src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_sse4_1_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse4_1_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse4_1_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_VF_sse4_1_single
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_VF_sse4_1_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  91     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  93     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     int              nvdwtype;
 106     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 107     int              *vdwtype;
 108     real             *vdwparam;
 109     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 110     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 111     __m128i          vfitab;
 112     __m128i          ifour       = _mm_set1_epi32(4);
 113     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 114     real             *vftab;
 115     __m128           dummy_mask,cutoff_mask;
 116     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 117     __m128           one     = _mm_set1_ps(1.0);
 118     __m128           two     = _mm_set1_ps(2.0);
 119     x                = xx[0];
 120     f                = ff[0];
 121
 122     nri              = nlist->nri;
 123     iinr             = nlist->iinr;
 124     jindex           = nlist->jindex;
 125     jjnr             = nlist->jjnr;
 126     shiftidx         = nlist->shift;
 127     gid              = nlist->gid;
 128     shiftvec         = fr->shift_vec[0];
 129     fshift           = fr->fshift[0];
 130     facel            = _mm_set1_ps(fr->ic->epsfac);
 131     charge           = mdatoms->chargeA;
 132     nvdwtype         = fr->ntype;
 133     vdwparam         = fr->nbfp;
 134     vdwtype          = mdatoms->typeA;
 135
 136     vftab            = kernel_data->table_vdw->data;
 137     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 138
 139     /* Setup water-specific parameters */
 140     inr              = nlist->iinr[0];
 141     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 142     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 143     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 144     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 145
 146     jq0              = _mm_set1_ps(charge[inr+0]);
 147     jq1              = _mm_set1_ps(charge[inr+1]);
 148     jq2              = _mm_set1_ps(charge[inr+2]);
 149     vdwjidx0A        = 2*vdwtype[inr+0];
 150     qq00             = _mm_mul_ps(iq0,jq0);
 151     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 152     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 153     qq01             = _mm_mul_ps(iq0,jq1);
 154     qq02             = _mm_mul_ps(iq0,jq2);
 155     qq10             = _mm_mul_ps(iq1,jq0);
 156     qq11             = _mm_mul_ps(iq1,jq1);
 157     qq12             = _mm_mul_ps(iq1,jq2);
 158     qq20             = _mm_mul_ps(iq2,jq0);
 159     qq21             = _mm_mul_ps(iq2,jq1);
 160     qq22             = _mm_mul_ps(iq2,jq2);
 161
 162     /* Avoid stupid compiler warnings */
 163     jnrA = jnrB = jnrC = jnrD = 0;
 164     j_coord_offsetA = 0;
 165     j_coord_offsetB = 0;
 166     j_coord_offsetC = 0;
 167     j_coord_offsetD = 0;
 168
 169     outeriter        = 0;
 170     inneriter        = 0;
 171
 172     for(iidx=0;iidx<4*DIM;iidx++)
 173     {
 174         scratch[iidx] = 0.0;
 175     }
 176
 177     /* Start outer loop over neighborlists */
 178     for(iidx=0; iidx<nri; iidx++)
 179     {
 180         /* Load shift vector for this list */
 181         i_shift_offset   = DIM*shiftidx[iidx];
 182
 183         /* Load limits for loop over neighbors */
 184         j_index_start    = jindex[iidx];
 185         j_index_end      = jindex[iidx+1];
 186
 187         /* Get outer coordinate index */
 188         inr              = iinr[iidx];
 189         i_coord_offset   = DIM*inr;
 190
 191         /* Load i particle coords and add shift vector */
 192         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 193                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 194
 195         fix0             = _mm_setzero_ps();
 196         fiy0             = _mm_setzero_ps();
 197         fiz0             = _mm_setzero_ps();
 198         fix1             = _mm_setzero_ps();
 199         fiy1             = _mm_setzero_ps();
 200         fiz1             = _mm_setzero_ps();
 201         fix2             = _mm_setzero_ps();
 202         fiy2             = _mm_setzero_ps();
 203         fiz2             = _mm_setzero_ps();
 204
 205         /* Reset potential sums */
 206         velecsum         = _mm_setzero_ps();
 207         vvdwsum          = _mm_setzero_ps();
 208
 209         /* Start inner kernel loop */
 210         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 211         {
 212
 213             /* Get j neighbor index, and coordinate index */
 214             jnrA             = jjnr[jidx];
 215             jnrB             = jjnr[jidx+1];
 216             jnrC             = jjnr[jidx+2];
 217             jnrD             = jjnr[jidx+3];
 218             j_coord_offsetA  = DIM*jnrA;
 219             j_coord_offsetB  = DIM*jnrB;
 220             j_coord_offsetC  = DIM*jnrC;
 221             j_coord_offsetD  = DIM*jnrD;
 222
 223             /* load j atom coordinates */
 224             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 225                                               x+j_coord_offsetC,x+j_coord_offsetD,
 226                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 227
 228             /* Calculate displacement vector */
 229             dx00             = _mm_sub_ps(ix0,jx0);
 230             dy00             = _mm_sub_ps(iy0,jy0);
 231             dz00             = _mm_sub_ps(iz0,jz0);
 232             dx01             = _mm_sub_ps(ix0,jx1);
 233             dy01             = _mm_sub_ps(iy0,jy1);
 234             dz01             = _mm_sub_ps(iz0,jz1);
 235             dx02             = _mm_sub_ps(ix0,jx2);
 236             dy02             = _mm_sub_ps(iy0,jy2);
 237             dz02             = _mm_sub_ps(iz0,jz2);
 238             dx10             = _mm_sub_ps(ix1,jx0);
 239             dy10             = _mm_sub_ps(iy1,jy0);
 240             dz10             = _mm_sub_ps(iz1,jz0);
 241             dx11             = _mm_sub_ps(ix1,jx1);
 242             dy11             = _mm_sub_ps(iy1,jy1);
 243             dz11             = _mm_sub_ps(iz1,jz1);
 244             dx12             = _mm_sub_ps(ix1,jx2);
 245             dy12             = _mm_sub_ps(iy1,jy2);
 246             dz12             = _mm_sub_ps(iz1,jz2);
 247             dx20             = _mm_sub_ps(ix2,jx0);
 248             dy20             = _mm_sub_ps(iy2,jy0);
 249             dz20             = _mm_sub_ps(iz2,jz0);
 250             dx21             = _mm_sub_ps(ix2,jx1);
 251             dy21             = _mm_sub_ps(iy2,jy1);
 252             dz21             = _mm_sub_ps(iz2,jz1);
 253             dx22             = _mm_sub_ps(ix2,jx2);
 254             dy22             = _mm_sub_ps(iy2,jy2);
 255             dz22             = _mm_sub_ps(iz2,jz2);
 256
 257             /* Calculate squared distance and things based on it */
 258             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 259             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 260             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 261             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 262             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 263             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 264             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 265             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 266             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 267
 268             rinv00           = sse41_invsqrt_f(rsq00);
 269             rinv01           = sse41_invsqrt_f(rsq01);
 270             rinv02           = sse41_invsqrt_f(rsq02);
 271             rinv10           = sse41_invsqrt_f(rsq10);
 272             rinv11           = sse41_invsqrt_f(rsq11);
 273             rinv12           = sse41_invsqrt_f(rsq12);
 274             rinv20           = sse41_invsqrt_f(rsq20);
 275             rinv21           = sse41_invsqrt_f(rsq21);
 276             rinv22           = sse41_invsqrt_f(rsq22);
 277
 278             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 279             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 280             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 281             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 282             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 283             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 284             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 285             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 286             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 287
 288             fjx0             = _mm_setzero_ps();
 289             fjy0             = _mm_setzero_ps();
 290             fjz0             = _mm_setzero_ps();
 291             fjx1             = _mm_setzero_ps();
 292             fjy1             = _mm_setzero_ps();
 293             fjz1             = _mm_setzero_ps();
 294             fjx2             = _mm_setzero_ps();
 295             fjy2             = _mm_setzero_ps();
 296             fjz2             = _mm_setzero_ps();
 297
 298             /**************************
 299              * CALCULATE INTERACTIONS *
 300              **************************/
 301
 302             r00              = _mm_mul_ps(rsq00,rinv00);
 303
 304             /* Calculate table index by multiplying r with table scale and truncate to integer */
 305             rt               = _mm_mul_ps(r00,vftabscale);
 306             vfitab           = _mm_cvttps_epi32(rt);
 307             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 308             vfitab           = _mm_slli_epi32(vfitab,3);
 309
 310             /* COULOMB ELECTROSTATICS */
 311             velec            = _mm_mul_ps(qq00,rinv00);
 312             felec            = _mm_mul_ps(velec,rinvsq00);
 313
 314             /* CUBIC SPLINE TABLE DISPERSION */
 315             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 316             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 317             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 318             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 319             _MM_TRANSPOSE4_PS(Y,F,G,H);
 320             Heps             = _mm_mul_ps(vfeps,H);
 321             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 322             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 323             vvdw6            = _mm_mul_ps(c6_00,VV);
 324             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 325             fvdw6            = _mm_mul_ps(c6_00,FF);
 326
 327             /* CUBIC SPLINE TABLE REPULSION */
 328             vfitab           = _mm_add_epi32(vfitab,ifour);
 329             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 330             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 331             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 332             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 333             _MM_TRANSPOSE4_PS(Y,F,G,H);
 334             Heps             = _mm_mul_ps(vfeps,H);
 335             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 336             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 337             vvdw12           = _mm_mul_ps(c12_00,VV);
 338             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 339             fvdw12           = _mm_mul_ps(c12_00,FF);
 340             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 341             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             velecsum         = _mm_add_ps(velecsum,velec);
 345             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 346
 347             fscal            = _mm_add_ps(felec,fvdw);
 348
 349             /* Calculate temporary vectorial force */
 350             tx               = _mm_mul_ps(fscal,dx00);
 351             ty               = _mm_mul_ps(fscal,dy00);
 352             tz               = _mm_mul_ps(fscal,dz00);
 353
 354             /* Update vectorial force */
 355             fix0             = _mm_add_ps(fix0,tx);
 356             fiy0             = _mm_add_ps(fiy0,ty);
 357             fiz0             = _mm_add_ps(fiz0,tz);
 358
 359             fjx0             = _mm_add_ps(fjx0,tx);
 360             fjy0             = _mm_add_ps(fjy0,ty);
 361             fjz0             = _mm_add_ps(fjz0,tz);
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             /* COULOMB ELECTROSTATICS */
 368             velec            = _mm_mul_ps(qq01,rinv01);
 369             felec            = _mm_mul_ps(velec,rinvsq01);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velecsum         = _mm_add_ps(velecsum,velec);
 373
 374             fscal            = felec;
 375
 376             /* Calculate temporary vectorial force */
 377             tx               = _mm_mul_ps(fscal,dx01);
 378             ty               = _mm_mul_ps(fscal,dy01);
 379             tz               = _mm_mul_ps(fscal,dz01);
 380
 381             /* Update vectorial force */
 382             fix0             = _mm_add_ps(fix0,tx);
 383             fiy0             = _mm_add_ps(fiy0,ty);
 384             fiz0             = _mm_add_ps(fiz0,tz);
 385
 386             fjx1             = _mm_add_ps(fjx1,tx);
 387             fjy1             = _mm_add_ps(fjy1,ty);
 388             fjz1             = _mm_add_ps(fjz1,tz);
 389
 390             /**************************
 391              * CALCULATE INTERACTIONS *
 392              **************************/
 393
 394             /* COULOMB ELECTROSTATICS */
 395             velec            = _mm_mul_ps(qq02,rinv02);
 396             felec            = _mm_mul_ps(velec,rinvsq02);
 397
 398             /* Update potential sum for this i atom from the interaction with this j atom. */
 399             velecsum         = _mm_add_ps(velecsum,velec);
 400
 401             fscal            = felec;
 402
 403             /* Calculate temporary vectorial force */
 404             tx               = _mm_mul_ps(fscal,dx02);
 405             ty               = _mm_mul_ps(fscal,dy02);
 406             tz               = _mm_mul_ps(fscal,dz02);
 407
 408             /* Update vectorial force */
 409             fix0             = _mm_add_ps(fix0,tx);
 410             fiy0             = _mm_add_ps(fiy0,ty);
 411             fiz0             = _mm_add_ps(fiz0,tz);
 412
 413             fjx2             = _mm_add_ps(fjx2,tx);
 414             fjy2             = _mm_add_ps(fjy2,ty);
 415             fjz2             = _mm_add_ps(fjz2,tz);
 416
 417             /**************************
 418              * CALCULATE INTERACTIONS *
 419              **************************/
 420
 421             /* COULOMB ELECTROSTATICS */
 422             velec            = _mm_mul_ps(qq10,rinv10);
 423             felec            = _mm_mul_ps(velec,rinvsq10);
 424
 425             /* Update potential sum for this i atom from the interaction with this j atom. */
 426             velecsum         = _mm_add_ps(velecsum,velec);
 427
 428             fscal            = felec;
 429
 430             /* Calculate temporary vectorial force */
 431             tx               = _mm_mul_ps(fscal,dx10);
 432             ty               = _mm_mul_ps(fscal,dy10);
 433             tz               = _mm_mul_ps(fscal,dz10);
 434
 435             /* Update vectorial force */
 436             fix1             = _mm_add_ps(fix1,tx);
 437             fiy1             = _mm_add_ps(fiy1,ty);
 438             fiz1             = _mm_add_ps(fiz1,tz);
 439
 440             fjx0             = _mm_add_ps(fjx0,tx);
 441             fjy0             = _mm_add_ps(fjy0,ty);
 442             fjz0             = _mm_add_ps(fjz0,tz);
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             /* COULOMB ELECTROSTATICS */
 449             velec            = _mm_mul_ps(qq11,rinv11);
 450             felec            = _mm_mul_ps(velec,rinvsq11);
 451
 452             /* Update potential sum for this i atom from the interaction with this j atom. */
 453             velecsum         = _mm_add_ps(velecsum,velec);
 454
 455             fscal            = felec;
 456
 457             /* Calculate temporary vectorial force */
 458             tx               = _mm_mul_ps(fscal,dx11);
 459             ty               = _mm_mul_ps(fscal,dy11);
 460             tz               = _mm_mul_ps(fscal,dz11);
 461
 462             /* Update vectorial force */
 463             fix1             = _mm_add_ps(fix1,tx);
 464             fiy1             = _mm_add_ps(fiy1,ty);
 465             fiz1             = _mm_add_ps(fiz1,tz);
 466
 467             fjx1             = _mm_add_ps(fjx1,tx);
 468             fjy1             = _mm_add_ps(fjy1,ty);
 469             fjz1             = _mm_add_ps(fjz1,tz);
 470
 471             /**************************
 472              * CALCULATE INTERACTIONS *
 473              **************************/
 474
 475             /* COULOMB ELECTROSTATICS */
 476             velec            = _mm_mul_ps(qq12,rinv12);
 477             felec            = _mm_mul_ps(velec,rinvsq12);
 478
 479             /* Update potential sum for this i atom from the interaction with this j atom. */
 480             velecsum         = _mm_add_ps(velecsum,velec);
 481
 482             fscal            = felec;
 483
 484             /* Calculate temporary vectorial force */
 485             tx               = _mm_mul_ps(fscal,dx12);
 486             ty               = _mm_mul_ps(fscal,dy12);
 487             tz               = _mm_mul_ps(fscal,dz12);
 488
 489             /* Update vectorial force */
 490             fix1             = _mm_add_ps(fix1,tx);
 491             fiy1             = _mm_add_ps(fiy1,ty);
 492             fiz1             = _mm_add_ps(fiz1,tz);
 493
 494             fjx2             = _mm_add_ps(fjx2,tx);
 495             fjy2             = _mm_add_ps(fjy2,ty);
 496             fjz2             = _mm_add_ps(fjz2,tz);
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             /* COULOMB ELECTROSTATICS */
 503             velec            = _mm_mul_ps(qq20,rinv20);
 504             felec            = _mm_mul_ps(velec,rinvsq20);
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velecsum         = _mm_add_ps(velecsum,velec);
 508
 509             fscal            = felec;
 510
 511             /* Calculate temporary vectorial force */
 512             tx               = _mm_mul_ps(fscal,dx20);
 513             ty               = _mm_mul_ps(fscal,dy20);
 514             tz               = _mm_mul_ps(fscal,dz20);
 515
 516             /* Update vectorial force */
 517             fix2             = _mm_add_ps(fix2,tx);
 518             fiy2             = _mm_add_ps(fiy2,ty);
 519             fiz2             = _mm_add_ps(fiz2,tz);
 520
 521             fjx0             = _mm_add_ps(fjx0,tx);
 522             fjy0             = _mm_add_ps(fjy0,ty);
 523             fjz0             = _mm_add_ps(fjz0,tz);
 524
 525             /**************************
 526              * CALCULATE INTERACTIONS *
 527              **************************/
 528
 529             /* COULOMB ELECTROSTATICS */
 530             velec            = _mm_mul_ps(qq21,rinv21);
 531             felec            = _mm_mul_ps(velec,rinvsq21);
 532
 533             /* Update potential sum for this i atom from the interaction with this j atom. */
 534             velecsum         = _mm_add_ps(velecsum,velec);
 535
 536             fscal            = felec;
 537
 538             /* Calculate temporary vectorial force */
 539             tx               = _mm_mul_ps(fscal,dx21);
 540             ty               = _mm_mul_ps(fscal,dy21);
 541             tz               = _mm_mul_ps(fscal,dz21);
 542
 543             /* Update vectorial force */
 544             fix2             = _mm_add_ps(fix2,tx);
 545             fiy2             = _mm_add_ps(fiy2,ty);
 546             fiz2             = _mm_add_ps(fiz2,tz);
 547
 548             fjx1             = _mm_add_ps(fjx1,tx);
 549             fjy1             = _mm_add_ps(fjy1,ty);
 550             fjz1             = _mm_add_ps(fjz1,tz);
 551
 552             /**************************
 553              * CALCULATE INTERACTIONS *
 554              **************************/
 555
 556             /* COULOMB ELECTROSTATICS */
 557             velec            = _mm_mul_ps(qq22,rinv22);
 558             felec            = _mm_mul_ps(velec,rinvsq22);
 559
 560             /* Update potential sum for this i atom from the interaction with this j atom. */
 561             velecsum         = _mm_add_ps(velecsum,velec);
 562
 563             fscal            = felec;
 564
 565             /* Calculate temporary vectorial force */
 566             tx               = _mm_mul_ps(fscal,dx22);
 567             ty               = _mm_mul_ps(fscal,dy22);
 568             tz               = _mm_mul_ps(fscal,dz22);
 569
 570             /* Update vectorial force */
 571             fix2             = _mm_add_ps(fix2,tx);
 572             fiy2             = _mm_add_ps(fiy2,ty);
 573             fiz2             = _mm_add_ps(fiz2,tz);
 574
 575             fjx2             = _mm_add_ps(fjx2,tx);
 576             fjy2             = _mm_add_ps(fjy2,ty);
 577             fjz2             = _mm_add_ps(fjz2,tz);
 578
 579             fjptrA             = f+j_coord_offsetA;
 580             fjptrB             = f+j_coord_offsetB;
 581             fjptrC             = f+j_coord_offsetC;
 582             fjptrD             = f+j_coord_offsetD;
 583
 584             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 585                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 586
 587             /* Inner loop uses 287 flops */
 588         }
 589
 590         if(jidx<j_index_end)
 591         {
 592
 593             /* Get j neighbor index, and coordinate index */
 594             jnrlistA         = jjnr[jidx];
 595             jnrlistB         = jjnr[jidx+1];
 596             jnrlistC         = jjnr[jidx+2];
 597             jnrlistD         = jjnr[jidx+3];
 598             /* Sign of each element will be negative for non-real atoms.
 599              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 600              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 601              */
 602             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 603             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 604             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 605             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 606             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 607             j_coord_offsetA  = DIM*jnrA;
 608             j_coord_offsetB  = DIM*jnrB;
 609             j_coord_offsetC  = DIM*jnrC;
 610             j_coord_offsetD  = DIM*jnrD;
 611
 612             /* load j atom coordinates */
 613             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 614                                               x+j_coord_offsetC,x+j_coord_offsetD,
 615                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 616
 617             /* Calculate displacement vector */
 618             dx00             = _mm_sub_ps(ix0,jx0);
 619             dy00             = _mm_sub_ps(iy0,jy0);
 620             dz00             = _mm_sub_ps(iz0,jz0);
 621             dx01             = _mm_sub_ps(ix0,jx1);
 622             dy01             = _mm_sub_ps(iy0,jy1);
 623             dz01             = _mm_sub_ps(iz0,jz1);
 624             dx02             = _mm_sub_ps(ix0,jx2);
 625             dy02             = _mm_sub_ps(iy0,jy2);
 626             dz02             = _mm_sub_ps(iz0,jz2);
 627             dx10             = _mm_sub_ps(ix1,jx0);
 628             dy10             = _mm_sub_ps(iy1,jy0);
 629             dz10             = _mm_sub_ps(iz1,jz0);
 630             dx11             = _mm_sub_ps(ix1,jx1);
 631             dy11             = _mm_sub_ps(iy1,jy1);
 632             dz11             = _mm_sub_ps(iz1,jz1);
 633             dx12             = _mm_sub_ps(ix1,jx2);
 634             dy12             = _mm_sub_ps(iy1,jy2);
 635             dz12             = _mm_sub_ps(iz1,jz2);
 636             dx20             = _mm_sub_ps(ix2,jx0);
 637             dy20             = _mm_sub_ps(iy2,jy0);
 638             dz20             = _mm_sub_ps(iz2,jz0);
 639             dx21             = _mm_sub_ps(ix2,jx1);
 640             dy21             = _mm_sub_ps(iy2,jy1);
 641             dz21             = _mm_sub_ps(iz2,jz1);
 642             dx22             = _mm_sub_ps(ix2,jx2);
 643             dy22             = _mm_sub_ps(iy2,jy2);
 644             dz22             = _mm_sub_ps(iz2,jz2);
 645
 646             /* Calculate squared distance and things based on it */
 647             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 648             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 649             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 650             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 651             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 652             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 653             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 654             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 655             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 656
 657             rinv00           = sse41_invsqrt_f(rsq00);
 658             rinv01           = sse41_invsqrt_f(rsq01);
 659             rinv02           = sse41_invsqrt_f(rsq02);
 660             rinv10           = sse41_invsqrt_f(rsq10);
 661             rinv11           = sse41_invsqrt_f(rsq11);
 662             rinv12           = sse41_invsqrt_f(rsq12);
 663             rinv20           = sse41_invsqrt_f(rsq20);
 664             rinv21           = sse41_invsqrt_f(rsq21);
 665             rinv22           = sse41_invsqrt_f(rsq22);
 666
 667             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 668             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 669             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 670             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 671             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 672             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 673             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 674             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 675             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 676
 677             fjx0             = _mm_setzero_ps();
 678             fjy0             = _mm_setzero_ps();
 679             fjz0             = _mm_setzero_ps();
 680             fjx1             = _mm_setzero_ps();
 681             fjy1             = _mm_setzero_ps();
 682             fjz1             = _mm_setzero_ps();
 683             fjx2             = _mm_setzero_ps();
 684             fjy2             = _mm_setzero_ps();
 685             fjz2             = _mm_setzero_ps();
 686
 687             /**************************
 688              * CALCULATE INTERACTIONS *
 689              **************************/
 690
 691             r00              = _mm_mul_ps(rsq00,rinv00);
 692             r00              = _mm_andnot_ps(dummy_mask,r00);
 693
 694             /* Calculate table index by multiplying r with table scale and truncate to integer */
 695             rt               = _mm_mul_ps(r00,vftabscale);
 696             vfitab           = _mm_cvttps_epi32(rt);
 697             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 698             vfitab           = _mm_slli_epi32(vfitab,3);
 699
 700             /* COULOMB ELECTROSTATICS */
 701             velec            = _mm_mul_ps(qq00,rinv00);
 702             felec            = _mm_mul_ps(velec,rinvsq00);
 703
 704             /* CUBIC SPLINE TABLE DISPERSION */
 705             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 706             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 707             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 708             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 709             _MM_TRANSPOSE4_PS(Y,F,G,H);
 710             Heps             = _mm_mul_ps(vfeps,H);
 711             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 712             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 713             vvdw6            = _mm_mul_ps(c6_00,VV);
 714             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 715             fvdw6            = _mm_mul_ps(c6_00,FF);
 716
 717             /* CUBIC SPLINE TABLE REPULSION */
 718             vfitab           = _mm_add_epi32(vfitab,ifour);
 719             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 720             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 721             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 722             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 723             _MM_TRANSPOSE4_PS(Y,F,G,H);
 724             Heps             = _mm_mul_ps(vfeps,H);
 725             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 726             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 727             vvdw12           = _mm_mul_ps(c12_00,VV);
 728             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 729             fvdw12           = _mm_mul_ps(c12_00,FF);
 730             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 731             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 732
 733             /* Update potential sum for this i atom from the interaction with this j atom. */
 734             velec            = _mm_andnot_ps(dummy_mask,velec);
 735             velecsum         = _mm_add_ps(velecsum,velec);
 736             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 737             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 738
 739             fscal            = _mm_add_ps(felec,fvdw);
 740
 741             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 742
 743             /* Calculate temporary vectorial force */
 744             tx               = _mm_mul_ps(fscal,dx00);
 745             ty               = _mm_mul_ps(fscal,dy00);
 746             tz               = _mm_mul_ps(fscal,dz00);
 747
 748             /* Update vectorial force */
 749             fix0             = _mm_add_ps(fix0,tx);
 750             fiy0             = _mm_add_ps(fiy0,ty);
 751             fiz0             = _mm_add_ps(fiz0,tz);
 752
 753             fjx0             = _mm_add_ps(fjx0,tx);
 754             fjy0             = _mm_add_ps(fjy0,ty);
 755             fjz0             = _mm_add_ps(fjz0,tz);
 756
 757             /**************************
 758              * CALCULATE INTERACTIONS *
 759              **************************/
 760
 761             /* COULOMB ELECTROSTATICS */
 762             velec            = _mm_mul_ps(qq01,rinv01);
 763             felec            = _mm_mul_ps(velec,rinvsq01);
 764
 765             /* Update potential sum for this i atom from the interaction with this j atom. */
 766             velec            = _mm_andnot_ps(dummy_mask,velec);
 767             velecsum         = _mm_add_ps(velecsum,velec);
 768
 769             fscal            = felec;
 770
 771             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 772
 773             /* Calculate temporary vectorial force */
 774             tx               = _mm_mul_ps(fscal,dx01);
 775             ty               = _mm_mul_ps(fscal,dy01);
 776             tz               = _mm_mul_ps(fscal,dz01);
 777
 778             /* Update vectorial force */
 779             fix0             = _mm_add_ps(fix0,tx);
 780             fiy0             = _mm_add_ps(fiy0,ty);
 781             fiz0             = _mm_add_ps(fiz0,tz);
 782
 783             fjx1             = _mm_add_ps(fjx1,tx);
 784             fjy1             = _mm_add_ps(fjy1,ty);
 785             fjz1             = _mm_add_ps(fjz1,tz);
 786
 787             /**************************
 788              * CALCULATE INTERACTIONS *
 789              **************************/
 790
 791             /* COULOMB ELECTROSTATICS */
 792             velec            = _mm_mul_ps(qq02,rinv02);
 793             felec            = _mm_mul_ps(velec,rinvsq02);
 794
 795             /* Update potential sum for this i atom from the interaction with this j atom. */
 796             velec            = _mm_andnot_ps(dummy_mask,velec);
 797             velecsum         = _mm_add_ps(velecsum,velec);
 798
 799             fscal            = felec;
 800
 801             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 802
 803             /* Calculate temporary vectorial force */
 804             tx               = _mm_mul_ps(fscal,dx02);
 805             ty               = _mm_mul_ps(fscal,dy02);
 806             tz               = _mm_mul_ps(fscal,dz02);
 807
 808             /* Update vectorial force */
 809             fix0             = _mm_add_ps(fix0,tx);
 810             fiy0             = _mm_add_ps(fiy0,ty);
 811             fiz0             = _mm_add_ps(fiz0,tz);
 812
 813             fjx2             = _mm_add_ps(fjx2,tx);
 814             fjy2             = _mm_add_ps(fjy2,ty);
 815             fjz2             = _mm_add_ps(fjz2,tz);
 816
 817             /**************************
 818              * CALCULATE INTERACTIONS *
 819              **************************/
 820
 821             /* COULOMB ELECTROSTATICS */
 822             velec            = _mm_mul_ps(qq10,rinv10);
 823             felec            = _mm_mul_ps(velec,rinvsq10);
 824
 825             /* Update potential sum for this i atom from the interaction with this j atom. */
 826             velec            = _mm_andnot_ps(dummy_mask,velec);
 827             velecsum         = _mm_add_ps(velecsum,velec);
 828
 829             fscal            = felec;
 830
 831             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 832
 833             /* Calculate temporary vectorial force */
 834             tx               = _mm_mul_ps(fscal,dx10);
 835             ty               = _mm_mul_ps(fscal,dy10);
 836             tz               = _mm_mul_ps(fscal,dz10);
 837
 838             /* Update vectorial force */
 839             fix1             = _mm_add_ps(fix1,tx);
 840             fiy1             = _mm_add_ps(fiy1,ty);
 841             fiz1             = _mm_add_ps(fiz1,tz);
 842
 843             fjx0             = _mm_add_ps(fjx0,tx);
 844             fjy0             = _mm_add_ps(fjy0,ty);
 845             fjz0             = _mm_add_ps(fjz0,tz);
 846
 847             /**************************
 848              * CALCULATE INTERACTIONS *
 849              **************************/
 850
 851             /* COULOMB ELECTROSTATICS */
 852             velec            = _mm_mul_ps(qq11,rinv11);
 853             felec            = _mm_mul_ps(velec,rinvsq11);
 854
 855             /* Update potential sum for this i atom from the interaction with this j atom. */
 856             velec            = _mm_andnot_ps(dummy_mask,velec);
 857             velecsum         = _mm_add_ps(velecsum,velec);
 858
 859             fscal            = felec;
 860
 861             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 862
 863             /* Calculate temporary vectorial force */
 864             tx               = _mm_mul_ps(fscal,dx11);
 865             ty               = _mm_mul_ps(fscal,dy11);
 866             tz               = _mm_mul_ps(fscal,dz11);
 867
 868             /* Update vectorial force */
 869             fix1             = _mm_add_ps(fix1,tx);
 870             fiy1             = _mm_add_ps(fiy1,ty);
 871             fiz1             = _mm_add_ps(fiz1,tz);
 872
 873             fjx1             = _mm_add_ps(fjx1,tx);
 874             fjy1             = _mm_add_ps(fjy1,ty);
 875             fjz1             = _mm_add_ps(fjz1,tz);
 876
 877             /**************************
 878              * CALCULATE INTERACTIONS *
 879              **************************/
 880
 881             /* COULOMB ELECTROSTATICS */
 882             velec            = _mm_mul_ps(qq12,rinv12);
 883             felec            = _mm_mul_ps(velec,rinvsq12);
 884
 885             /* Update potential sum for this i atom from the interaction with this j atom. */
 886             velec            = _mm_andnot_ps(dummy_mask,velec);
 887             velecsum         = _mm_add_ps(velecsum,velec);
 888
 889             fscal            = felec;
 890
 891             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 892
 893             /* Calculate temporary vectorial force */
 894             tx               = _mm_mul_ps(fscal,dx12);
 895             ty               = _mm_mul_ps(fscal,dy12);
 896             tz               = _mm_mul_ps(fscal,dz12);
 897
 898             /* Update vectorial force */
 899             fix1             = _mm_add_ps(fix1,tx);
 900             fiy1             = _mm_add_ps(fiy1,ty);
 901             fiz1             = _mm_add_ps(fiz1,tz);
 902
 903             fjx2             = _mm_add_ps(fjx2,tx);
 904             fjy2             = _mm_add_ps(fjy2,ty);
 905             fjz2             = _mm_add_ps(fjz2,tz);
 906
 907             /**************************
 908              * CALCULATE INTERACTIONS *
 909              **************************/
 910
 911             /* COULOMB ELECTROSTATICS */
 912             velec            = _mm_mul_ps(qq20,rinv20);
 913             felec            = _mm_mul_ps(velec,rinvsq20);
 914
 915             /* Update potential sum for this i atom from the interaction with this j atom. */
 916             velec            = _mm_andnot_ps(dummy_mask,velec);
 917             velecsum         = _mm_add_ps(velecsum,velec);
 918
 919             fscal            = felec;
 920
 921             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 922
 923             /* Calculate temporary vectorial force */
 924             tx               = _mm_mul_ps(fscal,dx20);
 925             ty               = _mm_mul_ps(fscal,dy20);
 926             tz               = _mm_mul_ps(fscal,dz20);
 927
 928             /* Update vectorial force */
 929             fix2             = _mm_add_ps(fix2,tx);
 930             fiy2             = _mm_add_ps(fiy2,ty);
 931             fiz2             = _mm_add_ps(fiz2,tz);
 932
 933             fjx0             = _mm_add_ps(fjx0,tx);
 934             fjy0             = _mm_add_ps(fjy0,ty);
 935             fjz0             = _mm_add_ps(fjz0,tz);
 936
 937             /**************************
 938              * CALCULATE INTERACTIONS *
 939              **************************/
 940
 941             /* COULOMB ELECTROSTATICS */
 942             velec            = _mm_mul_ps(qq21,rinv21);
 943             felec            = _mm_mul_ps(velec,rinvsq21);
 944
 945             /* Update potential sum for this i atom from the interaction with this j atom. */
 946             velec            = _mm_andnot_ps(dummy_mask,velec);
 947             velecsum         = _mm_add_ps(velecsum,velec);
 948
 949             fscal            = felec;
 950
 951             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 952
 953             /* Calculate temporary vectorial force */
 954             tx               = _mm_mul_ps(fscal,dx21);
 955             ty               = _mm_mul_ps(fscal,dy21);
 956             tz               = _mm_mul_ps(fscal,dz21);
 957
 958             /* Update vectorial force */
 959             fix2             = _mm_add_ps(fix2,tx);
 960             fiy2             = _mm_add_ps(fiy2,ty);
 961             fiz2             = _mm_add_ps(fiz2,tz);
 962
 963             fjx1             = _mm_add_ps(fjx1,tx);
 964             fjy1             = _mm_add_ps(fjy1,ty);
 965             fjz1             = _mm_add_ps(fjz1,tz);
 966
 967             /**************************
 968              * CALCULATE INTERACTIONS *
 969              **************************/
 970
 971             /* COULOMB ELECTROSTATICS */
 972             velec            = _mm_mul_ps(qq22,rinv22);
 973             felec            = _mm_mul_ps(velec,rinvsq22);
 974
 975             /* Update potential sum for this i atom from the interaction with this j atom. */
 976             velec            = _mm_andnot_ps(dummy_mask,velec);
 977             velecsum         = _mm_add_ps(velecsum,velec);
 978
 979             fscal            = felec;
 980
 981             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 982
 983             /* Calculate temporary vectorial force */
 984             tx               = _mm_mul_ps(fscal,dx22);
 985             ty               = _mm_mul_ps(fscal,dy22);
 986             tz               = _mm_mul_ps(fscal,dz22);
 987
 988             /* Update vectorial force */
 989             fix2             = _mm_add_ps(fix2,tx);
 990             fiy2             = _mm_add_ps(fiy2,ty);
 991             fiz2             = _mm_add_ps(fiz2,tz);
 992
 993             fjx2             = _mm_add_ps(fjx2,tx);
 994             fjy2             = _mm_add_ps(fjy2,ty);
 995             fjz2             = _mm_add_ps(fjz2,tz);
 996
 997             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 998             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 999             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1000             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1001
1002             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1003                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1004
1005             /* Inner loop uses 288 flops */
1006         }
1007
1008         /* End of innermost loop */
1009
1010         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1011                                               f+i_coord_offset,fshift+i_shift_offset);
1012
1013         ggid                        = gid[iidx];
1014         /* Update potential energies */
1015         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1016         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1017
1018         /* Increment number of inner iterations */
1019         inneriter                  += j_index_end - j_index_start;
1020
1021         /* Outer loop uses 20 flops */
1022     }
1023
1024     /* Increment number of outer iterations */
1025     outeriter        += nri;
1026
1027     /* Update outer/inner flops */
1028
1029     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*288);
1030 }
1031 /*
1032  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_F_sse4_1_single
1033  * Electrostatics interaction: Coulomb
1034  * VdW interaction:            CubicSplineTable
1035  * Geometry:                   Water3-Water3
1036  * Calculate force/pot:        Force
1037  */
1038 void
1039 nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_F_sse4_1_single
1040                     (t_nblist                    * gmx_restrict       nlist,
1041                      rvec                        * gmx_restrict          xx,
1042                      rvec                        * gmx_restrict          ff,
1043                      struct t_forcerec           * gmx_restrict          fr,
1044                      t_mdatoms                   * gmx_restrict     mdatoms,
1045                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1046                      t_nrnb                      * gmx_restrict        nrnb)
1047 {
1048     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1049      * just 0 for non-waters.
1050      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1051      * jnr indices corresponding to data put in the four positions in the SIMD register.
1052      */
1053     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1054     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1055     int              jnrA,jnrB,jnrC,jnrD;
1056     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1057     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1058     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1059     real             rcutoff_scalar;
1060     real             *shiftvec,*fshift,*x,*f;
1061     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1062     real             scratch[4*DIM];
1063     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1064     int              vdwioffset0;
1065     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1066     int              vdwioffset1;
1067     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1068     int              vdwioffset2;
1069     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1070     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1071     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1072     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1073     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1074     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1075     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1076     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1077     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1078     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1079     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1080     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1081     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1082     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1083     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1084     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1085     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1086     real             *charge;
1087     int              nvdwtype;
1088     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1089     int              *vdwtype;
1090     real             *vdwparam;
1091     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1092     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1093     __m128i          vfitab;
1094     __m128i          ifour       = _mm_set1_epi32(4);
1095     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1096     real             *vftab;
1097     __m128           dummy_mask,cutoff_mask;
1098     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1099     __m128           one     = _mm_set1_ps(1.0);
1100     __m128           two     = _mm_set1_ps(2.0);
1101     x                = xx[0];
1102     f                = ff[0];
1103
1104     nri              = nlist->nri;
1105     iinr             = nlist->iinr;
1106     jindex           = nlist->jindex;
1107     jjnr             = nlist->jjnr;
1108     shiftidx         = nlist->shift;
1109     gid              = nlist->gid;
1110     shiftvec         = fr->shift_vec[0];
1111     fshift           = fr->fshift[0];
1112     facel            = _mm_set1_ps(fr->ic->epsfac);
1113     charge           = mdatoms->chargeA;
1114     nvdwtype         = fr->ntype;
1115     vdwparam         = fr->nbfp;
1116     vdwtype          = mdatoms->typeA;
1117
1118     vftab            = kernel_data->table_vdw->data;
1119     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1120
1121     /* Setup water-specific parameters */
1122     inr              = nlist->iinr[0];
1123     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1124     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1125     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1126     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1127
1128     jq0              = _mm_set1_ps(charge[inr+0]);
1129     jq1              = _mm_set1_ps(charge[inr+1]);
1130     jq2              = _mm_set1_ps(charge[inr+2]);
1131     vdwjidx0A        = 2*vdwtype[inr+0];
1132     qq00             = _mm_mul_ps(iq0,jq0);
1133     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1134     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1135     qq01             = _mm_mul_ps(iq0,jq1);
1136     qq02             = _mm_mul_ps(iq0,jq2);
1137     qq10             = _mm_mul_ps(iq1,jq0);
1138     qq11             = _mm_mul_ps(iq1,jq1);
1139     qq12             = _mm_mul_ps(iq1,jq2);
1140     qq20             = _mm_mul_ps(iq2,jq0);
1141     qq21             = _mm_mul_ps(iq2,jq1);
1142     qq22             = _mm_mul_ps(iq2,jq2);
1143
1144     /* Avoid stupid compiler warnings */
1145     jnrA = jnrB = jnrC = jnrD = 0;
1146     j_coord_offsetA = 0;
1147     j_coord_offsetB = 0;
1148     j_coord_offsetC = 0;
1149     j_coord_offsetD = 0;
1150
1151     outeriter        = 0;
1152     inneriter        = 0;
1153
1154     for(iidx=0;iidx<4*DIM;iidx++)
1155     {
1156         scratch[iidx] = 0.0;
1157     }
1158
1159     /* Start outer loop over neighborlists */
1160     for(iidx=0; iidx<nri; iidx++)
1161     {
1162         /* Load shift vector for this list */
1163         i_shift_offset   = DIM*shiftidx[iidx];
1164
1165         /* Load limits for loop over neighbors */
1166         j_index_start    = jindex[iidx];
1167         j_index_end      = jindex[iidx+1];
1168
1169         /* Get outer coordinate index */
1170         inr              = iinr[iidx];
1171         i_coord_offset   = DIM*inr;
1172
1173         /* Load i particle coords and add shift vector */
1174         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1175                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1176
1177         fix0             = _mm_setzero_ps();
1178         fiy0             = _mm_setzero_ps();
1179         fiz0             = _mm_setzero_ps();
1180         fix1             = _mm_setzero_ps();
1181         fiy1             = _mm_setzero_ps();
1182         fiz1             = _mm_setzero_ps();
1183         fix2             = _mm_setzero_ps();
1184         fiy2             = _mm_setzero_ps();
1185         fiz2             = _mm_setzero_ps();
1186
1187         /* Start inner kernel loop */
1188         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1189         {
1190
1191             /* Get j neighbor index, and coordinate index */
1192             jnrA             = jjnr[jidx];
1193             jnrB             = jjnr[jidx+1];
1194             jnrC             = jjnr[jidx+2];
1195             jnrD             = jjnr[jidx+3];
1196             j_coord_offsetA  = DIM*jnrA;
1197             j_coord_offsetB  = DIM*jnrB;
1198             j_coord_offsetC  = DIM*jnrC;
1199             j_coord_offsetD  = DIM*jnrD;
1200
1201             /* load j atom coordinates */
1202             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1203                                               x+j_coord_offsetC,x+j_coord_offsetD,
1204                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1205
1206             /* Calculate displacement vector */
1207             dx00             = _mm_sub_ps(ix0,jx0);
1208             dy00             = _mm_sub_ps(iy0,jy0);
1209             dz00             = _mm_sub_ps(iz0,jz0);
1210             dx01             = _mm_sub_ps(ix0,jx1);
1211             dy01             = _mm_sub_ps(iy0,jy1);
1212             dz01             = _mm_sub_ps(iz0,jz1);
1213             dx02             = _mm_sub_ps(ix0,jx2);
1214             dy02             = _mm_sub_ps(iy0,jy2);
1215             dz02             = _mm_sub_ps(iz0,jz2);
1216             dx10             = _mm_sub_ps(ix1,jx0);
1217             dy10             = _mm_sub_ps(iy1,jy0);
1218             dz10             = _mm_sub_ps(iz1,jz0);
1219             dx11             = _mm_sub_ps(ix1,jx1);
1220             dy11             = _mm_sub_ps(iy1,jy1);
1221             dz11             = _mm_sub_ps(iz1,jz1);
1222             dx12             = _mm_sub_ps(ix1,jx2);
1223             dy12             = _mm_sub_ps(iy1,jy2);
1224             dz12             = _mm_sub_ps(iz1,jz2);
1225             dx20             = _mm_sub_ps(ix2,jx0);
1226             dy20             = _mm_sub_ps(iy2,jy0);
1227             dz20             = _mm_sub_ps(iz2,jz0);
1228             dx21             = _mm_sub_ps(ix2,jx1);
1229             dy21             = _mm_sub_ps(iy2,jy1);
1230             dz21             = _mm_sub_ps(iz2,jz1);
1231             dx22             = _mm_sub_ps(ix2,jx2);
1232             dy22             = _mm_sub_ps(iy2,jy2);
1233             dz22             = _mm_sub_ps(iz2,jz2);
1234
1235             /* Calculate squared distance and things based on it */
1236             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1237             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1238             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1239             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1240             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1241             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1242             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1243             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1244             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1245
1246             rinv00           = sse41_invsqrt_f(rsq00);
1247             rinv01           = sse41_invsqrt_f(rsq01);
1248             rinv02           = sse41_invsqrt_f(rsq02);
1249             rinv10           = sse41_invsqrt_f(rsq10);
1250             rinv11           = sse41_invsqrt_f(rsq11);
1251             rinv12           = sse41_invsqrt_f(rsq12);
1252             rinv20           = sse41_invsqrt_f(rsq20);
1253             rinv21           = sse41_invsqrt_f(rsq21);
1254             rinv22           = sse41_invsqrt_f(rsq22);
1255
1256             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1257             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1258             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1259             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1260             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1261             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1262             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1263             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1264             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1265
1266             fjx0             = _mm_setzero_ps();
1267             fjy0             = _mm_setzero_ps();
1268             fjz0             = _mm_setzero_ps();
1269             fjx1             = _mm_setzero_ps();
1270             fjy1             = _mm_setzero_ps();
1271             fjz1             = _mm_setzero_ps();
1272             fjx2             = _mm_setzero_ps();
1273             fjy2             = _mm_setzero_ps();
1274             fjz2             = _mm_setzero_ps();
1275
1276             /**************************
1277              * CALCULATE INTERACTIONS *
1278              **************************/
1279
1280             r00              = _mm_mul_ps(rsq00,rinv00);
1281
1282             /* Calculate table index by multiplying r with table scale and truncate to integer */
1283             rt               = _mm_mul_ps(r00,vftabscale);
1284             vfitab           = _mm_cvttps_epi32(rt);
1285             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1286             vfitab           = _mm_slli_epi32(vfitab,3);
1287
1288             /* COULOMB ELECTROSTATICS */
1289             velec            = _mm_mul_ps(qq00,rinv00);
1290             felec            = _mm_mul_ps(velec,rinvsq00);
1291
1292             /* CUBIC SPLINE TABLE DISPERSION */
1293             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1294             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1295             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1296             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1297             _MM_TRANSPOSE4_PS(Y,F,G,H);
1298             Heps             = _mm_mul_ps(vfeps,H);
1299             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1300             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1301             fvdw6            = _mm_mul_ps(c6_00,FF);
1302
1303             /* CUBIC SPLINE TABLE REPULSION */
1304             vfitab           = _mm_add_epi32(vfitab,ifour);
1305             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1306             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1307             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1308             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1309             _MM_TRANSPOSE4_PS(Y,F,G,H);
1310             Heps             = _mm_mul_ps(vfeps,H);
1311             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1312             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1313             fvdw12           = _mm_mul_ps(c12_00,FF);
1314             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1315
1316             fscal            = _mm_add_ps(felec,fvdw);
1317
1318             /* Calculate temporary vectorial force */
1319             tx               = _mm_mul_ps(fscal,dx00);
1320             ty               = _mm_mul_ps(fscal,dy00);
1321             tz               = _mm_mul_ps(fscal,dz00);
1322
1323             /* Update vectorial force */
1324             fix0             = _mm_add_ps(fix0,tx);
1325             fiy0             = _mm_add_ps(fiy0,ty);
1326             fiz0             = _mm_add_ps(fiz0,tz);
1327
1328             fjx0             = _mm_add_ps(fjx0,tx);
1329             fjy0             = _mm_add_ps(fjy0,ty);
1330             fjz0             = _mm_add_ps(fjz0,tz);
1331
1332             /**************************
1333              * CALCULATE INTERACTIONS *
1334              **************************/
1335
1336             /* COULOMB ELECTROSTATICS */
1337             velec            = _mm_mul_ps(qq01,rinv01);
1338             felec            = _mm_mul_ps(velec,rinvsq01);
1339
1340             fscal            = felec;
1341
1342             /* Calculate temporary vectorial force */
1343             tx               = _mm_mul_ps(fscal,dx01);
1344             ty               = _mm_mul_ps(fscal,dy01);
1345             tz               = _mm_mul_ps(fscal,dz01);
1346
1347             /* Update vectorial force */
1348             fix0             = _mm_add_ps(fix0,tx);
1349             fiy0             = _mm_add_ps(fiy0,ty);
1350             fiz0             = _mm_add_ps(fiz0,tz);
1351
1352             fjx1             = _mm_add_ps(fjx1,tx);
1353             fjy1             = _mm_add_ps(fjy1,ty);
1354             fjz1             = _mm_add_ps(fjz1,tz);
1355
1356             /**************************
1357              * CALCULATE INTERACTIONS *
1358              **************************/
1359
1360             /* COULOMB ELECTROSTATICS */
1361             velec            = _mm_mul_ps(qq02,rinv02);
1362             felec            = _mm_mul_ps(velec,rinvsq02);
1363
1364             fscal            = felec;
1365
1366             /* Calculate temporary vectorial force */
1367             tx               = _mm_mul_ps(fscal,dx02);
1368             ty               = _mm_mul_ps(fscal,dy02);
1369             tz               = _mm_mul_ps(fscal,dz02);
1370
1371             /* Update vectorial force */
1372             fix0             = _mm_add_ps(fix0,tx);
1373             fiy0             = _mm_add_ps(fiy0,ty);
1374             fiz0             = _mm_add_ps(fiz0,tz);
1375
1376             fjx2             = _mm_add_ps(fjx2,tx);
1377             fjy2             = _mm_add_ps(fjy2,ty);
1378             fjz2             = _mm_add_ps(fjz2,tz);
1379
1380             /**************************
1381              * CALCULATE INTERACTIONS *
1382              **************************/
1383
1384             /* COULOMB ELECTROSTATICS */
1385             velec            = _mm_mul_ps(qq10,rinv10);
1386             felec            = _mm_mul_ps(velec,rinvsq10);
1387
1388             fscal            = felec;
1389
1390             /* Calculate temporary vectorial force */
1391             tx               = _mm_mul_ps(fscal,dx10);
1392             ty               = _mm_mul_ps(fscal,dy10);
1393             tz               = _mm_mul_ps(fscal,dz10);
1394
1395             /* Update vectorial force */
1396             fix1             = _mm_add_ps(fix1,tx);
1397             fiy1             = _mm_add_ps(fiy1,ty);
1398             fiz1             = _mm_add_ps(fiz1,tz);
1399
1400             fjx0             = _mm_add_ps(fjx0,tx);
1401             fjy0             = _mm_add_ps(fjy0,ty);
1402             fjz0             = _mm_add_ps(fjz0,tz);
1403
1404             /**************************
1405              * CALCULATE INTERACTIONS *
1406              **************************/
1407
1408             /* COULOMB ELECTROSTATICS */
1409             velec            = _mm_mul_ps(qq11,rinv11);
1410             felec            = _mm_mul_ps(velec,rinvsq11);
1411
1412             fscal            = felec;
1413
1414             /* Calculate temporary vectorial force */
1415             tx               = _mm_mul_ps(fscal,dx11);
1416             ty               = _mm_mul_ps(fscal,dy11);
1417             tz               = _mm_mul_ps(fscal,dz11);
1418
1419             /* Update vectorial force */
1420             fix1             = _mm_add_ps(fix1,tx);
1421             fiy1             = _mm_add_ps(fiy1,ty);
1422             fiz1             = _mm_add_ps(fiz1,tz);
1423
1424             fjx1             = _mm_add_ps(fjx1,tx);
1425             fjy1             = _mm_add_ps(fjy1,ty);
1426             fjz1             = _mm_add_ps(fjz1,tz);
1427
1428             /**************************
1429              * CALCULATE INTERACTIONS *
1430              **************************/
1431
1432             /* COULOMB ELECTROSTATICS */
1433             velec            = _mm_mul_ps(qq12,rinv12);
1434             felec            = _mm_mul_ps(velec,rinvsq12);
1435
1436             fscal            = felec;
1437
1438             /* Calculate temporary vectorial force */
1439             tx               = _mm_mul_ps(fscal,dx12);
1440             ty               = _mm_mul_ps(fscal,dy12);
1441             tz               = _mm_mul_ps(fscal,dz12);
1442
1443             /* Update vectorial force */
1444             fix1             = _mm_add_ps(fix1,tx);
1445             fiy1             = _mm_add_ps(fiy1,ty);
1446             fiz1             = _mm_add_ps(fiz1,tz);
1447
1448             fjx2             = _mm_add_ps(fjx2,tx);
1449             fjy2             = _mm_add_ps(fjy2,ty);
1450             fjz2             = _mm_add_ps(fjz2,tz);
1451
1452             /**************************
1453              * CALCULATE INTERACTIONS *
1454              **************************/
1455
1456             /* COULOMB ELECTROSTATICS */
1457             velec            = _mm_mul_ps(qq20,rinv20);
1458             felec            = _mm_mul_ps(velec,rinvsq20);
1459
1460             fscal            = felec;
1461
1462             /* Calculate temporary vectorial force */
1463             tx               = _mm_mul_ps(fscal,dx20);
1464             ty               = _mm_mul_ps(fscal,dy20);
1465             tz               = _mm_mul_ps(fscal,dz20);
1466
1467             /* Update vectorial force */
1468             fix2             = _mm_add_ps(fix2,tx);
1469             fiy2             = _mm_add_ps(fiy2,ty);
1470             fiz2             = _mm_add_ps(fiz2,tz);
1471
1472             fjx0             = _mm_add_ps(fjx0,tx);
1473             fjy0             = _mm_add_ps(fjy0,ty);
1474             fjz0             = _mm_add_ps(fjz0,tz);
1475
1476             /**************************
1477              * CALCULATE INTERACTIONS *
1478              **************************/
1479
1480             /* COULOMB ELECTROSTATICS */
1481             velec            = _mm_mul_ps(qq21,rinv21);
1482             felec            = _mm_mul_ps(velec,rinvsq21);
1483
1484             fscal            = felec;
1485
1486             /* Calculate temporary vectorial force */
1487             tx               = _mm_mul_ps(fscal,dx21);
1488             ty               = _mm_mul_ps(fscal,dy21);
1489             tz               = _mm_mul_ps(fscal,dz21);
1490
1491             /* Update vectorial force */
1492             fix2             = _mm_add_ps(fix2,tx);
1493             fiy2             = _mm_add_ps(fiy2,ty);
1494             fiz2             = _mm_add_ps(fiz2,tz);
1495
1496             fjx1             = _mm_add_ps(fjx1,tx);
1497             fjy1             = _mm_add_ps(fjy1,ty);
1498             fjz1             = _mm_add_ps(fjz1,tz);
1499
1500             /**************************
1501              * CALCULATE INTERACTIONS *
1502              **************************/
1503
1504             /* COULOMB ELECTROSTATICS */
1505             velec            = _mm_mul_ps(qq22,rinv22);
1506             felec            = _mm_mul_ps(velec,rinvsq22);
1507
1508             fscal            = felec;
1509
1510             /* Calculate temporary vectorial force */
1511             tx               = _mm_mul_ps(fscal,dx22);
1512             ty               = _mm_mul_ps(fscal,dy22);
1513             tz               = _mm_mul_ps(fscal,dz22);
1514
1515             /* Update vectorial force */
1516             fix2             = _mm_add_ps(fix2,tx);
1517             fiy2             = _mm_add_ps(fiy2,ty);
1518             fiz2             = _mm_add_ps(fiz2,tz);
1519
1520             fjx2             = _mm_add_ps(fjx2,tx);
1521             fjy2             = _mm_add_ps(fjy2,ty);
1522             fjz2             = _mm_add_ps(fjz2,tz);
1523
1524             fjptrA             = f+j_coord_offsetA;
1525             fjptrB             = f+j_coord_offsetB;
1526             fjptrC             = f+j_coord_offsetC;
1527             fjptrD             = f+j_coord_offsetD;
1528
1529             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1530                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1531
1532             /* Inner loop uses 270 flops */
1533         }
1534
1535         if(jidx<j_index_end)
1536         {
1537
1538             /* Get j neighbor index, and coordinate index */
1539             jnrlistA         = jjnr[jidx];
1540             jnrlistB         = jjnr[jidx+1];
1541             jnrlistC         = jjnr[jidx+2];
1542             jnrlistD         = jjnr[jidx+3];
1543             /* Sign of each element will be negative for non-real atoms.
1544              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1545              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1546              */
1547             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1548             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1549             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1550             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1551             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1552             j_coord_offsetA  = DIM*jnrA;
1553             j_coord_offsetB  = DIM*jnrB;
1554             j_coord_offsetC  = DIM*jnrC;
1555             j_coord_offsetD  = DIM*jnrD;
1556
1557             /* load j atom coordinates */
1558             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1559                                               x+j_coord_offsetC,x+j_coord_offsetD,
1560                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1561
1562             /* Calculate displacement vector */
1563             dx00             = _mm_sub_ps(ix0,jx0);
1564             dy00             = _mm_sub_ps(iy0,jy0);
1565             dz00             = _mm_sub_ps(iz0,jz0);
1566             dx01             = _mm_sub_ps(ix0,jx1);
1567             dy01             = _mm_sub_ps(iy0,jy1);
1568             dz01             = _mm_sub_ps(iz0,jz1);
1569             dx02             = _mm_sub_ps(ix0,jx2);
1570             dy02             = _mm_sub_ps(iy0,jy2);
1571             dz02             = _mm_sub_ps(iz0,jz2);
1572             dx10             = _mm_sub_ps(ix1,jx0);
1573             dy10             = _mm_sub_ps(iy1,jy0);
1574             dz10             = _mm_sub_ps(iz1,jz0);
1575             dx11             = _mm_sub_ps(ix1,jx1);
1576             dy11             = _mm_sub_ps(iy1,jy1);
1577             dz11             = _mm_sub_ps(iz1,jz1);
1578             dx12             = _mm_sub_ps(ix1,jx2);
1579             dy12             = _mm_sub_ps(iy1,jy2);
1580             dz12             = _mm_sub_ps(iz1,jz2);
1581             dx20             = _mm_sub_ps(ix2,jx0);
1582             dy20             = _mm_sub_ps(iy2,jy0);
1583             dz20             = _mm_sub_ps(iz2,jz0);
1584             dx21             = _mm_sub_ps(ix2,jx1);
1585             dy21             = _mm_sub_ps(iy2,jy1);
1586             dz21             = _mm_sub_ps(iz2,jz1);
1587             dx22             = _mm_sub_ps(ix2,jx2);
1588             dy22             = _mm_sub_ps(iy2,jy2);
1589             dz22             = _mm_sub_ps(iz2,jz2);
1590
1591             /* Calculate squared distance and things based on it */
1592             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1593             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1594             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1595             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1596             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1597             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1598             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1599             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1600             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1601
1602             rinv00           = sse41_invsqrt_f(rsq00);
1603             rinv01           = sse41_invsqrt_f(rsq01);
1604             rinv02           = sse41_invsqrt_f(rsq02);
1605             rinv10           = sse41_invsqrt_f(rsq10);
1606             rinv11           = sse41_invsqrt_f(rsq11);
1607             rinv12           = sse41_invsqrt_f(rsq12);
1608             rinv20           = sse41_invsqrt_f(rsq20);
1609             rinv21           = sse41_invsqrt_f(rsq21);
1610             rinv22           = sse41_invsqrt_f(rsq22);
1611
1612             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1613             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1614             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1615             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1616             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1617             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1618             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1619             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1620             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1621
1622             fjx0             = _mm_setzero_ps();
1623             fjy0             = _mm_setzero_ps();
1624             fjz0             = _mm_setzero_ps();
1625             fjx1             = _mm_setzero_ps();
1626             fjy1             = _mm_setzero_ps();
1627             fjz1             = _mm_setzero_ps();
1628             fjx2             = _mm_setzero_ps();
1629             fjy2             = _mm_setzero_ps();
1630             fjz2             = _mm_setzero_ps();
1631
1632             /**************************
1633              * CALCULATE INTERACTIONS *
1634              **************************/
1635
1636             r00              = _mm_mul_ps(rsq00,rinv00);
1637             r00              = _mm_andnot_ps(dummy_mask,r00);
1638
1639             /* Calculate table index by multiplying r with table scale and truncate to integer */
1640             rt               = _mm_mul_ps(r00,vftabscale);
1641             vfitab           = _mm_cvttps_epi32(rt);
1642             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1643             vfitab           = _mm_slli_epi32(vfitab,3);
1644
1645             /* COULOMB ELECTROSTATICS */
1646             velec            = _mm_mul_ps(qq00,rinv00);
1647             felec            = _mm_mul_ps(velec,rinvsq00);
1648
1649             /* CUBIC SPLINE TABLE DISPERSION */
1650             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1651             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1652             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1653             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1654             _MM_TRANSPOSE4_PS(Y,F,G,H);
1655             Heps             = _mm_mul_ps(vfeps,H);
1656             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1657             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1658             fvdw6            = _mm_mul_ps(c6_00,FF);
1659
1660             /* CUBIC SPLINE TABLE REPULSION */
1661             vfitab           = _mm_add_epi32(vfitab,ifour);
1662             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1663             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1664             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1665             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1666             _MM_TRANSPOSE4_PS(Y,F,G,H);
1667             Heps             = _mm_mul_ps(vfeps,H);
1668             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1669             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1670             fvdw12           = _mm_mul_ps(c12_00,FF);
1671             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1672
1673             fscal            = _mm_add_ps(felec,fvdw);
1674
1675             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1676
1677             /* Calculate temporary vectorial force */
1678             tx               = _mm_mul_ps(fscal,dx00);
1679             ty               = _mm_mul_ps(fscal,dy00);
1680             tz               = _mm_mul_ps(fscal,dz00);
1681
1682             /* Update vectorial force */
1683             fix0             = _mm_add_ps(fix0,tx);
1684             fiy0             = _mm_add_ps(fiy0,ty);
1685             fiz0             = _mm_add_ps(fiz0,tz);
1686
1687             fjx0             = _mm_add_ps(fjx0,tx);
1688             fjy0             = _mm_add_ps(fjy0,ty);
1689             fjz0             = _mm_add_ps(fjz0,tz);
1690
1691             /**************************
1692              * CALCULATE INTERACTIONS *
1693              **************************/
1694
1695             /* COULOMB ELECTROSTATICS */
1696             velec            = _mm_mul_ps(qq01,rinv01);
1697             felec            = _mm_mul_ps(velec,rinvsq01);
1698
1699             fscal            = felec;
1700
1701             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1702
1703             /* Calculate temporary vectorial force */
1704             tx               = _mm_mul_ps(fscal,dx01);
1705             ty               = _mm_mul_ps(fscal,dy01);
1706             tz               = _mm_mul_ps(fscal,dz01);
1707
1708             /* Update vectorial force */
1709             fix0             = _mm_add_ps(fix0,tx);
1710             fiy0             = _mm_add_ps(fiy0,ty);
1711             fiz0             = _mm_add_ps(fiz0,tz);
1712
1713             fjx1             = _mm_add_ps(fjx1,tx);
1714             fjy1             = _mm_add_ps(fjy1,ty);
1715             fjz1             = _mm_add_ps(fjz1,tz);
1716
1717             /**************************
1718              * CALCULATE INTERACTIONS *
1719              **************************/
1720
1721             /* COULOMB ELECTROSTATICS */
1722             velec            = _mm_mul_ps(qq02,rinv02);
1723             felec            = _mm_mul_ps(velec,rinvsq02);
1724
1725             fscal            = felec;
1726
1727             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1728
1729             /* Calculate temporary vectorial force */
1730             tx               = _mm_mul_ps(fscal,dx02);
1731             ty               = _mm_mul_ps(fscal,dy02);
1732             tz               = _mm_mul_ps(fscal,dz02);
1733
1734             /* Update vectorial force */
1735             fix0             = _mm_add_ps(fix0,tx);
1736             fiy0             = _mm_add_ps(fiy0,ty);
1737             fiz0             = _mm_add_ps(fiz0,tz);
1738
1739             fjx2             = _mm_add_ps(fjx2,tx);
1740             fjy2             = _mm_add_ps(fjy2,ty);
1741             fjz2             = _mm_add_ps(fjz2,tz);
1742
1743             /**************************
1744              * CALCULATE INTERACTIONS *
1745              **************************/
1746
1747             /* COULOMB ELECTROSTATICS */
1748             velec            = _mm_mul_ps(qq10,rinv10);
1749             felec            = _mm_mul_ps(velec,rinvsq10);
1750
1751             fscal            = felec;
1752
1753             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1754
1755             /* Calculate temporary vectorial force */
1756             tx               = _mm_mul_ps(fscal,dx10);
1757             ty               = _mm_mul_ps(fscal,dy10);
1758             tz               = _mm_mul_ps(fscal,dz10);
1759
1760             /* Update vectorial force */
1761             fix1             = _mm_add_ps(fix1,tx);
1762             fiy1             = _mm_add_ps(fiy1,ty);
1763             fiz1             = _mm_add_ps(fiz1,tz);
1764
1765             fjx0             = _mm_add_ps(fjx0,tx);
1766             fjy0             = _mm_add_ps(fjy0,ty);
1767             fjz0             = _mm_add_ps(fjz0,tz);
1768
1769             /**************************
1770              * CALCULATE INTERACTIONS *
1771              **************************/
1772
1773             /* COULOMB ELECTROSTATICS */
1774             velec            = _mm_mul_ps(qq11,rinv11);
1775             felec            = _mm_mul_ps(velec,rinvsq11);
1776
1777             fscal            = felec;
1778
1779             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1780
1781             /* Calculate temporary vectorial force */
1782             tx               = _mm_mul_ps(fscal,dx11);
1783             ty               = _mm_mul_ps(fscal,dy11);
1784             tz               = _mm_mul_ps(fscal,dz11);
1785
1786             /* Update vectorial force */
1787             fix1             = _mm_add_ps(fix1,tx);
1788             fiy1             = _mm_add_ps(fiy1,ty);
1789             fiz1             = _mm_add_ps(fiz1,tz);
1790
1791             fjx1             = _mm_add_ps(fjx1,tx);
1792             fjy1             = _mm_add_ps(fjy1,ty);
1793             fjz1             = _mm_add_ps(fjz1,tz);
1794
1795             /**************************
1796              * CALCULATE INTERACTIONS *
1797              **************************/
1798
1799             /* COULOMB ELECTROSTATICS */
1800             velec            = _mm_mul_ps(qq12,rinv12);
1801             felec            = _mm_mul_ps(velec,rinvsq12);
1802
1803             fscal            = felec;
1804
1805             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1806
1807             /* Calculate temporary vectorial force */
1808             tx               = _mm_mul_ps(fscal,dx12);
1809             ty               = _mm_mul_ps(fscal,dy12);
1810             tz               = _mm_mul_ps(fscal,dz12);
1811
1812             /* Update vectorial force */
1813             fix1             = _mm_add_ps(fix1,tx);
1814             fiy1             = _mm_add_ps(fiy1,ty);
1815             fiz1             = _mm_add_ps(fiz1,tz);
1816
1817             fjx2             = _mm_add_ps(fjx2,tx);
1818             fjy2             = _mm_add_ps(fjy2,ty);
1819             fjz2             = _mm_add_ps(fjz2,tz);
1820
1821             /**************************
1822              * CALCULATE INTERACTIONS *
1823              **************************/
1824
1825             /* COULOMB ELECTROSTATICS */
1826             velec            = _mm_mul_ps(qq20,rinv20);
1827             felec            = _mm_mul_ps(velec,rinvsq20);
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,dx20);
1835             ty               = _mm_mul_ps(fscal,dy20);
1836             tz               = _mm_mul_ps(fscal,dz20);
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             fjx0             = _mm_add_ps(fjx0,tx);
1844             fjy0             = _mm_add_ps(fjy0,ty);
1845             fjz0             = _mm_add_ps(fjz0,tz);
1846
1847             /**************************
1848              * CALCULATE INTERACTIONS *
1849              **************************/
1850
1851             /* COULOMB ELECTROSTATICS */
1852             velec            = _mm_mul_ps(qq21,rinv21);
1853             felec            = _mm_mul_ps(velec,rinvsq21);
1854
1855             fscal            = felec;
1856
1857             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1858
1859             /* Calculate temporary vectorial force */
1860             tx               = _mm_mul_ps(fscal,dx21);
1861             ty               = _mm_mul_ps(fscal,dy21);
1862             tz               = _mm_mul_ps(fscal,dz21);
1863
1864             /* Update vectorial force */
1865             fix2             = _mm_add_ps(fix2,tx);
1866             fiy2             = _mm_add_ps(fiy2,ty);
1867             fiz2             = _mm_add_ps(fiz2,tz);
1868
1869             fjx1             = _mm_add_ps(fjx1,tx);
1870             fjy1             = _mm_add_ps(fjy1,ty);
1871             fjz1             = _mm_add_ps(fjz1,tz);
1872
1873             /**************************
1874              * CALCULATE INTERACTIONS *
1875              **************************/
1876
1877             /* COULOMB ELECTROSTATICS */
1878             velec            = _mm_mul_ps(qq22,rinv22);
1879             felec            = _mm_mul_ps(velec,rinvsq22);
1880
1881             fscal            = felec;
1882
1883             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1884
1885             /* Calculate temporary vectorial force */
1886             tx               = _mm_mul_ps(fscal,dx22);
1887             ty               = _mm_mul_ps(fscal,dy22);
1888             tz               = _mm_mul_ps(fscal,dz22);
1889
1890             /* Update vectorial force */
1891             fix2             = _mm_add_ps(fix2,tx);
1892             fiy2             = _mm_add_ps(fiy2,ty);
1893             fiz2             = _mm_add_ps(fiz2,tz);
1894
1895             fjx2             = _mm_add_ps(fjx2,tx);
1896             fjy2             = _mm_add_ps(fjy2,ty);
1897             fjz2             = _mm_add_ps(fjz2,tz);
1898
1899             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1900             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1901             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1902             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1903
1904             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1905                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1906
1907             /* Inner loop uses 271 flops */
1908         }
1909
1910         /* End of innermost loop */
1911
1912         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1913                                               f+i_coord_offset,fshift+i_shift_offset);
1914
1915         /* Increment number of inner iterations */
1916         inneriter                  += j_index_end - j_index_start;
1917
1918         /* Outer loop uses 18 flops */
1919     }
1920
1921     /* Increment number of outer iterations */
1922     outeriter        += nri;
1923
1924     /* Update outer/inner flops */
1925
1926     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*271);
1927 }