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