfitab,3);
1371
1372             /* CUBIC SPLINE TABLE DISPERSION */
1373             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1374             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1375             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1376             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1377             _MM_TRANSPOSE4_PS(Y,F,G,H);
1378             Heps             = _mm_mul_ps(vfeps,H);
1379             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1380             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1381             fvdw6            = _mm_mul_ps(c6_00,FF);
1382
1383             /* CUBIC SPLINE TABLE REPULSION */
1384             vfitab           = _mm_add_epi32(vfitab,ifour);
1385             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1386             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1387             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1388             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1389             _MM_TRANSPOSE4_PS(Y,F,G,H);
1390             Heps             = _mm_mul_ps(vfeps,H);
1391             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1392             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1393             fvdw12           = _mm_mul_ps(c12_00,FF);
1394             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1395
1396             fscal            = fvdw;
1397
1398             /* Calculate temporary vectorial force */
1399             tx               = _mm_mul_ps(fscal,dx00);
1400             ty               = _mm_mul_ps(fscal,dy00);
1401             tz               = _mm_mul_ps(fscal,dz00);
1402
1403             /* Update vectorial force */
1404             fix0             = _mm_add_ps(fix0,tx);
1405             fiy0             = _mm_add_ps(fiy0,ty);
1406             fiz0             = _mm_add_ps(fiz0,tz);
1407
1408             fjx0             = _mm_add_ps(fjx0,tx);
1409             fjy0             = _mm_add_ps(fjy0,ty);
1410             fjz0             = _mm_add_ps(fjz0,tz);
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             /* COULOMB ELECTROSTATICS */
1417             velec            = _mm_mul_ps(qq11,rinv11);
1418             felec            = _mm_mul_ps(velec,rinvsq11);
1419
1420             fscal            = felec;
1421
1422             /* Calculate temporary vectorial force */
1423             tx               = _mm_mul_ps(fscal,dx11);
1424             ty               = _mm_mul_ps(fscal,dy11);
1425             tz               = _mm_mul_ps(fscal,dz11);
1426
1427             /* Update vectorial force */
1428             fix1             = _mm_add_ps(fix1,tx);
1429             fiy1             = _mm_add_ps(fiy1,ty);
1430             fiz1             = _mm_add_ps(fiz1,tz);
1431
1432             fjx1             = _mm_add_ps(fjx1,tx);
1433             fjy1             = _mm_add_ps(fjy1,ty);
1434             fjz1             = _mm_add_ps(fjz1,tz);
1435
1436             /**************************
1437              * CALCULATE INTERACTIONS *
1438              **************************/
1439
1440             /* COULOMB ELECTROSTATICS */
1441             velec            = _mm_mul_ps(qq12,rinv12);
1442             felec            = _mm_mul_ps(velec,rinvsq12);
1443
1444             fscal            = felec;
1445
1446             /* Calculate temporary vectorial force */
1447             tx               = _mm_mul_ps(fscal,dx12);
1448             ty               = _mm_mul_ps(fscal,dy12);
1449             tz               = _mm_mul_ps(fscal,dz12);
1450
1451             /* Update vectorial force */
1452             fix1             = _mm_add_ps(fix1,tx);
1453             fiy1             = _mm_add_ps(fiy1,ty);
1454             fiz1             = _mm_add_ps(fiz1,tz);
1455
1456             fjx2             = _mm_add_ps(fjx2,tx);
1457             fjy2             = _mm_add_ps(fjy2,ty);
1458             fjz2             = _mm_add_ps(fjz2,tz);
1459
1460             /**************************
1461              * CALCULATE INTERACTIONS *
1462              **************************/
1463
1464             /* COULOMB ELECTROSTATICS */
1465             velec            = _mm_mul_ps(qq13,rinv13);
1466             felec            = _mm_mul_ps(velec,rinvsq13);
1467
1468             fscal            = felec;
1469
1470             /* Calculate temporary vectorial force */
1471             tx               = _mm_mul_ps(fscal,dx13);
1472             ty               = _mm_mul_ps(fscal,dy13);
1473             tz               = _mm_mul_ps(fscal,dz13);
1474
1475             /* Update vectorial force */
1476             fix1             = _mm_add_ps(fix1,tx);
1477             fiy1             = _mm_add_ps(fiy1,ty);
1478             fiz1             = _mm_add_ps(fiz1,tz);
1479
1480             fjx3             = _mm_add_ps(fjx3,tx);
1481             fjy3             = _mm_add_ps(fjy3,ty);
1482             fjz3             = _mm_add_ps(fjz3,tz);
1483
1484             /**************************
1485              * CALCULATE INTERACTIONS *
1486              **************************/
1487
1488             /* COULOMB ELECTROSTATICS */
1489             velec            = _mm_mul_ps(qq21,rinv21);
1490             felec            = _mm_mul_ps(velec,rinvsq21);
1491
1492             fscal            = felec;
1493
1494             /* Calculate temporary vectorial force */
1495             tx               = _mm_mul_ps(fscal,dx21);
1496             ty               = _mm_mul_ps(fscal,dy21);
1497             tz               = _mm_mul_ps(fscal,dz21);
1498
1499             /* Update vectorial force */
1500             fix2             = _mm_add_ps(fix2,tx);
1501             fiy2             = _mm_add_ps(fiy2,ty);
1502             fiz2             = _mm_add_ps(fiz2,tz);
1503
1504             fjx1             = _mm_add_ps(fjx1,tx);
1505             fjy1             = _mm_add_ps(fjy1,ty);
1506             fjz1             = _mm_add_ps(fjz1,tz);
1507
1508             /**************************
1509              * CALCULATE INTERACTIONS *
1510              **************************/
1511
1512             /* COULOMB ELECTROSTATICS */
1513             velec            = _mm_mul_ps(qq22,rinv22);
1514             felec            = _mm_mul_ps(velec,rinvsq22);
1515
1516             fscal            = felec;
1517
1518             /* Calculate temporary vectorial force */
1519             tx               = _mm_mul_ps(fscal,dx22);
1520             ty               = _mm_mul_ps(fscal,dy22);
1521             tz               = _mm_mul_ps(fscal,dz22);
1522
1523             /* Update vectorial force */
1524             fix2             = _mm_add_ps(fix2,tx);
1525             fiy2             = _mm_add_ps(fiy2,ty);
1526             fiz2             = _mm_add_ps(fiz2,tz);
1527
1528             fjx2             = _mm_add_ps(fjx2,tx);
1529             fjy2             = _mm_add_ps(fjy2,ty);
1530             fjz2             = _mm_add_ps(fjz2,tz);
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             /* COULOMB ELECTROSTATICS */
1537             velec            = _mm_mul_ps(qq23,rinv23);
1538             felec            = _mm_mul_ps(velec,rinvsq23);
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             /* COULOMB ELECTROSTATICS */
1561             velec            = _mm_mul_ps(qq31,rinv31);
1562             felec            = _mm_mul_ps(velec,rinvsq31);
1563
1564             fscal            = felec;
1565
1566             /* Calculate temporary vectorial force */
1567             tx               = _mm_mul_ps(fscal,dx31);
1568             ty               = _mm_mul_ps(fscal,dy31);
1569             tz               = _mm_mul_ps(fscal,dz31);
1570
1571             /* Update vectorial force */
1572             fix3             = _mm_add_ps(fix3,tx);
1573             fiy3             = _mm_add_ps(fiy3,ty);
1574             fiz3             = _mm_add_ps(fiz3,tz);
1575
1576             fjx1             = _mm_add_ps(fjx1,tx);
1577             fjy1             = _mm_add_ps(fjy1,ty);
1578             fjz1             = _mm_add_ps(fjz1,tz);
1579
1580             /**************************
1581              * CALCULATE INTERACTIONS *
1582              **************************/
1583
1584             /* COULOMB ELECTROSTATICS */
1585             velec            = _mm_mul_ps(qq32,rinv32);
1586             felec            = _mm_mul_ps(velec,rinvsq32);
1587
1588             fscal            = felec;
1589
1590             /* Calculate temporary vectorial force */
1591             tx               = _mm_mul_ps(fscal,dx32);
1592             ty               = _mm_mul_ps(fscal,dy32);
1593             tz               = _mm_mul_ps(fscal,dz32);
1594
1595             /* Update vectorial force */
1596             fix3             = _mm_add_ps(fix3,tx);
1597             fiy3             = _mm_add_ps(fiy3,ty);
1598             fiz3             = _mm_add_ps(fiz3,tz);
1599
1600             fjx2             = _mm_add_ps(fjx2,tx);
1601             fjy2             = _mm_add_ps(fjy2,ty);
1602             fjz2             = _mm_add_ps(fjz2,tz);
1603
1604             /**************************
1605              * CALCULATE INTERACTIONS *
1606              **************************/
1607
1608             /* COULOMB ELECTROSTATICS */
1609             velec            = _mm_mul_ps(qq33,rinv33);
1610             felec            = _mm_mul_ps(velec,rinvsq33);
1611
1612             fscal            = felec;
1613
1614             /* Calculate temporary vectorial force */
1615             tx               = _mm_mul_ps(fscal,dx33);
1616             ty               = _mm_mul_ps(fscal,dy33);
1617             tz               = _mm_mul_ps(fscal,dz33);
1618
1619             /* Update vectorial force */
1620             fix3             = _mm_add_ps(fix3,tx);
1621             fiy3             = _mm_add_ps(fiy3,ty);
1622             fiz3             = _mm_add_ps(fiz3,tz);
1623
1624             fjx3             = _mm_add_ps(fjx3,tx);
1625             fjy3             = _mm_add_ps(fjy3,ty);
1626             fjz3             = _mm_add_ps(fjz3,tz);
1627
1628             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1629                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1630                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1631                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1632
1633             /* Inner loop uses 294 flops */
1634         }
1635
1636         if(jidx<j_index_end)
1637         {
1638
1639             /* Get j neighbor index, and coordinate index */
1640             jnrA             = jjnr[jidx];
1641             jnrB             = jjnr[jidx+1];
1642             jnrC             = jjnr[jidx+2];
1643             jnrD             = jjnr[jidx+3];
1644
1645             /* Sign of each element will be negative for non-real atoms.
1646              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1647              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1648              */
1649             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1650             jnrA       = (jnrA>=0) ? jnrA : 0;
1651             jnrB       = (jnrB>=0) ? jnrB : 0;
1652             jnrC       = (jnrC>=0) ? jnrC : 0;
1653             jnrD       = (jnrD>=0) ? jnrD : 0;
1654
1655             j_coord_offsetA  = DIM*jnrA;
1656             j_coord_offsetB  = DIM*jnrB;
1657             j_coord_offsetC  = DIM*jnrC;
1658             j_coord_offsetD  = DIM*jnrD;
1659
1660             /* load j atom coordinates */
1661             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1662                                               x+j_coord_offsetC,x+j_coord_offsetD,
1663                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1664                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1665
1666             /* Calculate displacement vector */
1667             dx00             = _mm_sub_ps(ix0,jx0);
1668             dy00             = _mm_sub_ps(iy0,jy0);
1669             dz00             = _mm_sub_ps(iz0,jz0);
1670             dx11             = _mm_sub_ps(ix1,jx1);
1671             dy11             = _mm_sub_ps(iy1,jy1);
1672             dz11             = _mm_sub_ps(iz1,jz1);
1673             dx12             = _mm_sub_ps(ix1,jx2);
1674             dy12             = _mm_sub_ps(iy1,jy2);
1675             dz12             = _mm_sub_ps(iz1,jz2);
1676             dx13             = _mm_sub_ps(ix1,jx3);
1677             dy13             = _mm_sub_ps(iy1,jy3);
1678             dz13             = _mm_sub_ps(iz1,jz3);
1679             dx21             = _mm_sub_ps(ix2,jx1);
1680             dy21             = _mm_sub_ps(iy2,jy1);
1681             dz21             = _mm_sub_ps(iz2,jz1);
1682             dx22             = _mm_sub_ps(ix2,jx2);
1683             dy22             = _mm_sub_ps(iy2,jy2);
1684             dz22             = _mm_sub_ps(iz2,jz2);
1685             dx23             = _mm_sub_ps(ix2,jx3);
1686             dy23             = _mm_sub_ps(iy2,jy3);
1687             dz23             = _mm_sub_ps(iz2,jz3);
1688             dx31             = _mm_sub_ps(ix3,jx1);
1689             dy31             = _mm_sub_ps(iy3,jy1);
1690             dz31             = _mm_sub_ps(iz3,jz1);
1691             dx32             = _mm_sub_ps(ix3,jx2);
1692             dy32             = _mm_sub_ps(iy3,jy2);
1693             dz32             = _mm_sub_ps(iz3,jz2);
1694             dx33             = _mm_sub_ps(ix3,jx3);
1695             dy33             = _mm_sub_ps(iy3,jy3);
1696             dz33             = _mm_sub_ps(iz3,jz3);
1697
1698             /* Calculate squared distance and things based on it */
1699             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1700             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1701             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1702             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1703             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1704             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1705             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1706             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1707             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1708             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1709
1710             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1711             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1712             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1713             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1714             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1715             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1716             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1717             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1718             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1719             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1720
1721             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1722             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1723             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1724             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1725             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1726             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1727             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1728             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1729             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1730
1731             fjx0             = _mm_setzero_ps();
1732             fjy0             = _mm_setzero_ps();
1733             fjz0             = _mm_setzero_ps();
1734             fjx1             = _mm_setzero_ps();
1735             fjy1             = _mm_setzero_ps();
1736             fjz1             = _mm_setzero_ps();
1737             fjx2             = _mm_setzero_ps();
1738             fjy2             = _mm_setzero_ps();
1739             fjz2             = _mm_setzero_ps();
1740             fjx3             = _mm_setzero_ps();
1741             fjy3             = _mm_setzero_ps();
1742             fjz3             = _mm_setzero_ps();
1743
1744             /**************************
1745              * CALCULATE INTERACTIONS *
1746              **************************/
1747
1748             r00              = _mm_mul_ps(rsq00,rinv00);
1749             r00              = _mm_andnot_ps(dummy_mask,r00);
1750
1751             /* Calculate table index by multiplying r with table scale and truncate to integer */
1752             rt               = _mm_mul_ps(r00,vftabscale);
1753             vfitab           = _mm_cvttps_epi32(rt);
1754             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1755             vfitab           = _mm_slli_epi32(vfitab,3);
1756
1757             /* CUBIC SPLINE TABLE DISPERSION */
1758             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1759             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1760             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1761             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1762             _MM_TRANSPOSE4_PS(Y,F,G,H);
1763             Heps             = _mm_mul_ps(vfeps,H);
1764             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1765             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1766             fvdw6            = _mm_mul_ps(c6_00,FF);
1767
1768             /* CUBIC SPLINE TABLE REPULSION */
1769             vfitab           = _mm_add_epi32(vfitab,ifour);
1770             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1771             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1772             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1773             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1774             _MM_TRANSPOSE4_PS(Y,F,G,H);
1775             Heps             = _mm_mul_ps(vfeps,H);
1776             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1777             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1778             fvdw12           = _mm_mul_ps(c12_00,FF);
1779             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1780
1781             fscal            = fvdw;
1782
1783             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1784
1785             /* Calculate temporary vectorial force */
1786             tx               = _mm_mul_ps(fscal,dx00);
1787             ty               = _mm_mul_ps(fscal,dy00);
1788             tz               = _mm_mul_ps(fscal,dz00);
1789
1790             /* Update vectorial force */
1791             fix0             = _mm_add_ps(fix0,tx);
1792             fiy0             = _mm_add_ps(fiy0,ty);
1793             fiz0             = _mm_add_ps(fiz0,tz);
1794
1795             fjx0             = _mm_add_ps(fjx0,tx);
1796             fjy0             = _mm_add_ps(fjy0,ty);
1797             fjz0             = _mm_add_ps(fjz0,tz);
1798
1799             /**************************
1800              * CALCULATE INTERACTIONS *
1801              **************************/
1802
1803             /* COULOMB ELECTROSTATICS */
1804             velec            = _mm_mul_ps(qq11,rinv11);
1805             felec            = _mm_mul_ps(velec,rinvsq11);
1806
1807             fscal            = felec;
1808
1809             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1810
1811             /* Calculate temporary vectorial force */
1812             tx               = _mm_mul_ps(fscal,dx11);
1813             ty               = _mm_mul_ps(fscal,dy11);
1814             tz               = _mm_mul_ps(fscal,dz11);
1815
1816             /* Update vectorial force */
1817             fix1             = _mm_add_ps(fix1,tx);
1818             fiy1             = _mm_add_ps(fiy1,ty);
1819             fiz1             = _mm_add_ps(fiz1,tz);
1820
1821             fjx1             = _mm_add_ps(fjx1,tx);
1822             fjy1             = _mm_add_ps(fjy1,ty);
1823             fjz1             = _mm_add_ps(fjz1,tz);
1824
1825             /**************************
1826              * CALCULATE INTERACTIONS *
1827              **************************/
1828
1829             /* COULOMB ELECTROSTATICS */
1830             velec            = _mm_mul_ps(qq12,rinv12);
1831             felec            = _mm_mul_ps(velec,rinvsq12);
1832
1833             fscal            = felec;
1834
1835             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1836
1837             /* Calculate temporary vectorial force */
1838             tx               = _mm_mul_ps(fscal,dx12);
1839             ty               = _mm_mul_ps(fscal,dy12);
1840             tz               = _mm_mul_ps(fscal,dz12);
1841
1842             /* Update vectorial force */
1843             fix1             = _mm_add_ps(fix1,tx);
1844             fiy1             = _mm_add_ps(fiy1,ty);
1845             fiz1             = _mm_add_ps(fiz1,tz);
1846
1847             fjx2             = _mm_add_ps(fjx2,tx);
1848             fjy2             = _mm_add_ps(fjy2,ty);
1849             fjz2             = _mm_add_ps(fjz2,tz);
1850
1851             /**************************
1852              * CALCULATE INTERACTIONS *
1853              **************************/
1854
1855             /* COULOMB ELECTROSTATICS */
1856             velec            = _mm_mul_ps(qq13,rinv13);
1857             felec            = _mm_mul_ps(velec,rinvsq13);
1858
1859             fscal            = felec;
1860
1861             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1862
1863             /* Calculate temporary vectorial force */
1864             tx               = _mm_mul_ps(fscal,dx13);
1865             ty               = _mm_mul_ps(fscal,dy13);
1866             tz               = _mm_mul_ps(fscal,dz13);
1867
1868             /* Update vectorial force */
1869             fix1             = _mm_add_ps(fix1,tx);
1870             fiy1             = _mm_add_ps(fiy1,ty);
1871             fiz1             = _mm_add_ps(fiz1,tz);
1872
1873             fjx3             = _mm_add_ps(fjx3,tx);
1874             fjy3             = _mm_add_ps(fjy3,ty);
1875             fjz3             = _mm_add_ps(fjz3,tz);
1876
1877             /**************************
1878              * CALCULATE INTERACTIONS *
1879              **************************/
1880
1881             /* COULOMB ELECTROSTATICS */
1882             velec            = _mm_mul_ps(qq21,rinv21);
1883             felec            = _mm_mul_ps(velec,rinvsq21);
1884
1885             fscal            = felec;
1886
1887             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1888
1889             /* Calculate temporary vectorial force */
1890             tx               = _mm_mul_ps(fscal,dx21);
1891             ty               = _mm_mul_ps(fscal,dy21);
1892             tz               = _mm_mul_ps(fscal,dz21);
1893
1894             /* Update vectorial force */
1895             fix2             = _mm_add_ps(fix2,tx);
1896             fiy2             = _mm_add_ps(fiy2,ty);
1897             fiz2             = _mm_add_ps(fiz2,tz);
1898
1899             fjx1             = _mm_add_ps(fjx1,tx);
1900             fjy1             = _mm_add_ps(fjy1,ty);
1901             fjz1             = _mm_add_ps(fjz1,tz);
1902
1903             /**************************
1904              * CALCULATE INTERACTIONS *
1905              **************************/
1906
1907             /* COULOMB ELECTROSTATICS */
1908             velec            = _mm_mul_ps(qq22,rinv22);
1909             felec            = _mm_mul_ps(velec,rinvsq22);
1910
1911             fscal            = felec;
1912
1913             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1914
1915             /* Calculate temporary vectorial force */
1916             tx               = _mm_mul_ps(fscal,dx22);
1917             ty               = _mm_mul_ps(fscal,dy22);
1918             tz               = _mm_mul_ps(fscal,dz22);
1919
1920             /* Update vectorial force */
1921             fix2             = _mm_add_ps(fix2,tx);
1922             fiy2             = _mm_add_ps(fiy2,ty);
1923             fiz2             = _mm_add_ps(fiz2,tz);
1924
1925             fjx2             = _mm_add_ps(fjx2,tx);
1926             fjy2             = _mm_add_ps(fjy2,ty);
1927             fjz2             = _mm_add_ps(fjz2,tz);
1928
1929             /**************************
1930              * CALCULATE INTERACTIONS *
1931              **************************/
1932
1933             /* COULOMB ELECTROSTATICS */
1934             velec            = _mm_mul_ps(qq23,rinv23);
1935             felec            = _mm_mul_ps(velec,rinvsq23);
1936
1937             fscal            = felec;
1938
1939             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1940
1941             /* Calculate temporary vectorial force */
1942             tx               = _mm_mul_ps(fscal,dx23);
1943             ty               = _mm_mul_ps(fscal,dy23);
1944             tz               = _mm_mul_ps(fscal,dz23);
1945
1946             /* Update vectorial force */
1947             fix2             = _mm_add_ps(fix2,tx);
1948             fiy2             = _mm_add_ps(fiy2,ty);
1949             fiz2             = _mm_add_ps(fiz2,tz);
1950
1951             fjx3             = _mm_add_ps(fjx3,tx);
1952             fjy3             = _mm_add_ps(fjy3,ty);
1953             fjz3             = _mm_add_ps(fjz3,tz);
1954
1955             /**************************
1956              * CALCULATE INTERACTIONS *
1957              **************************/
1958
1959             /* COULOMB ELECTROSTATICS */
1960             velec            = _mm_mul_ps(qq31,rinv31);
1961             felec            = _mm_mul_ps(velec,rinvsq31);
1962
1963             fscal            = felec;
1964
1965             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1966
1967             /* Calculate temporary vectorial force */
1968             tx               = _mm_mul_ps(fscal,dx31);
1969             ty               = _mm_mul_ps(fscal,dy31);
1970             tz               = _mm_mul_ps(fscal,dz31);
1971
1972             /* Update vectorial force */
1973             fix3             = _mm_add_ps(fix3,tx);
1974             fiy3             = _mm_add_ps(fiy3,ty);
1975             fiz3             = _mm_add_ps(fiz3,tz);
1976
1977             fjx1             = _mm_add_ps(fjx1,tx);
1978             fjy1             = _mm_add_ps(fjy1,ty);
1979             fjz1             = _mm_add_ps(fjz1,tz);
1980
1981             /**************************
1982              * CALCULATE INTERACTIONS *
1983              **************************/
1984
1985             /* COULOMB ELECTROSTATICS */
1986             velec            = _mm_mul_ps(qq32,rinv32);
1987             felec            = _mm_mul_ps(velec,rinvsq32);
1988
1989             fscal            = felec;
1990
1991             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1992
1993             /* Calculate temporary vectorial force */
1994             tx               = _mm_mul_ps(fscal,dx32);
1995             ty               = _mm_mul_ps(fscal,dy32);
1996             tz               = _mm_mul_ps(fscal,dz32);
1997
1998             /* Update vectorial force */
1999             fix3             = _mm_add_ps(fix3,tx);
2000             fiy3             = _mm_add_ps(fiy3,ty);
2001             fiz3             = _mm_add_ps(fiz3,tz);
2002
2003             fjx2             = _mm_add_ps(fjx2,tx);
2004             fjy2             = _mm_add_ps(fjy2,ty);
2005             fjz2             = _mm_add_ps(fjz2,tz);
2006
2007             /**************************
2008              * CALCULATE INTERACTIONS *
2009              **************************/
2010
2011             /* COULOMB ELECTROSTATICS */
2012             velec            = _mm_mul_ps(qq33,rinv33);
2013             felec            = _mm_mul_ps(velec,rinvsq33);
2014
2015             fscal            = felec;
2016
2017             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2018
2019             /* Calculate temporary vectorial force */
2020             tx               = _mm_mul_ps(fscal,dx33);
2021             ty               = _mm_mul_ps(fscal,dy33);
2022             tz               = _mm_mul_ps(fscal,dz33);
2023
2024             /* Update vectorial force */
2025             fix3             = _mm_add_ps(fix3,tx);
2026             fiy3             = _mm_add_ps(fiy3,ty);
2027             fiz3             = _mm_add_ps(fiz3,tz);
2028
2029             fjx3             = _mm_add_ps(fjx3,tx);
2030             fjy3             = _mm_add_ps(fjy3,ty);
2031             fjz3             = _mm_add_ps(fjz3,tz);
2032
2033             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
2034                                                    f+j_coord_offsetC,f+j_coord_offsetD,
2035                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2036                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2037
2038             /* Inner loop uses 295 flops */
2039         }
2040
2041         /* End of innermost loop */
2042
2043         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2044                                               f+i_coord_offset,fshift+i_shift_offset);
2045
2046         /* Increment number of inner iterations */
2047         inneriter                  += j_index_end - j_index_start;
2048
2049         /* Outer loop uses 36 flops */
2050     }
2051
2052     /* Increment number of outer iterations */
2053     outeriter        += nri;
2054
2055     /* Update outer/inner flops */
2056
2057     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*36 + inneriter*295);
2058 }