src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRF_VdwCSTab_GeomW3W3_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_sse2_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_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              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  73     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  75     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  77     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  94     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  95     __m128i          vfitab;
  96     __m128i          ifour       = _mm_set1_epi32(4);
  97     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  98     real             *vftab;
  99     __m128           dummy_mask,cutoff_mask;
 100     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 101     __m128           one     = _mm_set1_ps(1.0);
 102     __m128           two     = _mm_set1_ps(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm_set1_ps(fr->epsfac);
 115     charge           = mdatoms->chargeA;
 116     krf              = _mm_set1_ps(fr->ic->k_rf);
 117     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 118     crf              = _mm_set1_ps(fr->ic->c_rf);
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_vdw->data;
 124     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     jq0              = _mm_set1_ps(charge[inr+0]);
 134     jq1              = _mm_set1_ps(charge[inr+1]);
 135     jq2              = _mm_set1_ps(charge[inr+2]);
 136     vdwjidx0A        = 2*vdwtype[inr+0];
 137     qq00             = _mm_mul_ps(iq0,jq0);
 138     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 139     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 140     qq01             = _mm_mul_ps(iq0,jq1);
 141     qq02             = _mm_mul_ps(iq0,jq2);
 142     qq10             = _mm_mul_ps(iq1,jq0);
 143     qq11             = _mm_mul_ps(iq1,jq1);
 144     qq12             = _mm_mul_ps(iq1,jq2);
 145     qq20             = _mm_mul_ps(iq2,jq0);
 146     qq21             = _mm_mul_ps(iq2,jq1);
 147     qq22             = _mm_mul_ps(iq2,jq2);
 148
 149     /* Avoid stupid compiler warnings */
 150     jnrA = jnrB = jnrC = jnrD = 0;
 151     j_coord_offsetA = 0;
 152     j_coord_offsetB = 0;
 153     j_coord_offsetC = 0;
 154     j_coord_offsetD = 0;
 155
 156     outeriter        = 0;
 157     inneriter        = 0;
 158
 159     /* Start outer loop over neighborlists */
 160     for(iidx=0; iidx<nri; iidx++)
 161     {
 162         /* Load shift vector for this list */
 163         i_shift_offset   = DIM*shiftidx[iidx];
 164         shX              = shiftvec[i_shift_offset+XX];
 165         shY              = shiftvec[i_shift_offset+YY];
 166         shZ              = shiftvec[i_shift_offset+ZZ];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 178         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 179         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 180         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 181         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 182         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 183         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 184         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 185         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 186
 187         fix0             = _mm_setzero_ps();
 188         fiy0             = _mm_setzero_ps();
 189         fiz0             = _mm_setzero_ps();
 190         fix1             = _mm_setzero_ps();
 191         fiy1             = _mm_setzero_ps();
 192         fiz1             = _mm_setzero_ps();
 193         fix2             = _mm_setzero_ps();
 194         fiy2             = _mm_setzero_ps();
 195         fiz2             = _mm_setzero_ps();
 196
 197         /* Reset potential sums */
 198         velecsum         = _mm_setzero_ps();
 199         vvdwsum          = _mm_setzero_ps();
 200
 201         /* Start inner kernel loop */
 202         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 203         {
 204
 205             /* Get j neighbor index, and coordinate index */
 206             jnrA             = jjnr[jidx];
 207             jnrB             = jjnr[jidx+1];
 208             jnrC             = jjnr[jidx+2];
 209             jnrD             = jjnr[jidx+3];
 210
 211             j_coord_offsetA  = DIM*jnrA;
 212             j_coord_offsetB  = DIM*jnrB;
 213             j_coord_offsetC  = DIM*jnrC;
 214             j_coord_offsetD  = DIM*jnrD;
 215
 216             /* load j atom coordinates */
 217             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 218                                               x+j_coord_offsetC,x+j_coord_offsetD,
 219                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 220
 221             /* Calculate displacement vector */
 222             dx00             = _mm_sub_ps(ix0,jx0);
 223             dy00             = _mm_sub_ps(iy0,jy0);
 224             dz00             = _mm_sub_ps(iz0,jz0);
 225             dx01             = _mm_sub_ps(ix0,jx1);
 226             dy01             = _mm_sub_ps(iy0,jy1);
 227             dz01             = _mm_sub_ps(iz0,jz1);
 228             dx02             = _mm_sub_ps(ix0,jx2);
 229             dy02             = _mm_sub_ps(iy0,jy2);
 230             dz02             = _mm_sub_ps(iz0,jz2);
 231             dx10             = _mm_sub_ps(ix1,jx0);
 232             dy10             = _mm_sub_ps(iy1,jy0);
 233             dz10             = _mm_sub_ps(iz1,jz0);
 234             dx11             = _mm_sub_ps(ix1,jx1);
 235             dy11             = _mm_sub_ps(iy1,jy1);
 236             dz11             = _mm_sub_ps(iz1,jz1);
 237             dx12             = _mm_sub_ps(ix1,jx2);
 238             dy12             = _mm_sub_ps(iy1,jy2);
 239             dz12             = _mm_sub_ps(iz1,jz2);
 240             dx20             = _mm_sub_ps(ix2,jx0);
 241             dy20             = _mm_sub_ps(iy2,jy0);
 242             dz20             = _mm_sub_ps(iz2,jz0);
 243             dx21             = _mm_sub_ps(ix2,jx1);
 244             dy21             = _mm_sub_ps(iy2,jy1);
 245             dz21             = _mm_sub_ps(iz2,jz1);
 246             dx22             = _mm_sub_ps(ix2,jx2);
 247             dy22             = _mm_sub_ps(iy2,jy2);
 248             dz22             = _mm_sub_ps(iz2,jz2);
 249
 250             /* Calculate squared distance and things based on it */
 251             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 252             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 253             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 254             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 255             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 256             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 257             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 258             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 259             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 260
 261             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 262             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 263             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 264             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 265             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 266             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 267             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 268             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 269             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 270
 271             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 272             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 273             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 274             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 275             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 276             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 277             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 278             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 279             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 280
 281             fjx0             = _mm_setzero_ps();
 282             fjy0             = _mm_setzero_ps();
 283             fjz0             = _mm_setzero_ps();
 284             fjx1             = _mm_setzero_ps();
 285             fjy1             = _mm_setzero_ps();
 286             fjz1             = _mm_setzero_ps();
 287             fjx2             = _mm_setzero_ps();
 288             fjy2             = _mm_setzero_ps();
 289             fjz2             = _mm_setzero_ps();
 290
 291             /**************************
 292              * CALCULATE INTERACTIONS *
 293              **************************/
 294
 295             r00              = _mm_mul_ps(rsq00,rinv00);
 296
 297             /* Calculate table index by multiplying r with table scale and truncate to integer */
 298             rt               = _mm_mul_ps(r00,vftabscale);
 299             vfitab           = _mm_cvttps_epi32(rt);
 300             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 301             vfitab           = _mm_slli_epi32(vfitab,3);
 302
 303             /* REACTION-FIELD ELECTROSTATICS */
 304             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 305             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 306
 307             /* CUBIC SPLINE TABLE DISPERSION */
 308             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 309             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 310             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 311             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 312             _MM_TRANSPOSE4_PS(Y,F,G,H);
 313             Heps             = _mm_mul_ps(vfeps,H);
 314             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 315             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 316             vvdw6            = _mm_mul_ps(c6_00,VV);
 317             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 318             fvdw6            = _mm_mul_ps(c6_00,FF);
 319
 320             /* CUBIC SPLINE TABLE REPULSION */
 321             vfitab           = _mm_add_epi32(vfitab,ifour);
 322             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 323             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 324             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 325             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 326             _MM_TRANSPOSE4_PS(Y,F,G,H);
 327             Heps             = _mm_mul_ps(vfeps,H);
 328             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 329             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 330             vvdw12           = _mm_mul_ps(c12_00,VV);
 331             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 332             fvdw12           = _mm_mul_ps(c12_00,FF);
 333             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 334             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 335
 336             /* Update potential sum for this i atom from the interaction with this j atom. */
 337             velecsum         = _mm_add_ps(velecsum,velec);
 338             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 339
 340             fscal            = _mm_add_ps(felec,fvdw);
 341
 342             /* Calculate temporary vectorial force */
 343             tx               = _mm_mul_ps(fscal,dx00);
 344             ty               = _mm_mul_ps(fscal,dy00);
 345             tz               = _mm_mul_ps(fscal,dz00);
 346
 347             /* Update vectorial force */
 348             fix0             = _mm_add_ps(fix0,tx);
 349             fiy0             = _mm_add_ps(fiy0,ty);
 350             fiz0             = _mm_add_ps(fiz0,tz);
 351
 352             fjx0             = _mm_add_ps(fjx0,tx);
 353             fjy0             = _mm_add_ps(fjy0,ty);
 354             fjz0             = _mm_add_ps(fjz0,tz);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 362             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 363
 364             /* Update potential sum for this i atom from the interaction with this j atom. */
 365             velecsum         = _mm_add_ps(velecsum,velec);
 366
 367             fscal            = felec;
 368
 369             /* Calculate temporary vectorial force */
 370             tx               = _mm_mul_ps(fscal,dx01);
 371             ty               = _mm_mul_ps(fscal,dy01);
 372             tz               = _mm_mul_ps(fscal,dz01);
 373
 374             /* Update vectorial force */
 375             fix0             = _mm_add_ps(fix0,tx);
 376             fiy0             = _mm_add_ps(fiy0,ty);
 377             fiz0             = _mm_add_ps(fiz0,tz);
 378
 379             fjx1             = _mm_add_ps(fjx1,tx);
 380             fjy1             = _mm_add_ps(fjy1,ty);
 381             fjz1             = _mm_add_ps(fjz1,tz);
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             /* REACTION-FIELD ELECTROSTATICS */
 388             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 389             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 390
 391             /* Update potential sum for this i atom from the interaction with this j atom. */
 392             velecsum         = _mm_add_ps(velecsum,velec);
 393
 394             fscal            = felec;
 395
 396             /* Calculate temporary vectorial force */
 397             tx               = _mm_mul_ps(fscal,dx02);
 398             ty               = _mm_mul_ps(fscal,dy02);
 399             tz               = _mm_mul_ps(fscal,dz02);
 400
 401             /* Update vectorial force */
 402             fix0             = _mm_add_ps(fix0,tx);
 403             fiy0             = _mm_add_ps(fiy0,ty);
 404             fiz0             = _mm_add_ps(fiz0,tz);
 405
 406             fjx2             = _mm_add_ps(fjx2,tx);
 407             fjy2             = _mm_add_ps(fjy2,ty);
 408             fjz2             = _mm_add_ps(fjz2,tz);
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 416             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velecsum         = _mm_add_ps(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423             /* Calculate temporary vectorial force */
 424             tx               = _mm_mul_ps(fscal,dx10);
 425             ty               = _mm_mul_ps(fscal,dy10);
 426             tz               = _mm_mul_ps(fscal,dz10);
 427
 428             /* Update vectorial force */
 429             fix1             = _mm_add_ps(fix1,tx);
 430             fiy1             = _mm_add_ps(fiy1,ty);
 431             fiz1             = _mm_add_ps(fiz1,tz);
 432
 433             fjx0             = _mm_add_ps(fjx0,tx);
 434             fjy0             = _mm_add_ps(fjy0,ty);
 435             fjz0             = _mm_add_ps(fjz0,tz);
 436
 437             /**************************
 438              * CALCULATE INTERACTIONS *
 439              **************************/
 440
 441             /* REACTION-FIELD ELECTROSTATICS */
 442             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 443             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velecsum         = _mm_add_ps(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             /* Calculate temporary vectorial force */
 451             tx               = _mm_mul_ps(fscal,dx11);
 452             ty               = _mm_mul_ps(fscal,dy11);
 453             tz               = _mm_mul_ps(fscal,dz11);
 454
 455             /* Update vectorial force */
 456             fix1             = _mm_add_ps(fix1,tx);
 457             fiy1             = _mm_add_ps(fiy1,ty);
 458             fiz1             = _mm_add_ps(fiz1,tz);
 459
 460             fjx1             = _mm_add_ps(fjx1,tx);
 461             fjy1             = _mm_add_ps(fjy1,ty);
 462             fjz1             = _mm_add_ps(fjz1,tz);
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 470             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velecsum         = _mm_add_ps(velecsum,velec);
 474
 475             fscal            = felec;
 476
 477             /* Calculate temporary vectorial force */
 478             tx               = _mm_mul_ps(fscal,dx12);
 479             ty               = _mm_mul_ps(fscal,dy12);
 480             tz               = _mm_mul_ps(fscal,dz12);
 481
 482             /* Update vectorial force */
 483             fix1             = _mm_add_ps(fix1,tx);
 484             fiy1             = _mm_add_ps(fiy1,ty);
 485             fiz1             = _mm_add_ps(fiz1,tz);
 486
 487             fjx2             = _mm_add_ps(fjx2,tx);
 488             fjy2             = _mm_add_ps(fjy2,ty);
 489             fjz2             = _mm_add_ps(fjz2,tz);
 490
 491             /**************************
 492              * CALCULATE INTERACTIONS *
 493              **************************/
 494
 495             /* REACTION-FIELD ELECTROSTATICS */
 496             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 497             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 498
 499             /* Update potential sum for this i atom from the interaction with this j atom. */
 500             velecsum         = _mm_add_ps(velecsum,velec);
 501
 502             fscal            = felec;
 503
 504             /* Calculate temporary vectorial force */
 505             tx               = _mm_mul_ps(fscal,dx20);
 506             ty               = _mm_mul_ps(fscal,dy20);
 507             tz               = _mm_mul_ps(fscal,dz20);
 508
 509             /* Update vectorial force */
 510             fix2             = _mm_add_ps(fix2,tx);
 511             fiy2             = _mm_add_ps(fiy2,ty);
 512             fiz2             = _mm_add_ps(fiz2,tz);
 513
 514             fjx0             = _mm_add_ps(fjx0,tx);
 515             fjy0             = _mm_add_ps(fjy0,ty);
 516             fjz0             = _mm_add_ps(fjz0,tz);
 517
 518             /**************************
 519              * CALCULATE INTERACTIONS *
 520              **************************/
 521
 522             /* REACTION-FIELD ELECTROSTATICS */
 523             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 524             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 525
 526             /* Update potential sum for this i atom from the interaction with this j atom. */
 527             velecsum         = _mm_add_ps(velecsum,velec);
 528
 529             fscal            = felec;
 530
 531             /* Calculate temporary vectorial force */
 532             tx               = _mm_mul_ps(fscal,dx21);
 533             ty               = _mm_mul_ps(fscal,dy21);
 534             tz               = _mm_mul_ps(fscal,dz21);
 535
 536             /* Update vectorial force */
 537             fix2             = _mm_add_ps(fix2,tx);
 538             fiy2             = _mm_add_ps(fiy2,ty);
 539             fiz2             = _mm_add_ps(fiz2,tz);
 540
 541             fjx1             = _mm_add_ps(fjx1,tx);
 542             fjy1             = _mm_add_ps(fjy1,ty);
 543             fjz1             = _mm_add_ps(fjz1,tz);
 544
 545             /**************************
 546              * CALCULATE INTERACTIONS *
 547              **************************/
 548
 549             /* REACTION-FIELD ELECTROSTATICS */
 550             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 551             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 552
 553             /* Update potential sum for this i atom from the interaction with this j atom. */
 554             velecsum         = _mm_add_ps(velecsum,velec);
 555
 556             fscal            = felec;
 557
 558             /* Calculate temporary vectorial force */
 559             tx               = _mm_mul_ps(fscal,dx22);
 560             ty               = _mm_mul_ps(fscal,dy22);
 561             tz               = _mm_mul_ps(fscal,dz22);
 562
 563             /* Update vectorial force */
 564             fix2             = _mm_add_ps(fix2,tx);
 565             fiy2             = _mm_add_ps(fiy2,ty);
 566             fiz2             = _mm_add_ps(fiz2,tz);
 567
 568             fjx2             = _mm_add_ps(fjx2,tx);
 569             fjy2             = _mm_add_ps(fjy2,ty);
 570             fjz2             = _mm_add_ps(fjz2,tz);
 571
 572             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 573                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 574                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 575
 576             /* Inner loop uses 323 flops */
 577         }
 578
 579         if(jidx<j_index_end)
 580         {
 581
 582             /* Get j neighbor index, and coordinate index */
 583             jnrA             = jjnr[jidx];
 584             jnrB             = jjnr[jidx+1];
 585             jnrC             = jjnr[jidx+2];
 586             jnrD             = jjnr[jidx+3];
 587
 588             /* Sign of each element will be negative for non-real atoms.
 589              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 590              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 591              */
 592             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 593             jnrA       = (jnrA>=0) ? jnrA : 0;
 594             jnrB       = (jnrB>=0) ? jnrB : 0;
 595             jnrC       = (jnrC>=0) ? jnrC : 0;
 596             jnrD       = (jnrD>=0) ? jnrD : 0;
 597
 598             j_coord_offsetA  = DIM*jnrA;
 599             j_coord_offsetB  = DIM*jnrB;
 600             j_coord_offsetC  = DIM*jnrC;
 601             j_coord_offsetD  = DIM*jnrD;
 602
 603             /* load j atom coordinates */
 604             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 605                                               x+j_coord_offsetC,x+j_coord_offsetD,
 606                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 607
 608             /* Calculate displacement vector */
 609             dx00             = _mm_sub_ps(ix0,jx0);
 610             dy00             = _mm_sub_ps(iy0,jy0);
 611             dz00             = _mm_sub_ps(iz0,jz0);
 612             dx01             = _mm_sub_ps(ix0,jx1);
 613             dy01             = _mm_sub_ps(iy0,jy1);
 614             dz01             = _mm_sub_ps(iz0,jz1);
 615             dx02             = _mm_sub_ps(ix0,jx2);
 616             dy02             = _mm_sub_ps(iy0,jy2);
 617             dz02             = _mm_sub_ps(iz0,jz2);
 618             dx10             = _mm_sub_ps(ix1,jx0);
 619             dy10             = _mm_sub_ps(iy1,jy0);
 620             dz10             = _mm_sub_ps(iz1,jz0);
 621             dx11             = _mm_sub_ps(ix1,jx1);
 622             dy11             = _mm_sub_ps(iy1,jy1);
 623             dz11             = _mm_sub_ps(iz1,jz1);
 624             dx12             = _mm_sub_ps(ix1,jx2);
 625             dy12             = _mm_sub_ps(iy1,jy2);
 626             dz12             = _mm_sub_ps(iz1,jz2);
 627             dx20             = _mm_sub_ps(ix2,jx0);
 628             dy20             = _mm_sub_ps(iy2,jy0);
 629             dz20             = _mm_sub_ps(iz2,jz0);
 630             dx21             = _mm_sub_ps(ix2,jx1);
 631             dy21             = _mm_sub_ps(iy2,jy1);
 632             dz21             = _mm_sub_ps(iz2,jz1);
 633             dx22             = _mm_sub_ps(ix2,jx2);
 634             dy22             = _mm_sub_ps(iy2,jy2);
 635             dz22             = _mm_sub_ps(iz2,jz2);
 636
 637             /* Calculate squared distance and things based on it */
 638             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 639             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 640             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 641             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 642             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 643             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 644             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 645             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 646             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 647
 648             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 649             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 650             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 651             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 652             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 653             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 654             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 655             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 656             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 657
 658             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 659             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 660             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 661             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 662             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 663             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 664             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 665             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 666             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 667
 668             fjx0             = _mm_setzero_ps();
 669             fjy0             = _mm_setzero_ps();
 670             fjz0             = _mm_setzero_ps();
 671             fjx1             = _mm_setzero_ps();
 672             fjy1             = _mm_setzero_ps();
 673             fjz1             = _mm_setzero_ps();
 674             fjx2             = _mm_setzero_ps();
 675             fjy2             = _mm_setzero_ps();
 676             fjz2             = _mm_setzero_ps();
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             r00              = _mm_mul_ps(rsq00,rinv00);
 683             r00              = _mm_andnot_ps(dummy_mask,r00);
 684
 685             /* Calculate table index by multiplying r with table scale and truncate to integer */
 686             rt               = _mm_mul_ps(r00,vftabscale);
 687             vfitab           = _mm_cvttps_epi32(rt);
 688             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 689             vfitab           = _mm_slli_epi32(vfitab,3);
 690
 691             /* REACTION-FIELD ELECTROSTATICS */
 692             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 693             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 694
 695             /* CUBIC SPLINE TABLE DISPERSION */
 696             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 697             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 698             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 699             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 700             _MM_TRANSPOSE4_PS(Y,F,G,H);
 701             Heps             = _mm_mul_ps(vfeps,H);
 702             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 703             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 704             vvdw6            = _mm_mul_ps(c6_00,VV);
 705             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 706             fvdw6            = _mm_mul_ps(c6_00,FF);
 707
 708             /* CUBIC SPLINE TABLE REPULSION */
 709             vfitab           = _mm_add_epi32(vfitab,ifour);
 710             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 711             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 712             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 713             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 714             _MM_TRANSPOSE4_PS(Y,F,G,H);
 715             Heps             = _mm_mul_ps(vfeps,H);
 716             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 717             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 718             vvdw12           = _mm_mul_ps(c12_00,VV);
 719             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 720             fvdw12           = _mm_mul_ps(c12_00,FF);
 721             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 722             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 723
 724             /* Update potential sum for this i atom from the interaction with this j atom. */
 725             velec            = _mm_andnot_ps(dummy_mask,velec);
 726             velecsum         = _mm_add_ps(velecsum,velec);
 727             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 728             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 729
 730             fscal            = _mm_add_ps(felec,fvdw);
 731
 732             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 733
 734             /* Calculate temporary vectorial force */
 735             tx               = _mm_mul_ps(fscal,dx00);
 736             ty               = _mm_mul_ps(fscal,dy00);
 737             tz               = _mm_mul_ps(fscal,dz00);
 738
 739             /* Update vectorial force */
 740             fix0             = _mm_add_ps(fix0,tx);
 741             fiy0             = _mm_add_ps(fiy0,ty);
 742             fiz0             = _mm_add_ps(fiz0,tz);
 743
 744             fjx0             = _mm_add_ps(fjx0,tx);
 745             fjy0             = _mm_add_ps(fjy0,ty);
 746             fjz0             = _mm_add_ps(fjz0,tz);
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             /* REACTION-FIELD ELECTROSTATICS */
 753             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 754             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 755
 756             /* Update potential sum for this i atom from the interaction with this j atom. */
 757             velec            = _mm_andnot_ps(dummy_mask,velec);
 758             velecsum         = _mm_add_ps(velecsum,velec);
 759
 760             fscal            = felec;
 761
 762             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 763
 764             /* Calculate temporary vectorial force */
 765             tx               = _mm_mul_ps(fscal,dx01);
 766             ty               = _mm_mul_ps(fscal,dy01);
 767             tz               = _mm_mul_ps(fscal,dz01);
 768
 769             /* Update vectorial force */
 770             fix0             = _mm_add_ps(fix0,tx);
 771             fiy0             = _mm_add_ps(fiy0,ty);
 772             fiz0             = _mm_add_ps(fiz0,tz);
 773
 774             fjx1             = _mm_add_ps(fjx1,tx);
 775             fjy1             = _mm_add_ps(fjy1,ty);
 776             fjz1             = _mm_add_ps(fjz1,tz);
 777
 778             /**************************
 779              * CALCULATE INTERACTIONS *
 780              **************************/
 781
 782             /* REACTION-FIELD ELECTROSTATICS */
 783             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 784             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 785
 786             /* Update potential sum for this i atom from the interaction with this j atom. */
 787             velec            = _mm_andnot_ps(dummy_mask,velec);
 788             velecsum         = _mm_add_ps(velecsum,velec);
 789
 790             fscal            = felec;
 791
 792             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 793
 794             /* Calculate temporary vectorial force */
 795             tx               = _mm_mul_ps(fscal,dx02);
 796             ty               = _mm_mul_ps(fscal,dy02);
 797             tz               = _mm_mul_ps(fscal,dz02);
 798
 799             /* Update vectorial force */
 800             fix0             = _mm_add_ps(fix0,tx);
 801             fiy0             = _mm_add_ps(fiy0,ty);
 802             fiz0             = _mm_add_ps(fiz0,tz);
 803
 804             fjx2             = _mm_add_ps(fjx2,tx);
 805             fjy2             = _mm_add_ps(fjy2,ty);
 806             fjz2             = _mm_add_ps(fjz2,tz);
 807
 808             /**************************
 809              * CALCULATE INTERACTIONS *
 810              **************************/
 811
 812             /* REACTION-FIELD ELECTROSTATICS */
 813             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 814             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 815
 816             /* Update potential sum for this i atom from the interaction with this j atom. */
 817             velec            = _mm_andnot_ps(dummy_mask,velec);
 818             velecsum         = _mm_add_ps(velecsum,velec);
 819
 820             fscal            = felec;
 821
 822             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 823
 824             /* Calculate temporary vectorial force */
 825             tx               = _mm_mul_ps(fscal,dx10);
 826             ty               = _mm_mul_ps(fscal,dy10);
 827             tz               = _mm_mul_ps(fscal,dz10);
 828
 829             /* Update vectorial force */
 830             fix1             = _mm_add_ps(fix1,tx);
 831             fiy1             = _mm_add_ps(fiy1,ty);
 832             fiz1             = _mm_add_ps(fiz1,tz);
 833
 834             fjx0             = _mm_add_ps(fjx0,tx);
 835             fjy0             = _mm_add_ps(fjy0,ty);
 836             fjz0             = _mm_add_ps(fjz0,tz);
 837
 838             /**************************
 839              * CALCULATE INTERACTIONS *
 840              **************************/
 841
 842             /* REACTION-FIELD ELECTROSTATICS */
 843             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 844             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 845
 846             /* Update potential sum for this i atom from the interaction with this j atom. */
 847             velec            = _mm_andnot_ps(dummy_mask,velec);
 848             velecsum         = _mm_add_ps(velecsum,velec);
 849
 850             fscal            = felec;
 851
 852             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 853
 854             /* Calculate temporary vectorial force */
 855             tx               = _mm_mul_ps(fscal,dx11);
 856             ty               = _mm_mul_ps(fscal,dy11);
 857             tz               = _mm_mul_ps(fscal,dz11);
 858
 859             /* Update vectorial force */
 860             fix1             = _mm_add_ps(fix1,tx);
 861             fiy1             = _mm_add_ps(fiy1,ty);
 862             fiz1             = _mm_add_ps(fiz1,tz);
 863
 864             fjx1             = _mm_add_ps(fjx1,tx);
 865             fjy1             = _mm_add_ps(fjy1,ty);
 866             fjz1             = _mm_add_ps(fjz1,tz);
 867
 868             /**************************
 869              * CALCULATE INTERACTIONS *
 870              **************************/
 871
 872             /* REACTION-FIELD ELECTROSTATICS */
 873             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 874             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 875
 876             /* Update potential sum for this i atom from the interaction with this j atom. */
 877             velec            = _mm_andnot_ps(dummy_mask,velec);
 878             velecsum         = _mm_add_ps(velecsum,velec);
 879
 880             fscal            = felec;
 881
 882             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 883
 884             /* Calculate temporary vectorial force */
 885             tx               = _mm_mul_ps(fscal,dx12);
 886             ty               = _mm_mul_ps(fscal,dy12);
 887             tz               = _mm_mul_ps(fscal,dz12);
 888
 889             /* Update vectorial force */
 890             fix1             = _mm_add_ps(fix1,tx);
 891             fiy1             = _mm_add_ps(fiy1,ty);
 892             fiz1             = _mm_add_ps(fiz1,tz);
 893
 894             fjx2             = _mm_add_ps(fjx2,tx);
 895             fjy2             = _mm_add_ps(fjy2,ty);
 896             fjz2             = _mm_add_ps(fjz2,tz);
 897
 898             /**************************
 899              * CALCULATE INTERACTIONS *
 900              **************************/
 901
 902             /* REACTION-FIELD ELECTROSTATICS */
 903             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 904             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 905
 906             /* Update potential sum for this i atom from the interaction with this j atom. */
 907             velec            = _mm_andnot_ps(dummy_mask,velec);
 908             velecsum         = _mm_add_ps(velecsum,velec);
 909
 910             fscal            = felec;
 911
 912             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 913
 914             /* Calculate temporary vectorial force */
 915             tx               = _mm_mul_ps(fscal,dx20);
 916             ty               = _mm_mul_ps(fscal,dy20);
 917             tz               = _mm_mul_ps(fscal,dz20);
 918
 919             /* Update vectorial force */
 920             fix2             = _mm_add_ps(fix2,tx);
 921             fiy2             = _mm_add_ps(fiy2,ty);
 922             fiz2             = _mm_add_ps(fiz2,tz);
 923
 924             fjx0             = _mm_add_ps(fjx0,tx);
 925             fjy0             = _mm_add_ps(fjy0,ty);
 926             fjz0             = _mm_add_ps(fjz0,tz);
 927
 928             /**************************
 929              * CALCULATE INTERACTIONS *
 930              **************************/
 931
 932             /* REACTION-FIELD ELECTROSTATICS */
 933             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 934             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 935
 936             /* Update potential sum for this i atom from the interaction with this j atom. */
 937             velec            = _mm_andnot_ps(dummy_mask,velec);
 938             velecsum         = _mm_add_ps(velecsum,velec);
 939
 940             fscal            = felec;
 941
 942             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 943
 944             /* Calculate temporary vectorial force */
 945             tx               = _mm_mul_ps(fscal,dx21);
 946             ty               = _mm_mul_ps(fscal,dy21);
 947             tz               = _mm_mul_ps(fscal,dz21);
 948
 949             /* Update vectorial force */
 950             fix2             = _mm_add_ps(fix2,tx);
 951             fiy2             = _mm_add_ps(fiy2,ty);
 952             fiz2             = _mm_add_ps(fiz2,tz);
 953
 954             fjx1             = _mm_add_ps(fjx1,tx);
 955             fjy1             = _mm_add_ps(fjy1,ty);
 956             fjz1             = _mm_add_ps(fjz1,tz);
 957
 958             /**************************
 959              * CALCULATE INTERACTIONS *
 960              **************************/
 961
 962             /* REACTION-FIELD ELECTROSTATICS */
 963             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 964             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 965
 966             /* Update potential sum for this i atom from the interaction with this j atom. */
 967             velec            = _mm_andnot_ps(dummy_mask,velec);
 968             velecsum         = _mm_add_ps(velecsum,velec);
 969
 970             fscal            = felec;
 971
 972             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 973
 974             /* Calculate temporary vectorial force */
 975             tx               = _mm_mul_ps(fscal,dx22);
 976             ty               = _mm_mul_ps(fscal,dy22);
 977             tz               = _mm_mul_ps(fscal,dz22);
 978
 979             /* Update vectorial force */
 980             fix2             = _mm_add_ps(fix2,tx);
 981             fiy2             = _mm_add_ps(fiy2,ty);
 982             fiz2             = _mm_add_ps(fiz2,tz);
 983
 984             fjx2             = _mm_add_ps(fjx2,tx);
 985             fjy2             = _mm_add_ps(fjy2,ty);
 986             fjz2             = _mm_add_ps(fjz2,tz);
 987
 988             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 989                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 990                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 991
 992             /* Inner loop uses 324 flops */
 993         }
 994
 995         /* End of innermost loop */
 996
 997         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 998                                               f+i_coord_offset,fshift+i_shift_offset);
 999
1000         ggid                        = gid[iidx];
1001         /* Update potential energies */
1002         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1003         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1004
1005         /* Increment number of inner iterations */
1006         inneriter                  += j_index_end - j_index_start;
1007
1008         /* Outer loop uses 29 flops */
1009     }
1010
1011     /* Increment number of outer iterations */
1012     outeriter        += nri;
1013
1014     /* Update outer/inner flops */
1015
1016     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*29 + inneriter*324);
1017 }
1018 /*
1019  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_sse2_single
1020  * Electrostatics interaction: ReactionField
1021  * VdW interaction:            CubicSplineTable
1022  * Geometry:                   Water3-Water3
1023  * Calculate force/pot:        Force
1024  */
1025 void
1026 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_sse2_single
1027                     (t_nblist * gmx_restrict                nlist,
1028                      rvec * gmx_restrict                    xx,
1029                      rvec * gmx_restrict                    ff,
1030                      t_forcerec * gmx_restrict              fr,
1031                      t_mdatoms * gmx_restrict               mdatoms,
1032                      nb_kernel_data_t * gmx_restrict        kernel_data,
1033                      t_nrnb * gmx_restrict                  nrnb)
1034 {
1035     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1036      * just 0 for non-waters.
1037      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1038      * jnr indices corresponding to data put in the four positions in the SIMD register.
1039      */
1040     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1041     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1042     int              jnrA,jnrB,jnrC,jnrD;
1043     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1044     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1045     real             shX,shY,shZ,rcutoff_scalar;
1046     real             *shiftvec,*fshift,*x,*f;
1047     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1048     int              vdwioffset0;
1049     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1050     int              vdwioffset1;
1051     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1052     int              vdwioffset2;
1053     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1054     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1055     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1056     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1057     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1058     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1059     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1060     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1061     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1062     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1063     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1064     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1065     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1066     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1067     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1068     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1069     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1070     real             *charge;
1071     int              nvdwtype;
1072     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1073     int              *vdwtype;
1074     real             *vdwparam;
1075     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1076     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1077     __m128i          vfitab;
1078     __m128i          ifour       = _mm_set1_epi32(4);
1079     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1080     real             *vftab;
1081     __m128           dummy_mask,cutoff_mask;
1082     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1083     __m128           one     = _mm_set1_ps(1.0);
1084     __m128           two     = _mm_set1_ps(2.0);
1085     x                = xx[0];
1086     f                = ff[0];
1087
1088     nri              = nlist->nri;
1089     iinr             = nlist->iinr;
1090     jindex           = nlist->jindex;
1091     jjnr             = nlist->jjnr;
1092     shiftidx         = nlist->shift;
1093     gid              = nlist->gid;
1094     shiftvec         = fr->shift_vec[0];
1095     fshift           = fr->fshift[0];
1096     facel            = _mm_set1_ps(fr->epsfac);
1097     charge           = mdatoms->chargeA;
1098     krf              = _mm_set1_ps(fr->ic->k_rf);
1099     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1100     crf              = _mm_set1_ps(fr->ic->c_rf);
1101     nvdwtype         = fr->ntype;
1102     vdwparam         = fr->nbfp;
1103     vdwtype          = mdatoms->typeA;
1104
1105     vftab            = kernel_data->table_vdw->data;
1106     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1107
1108     /* Setup water-specific parameters */
1109     inr              = nlist->iinr[0];
1110     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1111     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1112     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1113     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1114
1115     jq0              = _mm_set1_ps(charge[inr+0]);
1116     jq1              = _mm_set1_ps(charge[inr+1]);
1117     jq2              = _mm_set1_ps(charge[inr+2]);
1118     vdwjidx0A        = 2*vdwtype[inr+0];
1119     qq00             = _mm_mul_ps(iq0,jq0);
1120     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1121     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1122     qq01             = _mm_mul_ps(iq0,jq1);
1123     qq02             = _mm_mul_ps(iq0,jq2);
1124     qq10             = _mm_mul_ps(iq1,jq0);
1125     qq11             = _mm_mul_ps(iq1,jq1);
1126     qq12             = _mm_mul_ps(iq1,jq2);
1127     qq20             = _mm_mul_ps(iq2,jq0);
1128     qq21             = _mm_mul_ps(iq2,jq1);
1129     qq22             = _mm_mul_ps(iq2,jq2);
1130
1131     /* Avoid stupid compiler warnings */
1132     jnrA = jnrB = jnrC = jnrD = 0;
1133     j_coord_offsetA = 0;
1134     j_coord_offsetB = 0;
1135     j_coord_offsetC = 0;
1136     j_coord_offsetD = 0;
1137
1138     outeriter        = 0;
1139     inneriter        = 0;
1140
1141     /* Start outer loop over neighborlists */
1142     for(iidx=0; iidx<nri; iidx++)
1143     {
1144         /* Load shift vector for this list */
1145         i_shift_offset   = DIM*shiftidx[iidx];
1146         shX              = shiftvec[i_shift_offset+XX];
1147         shY              = shiftvec[i_shift_offset+YY];
1148         shZ              = shiftvec[i_shift_offset+ZZ];
1149
1150         /* Load limits for loop over neighbors */
1151         j_index_start    = jindex[iidx];
1152         j_index_end      = jindex[iidx+1];
1153
1154         /* Get outer coordinate index */
1155         inr              = iinr[iidx];
1156         i_coord_offset   = DIM*inr;
1157
1158         /* Load i particle coords and add shift vector */
1159         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1160         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1161         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1162         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1163         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1164         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1165         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1166         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1167         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1168
1169         fix0             = _mm_setzero_ps();
1170         fiy0             = _mm_setzero_ps();
1171         fiz0             = _mm_setzero_ps();
1172         fix1             = _mm_setzero_ps();
1173         fiy1             = _mm_setzero_ps();
1174         fiz1             = _mm_setzero_ps();
1175         fix2             = _mm_setzero_ps();
1176         fiy2             = _mm_setzero_ps();
1177         fiz2             = _mm_setzero_ps();
1178
1179         /* Start inner kernel loop */
1180         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1181         {
1182
1183             /* Get j neighbor index, and coordinate index */
1184             jnrA             = jjnr[jidx];
1185             jnrB             = jjnr[jidx+1];
1186             jnrC             = jjnr[jidx+2];
1187             jnrD             = jjnr[jidx+3];
1188
1189             j_coord_offsetA  = DIM*jnrA;
1190             j_coord_offsetB  = DIM*jnrB;
1191             j_coord_offsetC  = DIM*jnrC;
1192             j_coord_offsetD  = DIM*jnrD;
1193
1194             /* load j atom coordinates */
1195             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1196                                               x+j_coord_offsetC,x+j_coord_offsetD,
1197                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1198
1199             /* Calculate displacement vector */
1200             dx00             = _mm_sub_ps(ix0,jx0);
1201             dy00             = _mm_sub_ps(iy0,jy0);
1202             dz00             = _mm_sub_ps(iz0,jz0);
1203             dx01             = _mm_sub_ps(ix0,jx1);
1204             dy01             = _mm_sub_ps(iy0,jy1);
1205             dz01             = _mm_sub_ps(iz0,jz1);
1206             dx02             = _mm_sub_ps(ix0,jx2);
1207             dy02             = _mm_sub_ps(iy0,jy2);
1208             dz02             = _mm_sub_ps(iz0,jz2);
1209             dx10             = _mm_sub_ps(ix1,jx0);
1210             dy10             = _mm_sub_ps(iy1,jy0);
1211             dz10             = _mm_sub_ps(iz1,jz0);
1212             dx11             = _mm_sub_ps(ix1,jx1);
1213             dy11             = _mm_sub_ps(iy1,jy1);
1214             dz11             = _mm_sub_ps(iz1,jz1);
1215             dx12             = _mm_sub_ps(ix1,jx2);
1216             dy12             = _mm_sub_ps(iy1,jy2);
1217             dz12             = _mm_sub_ps(iz1,jz2);
1218             dx20             = _mm_sub_ps(ix2,jx0);
1219             dy20             = _mm_sub_ps(iy2,jy0);
1220             dz20             = _mm_sub_ps(iz2,jz0);
1221             dx21             = _mm_sub_ps(ix2,jx1);
1222             dy21             = _mm_sub_ps(iy2,jy1);
1223             dz21             = _mm_sub_ps(iz2,jz1);
1224             dx22             = _mm_sub_ps(ix2,jx2);
1225             dy22             = _mm_sub_ps(iy2,jy2);
1226             dz22             = _mm_sub_ps(iz2,jz2);
1227
1228             /* Calculate squared distance and things based on it */
1229             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1230             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1231             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1232             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1233             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1234             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1235             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1236             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1237             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1238
1239             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1240             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1241             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1242             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1243             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1244             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1245             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1246             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1247             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1248
1249             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1250             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1251             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1252             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1253             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1254             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1255             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1256             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1257             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1258
1259             fjx0             = _mm_setzero_ps();
1260             fjy0             = _mm_setzero_ps();
1261             fjz0             = _mm_setzero_ps();
1262             fjx1             = _mm_setzero_ps();
1263             fjy1             = _mm_setzero_ps();
1264             fjz1             = _mm_setzero_ps();
1265             fjx2             = _mm_setzero_ps();
1266             fjy2             = _mm_setzero_ps();
1267             fjz2             = _mm_setzero_ps();
1268
1269             /**************************
1270              * CALCULATE INTERACTIONS *
1271              **************************/
1272
1273             r00              = _mm_mul_ps(rsq00,rinv00);
1274
1275             /* Calculate table index by multiplying r with table scale and truncate to integer */
1276             rt               = _mm_mul_ps(r00,vftabscale);
1277             vfitab           = _mm_cvttps_epi32(rt);
1278             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1279             vfitab           = _mm_slli_epi32(vfitab,3);
1280
1281             /* REACTION-FIELD ELECTROSTATICS */
1282             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1283
1284             /* CUBIC SPLINE TABLE DISPERSION */
1285             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1286             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1287             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1288             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1289             _MM_TRANSPOSE4_PS(Y,F,G,H);
1290             Heps             = _mm_mul_ps(vfeps,H);
1291             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1292             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1293             fvdw6            = _mm_mul_ps(c6_00,FF);
1294
1295             /* CUBIC SPLINE TABLE REPULSION */
1296             vfitab           = _mm_add_epi32(vfitab,ifour);
1297             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1298             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1299             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1300             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1301             _MM_TRANSPOSE4_PS(Y,F,G,H);
1302             Heps             = _mm_mul_ps(vfeps,H);
1303             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1304             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1305             fvdw12           = _mm_mul_ps(c12_00,FF);
1306             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1307
1308             fscal            = _mm_add_ps(felec,fvdw);
1309
1310             /* Calculate temporary vectorial force */
1311             tx               = _mm_mul_ps(fscal,dx00);
1312             ty               = _mm_mul_ps(fscal,dy00);
1313             tz               = _mm_mul_ps(fscal,dz00);
1314
1315             /* Update vectorial force */
1316             fix0             = _mm_add_ps(fix0,tx);
1317             fiy0             = _mm_add_ps(fiy0,ty);
1318             fiz0             = _mm_add_ps(fiz0,tz);
1319
1320             fjx0             = _mm_add_ps(fjx0,tx);
1321             fjy0             = _mm_add_ps(fjy0,ty);
1322             fjz0             = _mm_add_ps(fjz0,tz);
1323
1324             /**************************
1325              * CALCULATE INTERACTIONS *
1326              **************************/
1327
1328             /* REACTION-FIELD ELECTROSTATICS */
1329             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1330
1331             fscal            = felec;
1332
1333             /* Calculate temporary vectorial force */
1334             tx               = _mm_mul_ps(fscal,dx01);
1335             ty               = _mm_mul_ps(fscal,dy01);
1336             tz               = _mm_mul_ps(fscal,dz01);
1337
1338             /* Update vectorial force */
1339             fix0             = _mm_add_ps(fix0,tx);
1340             fiy0             = _mm_add_ps(fiy0,ty);
1341             fiz0             = _mm_add_ps(fiz0,tz);
1342
1343             fjx1             = _mm_add_ps(fjx1,tx);
1344             fjy1             = _mm_add_ps(fjy1,ty);
1345             fjz1             = _mm_add_ps(fjz1,tz);
1346
1347             /**************************
1348              * CALCULATE INTERACTIONS *
1349              **************************/
1350
1351             /* REACTION-FIELD ELECTROSTATICS */
1352             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1353
1354             fscal            = felec;
1355
1356             /* Calculate temporary vectorial force */
1357             tx               = _mm_mul_ps(fscal,dx02);
1358             ty               = _mm_mul_ps(fscal,dy02);
1359             tz               = _mm_mul_ps(fscal,dz02);
1360
1361             /* Update vectorial force */
1362             fix0             = _mm_add_ps(fix0,tx);
1363             fiy0             = _mm_add_ps(fiy0,ty);
1364             fiz0             = _mm_add_ps(fiz0,tz);
1365
1366             fjx2             = _mm_add_ps(fjx2,tx);
1367             fjy2             = _mm_add_ps(fjy2,ty);
1368             fjz2             = _mm_add_ps(fjz2,tz);
1369
1370             /**************************
1371              * CALCULATE INTERACTIONS *
1372              **************************/
1373
1374             /* REACTION-FIELD ELECTROSTATICS */
1375             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1376
1377             fscal            = felec;
1378
1379             /* Calculate temporary vectorial force */
1380             tx               = _mm_mul_ps(fscal,dx10);
1381             ty               = _mm_mul_ps(fscal,dy10);
1382             tz               = _mm_mul_ps(fscal,dz10);
1383
1384             /* Update vectorial force */
1385             fix1             = _mm_add_ps(fix1,tx);
1386             fiy1             = _mm_add_ps(fiy1,ty);
1387             fiz1             = _mm_add_ps(fiz1,tz);
1388
1389             fjx0             = _mm_add_ps(fjx0,tx);
1390             fjy0             = _mm_add_ps(fjy0,ty);
1391             fjz0             = _mm_add_ps(fjz0,tz);
1392
1393             /**************************
1394              * CALCULATE INTERACTIONS *
1395              **************************/
1396
1397             /* REACTION-FIELD ELECTROSTATICS */
1398             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1399
1400             fscal            = felec;
1401
1402             /* Calculate temporary vectorial force */
1403             tx               = _mm_mul_ps(fscal,dx11);
1404             ty               = _mm_mul_ps(fscal,dy11);
1405             tz               = _mm_mul_ps(fscal,dz11);
1406
1407             /* Update vectorial force */
1408             fix1             = _mm_add_ps(fix1,tx);
1409             fiy1             = _mm_add_ps(fiy1,ty);
1410             fiz1             = _mm_add_ps(fiz1,tz);
1411
1412             fjx1             = _mm_add_ps(fjx1,tx);
1413             fjy1             = _mm_add_ps(fjy1,ty);
1414             fjz1             = _mm_add_ps(fjz1,tz);
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             /* REACTION-FIELD ELECTROSTATICS */
1421             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1422
1423             fscal            = felec;
1424
1425             /* Calculate temporary vectorial force */
1426             tx               = _mm_mul_ps(fscal,dx12);
1427             ty               = _mm_mul_ps(fscal,dy12);
1428             tz               = _mm_mul_ps(fscal,dz12);
1429
1430             /* Update vectorial force */
1431             fix1             = _mm_add_ps(fix1,tx);
1432             fiy1             = _mm_add_ps(fiy1,ty);
1433             fiz1             = _mm_add_ps(fiz1,tz);
1434
1435             fjx2             = _mm_add_ps(fjx2,tx);
1436             fjy2             = _mm_add_ps(fjy2,ty);
1437             fjz2             = _mm_add_ps(fjz2,tz);
1438
1439             /**************************
1440              * CALCULATE INTERACTIONS *
1441              **************************/
1442
1443             /* REACTION-FIELD ELECTROSTATICS */
1444             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1445
1446             fscal            = felec;
1447
1448             /* Calculate temporary vectorial force */
1449             tx               = _mm_mul_ps(fscal,dx20);
1450             ty               = _mm_mul_ps(fscal,dy20);
1451             tz               = _mm_mul_ps(fscal,dz20);
1452
1453             /* Update vectorial force */
1454             fix2             = _mm_add_ps(fix2,tx);
1455             fiy2             = _mm_add_ps(fiy2,ty);
1456             fiz2             = _mm_add_ps(fiz2,tz);
1457
1458             fjx0             = _mm_add_ps(fjx0,tx);
1459             fjy0             = _mm_add_ps(fjy0,ty);
1460             fjz0             = _mm_add_ps(fjz0,tz);
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             /* REACTION-FIELD ELECTROSTATICS */
1467             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1468
1469             fscal            = felec;
1470
1471             /* Calculate temporary vectorial force */
1472             tx               = _mm_mul_ps(fscal,dx21);
1473             ty               = _mm_mul_ps(fscal,dy21);
1474             tz               = _mm_mul_ps(fscal,dz21);
1475
1476             /* Update vectorial force */
1477             fix2             = _mm_add_ps(fix2,tx);
1478             fiy2             = _mm_add_ps(fiy2,ty);
1479             fiz2             = _mm_add_ps(fiz2,tz);
1480
1481             fjx1             = _mm_add_ps(fjx1,tx);
1482             fjy1             = _mm_add_ps(fjy1,ty);
1483             fjz1             = _mm_add_ps(fjz1,tz);
1484
1485             /**************************
1486              * CALCULATE INTERACTIONS *
1487              **************************/
1488
1489             /* REACTION-FIELD ELECTROSTATICS */
1490             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1491
1492             fscal            = felec;
1493
1494             /* Calculate temporary vectorial force */
1495             tx               = _mm_mul_ps(fscal,dx22);
1496             ty               = _mm_mul_ps(fscal,dy22);
1497             tz               = _mm_mul_ps(fscal,dz22);
1498
1499             /* Update vectorial force */
1500             fix2             = _mm_add_ps(fix2,tx);
1501             fiy2             = _mm_add_ps(fiy2,ty);
1502             fiz2             = _mm_add_ps(fiz2,tz);
1503
1504             fjx2             = _mm_add_ps(fjx2,tx);
1505             fjy2             = _mm_add_ps(fjy2,ty);
1506             fjz2             = _mm_add_ps(fjz2,tz);
1507
1508             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1509                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1510                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1511
1512             /* Inner loop uses 270 flops */
1513         }
1514
1515         if(jidx<j_index_end)
1516         {
1517
1518             /* Get j neighbor index, and coordinate index */
1519             jnrA             = jjnr[jidx];
1520             jnrB             = jjnr[jidx+1];
1521             jnrC             = jjnr[jidx+2];
1522             jnrD             = jjnr[jidx+3];
1523
1524             /* Sign of each element will be negative for non-real atoms.
1525              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1526              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1527              */
1528             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1529             jnrA       = (jnrA>=0) ? jnrA : 0;
1530             jnrB       = (jnrB>=0) ? jnrB : 0;
1531             jnrC       = (jnrC>=0) ? jnrC : 0;
1532             jnrD       = (jnrD>=0) ? jnrD : 0;
1533
1534             j_coord_offsetA  = DIM*jnrA;
1535             j_coord_offsetB  = DIM*jnrB;
1536             j_coord_offsetC  = DIM*jnrC;
1537             j_coord_offsetD  = DIM*jnrD;
1538
1539             /* load j atom coordinates */
1540             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1541                                               x+j_coord_offsetC,x+j_coord_offsetD,
1542                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1543
1544             /* Calculate displacement vector */
1545             dx00             = _mm_sub_ps(ix0,jx0);
1546             dy00             = _mm_sub_ps(iy0,jy0);
1547             dz00             = _mm_sub_ps(iz0,jz0);
1548             dx01             = _mm_sub_ps(ix0,jx1);
1549             dy01             = _mm_sub_ps(iy0,jy1);
1550             dz01             = _mm_sub_ps(iz0,jz1);
1551             dx02             = _mm_sub_ps(ix0,jx2);
1552             dy02             = _mm_sub_ps(iy0,jy2);
1553             dz02             = _mm_sub_ps(iz0,jz2);
1554             dx10             = _mm_sub_ps(ix1,jx0);
1555             dy10             = _mm_sub_ps(iy1,jy0);
1556             dz10             = _mm_sub_ps(iz1,jz0);
1557             dx11             = _mm_sub_ps(ix1,jx1);
1558             dy11             = _mm_sub_ps(iy1,jy1);
1559             dz11             = _mm_sub_ps(iz1,jz1);
1560             dx12             = _mm_sub_ps(ix1,jx2);
1561             dy12             = _mm_sub_ps(iy1,jy2);
1562             dz12             = _mm_sub_ps(iz1,jz2);
1563             dx20             = _mm_sub_ps(ix2,jx0);
1564             dy20             = _mm_sub_ps(iy2,jy0);
1565             dz20             = _mm_sub_ps(iz2,jz0);
1566             dx21             = _mm_sub_ps(ix2,jx1);
1567             dy21             = _mm_sub_ps(iy2,jy1);
1568             dz21             = _mm_sub_ps(iz2,jz1);
1569             dx22             = _mm_sub_ps(ix2,jx2);
1570             dy22             = _mm_sub_ps(iy2,jy2);
1571             dz22             = _mm_sub_ps(iz2,jz2);
1572
1573             /* Calculate squared distance and things based on it */
1574             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1575             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1576             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1577             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1578             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1579             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1580             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1581             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1582             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1583
1584             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1585             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1586             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1587             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1588             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1589             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1590             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1591             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1592             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1593
1594             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1595             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1596             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1597             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1598             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1599             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1600             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1601             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1602             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1603
1604             fjx0             = _mm_setzero_ps();
1605             fjy0             = _mm_setzero_ps();
1606             fjz0             = _mm_setzero_ps();
1607             fjx1             = _mm_setzero_ps();
1608             fjy1             = _mm_setzero_ps();
1609             fjz1             = _mm_setzero_ps();
1610             fjx2             = _mm_setzero_ps();
1611             fjy2             = _mm_setzero_ps();
1612             fjz2             = _mm_setzero_ps();
1613
1614             /**************************
1615              * CALCULATE INTERACTIONS *
1616              **************************/
1617
1618             r00              = _mm_mul_ps(rsq00,rinv00);
1619             r00              = _mm_andnot_ps(dummy_mask,r00);
1620
1621             /* Calculate table index by multiplying r with table scale and truncate to integer */
1622             rt               = _mm_mul_ps(r00,vftabscale);
1623             vfitab           = _mm_cvttps_epi32(rt);
1624             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1625             vfitab           = _mm_slli_epi32(vfitab,3);
1626
1627             /* REACTION-FIELD ELECTROSTATICS */
1628             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1629
1630             /* CUBIC SPLINE TABLE DISPERSION */
1631             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1632             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1633             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1634             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1635             _MM_TRANSPOSE4_PS(Y,F,G,H);
1636             Heps             = _mm_mul_ps(vfeps,H);
1637             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1638             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1639             fvdw6            = _mm_mul_ps(c6_00,FF);
1640
1641             /* CUBIC SPLINE TABLE REPULSION */
1642             vfitab           = _mm_add_epi32(vfitab,ifour);
1643             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1644             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1645             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1646             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1647             _MM_TRANSPOSE4_PS(Y,F,G,H);
1648             Heps             = _mm_mul_ps(vfeps,H);
1649             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1650             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1651             fvdw12           = _mm_mul_ps(c12_00,FF);
1652             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1653
1654             fscal            = _mm_add_ps(felec,fvdw);
1655
1656             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1657
1658             /* Calculate temporary vectorial force */
1659             tx               = _mm_mul_ps(fscal,dx00);
1660             ty               = _mm_mul_ps(fscal,dy00);
1661             tz               = _mm_mul_ps(fscal,dz00);
1662
1663             /* Update vectorial force */
1664             fix0             = _mm_add_ps(fix0,tx);
1665             fiy0             = _mm_add_ps(fiy0,ty);
1666             fiz0             = _mm_add_ps(fiz0,tz);
1667
1668             fjx0             = _mm_add_ps(fjx0,tx);
1669             fjy0             = _mm_add_ps(fjy0,ty);
1670             fjz0             = _mm_add_ps(fjz0,tz);
1671
1672             /**************************
1673              * CALCULATE INTERACTIONS *
1674              **************************/
1675
1676             /* REACTION-FIELD ELECTROSTATICS */
1677             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1678
1679             fscal            = felec;
1680
1681             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1682
1683             /* Calculate temporary vectorial force */
1684             tx               = _mm_mul_ps(fscal,dx01);
1685             ty               = _mm_mul_ps(fscal,dy01);
1686             tz               = _mm_mul_ps(fscal,dz01);
1687
1688             /* Update vectorial force */
1689             fix0             = _mm_add_ps(fix0,tx);
1690             fiy0             = _mm_add_ps(fiy0,ty);
1691             fiz0             = _mm_add_ps(fiz0,tz);
1692
1693             fjx1             = _mm_add_ps(fjx1,tx);
1694             fjy1             = _mm_add_ps(fjy1,ty);
1695             fjz1             = _mm_add_ps(fjz1,tz);
1696
1697             /**************************
1698              * CALCULATE INTERACTIONS *
1699              **************************/
1700
1701             /* REACTION-FIELD ELECTROSTATICS */
1702             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1703
1704             fscal            = felec;
1705
1706             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1707
1708             /* Calculate temporary vectorial force */
1709             tx               = _mm_mul_ps(fscal,dx02);
1710             ty               = _mm_mul_ps(fscal,dy02);
1711             tz               = _mm_mul_ps(fscal,dz02);
1712
1713             /* Update vectorial force */
1714             fix0             = _mm_add_ps(fix0,tx);
1715             fiy0             = _mm_add_ps(fiy0,ty);
1716             fiz0             = _mm_add_ps(fiz0,tz);
1717
1718             fjx2             = _mm_add_ps(fjx2,tx);
1719             fjy2             = _mm_add_ps(fjy2,ty);
1720             fjz2             = _mm_add_ps(fjz2,tz);
1721
1722             /**************************
1723              * CALCULATE INTERACTIONS *
1724              **************************/
1725
1726             /* REACTION-FIELD ELECTROSTATICS */
1727             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1728
1729             fscal            = felec;
1730
1731             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1732
1733             /* Calculate temporary vectorial force */
1734             tx               = _mm_mul_ps(fscal,dx10);
1735             ty               = _mm_mul_ps(fscal,dy10);
1736             tz               = _mm_mul_ps(fscal,dz10);
1737
1738             /* Update vectorial force */
1739             fix1             = _mm_add_ps(fix1,tx);
1740             fiy1             = _mm_add_ps(fiy1,ty);
1741             fiz1             = _mm_add_ps(fiz1,tz);
1742
1743             fjx0             = _mm_add_ps(fjx0,tx);
1744             fjy0             = _mm_add_ps(fjy0,ty);
1745             fjz0             = _mm_add_ps(fjz0,tz);
1746
1747             /**************************
1748              * CALCULATE INTERACTIONS *
1749              **************************/
1750
1751             /* REACTION-FIELD ELECTROSTATICS */
1752             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1753
1754             fscal            = felec;
1755
1756             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1757
1758             /* Calculate temporary vectorial force */
1759             tx               = _mm_mul_ps(fscal,dx11);
1760             ty               = _mm_mul_ps(fscal,dy11);
1761             tz               = _mm_mul_ps(fscal,dz11);
1762
1763             /* Update vectorial force */
1764             fix1             = _mm_add_ps(fix1,tx);
1765             fiy1             = _mm_add_ps(fiy1,ty);
1766             fiz1             = _mm_add_ps(fiz1,tz);
1767
1768             fjx1             = _mm_add_ps(fjx1,tx);
1769             fjy1             = _mm_add_ps(fjy1,ty);
1770             fjz1             = _mm_add_ps(fjz1,tz);
1771
1772             /**************************
1773              * CALCULATE INTERACTIONS *
1774              **************************/
1775
1776             /* REACTION-FIELD ELECTROSTATICS */
1777             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1778
1779             fscal            = felec;
1780
1781             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1782
1783             /* Calculate temporary vectorial force */
1784             tx               = _mm_mul_ps(fscal,dx12);
1785             ty               = _mm_mul_ps(fscal,dy12);
1786             tz               = _mm_mul_ps(fscal,dz12);
1787
1788             /* Update vectorial force */
1789             fix1             = _mm_add_ps(fix1,tx);
1790             fiy1             = _mm_add_ps(fiy1,ty);
1791             fiz1             = _mm_add_ps(fiz1,tz);
1792
1793             fjx2             = _mm_add_ps(fjx2,tx);
1794             fjy2             = _mm_add_ps(fjy2,ty);
1795             fjz2             = _mm_add_ps(fjz2,tz);
1796
1797             /**************************
1798              * CALCULATE INTERACTIONS *
1799              **************************/
1800
1801             /* REACTION-FIELD ELECTROSTATICS */
1802             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1803
1804             fscal            = felec;
1805
1806             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1807
1808             /* Calculate temporary vectorial force */
1809             tx               = _mm_mul_ps(fscal,dx20);
1810             ty               = _mm_mul_ps(fscal,dy20);
1811             tz               = _mm_mul_ps(fscal,dz20);
1812
1813             /* Update vectorial force */
1814             fix2             = _mm_add_ps(fix2,tx);
1815             fiy2             = _mm_add_ps(fiy2,ty);
1816             fiz2             = _mm_add_ps(fiz2,tz);
1817
1818             fjx0             = _mm_add_ps(fjx0,tx);
1819             fjy0             = _mm_add_ps(fjy0,ty);
1820             fjz0             = _mm_add_ps(fjz0,tz);
1821
1822             /**************************
1823              * CALCULATE INTERACTIONS *
1824              **************************/
1825
1826             /* REACTION-FIELD ELECTROSTATICS */
1827             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1828
1829             fscal            = felec;
1830
1831             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1832
1833             /* Calculate temporary vectorial force */
1834             tx               = _mm_mul_ps(fscal,dx21);
1835             ty               = _mm_mul_ps(fscal,dy21);
1836             tz               = _mm_mul_ps(fscal,dz21);
1837
1838             /* Update vectorial force */
1839             fix2             = _mm_add_ps(fix2,tx);
1840             fiy2             = _mm_add_ps(fiy2,ty);
1841             fiz2             = _mm_add_ps(fiz2,tz);
1842
1843             fjx1             = _mm_add_ps(fjx1,tx);
1844             fjy1             = _mm_add_ps(fjy1,ty);
1845             fjz1             = _mm_add_ps(fjz1,tz);
1846
1847             /**************************
1848              * CALCULATE INTERACTIONS *
1849              **************************/
1850
1851             /* REACTION-FIELD ELECTROSTATICS */
1852             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1853
1854             fscal            = felec;
1855
1856             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1857
1858             /* Calculate temporary vectorial force */
1859             tx               = _mm_mul_ps(fscal,dx22);
1860             ty               = _mm_mul_ps(fscal,dy22);
1861             tz               = _mm_mul_ps(fscal,dz22);
1862
1863             /* Update vectorial force */
1864             fix2             = _mm_add_ps(fix2,tx);
1865             fiy2             = _mm_add_ps(fiy2,ty);
1866             fiz2             = _mm_add_ps(fiz2,tz);
1867
1868             fjx2             = _mm_add_ps(fjx2,tx);
1869             fjy2             = _mm_add_ps(fjy2,ty);
1870             fjz2             = _mm_add_ps(fjz2,tz);
1871
1872             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1873                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1874                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1875
1876             /* Inner loop uses 271 flops */
1877         }
1878
1879         /* End of innermost loop */
1880
1881         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1882                                               f+i_coord_offset,fshift+i_shift_offset);
1883
1884         /* Increment number of inner iterations */
1885         inneriter                  += j_index_end - j_index_start;
1886
1887         /* Outer loop uses 27 flops */
1888     }
1889
1890     /* Increment number of outer iterations */
1891     outeriter        += nri;
1892
1893     /* Update outer/inner flops */
1894
1895     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*27 + inneriter*271);
1896 }