src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRF_VdwLJ_GeomW4W4_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwLJ_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              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwioffset3;
  76     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  77     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  78     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  79     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  80     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  81     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  82     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  83     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  84     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  85     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  86     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  87     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  88     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  89     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  90     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  91     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  92     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  93     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  94     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  95     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 102     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 103     __m128           dummy_mask,cutoff_mask;
 104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 105     __m128           one     = _mm_set1_ps(1.0);
 106     __m128           two     = _mm_set1_ps(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_ps(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     krf              = _mm_set1_ps(fr->ic->k_rf);
 121     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 122     crf              = _mm_set1_ps(fr->ic->c_rf);
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 132     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 133
 134     jq1              = _mm_set1_ps(charge[inr+1]);
 135     jq2              = _mm_set1_ps(charge[inr+2]);
 136     jq3              = _mm_set1_ps(charge[inr+3]);
 137     vdwjidx0A        = 2*vdwtype[inr+0];
 138     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 139     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 140     qq11             = _mm_mul_ps(iq1,jq1);
 141     qq12             = _mm_mul_ps(iq1,jq2);
 142     qq13             = _mm_mul_ps(iq1,jq3);
 143     qq21             = _mm_mul_ps(iq2,jq1);
 144     qq22             = _mm_mul_ps(iq2,jq2);
 145     qq23             = _mm_mul_ps(iq2,jq3);
 146     qq31             = _mm_mul_ps(iq3,jq1);
 147     qq32             = _mm_mul_ps(iq3,jq2);
 148     qq33             = _mm_mul_ps(iq3,jq3);
 149
 150     /* Avoid stupid compiler warnings */
 151     jnrA = jnrB = jnrC = jnrD = 0;
 152     j_coord_offsetA = 0;
 153     j_coord_offsetB = 0;
 154     j_coord_offsetC = 0;
 155     j_coord_offsetD = 0;
 156
 157     outeriter        = 0;
 158     inneriter        = 0;
 159
 160     for(iidx=0;iidx<4*DIM;iidx++)
 161     {
 162         scratch[iidx] = 0.0;
 163     }
 164
 165     /* Start outer loop over neighborlists */
 166     for(iidx=0; iidx<nri; iidx++)
 167     {
 168         /* Load shift vector for this list */
 169         i_shift_offset   = DIM*shiftidx[iidx];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 181                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 182
 183         fix0             = _mm_setzero_ps();
 184         fiy0             = _mm_setzero_ps();
 185         fiz0             = _mm_setzero_ps();
 186         fix1             = _mm_setzero_ps();
 187         fiy1             = _mm_setzero_ps();
 188         fiz1             = _mm_setzero_ps();
 189         fix2             = _mm_setzero_ps();
 190         fiy2             = _mm_setzero_ps();
 191         fiz2             = _mm_setzero_ps();
 192         fix3             = _mm_setzero_ps();
 193         fiy3             = _mm_setzero_ps();
 194         fiz3             = _mm_setzero_ps();
 195
 196         /* Reset potential sums */
 197         velecsum         = _mm_setzero_ps();
 198         vvdwsum          = _mm_setzero_ps();
 199
 200         /* Start inner kernel loop */
 201         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 202         {
 203
 204             /* Get j neighbor index, and coordinate index */
 205             jnrA             = jjnr[jidx];
 206             jnrB             = jjnr[jidx+1];
 207             jnrC             = jjnr[jidx+2];
 208             jnrD             = jjnr[jidx+3];
 209             j_coord_offsetA  = DIM*jnrA;
 210             j_coord_offsetB  = DIM*jnrB;
 211             j_coord_offsetC  = DIM*jnrC;
 212             j_coord_offsetD  = DIM*jnrD;
 213
 214             /* load j atom coordinates */
 215             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 216                                               x+j_coord_offsetC,x+j_coord_offsetD,
 217                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 218                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 219
 220             /* Calculate displacement vector */
 221             dx00             = _mm_sub_ps(ix0,jx0);
 222             dy00             = _mm_sub_ps(iy0,jy0);
 223             dz00             = _mm_sub_ps(iz0,jz0);
 224             dx11             = _mm_sub_ps(ix1,jx1);
 225             dy11             = _mm_sub_ps(iy1,jy1);
 226             dz11             = _mm_sub_ps(iz1,jz1);
 227             dx12             = _mm_sub_ps(ix1,jx2);
 228             dy12             = _mm_sub_ps(iy1,jy2);
 229             dz12             = _mm_sub_ps(iz1,jz2);
 230             dx13             = _mm_sub_ps(ix1,jx3);
 231             dy13             = _mm_sub_ps(iy1,jy3);
 232             dz13             = _mm_sub_ps(iz1,jz3);
 233             dx21             = _mm_sub_ps(ix2,jx1);
 234             dy21             = _mm_sub_ps(iy2,jy1);
 235             dz21             = _mm_sub_ps(iz2,jz1);
 236             dx22             = _mm_sub_ps(ix2,jx2);
 237             dy22             = _mm_sub_ps(iy2,jy2);
 238             dz22             = _mm_sub_ps(iz2,jz2);
 239             dx23             = _mm_sub_ps(ix2,jx3);
 240             dy23             = _mm_sub_ps(iy2,jy3);
 241             dz23             = _mm_sub_ps(iz2,jz3);
 242             dx31             = _mm_sub_ps(ix3,jx1);
 243             dy31             = _mm_sub_ps(iy3,jy1);
 244             dz31             = _mm_sub_ps(iz3,jz1);
 245             dx32             = _mm_sub_ps(ix3,jx2);
 246             dy32             = _mm_sub_ps(iy3,jy2);
 247             dz32             = _mm_sub_ps(iz3,jz2);
 248             dx33             = _mm_sub_ps(ix3,jx3);
 249             dy33             = _mm_sub_ps(iy3,jy3);
 250             dz33             = _mm_sub_ps(iz3,jz3);
 251
 252             /* Calculate squared distance and things based on it */
 253             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 254             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 255             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 256             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 257             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 258             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 259             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 260             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 261             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 262             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 263
 264             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 265             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 266             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 267             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 268             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 269             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 270             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 271             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 272             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 273
 274             rinvsq00         = gmx_mm_inv_ps(rsq00);
 275             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 276             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 277             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 278             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 279             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 280             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 281             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 282             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 283             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 284
 285             fjx0             = _mm_setzero_ps();
 286             fjy0             = _mm_setzero_ps();
 287             fjz0             = _mm_setzero_ps();
 288             fjx1             = _mm_setzero_ps();
 289             fjy1             = _mm_setzero_ps();
 290             fjz1             = _mm_setzero_ps();
 291             fjx2             = _mm_setzero_ps();
 292             fjy2             = _mm_setzero_ps();
 293             fjz2             = _mm_setzero_ps();
 294             fjx3             = _mm_setzero_ps();
 295             fjy3             = _mm_setzero_ps();
 296             fjz3             = _mm_setzero_ps();
 297
 298             /**************************
 299              * CALCULATE INTERACTIONS *
 300              **************************/
 301
 302             /* LENNARD-JONES DISPERSION/REPULSION */
 303
 304             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 305             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 306             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 307             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 308             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 312
 313             fscal            = fvdw;
 314
 315             /* Calculate temporary vectorial force */
 316             tx               = _mm_mul_ps(fscal,dx00);
 317             ty               = _mm_mul_ps(fscal,dy00);
 318             tz               = _mm_mul_ps(fscal,dz00);
 319
 320             /* Update vectorial force */
 321             fix0             = _mm_add_ps(fix0,tx);
 322             fiy0             = _mm_add_ps(fiy0,ty);
 323             fiz0             = _mm_add_ps(fiz0,tz);
 324
 325             fjx0             = _mm_add_ps(fjx0,tx);
 326             fjy0             = _mm_add_ps(fjy0,ty);
 327             fjz0             = _mm_add_ps(fjz0,tz);
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             /* REACTION-FIELD ELECTROSTATICS */
 334             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 335             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 336
 337             /* Update potential sum for this i atom from the interaction with this j atom. */
 338             velecsum         = _mm_add_ps(velecsum,velec);
 339
 340             fscal            = felec;
 341
 342             /* Calculate temporary vectorial force */
 343             tx               = _mm_mul_ps(fscal,dx11);
 344             ty               = _mm_mul_ps(fscal,dy11);
 345             tz               = _mm_mul_ps(fscal,dz11);
 346
 347             /* Update vectorial force */
 348             fix1             = _mm_add_ps(fix1,tx);
 349             fiy1             = _mm_add_ps(fiy1,ty);
 350             fiz1             = _mm_add_ps(fiz1,tz);
 351
 352             fjx1             = _mm_add_ps(fjx1,tx);
 353             fjy1             = _mm_add_ps(fjy1,ty);
 354             fjz1             = _mm_add_ps(fjz1,tz);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 362             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 363
 364             /* Update potential sum for this i atom from the interaction with this j atom. */
 365             velecsum         = _mm_add_ps(velecsum,velec);
 366
 367             fscal            = felec;
 368
 369             /* Calculate temporary vectorial force */
 370             tx               = _mm_mul_ps(fscal,dx12);
 371             ty               = _mm_mul_ps(fscal,dy12);
 372             tz               = _mm_mul_ps(fscal,dz12);
 373
 374             /* Update vectorial force */
 375             fix1             = _mm_add_ps(fix1,tx);
 376             fiy1             = _mm_add_ps(fiy1,ty);
 377             fiz1             = _mm_add_ps(fiz1,tz);
 378
 379             fjx2             = _mm_add_ps(fjx2,tx);
 380             fjy2             = _mm_add_ps(fjy2,ty);
 381             fjz2             = _mm_add_ps(fjz2,tz);
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             /* REACTION-FIELD ELECTROSTATICS */
 388             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 389             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 390
 391             /* Update potential sum for this i atom from the interaction with this j atom. */
 392             velecsum         = _mm_add_ps(velecsum,velec);
 393
 394             fscal            = felec;
 395
 396             /* Calculate temporary vectorial force */
 397             tx               = _mm_mul_ps(fscal,dx13);
 398             ty               = _mm_mul_ps(fscal,dy13);
 399             tz               = _mm_mul_ps(fscal,dz13);
 400
 401             /* Update vectorial force */
 402             fix1             = _mm_add_ps(fix1,tx);
 403             fiy1             = _mm_add_ps(fiy1,ty);
 404             fiz1             = _mm_add_ps(fiz1,tz);
 405
 406             fjx3             = _mm_add_ps(fjx3,tx);
 407             fjy3             = _mm_add_ps(fjy3,ty);
 408             fjz3             = _mm_add_ps(fjz3,tz);
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 416             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velecsum         = _mm_add_ps(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423             /* Calculate temporary vectorial force */
 424             tx               = _mm_mul_ps(fscal,dx21);
 425             ty               = _mm_mul_ps(fscal,dy21);
 426             tz               = _mm_mul_ps(fscal,dz21);
 427
 428             /* Update vectorial force */
 429             fix2             = _mm_add_ps(fix2,tx);
 430             fiy2             = _mm_add_ps(fiy2,ty);
 431             fiz2             = _mm_add_ps(fiz2,tz);
 432
 433             fjx1             = _mm_add_ps(fjx1,tx);
 434             fjy1             = _mm_add_ps(fjy1,ty);
 435             fjz1             = _mm_add_ps(fjz1,tz);
 436
 437             /**************************
 438              * CALCULATE INTERACTIONS *
 439              **************************/
 440
 441             /* REACTION-FIELD ELECTROSTATICS */
 442             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 443             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velecsum         = _mm_add_ps(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             /* Calculate temporary vectorial force */
 451             tx               = _mm_mul_ps(fscal,dx22);
 452             ty               = _mm_mul_ps(fscal,dy22);
 453             tz               = _mm_mul_ps(fscal,dz22);
 454
 455             /* Update vectorial force */
 456             fix2             = _mm_add_ps(fix2,tx);
 457             fiy2             = _mm_add_ps(fiy2,ty);
 458             fiz2             = _mm_add_ps(fiz2,tz);
 459
 460             fjx2             = _mm_add_ps(fjx2,tx);
 461             fjy2             = _mm_add_ps(fjy2,ty);
 462             fjz2             = _mm_add_ps(fjz2,tz);
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 470             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velecsum         = _mm_add_ps(velecsum,velec);
 474
 475             fscal            = felec;
 476
 477             /* Calculate temporary vectorial force */
 478             tx               = _mm_mul_ps(fscal,dx23);
 479             ty               = _mm_mul_ps(fscal,dy23);
 480             tz               = _mm_mul_ps(fscal,dz23);
 481
 482             /* Update vectorial force */
 483             fix2             = _mm_add_ps(fix2,tx);
 484             fiy2             = _mm_add_ps(fiy2,ty);
 485             fiz2             = _mm_add_ps(fiz2,tz);
 486
 487             fjx3             = _mm_add_ps(fjx3,tx);
 488             fjy3             = _mm_add_ps(fjy3,ty);
 489             fjz3             = _mm_add_ps(fjz3,tz);
 490
 491             /**************************
 492              * CALCULATE INTERACTIONS *
 493              **************************/
 494
 495             /* REACTION-FIELD ELECTROSTATICS */
 496             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 497             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 498
 499             /* Update potential sum for this i atom from the interaction with this j atom. */
 500             velecsum         = _mm_add_ps(velecsum,velec);
 501
 502             fscal            = felec;
 503
 504             /* Calculate temporary vectorial force */
 505             tx               = _mm_mul_ps(fscal,dx31);
 506             ty               = _mm_mul_ps(fscal,dy31);
 507             tz               = _mm_mul_ps(fscal,dz31);
 508
 509             /* Update vectorial force */
 510             fix3             = _mm_add_ps(fix3,tx);
 511             fiy3             = _mm_add_ps(fiy3,ty);
 512             fiz3             = _mm_add_ps(fiz3,tz);
 513
 514             fjx1             = _mm_add_ps(fjx1,tx);
 515             fjy1             = _mm_add_ps(fjy1,ty);
 516             fjz1             = _mm_add_ps(fjz1,tz);
 517
 518             /**************************
 519              * CALCULATE INTERACTIONS *
 520              **************************/
 521
 522             /* REACTION-FIELD ELECTROSTATICS */
 523             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 524             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 525
 526             /* Update potential sum for this i atom from the interaction with this j atom. */
 527             velecsum         = _mm_add_ps(velecsum,velec);
 528
 529             fscal            = felec;
 530
 531             /* Calculate temporary vectorial force */
 532             tx               = _mm_mul_ps(fscal,dx32);
 533             ty               = _mm_mul_ps(fscal,dy32);
 534             tz               = _mm_mul_ps(fscal,dz32);
 535
 536             /* Update vectorial force */
 537             fix3             = _mm_add_ps(fix3,tx);
 538             fiy3             = _mm_add_ps(fiy3,ty);
 539             fiz3             = _mm_add_ps(fiz3,tz);
 540
 541             fjx2             = _mm_add_ps(fjx2,tx);
 542             fjy2             = _mm_add_ps(fjy2,ty);
 543             fjz2             = _mm_add_ps(fjz2,tz);
 544
 545             /**************************
 546              * CALCULATE INTERACTIONS *
 547              **************************/
 548
 549             /* REACTION-FIELD ELECTROSTATICS */
 550             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 551             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 552
 553             /* Update potential sum for this i atom from the interaction with this j atom. */
 554             velecsum         = _mm_add_ps(velecsum,velec);
 555
 556             fscal            = felec;
 557
 558             /* Calculate temporary vectorial force */
 559             tx               = _mm_mul_ps(fscal,dx33);
 560             ty               = _mm_mul_ps(fscal,dy33);
 561             tz               = _mm_mul_ps(fscal,dz33);
 562
 563             /* Update vectorial force */
 564             fix3             = _mm_add_ps(fix3,tx);
 565             fiy3             = _mm_add_ps(fiy3,ty);
 566             fiz3             = _mm_add_ps(fiz3,tz);
 567
 568             fjx3             = _mm_add_ps(fjx3,tx);
 569             fjy3             = _mm_add_ps(fjy3,ty);
 570             fjz3             = _mm_add_ps(fjz3,tz);
 571
 572             fjptrA             = f+j_coord_offsetA;
 573             fjptrB             = f+j_coord_offsetB;
 574             fjptrC             = f+j_coord_offsetC;
 575             fjptrD             = f+j_coord_offsetD;
 576
 577             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 578                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 579                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 580
 581             /* Inner loop uses 323 flops */
 582         }
 583
 584         if(jidx<j_index_end)
 585         {
 586
 587             /* Get j neighbor index, and coordinate index */
 588             jnrlistA         = jjnr[jidx];
 589             jnrlistB         = jjnr[jidx+1];
 590             jnrlistC         = jjnr[jidx+2];
 591             jnrlistD         = jjnr[jidx+3];
 592             /* Sign of each element will be negative for non-real atoms.
 593              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 594              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 595              */
 596             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 597             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 598             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 599             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 600             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 601             j_coord_offsetA  = DIM*jnrA;
 602             j_coord_offsetB  = DIM*jnrB;
 603             j_coord_offsetC  = DIM*jnrC;
 604             j_coord_offsetD  = DIM*jnrD;
 605
 606             /* load j atom coordinates */
 607             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 608                                               x+j_coord_offsetC,x+j_coord_offsetD,
 609                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 610                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 611
 612             /* Calculate displacement vector */
 613             dx00             = _mm_sub_ps(ix0,jx0);
 614             dy00             = _mm_sub_ps(iy0,jy0);
 615             dz00             = _mm_sub_ps(iz0,jz0);
 616             dx11             = _mm_sub_ps(ix1,jx1);
 617             dy11             = _mm_sub_ps(iy1,jy1);
 618             dz11             = _mm_sub_ps(iz1,jz1);
 619             dx12             = _mm_sub_ps(ix1,jx2);
 620             dy12             = _mm_sub_ps(iy1,jy2);
 621             dz12             = _mm_sub_ps(iz1,jz2);
 622             dx13             = _mm_sub_ps(ix1,jx3);
 623             dy13             = _mm_sub_ps(iy1,jy3);
 624             dz13             = _mm_sub_ps(iz1,jz3);
 625             dx21             = _mm_sub_ps(ix2,jx1);
 626             dy21             = _mm_sub_ps(iy2,jy1);
 627             dz21             = _mm_sub_ps(iz2,jz1);
 628             dx22             = _mm_sub_ps(ix2,jx2);
 629             dy22             = _mm_sub_ps(iy2,jy2);
 630             dz22             = _mm_sub_ps(iz2,jz2);
 631             dx23             = _mm_sub_ps(ix2,jx3);
 632             dy23             = _mm_sub_ps(iy2,jy3);
 633             dz23             = _mm_sub_ps(iz2,jz3);
 634             dx31             = _mm_sub_ps(ix3,jx1);
 635             dy31             = _mm_sub_ps(iy3,jy1);
 636             dz31             = _mm_sub_ps(iz3,jz1);
 637             dx32             = _mm_sub_ps(ix3,jx2);
 638             dy32             = _mm_sub_ps(iy3,jy2);
 639             dz32             = _mm_sub_ps(iz3,jz2);
 640             dx33             = _mm_sub_ps(ix3,jx3);
 641             dy33             = _mm_sub_ps(iy3,jy3);
 642             dz33             = _mm_sub_ps(iz3,jz3);
 643
 644             /* Calculate squared distance and things based on it */
 645             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 646             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 647             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 648             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 649             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 650             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 651             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 652             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 653             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 654             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 655
 656             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 657             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 658             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 659             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 660             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 661             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 662             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 663             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 664             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 665
 666             rinvsq00         = gmx_mm_inv_ps(rsq00);
 667             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 668             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 669             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 670             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 671             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 672             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 673             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 674             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 675             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 676
 677             fjx0             = _mm_setzero_ps();
 678             fjy0             = _mm_setzero_ps();
 679             fjz0             = _mm_setzero_ps();
 680             fjx1             = _mm_setzero_ps();
 681             fjy1             = _mm_setzero_ps();
 682             fjz1             = _mm_setzero_ps();
 683             fjx2             = _mm_setzero_ps();
 684             fjy2             = _mm_setzero_ps();
 685             fjz2             = _mm_setzero_ps();
 686             fjx3             = _mm_setzero_ps();
 687             fjy3             = _mm_setzero_ps();
 688             fjz3             = _mm_setzero_ps();
 689
 690             /**************************
 691              * CALCULATE INTERACTIONS *
 692              **************************/
 693
 694             /* LENNARD-JONES DISPERSION/REPULSION */
 695
 696             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 697             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 698             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 699             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 700             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 701
 702             /* Update potential sum for this i atom from the interaction with this j atom. */
 703             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 704             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 705
 706             fscal            = fvdw;
 707
 708             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 709
 710             /* Calculate temporary vectorial force */
 711             tx               = _mm_mul_ps(fscal,dx00);
 712             ty               = _mm_mul_ps(fscal,dy00);
 713             tz               = _mm_mul_ps(fscal,dz00);
 714
 715             /* Update vectorial force */
 716             fix0             = _mm_add_ps(fix0,tx);
 717             fiy0             = _mm_add_ps(fiy0,ty);
 718             fiz0             = _mm_add_ps(fiz0,tz);
 719
 720             fjx0             = _mm_add_ps(fjx0,tx);
 721             fjy0             = _mm_add_ps(fjy0,ty);
 722             fjz0             = _mm_add_ps(fjz0,tz);
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             /* REACTION-FIELD ELECTROSTATICS */
 729             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 730             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 731
 732             /* Update potential sum for this i atom from the interaction with this j atom. */
 733             velec            = _mm_andnot_ps(dummy_mask,velec);
 734             velecsum         = _mm_add_ps(velecsum,velec);
 735
 736             fscal            = felec;
 737
 738             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 739
 740             /* Calculate temporary vectorial force */
 741             tx               = _mm_mul_ps(fscal,dx11);
 742             ty               = _mm_mul_ps(fscal,dy11);
 743             tz               = _mm_mul_ps(fscal,dz11);
 744
 745             /* Update vectorial force */
 746             fix1             = _mm_add_ps(fix1,tx);
 747             fiy1             = _mm_add_ps(fiy1,ty);
 748             fiz1             = _mm_add_ps(fiz1,tz);
 749
 750             fjx1             = _mm_add_ps(fjx1,tx);
 751             fjy1             = _mm_add_ps(fjy1,ty);
 752             fjz1             = _mm_add_ps(fjz1,tz);
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             /* REACTION-FIELD ELECTROSTATICS */
 759             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 760             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 761
 762             /* Update potential sum for this i atom from the interaction with this j atom. */
 763             velec            = _mm_andnot_ps(dummy_mask,velec);
 764             velecsum         = _mm_add_ps(velecsum,velec);
 765
 766             fscal            = felec;
 767
 768             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 769
 770             /* Calculate temporary vectorial force */
 771             tx               = _mm_mul_ps(fscal,dx12);
 772             ty               = _mm_mul_ps(fscal,dy12);
 773             tz               = _mm_mul_ps(fscal,dz12);
 774
 775             /* Update vectorial force */
 776             fix1             = _mm_add_ps(fix1,tx);
 777             fiy1             = _mm_add_ps(fiy1,ty);
 778             fiz1             = _mm_add_ps(fiz1,tz);
 779
 780             fjx2             = _mm_add_ps(fjx2,tx);
 781             fjy2             = _mm_add_ps(fjy2,ty);
 782             fjz2             = _mm_add_ps(fjz2,tz);
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             /* REACTION-FIELD ELECTROSTATICS */
 789             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 790             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 791
 792             /* Update potential sum for this i atom from the interaction with this j atom. */
 793             velec            = _mm_andnot_ps(dummy_mask,velec);
 794             velecsum         = _mm_add_ps(velecsum,velec);
 795
 796             fscal            = felec;
 797
 798             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 799
 800             /* Calculate temporary vectorial force */
 801             tx               = _mm_mul_ps(fscal,dx13);
 802             ty               = _mm_mul_ps(fscal,dy13);
 803             tz               = _mm_mul_ps(fscal,dz13);
 804
 805             /* Update vectorial force */
 806             fix1             = _mm_add_ps(fix1,tx);
 807             fiy1             = _mm_add_ps(fiy1,ty);
 808             fiz1             = _mm_add_ps(fiz1,tz);
 809
 810             fjx3             = _mm_add_ps(fjx3,tx);
 811             fjy3             = _mm_add_ps(fjy3,ty);
 812             fjz3             = _mm_add_ps(fjz3,tz);
 813
 814             /**************************
 815              * CALCULATE INTERACTIONS *
 816              **************************/
 817
 818             /* REACTION-FIELD ELECTROSTATICS */
 819             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 820             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 821
 822             /* Update potential sum for this i atom from the interaction with this j atom. */
 823             velec            = _mm_andnot_ps(dummy_mask,velec);
 824             velecsum         = _mm_add_ps(velecsum,velec);
 825
 826             fscal            = felec;
 827
 828             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 829
 830             /* Calculate temporary vectorial force */
 831             tx               = _mm_mul_ps(fscal,dx21);
 832             ty               = _mm_mul_ps(fscal,dy21);
 833             tz               = _mm_mul_ps(fscal,dz21);
 834
 835             /* Update vectorial force */
 836             fix2             = _mm_add_ps(fix2,tx);
 837             fiy2             = _mm_add_ps(fiy2,ty);
 838             fiz2             = _mm_add_ps(fiz2,tz);
 839
 840             fjx1             = _mm_add_ps(fjx1,tx);
 841             fjy1             = _mm_add_ps(fjy1,ty);
 842             fjz1             = _mm_add_ps(fjz1,tz);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* REACTION-FIELD ELECTROSTATICS */
 849             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 850             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 851
 852             /* Update potential sum for this i atom from the interaction with this j atom. */
 853             velec            = _mm_andnot_ps(dummy_mask,velec);
 854             velecsum         = _mm_add_ps(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 859
 860             /* Calculate temporary vectorial force */
 861             tx               = _mm_mul_ps(fscal,dx22);
 862             ty               = _mm_mul_ps(fscal,dy22);
 863             tz               = _mm_mul_ps(fscal,dz22);
 864
 865             /* Update vectorial force */
 866             fix2             = _mm_add_ps(fix2,tx);
 867             fiy2             = _mm_add_ps(fiy2,ty);
 868             fiz2             = _mm_add_ps(fiz2,tz);
 869
 870             fjx2             = _mm_add_ps(fjx2,tx);
 871             fjy2             = _mm_add_ps(fjy2,ty);
 872             fjz2             = _mm_add_ps(fjz2,tz);
 873
 874             /**************************
 875              * CALCULATE INTERACTIONS *
 876              **************************/
 877
 878             /* REACTION-FIELD ELECTROSTATICS */
 879             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 880             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 881
 882             /* Update potential sum for this i atom from the interaction with this j atom. */
 883             velec            = _mm_andnot_ps(dummy_mask,velec);
 884             velecsum         = _mm_add_ps(velecsum,velec);
 885
 886             fscal            = felec;
 887
 888             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 889
 890             /* Calculate temporary vectorial force */
 891             tx               = _mm_mul_ps(fscal,dx23);
 892             ty               = _mm_mul_ps(fscal,dy23);
 893             tz               = _mm_mul_ps(fscal,dz23);
 894
 895             /* Update vectorial force */
 896             fix2             = _mm_add_ps(fix2,tx);
 897             fiy2             = _mm_add_ps(fiy2,ty);
 898             fiz2             = _mm_add_ps(fiz2,tz);
 899
 900             fjx3             = _mm_add_ps(fjx3,tx);
 901             fjy3             = _mm_add_ps(fjy3,ty);
 902             fjz3             = _mm_add_ps(fjz3,tz);
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             /* REACTION-FIELD ELECTROSTATICS */
 909             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 910             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 911
 912             /* Update potential sum for this i atom from the interaction with this j atom. */
 913             velec            = _mm_andnot_ps(dummy_mask,velec);
 914             velecsum         = _mm_add_ps(velecsum,velec);
 915
 916             fscal            = felec;
 917
 918             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 919
 920             /* Calculate temporary vectorial force */
 921             tx               = _mm_mul_ps(fscal,dx31);
 922             ty               = _mm_mul_ps(fscal,dy31);
 923             tz               = _mm_mul_ps(fscal,dz31);
 924
 925             /* Update vectorial force */
 926             fix3             = _mm_add_ps(fix3,tx);
 927             fiy3             = _mm_add_ps(fiy3,ty);
 928             fiz3             = _mm_add_ps(fiz3,tz);
 929
 930             fjx1             = _mm_add_ps(fjx1,tx);
 931             fjy1             = _mm_add_ps(fjy1,ty);
 932             fjz1             = _mm_add_ps(fjz1,tz);
 933
 934             /**************************
 935              * CALCULATE INTERACTIONS *
 936              **************************/
 937
 938             /* REACTION-FIELD ELECTROSTATICS */
 939             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 940             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 941
 942             /* Update potential sum for this i atom from the interaction with this j atom. */
 943             velec            = _mm_andnot_ps(dummy_mask,velec);
 944             velecsum         = _mm_add_ps(velecsum,velec);
 945
 946             fscal            = felec;
 947
 948             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 949
 950             /* Calculate temporary vectorial force */
 951             tx               = _mm_mul_ps(fscal,dx32);
 952             ty               = _mm_mul_ps(fscal,dy32);
 953             tz               = _mm_mul_ps(fscal,dz32);
 954
 955             /* Update vectorial force */
 956             fix3             = _mm_add_ps(fix3,tx);
 957             fiy3             = _mm_add_ps(fiy3,ty);
 958             fiz3             = _mm_add_ps(fiz3,tz);
 959
 960             fjx2             = _mm_add_ps(fjx2,tx);
 961             fjy2             = _mm_add_ps(fjy2,ty);
 962             fjz2             = _mm_add_ps(fjz2,tz);
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             /* REACTION-FIELD ELECTROSTATICS */
 969             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 970             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 971
 972             /* Update potential sum for this i atom from the interaction with this j atom. */
 973             velec            = _mm_andnot_ps(dummy_mask,velec);
 974             velecsum         = _mm_add_ps(velecsum,velec);
 975
 976             fscal            = felec;
 977
 978             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 979
 980             /* Calculate temporary vectorial force */
 981             tx               = _mm_mul_ps(fscal,dx33);
 982             ty               = _mm_mul_ps(fscal,dy33);
 983             tz               = _mm_mul_ps(fscal,dz33);
 984
 985             /* Update vectorial force */
 986             fix3             = _mm_add_ps(fix3,tx);
 987             fiy3             = _mm_add_ps(fiy3,ty);
 988             fiz3             = _mm_add_ps(fiz3,tz);
 989
 990             fjx3             = _mm_add_ps(fjx3,tx);
 991             fjy3             = _mm_add_ps(fjy3,ty);
 992             fjz3             = _mm_add_ps(fjz3,tz);
 993
 994             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 995             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 996             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 997             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 998
 999             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1000                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1001                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1002
1003             /* Inner loop uses 323 flops */
1004         }
1005
1006         /* End of innermost loop */
1007
1008         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1009                                               f+i_coord_offset,fshift+i_shift_offset);
1010
1011         ggid                        = gid[iidx];
1012         /* Update potential energies */
1013         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1014         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1015
1016         /* Increment number of inner iterations */
1017         inneriter                  += j_index_end - j_index_start;
1018
1019         /* Outer loop uses 26 flops */
1020     }
1021
1022     /* Increment number of outer iterations */
1023     outeriter        += nri;
1024
1025     /* Update outer/inner flops */
1026
1027     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*323);
1028 }
1029 /*
1030  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_single
1031  * Electrostatics interaction: ReactionField
1032  * VdW interaction:            LennardJones
1033  * Geometry:                   Water4-Water4
1034  * Calculate force/pot:        Force
1035  */
1036 void
1037 nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_single
1038                     (t_nblist * gmx_restrict                nlist,
1039                      rvec * gmx_restrict                    xx,
1040                      rvec * gmx_restrict                    ff,
1041                      t_forcerec * gmx_restrict              fr,
1042                      t_mdatoms * gmx_restrict               mdatoms,
1043                      nb_kernel_data_t * gmx_restrict        kernel_data,
1044                      t_nrnb * gmx_restrict                  nrnb)
1045 {
1046     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1047      * just 0 for non-waters.
1048      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1049      * jnr indices corresponding to data put in the four positions in the SIMD register.
1050      */
1051     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1052     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1053     int              jnrA,jnrB,jnrC,jnrD;
1054     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1055     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1056     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1057     real             rcutoff_scalar;
1058     real             *shiftvec,*fshift,*x,*f;
1059     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1060     real             scratch[4*DIM];
1061     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1062     int              vdwioffset0;
1063     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1064     int              vdwioffset1;
1065     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1066     int              vdwioffset2;
1067     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1068     int              vdwioffset3;
1069     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1070     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1071     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1072     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1073     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1074     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1075     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1076     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1077     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1078     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1079     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1080     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1081     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1082     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1083     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1084     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1085     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1086     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1087     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1088     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1089     real             *charge;
1090     int              nvdwtype;
1091     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1092     int              *vdwtype;
1093     real             *vdwparam;
1094     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1095     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1096     __m128           dummy_mask,cutoff_mask;
1097     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1098     __m128           one     = _mm_set1_ps(1.0);
1099     __m128           two     = _mm_set1_ps(2.0);
1100     x                = xx[0];
1101     f                = ff[0];
1102
1103     nri              = nlist->nri;
1104     iinr             = nlist->iinr;
1105     jindex           = nlist->jindex;
1106     jjnr             = nlist->jjnr;
1107     shiftidx         = nlist->shift;
1108     gid              = nlist->gid;
1109     shiftvec         = fr->shift_vec[0];
1110     fshift           = fr->fshift[0];
1111     facel            = _mm_set1_ps(fr->epsfac);
1112     charge           = mdatoms->chargeA;
1113     krf              = _mm_set1_ps(fr->ic->k_rf);
1114     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1115     crf              = _mm_set1_ps(fr->ic->c_rf);
1116     nvdwtype         = fr->ntype;
1117     vdwparam         = fr->nbfp;
1118     vdwtype          = mdatoms->typeA;
1119
1120     /* Setup water-specific parameters */
1121     inr              = nlist->iinr[0];
1122     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1123     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1124     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1125     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1126
1127     jq1              = _mm_set1_ps(charge[inr+1]);
1128     jq2              = _mm_set1_ps(charge[inr+2]);
1129     jq3              = _mm_set1_ps(charge[inr+3]);
1130     vdwjidx0A        = 2*vdwtype[inr+0];
1131     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1132     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1133     qq11             = _mm_mul_ps(iq1,jq1);
1134     qq12             = _mm_mul_ps(iq1,jq2);
1135     qq13             = _mm_mul_ps(iq1,jq3);
1136     qq21             = _mm_mul_ps(iq2,jq1);
1137     qq22             = _mm_mul_ps(iq2,jq2);
1138     qq23             = _mm_mul_ps(iq2,jq3);
1139     qq31             = _mm_mul_ps(iq3,jq1);
1140     qq32             = _mm_mul_ps(iq3,jq2);
1141     qq33             = _mm_mul_ps(iq3,jq3);
1142
1143     /* Avoid stupid compiler warnings */
1144     jnrA = jnrB = jnrC = jnrD = 0;
1145     j_coord_offsetA = 0;
1146     j_coord_offsetB = 0;
1147     j_coord_offsetC = 0;
1148     j_coord_offsetD = 0;
1149
1150     outeriter        = 0;
1151     inneriter        = 0;
1152
1153     for(iidx=0;iidx<4*DIM;iidx++)
1154     {
1155         scratch[iidx] = 0.0;
1156     }
1157
1158     /* Start outer loop over neighborlists */
1159     for(iidx=0; iidx<nri; iidx++)
1160     {
1161         /* Load shift vector for this list */
1162         i_shift_offset   = DIM*shiftidx[iidx];
1163
1164         /* Load limits for loop over neighbors */
1165         j_index_start    = jindex[iidx];
1166         j_index_end      = jindex[iidx+1];
1167
1168         /* Get outer coordinate index */
1169         inr              = iinr[iidx];
1170         i_coord_offset   = DIM*inr;
1171
1172         /* Load i particle coords and add shift vector */
1173         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1174                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1175
1176         fix0             = _mm_setzero_ps();
1177         fiy0             = _mm_setzero_ps();
1178         fiz0             = _mm_setzero_ps();
1179         fix1             = _mm_setzero_ps();
1180         fiy1             = _mm_setzero_ps();
1181         fiz1             = _mm_setzero_ps();
1182         fix2             = _mm_setzero_ps();
1183         fiy2             = _mm_setzero_ps();
1184         fiz2             = _mm_setzero_ps();
1185         fix3             = _mm_setzero_ps();
1186         fiy3             = _mm_setzero_ps();
1187         fiz3             = _mm_setzero_ps();
1188
1189         /* Start inner kernel loop */
1190         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1191         {
1192
1193             /* Get j neighbor index, and coordinate index */
1194             jnrA             = jjnr[jidx];
1195             jnrB             = jjnr[jidx+1];
1196             jnrC             = jjnr[jidx+2];
1197             jnrD             = jjnr[jidx+3];
1198             j_coord_offsetA  = DIM*jnrA;
1199             j_coord_offsetB  = DIM*jnrB;
1200             j_coord_offsetC  = DIM*jnrC;
1201             j_coord_offsetD  = DIM*jnrD;
1202
1203             /* load j atom coordinates */
1204             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1205                                               x+j_coord_offsetC,x+j_coord_offsetD,
1206                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1207                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1208
1209             /* Calculate displacement vector */
1210             dx00             = _mm_sub_ps(ix0,jx0);
1211             dy00             = _mm_sub_ps(iy0,jy0);
1212             dz00             = _mm_sub_ps(iz0,jz0);
1213             dx11             = _mm_sub_ps(ix1,jx1);
1214             dy11             = _mm_sub_ps(iy1,jy1);
1215             dz11             = _mm_sub_ps(iz1,jz1);
1216             dx12             = _mm_sub_ps(ix1,jx2);
1217             dy12             = _mm_sub_ps(iy1,jy2);
1218             dz12             = _mm_sub_ps(iz1,jz2);
1219             dx13             = _mm_sub_ps(ix1,jx3);
1220             dy13             = _mm_sub_ps(iy1,jy3);
1221             dz13             = _mm_sub_ps(iz1,jz3);
1222             dx21             = _mm_sub_ps(ix2,jx1);
1223             dy21             = _mm_sub_ps(iy2,jy1);
1224             dz21             = _mm_sub_ps(iz2,jz1);
1225             dx22             = _mm_sub_ps(ix2,jx2);
1226             dy22             = _mm_sub_ps(iy2,jy2);
1227             dz22             = _mm_sub_ps(iz2,jz2);
1228             dx23             = _mm_sub_ps(ix2,jx3);
1229             dy23             = _mm_sub_ps(iy2,jy3);
1230             dz23             = _mm_sub_ps(iz2,jz3);
1231             dx31             = _mm_sub_ps(ix3,jx1);
1232             dy31             = _mm_sub_ps(iy3,jy1);
1233             dz31             = _mm_sub_ps(iz3,jz1);
1234             dx32             = _mm_sub_ps(ix3,jx2);
1235             dy32             = _mm_sub_ps(iy3,jy2);
1236             dz32             = _mm_sub_ps(iz3,jz2);
1237             dx33             = _mm_sub_ps(ix3,jx3);
1238             dy33             = _mm_sub_ps(iy3,jy3);
1239             dz33             = _mm_sub_ps(iz3,jz3);
1240
1241             /* Calculate squared distance and things based on it */
1242             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1243             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1244             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1245             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1246             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1247             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1248             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1249             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1250             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1251             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1252
1253             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1254             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1255             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1256             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1257             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1258             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1259             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1260             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1261             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1262
1263             rinvsq00         = gmx_mm_inv_ps(rsq00);
1264             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1265             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1266             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1267             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1268             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1269             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1270             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1271             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1272             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1273
1274             fjx0             = _mm_setzero_ps();
1275             fjy0             = _mm_setzero_ps();
1276             fjz0             = _mm_setzero_ps();
1277             fjx1             = _mm_setzero_ps();
1278             fjy1             = _mm_setzero_ps();
1279             fjz1             = _mm_setzero_ps();
1280             fjx2             = _mm_setzero_ps();
1281             fjy2             = _mm_setzero_ps();
1282             fjz2             = _mm_setzero_ps();
1283             fjx3             = _mm_setzero_ps();
1284             fjy3             = _mm_setzero_ps();
1285             fjz3             = _mm_setzero_ps();
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             /* LENNARD-JONES DISPERSION/REPULSION */
1292
1293             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1294             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1295
1296             fscal            = fvdw;
1297
1298             /* Calculate temporary vectorial force */
1299             tx               = _mm_mul_ps(fscal,dx00);
1300             ty               = _mm_mul_ps(fscal,dy00);
1301             tz               = _mm_mul_ps(fscal,dz00);
1302
1303             /* Update vectorial force */
1304             fix0             = _mm_add_ps(fix0,tx);
1305             fiy0             = _mm_add_ps(fiy0,ty);
1306             fiz0             = _mm_add_ps(fiz0,tz);
1307
1308             fjx0             = _mm_add_ps(fjx0,tx);
1309             fjy0             = _mm_add_ps(fjy0,ty);
1310             fjz0             = _mm_add_ps(fjz0,tz);
1311
1312             /**************************
1313              * CALCULATE INTERACTIONS *
1314              **************************/
1315
1316             /* REACTION-FIELD ELECTROSTATICS */
1317             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1318
1319             fscal            = felec;
1320
1321             /* Calculate temporary vectorial force */
1322             tx               = _mm_mul_ps(fscal,dx11);
1323             ty               = _mm_mul_ps(fscal,dy11);
1324             tz               = _mm_mul_ps(fscal,dz11);
1325
1326             /* Update vectorial force */
1327             fix1             = _mm_add_ps(fix1,tx);
1328             fiy1             = _mm_add_ps(fiy1,ty);
1329             fiz1             = _mm_add_ps(fiz1,tz);
1330
1331             fjx1             = _mm_add_ps(fjx1,tx);
1332             fjy1             = _mm_add_ps(fjy1,ty);
1333             fjz1             = _mm_add_ps(fjz1,tz);
1334
1335             /**************************
1336              * CALCULATE INTERACTIONS *
1337              **************************/
1338
1339             /* REACTION-FIELD ELECTROSTATICS */
1340             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1341
1342             fscal            = felec;
1343
1344             /* Calculate temporary vectorial force */
1345             tx               = _mm_mul_ps(fscal,dx12);
1346             ty               = _mm_mul_ps(fscal,dy12);
1347             tz               = _mm_mul_ps(fscal,dz12);
1348
1349             /* Update vectorial force */
1350             fix1             = _mm_add_ps(fix1,tx);
1351             fiy1             = _mm_add_ps(fiy1,ty);
1352             fiz1             = _mm_add_ps(fiz1,tz);
1353
1354             fjx2             = _mm_add_ps(fjx2,tx);
1355             fjy2             = _mm_add_ps(fjy2,ty);
1356             fjz2             = _mm_add_ps(fjz2,tz);
1357
1358             /**************************
1359              * CALCULATE INTERACTIONS *
1360              **************************/
1361
1362             /* REACTION-FIELD ELECTROSTATICS */
1363             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1364
1365             fscal            = felec;
1366
1367             /* Calculate temporary vectorial force */
1368             tx               = _mm_mul_ps(fscal,dx13);
1369             ty               = _mm_mul_ps(fscal,dy13);
1370             tz               = _mm_mul_ps(fscal,dz13);
1371
1372             /* Update vectorial force */
1373             fix1             = _mm_add_ps(fix1,tx);
1374             fiy1             = _mm_add_ps(fiy1,ty);
1375             fiz1             = _mm_add_ps(fiz1,tz);
1376
1377             fjx3             = _mm_add_ps(fjx3,tx);
1378             fjy3             = _mm_add_ps(fjy3,ty);
1379             fjz3             = _mm_add_ps(fjz3,tz);
1380
1381             /**************************
1382              * CALCULATE INTERACTIONS *
1383              **************************/
1384
1385             /* REACTION-FIELD ELECTROSTATICS */
1386             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1387
1388             fscal            = felec;
1389
1390             /* Calculate temporary vectorial force */
1391             tx               = _mm_mul_ps(fscal,dx21);
1392             ty               = _mm_mul_ps(fscal,dy21);
1393             tz               = _mm_mul_ps(fscal,dz21);
1394
1395             /* Update vectorial force */
1396             fix2             = _mm_add_ps(fix2,tx);
1397             fiy2             = _mm_add_ps(fiy2,ty);
1398             fiz2             = _mm_add_ps(fiz2,tz);
1399
1400             fjx1             = _mm_add_ps(fjx1,tx);
1401             fjy1             = _mm_add_ps(fjy1,ty);
1402             fjz1             = _mm_add_ps(fjz1,tz);
1403
1404             /**************************
1405              * CALCULATE INTERACTIONS *
1406              **************************/
1407
1408             /* REACTION-FIELD ELECTROSTATICS */
1409             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1410
1411             fscal            = felec;
1412
1413             /* Calculate temporary vectorial force */
1414             tx               = _mm_mul_ps(fscal,dx22);
1415             ty               = _mm_mul_ps(fscal,dy22);
1416             tz               = _mm_mul_ps(fscal,dz22);
1417
1418             /* Update vectorial force */
1419             fix2             = _mm_add_ps(fix2,tx);
1420             fiy2             = _mm_add_ps(fiy2,ty);
1421             fiz2             = _mm_add_ps(fiz2,tz);
1422
1423             fjx2             = _mm_add_ps(fjx2,tx);
1424             fjy2             = _mm_add_ps(fjy2,ty);
1425             fjz2             = _mm_add_ps(fjz2,tz);
1426
1427             /**************************
1428              * CALCULATE INTERACTIONS *
1429              **************************/
1430
1431             /* REACTION-FIELD ELECTROSTATICS */
1432             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1433
1434             fscal            = felec;
1435
1436             /* Calculate temporary vectorial force */
1437             tx               = _mm_mul_ps(fscal,dx23);
1438             ty               = _mm_mul_ps(fscal,dy23);
1439             tz               = _mm_mul_ps(fscal,dz23);
1440
1441             /* Update vectorial force */
1442             fix2             = _mm_add_ps(fix2,tx);
1443             fiy2             = _mm_add_ps(fiy2,ty);
1444             fiz2             = _mm_add_ps(fiz2,tz);
1445
1446             fjx3             = _mm_add_ps(fjx3,tx);
1447             fjy3             = _mm_add_ps(fjy3,ty);
1448             fjz3             = _mm_add_ps(fjz3,tz);
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             /* REACTION-FIELD ELECTROSTATICS */
1455             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1456
1457             fscal            = felec;
1458
1459             /* Calculate temporary vectorial force */
1460             tx               = _mm_mul_ps(fscal,dx31);
1461             ty               = _mm_mul_ps(fscal,dy31);
1462             tz               = _mm_mul_ps(fscal,dz31);
1463
1464             /* Update vectorial force */
1465             fix3             = _mm_add_ps(fix3,tx);
1466             fiy3             = _mm_add_ps(fiy3,ty);
1467             fiz3             = _mm_add_ps(fiz3,tz);
1468
1469             fjx1             = _mm_add_ps(fjx1,tx);
1470             fjy1             = _mm_add_ps(fjy1,ty);
1471             fjz1             = _mm_add_ps(fjz1,tz);
1472
1473             /**************************
1474              * CALCULATE INTERACTIONS *
1475              **************************/
1476
1477             /* REACTION-FIELD ELECTROSTATICS */
1478             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1479
1480             fscal            = felec;
1481
1482             /* Calculate temporary vectorial force */
1483             tx               = _mm_mul_ps(fscal,dx32);
1484             ty               = _mm_mul_ps(fscal,dy32);
1485             tz               = _mm_mul_ps(fscal,dz32);
1486
1487             /* Update vectorial force */
1488             fix3             = _mm_add_ps(fix3,tx);
1489             fiy3             = _mm_add_ps(fiy3,ty);
1490             fiz3             = _mm_add_ps(fiz3,tz);
1491
1492             fjx2             = _mm_add_ps(fjx2,tx);
1493             fjy2             = _mm_add_ps(fjy2,ty);
1494             fjz2             = _mm_add_ps(fjz2,tz);
1495
1496             /**************************
1497              * CALCULATE INTERACTIONS *
1498              **************************/
1499
1500             /* REACTION-FIELD ELECTROSTATICS */
1501             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1502
1503             fscal            = felec;
1504
1505             /* Calculate temporary vectorial force */
1506             tx               = _mm_mul_ps(fscal,dx33);
1507             ty               = _mm_mul_ps(fscal,dy33);
1508             tz               = _mm_mul_ps(fscal,dz33);
1509
1510             /* Update vectorial force */
1511             fix3             = _mm_add_ps(fix3,tx);
1512             fiy3             = _mm_add_ps(fiy3,ty);
1513             fiz3             = _mm_add_ps(fiz3,tz);
1514
1515             fjx3             = _mm_add_ps(fjx3,tx);
1516             fjy3             = _mm_add_ps(fjy3,ty);
1517             fjz3             = _mm_add_ps(fjz3,tz);
1518
1519             fjptrA             = f+j_coord_offsetA;
1520             fjptrB             = f+j_coord_offsetB;
1521             fjptrC             = f+j_coord_offsetC;
1522             fjptrD             = f+j_coord_offsetD;
1523
1524             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1525                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1526                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1527
1528             /* Inner loop uses 273 flops */
1529         }
1530
1531         if(jidx<j_index_end)
1532         {
1533
1534             /* Get j neighbor index, and coordinate index */
1535             jnrlistA         = jjnr[jidx];
1536             jnrlistB         = jjnr[jidx+1];
1537             jnrlistC         = jjnr[jidx+2];
1538             jnrlistD         = jjnr[jidx+3];
1539             /* Sign of each element will be negative for non-real atoms.
1540              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1541              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1542              */
1543             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1544             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1545             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1546             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1547             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1548             j_coord_offsetA  = DIM*jnrA;
1549             j_coord_offsetB  = DIM*jnrB;
1550             j_coord_offsetC  = DIM*jnrC;
1551             j_coord_offsetD  = DIM*jnrD;
1552
1553             /* load j atom coordinates */
1554             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1555                                               x+j_coord_offsetC,x+j_coord_offsetD,
1556                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1557                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1558
1559             /* Calculate displacement vector */
1560             dx00             = _mm_sub_ps(ix0,jx0);
1561             dy00             = _mm_sub_ps(iy0,jy0);
1562             dz00             = _mm_sub_ps(iz0,jz0);
1563             dx11             = _mm_sub_ps(ix1,jx1);
1564             dy11             = _mm_sub_ps(iy1,jy1);
1565             dz11             = _mm_sub_ps(iz1,jz1);
1566             dx12             = _mm_sub_ps(ix1,jx2);
1567             dy12             = _mm_sub_ps(iy1,jy2);
1568             dz12             = _mm_sub_ps(iz1,jz2);
1569             dx13             = _mm_sub_ps(ix1,jx3);
1570             dy13             = _mm_sub_ps(iy1,jy3);
1571             dz13             = _mm_sub_ps(iz1,jz3);
1572             dx21             = _mm_sub_ps(ix2,jx1);
1573             dy21             = _mm_sub_ps(iy2,jy1);
1574             dz21             = _mm_sub_ps(iz2,jz1);
1575             dx22             = _mm_sub_ps(ix2,jx2);
1576             dy22             = _mm_sub_ps(iy2,jy2);
1577             dz22             = _mm_sub_ps(iz2,jz2);
1578             dx23             = _mm_sub_ps(ix2,jx3);
1579             dy23             = _mm_sub_ps(iy2,jy3);
1580             dz23             = _mm_sub_ps(iz2,jz3);
1581             dx31             = _mm_sub_ps(ix3,jx1);
1582             dy31             = _mm_sub_ps(iy3,jy1);
1583             dz31             = _mm_sub_ps(iz3,jz1);
1584             dx32             = _mm_sub_ps(ix3,jx2);
1585             dy32             = _mm_sub_ps(iy3,jy2);
1586             dz32             = _mm_sub_ps(iz3,jz2);
1587             dx33             = _mm_sub_ps(ix3,jx3);
1588             dy33             = _mm_sub_ps(iy3,jy3);
1589             dz33             = _mm_sub_ps(iz3,jz3);
1590
1591             /* Calculate squared distance and things based on it */
1592             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1593             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1594             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1595             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1596             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1597             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1598             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1599             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1600             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1601             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1602
1603             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1604             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1605             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1606             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1607             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1608             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1609             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1610             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1611             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1612
1613             rinvsq00         = gmx_mm_inv_ps(rsq00);
1614             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1615             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1616             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1617             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1618             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1619             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1620             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1621             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1622             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1623
1624             fjx0             = _mm_setzero_ps();
1625             fjy0             = _mm_setzero_ps();
1626             fjz0             = _mm_setzero_ps();
1627             fjx1             = _mm_setzero_ps();
1628             fjy1             = _mm_setzero_ps();
1629             fjz1             = _mm_setzero_ps();
1630             fjx2             = _mm_setzero_ps();
1631             fjy2             = _mm_setzero_ps();
1632             fjz2             = _mm_setzero_ps();
1633             fjx3             = _mm_setzero_ps();
1634             fjy3             = _mm_setzero_ps();
1635             fjz3             = _mm_setzero_ps();
1636
1637             /**************************
1638              * CALCULATE INTERACTIONS *
1639              **************************/
1640
1641             /* LENNARD-JONES DISPERSION/REPULSION */
1642
1643             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1644             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1645
1646             fscal            = fvdw;
1647
1648             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1649
1650             /* Calculate temporary vectorial force */
1651             tx               = _mm_mul_ps(fscal,dx00);
1652             ty               = _mm_mul_ps(fscal,dy00);
1653             tz               = _mm_mul_ps(fscal,dz00);
1654
1655             /* Update vectorial force */
1656             fix0             = _mm_add_ps(fix0,tx);
1657             fiy0             = _mm_add_ps(fiy0,ty);
1658             fiz0             = _mm_add_ps(fiz0,tz);
1659
1660             fjx0             = _mm_add_ps(fjx0,tx);
1661             fjy0             = _mm_add_ps(fjy0,ty);
1662             fjz0             = _mm_add_ps(fjz0,tz);
1663
1664             /**************************
1665              * CALCULATE INTERACTIONS *
1666              **************************/
1667
1668             /* REACTION-FIELD ELECTROSTATICS */
1669             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1670
1671             fscal            = felec;
1672
1673             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1674
1675             /* Calculate temporary vectorial force */
1676             tx               = _mm_mul_ps(fscal,dx11);
1677             ty               = _mm_mul_ps(fscal,dy11);
1678             tz               = _mm_mul_ps(fscal,dz11);
1679
1680             /* Update vectorial force */
1681             fix1             = _mm_add_ps(fix1,tx);
1682             fiy1             = _mm_add_ps(fiy1,ty);
1683             fiz1             = _mm_add_ps(fiz1,tz);
1684
1685             fjx1             = _mm_add_ps(fjx1,tx);
1686             fjy1             = _mm_add_ps(fjy1,ty);
1687             fjz1             = _mm_add_ps(fjz1,tz);
1688
1689             /**************************
1690              * CALCULATE INTERACTIONS *
1691              **************************/
1692
1693             /* REACTION-FIELD ELECTROSTATICS */
1694             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1695
1696             fscal            = felec;
1697
1698             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1699
1700             /* Calculate temporary vectorial force */
1701             tx               = _mm_mul_ps(fscal,dx12);
1702             ty               = _mm_mul_ps(fscal,dy12);
1703             tz               = _mm_mul_ps(fscal,dz12);
1704
1705             /* Update vectorial force */
1706             fix1             = _mm_add_ps(fix1,tx);
1707             fiy1             = _mm_add_ps(fiy1,ty);
1708             fiz1             = _mm_add_ps(fiz1,tz);
1709
1710             fjx2             = _mm_add_ps(fjx2,tx);
1711             fjy2             = _mm_add_ps(fjy2,ty);
1712             fjz2             = _mm_add_ps(fjz2,tz);
1713
1714             /**************************
1715              * CALCULATE INTERACTIONS *
1716              **************************/
1717
1718             /* REACTION-FIELD ELECTROSTATICS */
1719             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1720
1721             fscal            = felec;
1722
1723             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1724
1725             /* Calculate temporary vectorial force */
1726             tx               = _mm_mul_ps(fscal,dx13);
1727             ty               = _mm_mul_ps(fscal,dy13);
1728             tz               = _mm_mul_ps(fscal,dz13);
1729
1730             /* Update vectorial force */
1731             fix1             = _mm_add_ps(fix1,tx);
1732             fiy1             = _mm_add_ps(fiy1,ty);
1733             fiz1             = _mm_add_ps(fiz1,tz);
1734
1735             fjx3             = _mm_add_ps(fjx3,tx);
1736             fjy3             = _mm_add_ps(fjy3,ty);
1737             fjz3             = _mm_add_ps(fjz3,tz);
1738
1739             /**************************
1740              * CALCULATE INTERACTIONS *
1741              **************************/
1742
1743             /* REACTION-FIELD ELECTROSTATICS */
1744             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1745
1746             fscal            = felec;
1747
1748             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1749
1750             /* Calculate temporary vectorial force */
1751             tx               = _mm_mul_ps(fscal,dx21);
1752             ty               = _mm_mul_ps(fscal,dy21);
1753             tz               = _mm_mul_ps(fscal,dz21);
1754
1755             /* Update vectorial force */
1756             fix2             = _mm_add_ps(fix2,tx);
1757             fiy2             = _mm_add_ps(fiy2,ty);
1758             fiz2             = _mm_add_ps(fiz2,tz);
1759
1760             fjx1             = _mm_add_ps(fjx1,tx);
1761             fjy1             = _mm_add_ps(fjy1,ty);
1762             fjz1             = _mm_add_ps(fjz1,tz);
1763
1764             /**************************
1765              * CALCULATE INTERACTIONS *
1766              **************************/
1767
1768             /* REACTION-FIELD ELECTROSTATICS */
1769             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1770
1771             fscal            = felec;
1772
1773             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1774
1775             /* Calculate temporary vectorial force */
1776             tx               = _mm_mul_ps(fscal,dx22);
1777             ty               = _mm_mul_ps(fscal,dy22);
1778             tz               = _mm_mul_ps(fscal,dz22);
1779
1780             /* Update vectorial force */
1781             fix2             = _mm_add_ps(fix2,tx);
1782             fiy2             = _mm_add_ps(fiy2,ty);
1783             fiz2             = _mm_add_ps(fiz2,tz);
1784
1785             fjx2             = _mm_add_ps(fjx2,tx);
1786             fjy2             = _mm_add_ps(fjy2,ty);
1787             fjz2             = _mm_add_ps(fjz2,tz);
1788
1789             /**************************
1790              * CALCULATE INTERACTIONS *
1791              **************************/
1792
1793             /* REACTION-FIELD ELECTROSTATICS */
1794             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1795
1796             fscal            = felec;
1797
1798             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1799
1800             /* Calculate temporary vectorial force */
1801             tx               = _mm_mul_ps(fscal,dx23);
1802             ty               = _mm_mul_ps(fscal,dy23);
1803             tz               = _mm_mul_ps(fscal,dz23);
1804
1805             /* Update vectorial force */
1806             fix2             = _mm_add_ps(fix2,tx);
1807             fiy2             = _mm_add_ps(fiy2,ty);
1808             fiz2             = _mm_add_ps(fiz2,tz);
1809
1810             fjx3             = _mm_add_ps(fjx3,tx);
1811             fjy3             = _mm_add_ps(fjy3,ty);
1812             fjz3             = _mm_add_ps(fjz3,tz);
1813
1814             /**************************
1815              * CALCULATE INTERACTIONS *
1816              **************************/
1817
1818             /* REACTION-FIELD ELECTROSTATICS */
1819             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1820
1821             fscal            = felec;
1822
1823             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1824
1825             /* Calculate temporary vectorial force */
1826             tx               = _mm_mul_ps(fscal,dx31);
1827             ty               = _mm_mul_ps(fscal,dy31);
1828             tz               = _mm_mul_ps(fscal,dz31);
1829
1830             /* Update vectorial force */
1831             fix3             = _mm_add_ps(fix3,tx);
1832             fiy3             = _mm_add_ps(fiy3,ty);
1833             fiz3             = _mm_add_ps(fiz3,tz);
1834
1835             fjx1             = _mm_add_ps(fjx1,tx);
1836             fjy1             = _mm_add_ps(fjy1,ty);
1837             fjz1             = _mm_add_ps(fjz1,tz);
1838
1839             /**************************
1840              * CALCULATE INTERACTIONS *
1841              **************************/
1842
1843             /* REACTION-FIELD ELECTROSTATICS */
1844             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1845
1846             fscal            = felec;
1847
1848             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1849
1850             /* Calculate temporary vectorial force */
1851             tx               = _mm_mul_ps(fscal,dx32);
1852             ty               = _mm_mul_ps(fscal,dy32);
1853             tz               = _mm_mul_ps(fscal,dz32);
1854
1855             /* Update vectorial force */
1856             fix3             = _mm_add_ps(fix3,tx);
1857             fiy3             = _mm_add_ps(fiy3,ty);
1858             fiz3             = _mm_add_ps(fiz3,tz);
1859
1860             fjx2             = _mm_add_ps(fjx2,tx);
1861             fjy2             = _mm_add_ps(fjy2,ty);
1862             fjz2             = _mm_add_ps(fjz2,tz);
1863
1864             /**************************
1865              * CALCULATE INTERACTIONS *
1866              **************************/
1867
1868             /* REACTION-FIELD ELECTROSTATICS */
1869             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1870
1871             fscal            = felec;
1872
1873             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1874
1875             /* Calculate temporary vectorial force */
1876             tx               = _mm_mul_ps(fscal,dx33);
1877             ty               = _mm_mul_ps(fscal,dy33);
1878             tz               = _mm_mul_ps(fscal,dz33);
1879
1880             /* Update vectorial force */
1881             fix3             = _mm_add_ps(fix3,tx);
1882             fiy3             = _mm_add_ps(fiy3,ty);
1883             fiz3             = _mm_add_ps(fiz3,tz);
1884
1885             fjx3             = _mm_add_ps(fjx3,tx);
1886             fjy3             = _mm_add_ps(fjy3,ty);
1887             fjz3             = _mm_add_ps(fjz3,tz);
1888
1889             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1890             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1891             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1892             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1893
1894             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1895                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1896                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1897
1898             /* Inner loop uses 273 flops */
1899         }
1900
1901         /* End of innermost loop */
1902
1903         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1904                                               f+i_coord_offset,fshift+i_shift_offset);
1905
1906         /* Increment number of inner iterations */
1907         inneriter                  += j_index_end - j_index_start;
1908
1909         /* Outer loop uses 24 flops */
1910     }
1911
1912     /* Increment number of outer iterations */
1913     outeriter        += nri;
1914
1915     /* Update outer/inner flops */
1916
1917     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1918 }