src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRF_VdwCSTab_GeomW4W4_sse4_1_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_sse4_1_single.h"
  48 #include "kernelutil_x86_sse4_1_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_sse4_1_single
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_sse4_1_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  98     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  99     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 100     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 101     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 102     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 103     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 104     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 105     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 106     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 107     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 108     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 109     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 110     real             *charge;
 111     int              nvdwtype;
 112     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 113     int              *vdwtype;
 114     real             *vdwparam;
 115     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 116     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 117     __m128i          vfitab;
 118     __m128i          ifour       = _mm_set1_epi32(4);
 119     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 120     real             *vftab;
 121     __m128           dummy_mask,cutoff_mask;
 122     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 123     __m128           one     = _mm_set1_ps(1.0);
 124     __m128           two     = _mm_set1_ps(2.0);
 125     x                = xx[0];
 126     f                = ff[0];
 127
 128     nri              = nlist->nri;
 129     iinr             = nlist->iinr;
 130     jindex           = nlist->jindex;
 131     jjnr             = nlist->jjnr;
 132     shiftidx         = nlist->shift;
 133     gid              = nlist->gid;
 134     shiftvec         = fr->shift_vec[0];
 135     fshift           = fr->fshift[0];
 136     facel            = _mm_set1_ps(fr->epsfac);
 137     charge           = mdatoms->chargeA;
 138     krf              = _mm_set1_ps(fr->ic->k_rf);
 139     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 140     crf              = _mm_set1_ps(fr->ic->c_rf);
 141     nvdwtype         = fr->ntype;
 142     vdwparam         = fr->nbfp;
 143     vdwtype          = mdatoms->typeA;
 144
 145     vftab            = kernel_data->table_vdw->data;
 146     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 147
 148     /* Setup water-specific parameters */
 149     inr              = nlist->iinr[0];
 150     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 151     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 152     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 153     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 154
 155     jq1              = _mm_set1_ps(charge[inr+1]);
 156     jq2              = _mm_set1_ps(charge[inr+2]);
 157     jq3              = _mm_set1_ps(charge[inr+3]);
 158     vdwjidx0A        = 2*vdwtype[inr+0];
 159     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 160     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 161     qq11             = _mm_mul_ps(iq1,jq1);
 162     qq12             = _mm_mul_ps(iq1,jq2);
 163     qq13             = _mm_mul_ps(iq1,jq3);
 164     qq21             = _mm_mul_ps(iq2,jq1);
 165     qq22             = _mm_mul_ps(iq2,jq2);
 166     qq23             = _mm_mul_ps(iq2,jq3);
 167     qq31             = _mm_mul_ps(iq3,jq1);
 168     qq32             = _mm_mul_ps(iq3,jq2);
 169     qq33             = _mm_mul_ps(iq3,jq3);
 170
 171     /* Avoid stupid compiler warnings */
 172     jnrA = jnrB = jnrC = jnrD = 0;
 173     j_coord_offsetA = 0;
 174     j_coord_offsetB = 0;
 175     j_coord_offsetC = 0;
 176     j_coord_offsetD = 0;
 177
 178     outeriter        = 0;
 179     inneriter        = 0;
 180
 181     for(iidx=0;iidx<4*DIM;iidx++)
 182     {
 183         scratch[iidx] = 0.0;
 184     }
 185
 186     /* Start outer loop over neighborlists */
 187     for(iidx=0; iidx<nri; iidx++)
 188     {
 189         /* Load shift vector for this list */
 190         i_shift_offset   = DIM*shiftidx[iidx];
 191
 192         /* Load limits for loop over neighbors */
 193         j_index_start    = jindex[iidx];
 194         j_index_end      = jindex[iidx+1];
 195
 196         /* Get outer coordinate index */
 197         inr              = iinr[iidx];
 198         i_coord_offset   = DIM*inr;
 199
 200         /* Load i particle coords and add shift vector */
 201         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 202                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 203
 204         fix0             = _mm_setzero_ps();
 205         fiy0             = _mm_setzero_ps();
 206         fiz0             = _mm_setzero_ps();
 207         fix1             = _mm_setzero_ps();
 208         fiy1             = _mm_setzero_ps();
 209         fiz1             = _mm_setzero_ps();
 210         fix2             = _mm_setzero_ps();
 211         fiy2             = _mm_setzero_ps();
 212         fiz2             = _mm_setzero_ps();
 213         fix3             = _mm_setzero_ps();
 214         fiy3             = _mm_setzero_ps();
 215         fiz3             = _mm_setzero_ps();
 216
 217         /* Reset potential sums */
 218         velecsum         = _mm_setzero_ps();
 219         vvdwsum          = _mm_setzero_ps();
 220
 221         /* Start inner kernel loop */
 222         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 223         {
 224
 225             /* Get j neighbor index, and coordinate index */
 226             jnrA             = jjnr[jidx];
 227             jnrB             = jjnr[jidx+1];
 228             jnrC             = jjnr[jidx+2];
 229             jnrD             = jjnr[jidx+3];
 230             j_coord_offsetA  = DIM*jnrA;
 231             j_coord_offsetB  = DIM*jnrB;
 232             j_coord_offsetC  = DIM*jnrC;
 233             j_coord_offsetD  = DIM*jnrD;
 234
 235             /* load j atom coordinates */
 236             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 237                                               x+j_coord_offsetC,x+j_coord_offsetD,
 238                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 239                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 240
 241             /* Calculate displacement vector */
 242             dx00             = _mm_sub_ps(ix0,jx0);
 243             dy00             = _mm_sub_ps(iy0,jy0);
 244             dz00             = _mm_sub_ps(iz0,jz0);
 245             dx11             = _mm_sub_ps(ix1,jx1);
 246             dy11             = _mm_sub_ps(iy1,jy1);
 247             dz11             = _mm_sub_ps(iz1,jz1);
 248             dx12             = _mm_sub_ps(ix1,jx2);
 249             dy12             = _mm_sub_ps(iy1,jy2);
 250             dz12             = _mm_sub_ps(iz1,jz2);
 251             dx13             = _mm_sub_ps(ix1,jx3);
 252             dy13             = _mm_sub_ps(iy1,jy3);
 253             dz13             = _mm_sub_ps(iz1,jz3);
 254             dx21             = _mm_sub_ps(ix2,jx1);
 255             dy21             = _mm_sub_ps(iy2,jy1);
 256             dz21             = _mm_sub_ps(iz2,jz1);
 257             dx22             = _mm_sub_ps(ix2,jx2);
 258             dy22             = _mm_sub_ps(iy2,jy2);
 259             dz22             = _mm_sub_ps(iz2,jz2);
 260             dx23             = _mm_sub_ps(ix2,jx3);
 261             dy23             = _mm_sub_ps(iy2,jy3);
 262             dz23             = _mm_sub_ps(iz2,jz3);
 263             dx31             = _mm_sub_ps(ix3,jx1);
 264             dy31             = _mm_sub_ps(iy3,jy1);
 265             dz31             = _mm_sub_ps(iz3,jz1);
 266             dx32             = _mm_sub_ps(ix3,jx2);
 267             dy32             = _mm_sub_ps(iy3,jy2);
 268             dz32             = _mm_sub_ps(iz3,jz2);
 269             dx33             = _mm_sub_ps(ix3,jx3);
 270             dy33             = _mm_sub_ps(iy3,jy3);
 271             dz33             = _mm_sub_ps(iz3,jz3);
 272
 273             /* Calculate squared distance and things based on it */
 274             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 275             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 276             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 277             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 278             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 279             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 280             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 281             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 282             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 283             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 284
 285             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 286             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 287             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 288             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 289             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 290             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 291             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 292             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 293             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 294             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 295
 296             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 297             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 298             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 299             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 300             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 301             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 302             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 303             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 304             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 305
 306             fjx0             = _mm_setzero_ps();
 307             fjy0             = _mm_setzero_ps();
 308             fjz0             = _mm_setzero_ps();
 309             fjx1             = _mm_setzero_ps();
 310             fjy1             = _mm_setzero_ps();
 311             fjz1             = _mm_setzero_ps();
 312             fjx2             = _mm_setzero_ps();
 313             fjy2             = _mm_setzero_ps();
 314             fjz2             = _mm_setzero_ps();
 315             fjx3             = _mm_setzero_ps();
 316             fjy3             = _mm_setzero_ps();
 317             fjz3             = _mm_setzero_ps();
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             r00              = _mm_mul_ps(rsq00,rinv00);
 324
 325             /* Calculate table index by multiplying r with table scale and truncate to integer */
 326             rt               = _mm_mul_ps(r00,vftabscale);
 327             vfitab           = _mm_cvttps_epi32(rt);
 328             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 329             vfitab           = _mm_slli_epi32(vfitab,3);
 330
 331             /* CUBIC SPLINE TABLE DISPERSION */
 332             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 333             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 334             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 335             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 336             _MM_TRANSPOSE4_PS(Y,F,G,H);
 337             Heps             = _mm_mul_ps(vfeps,H);
 338             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 339             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 340             vvdw6            = _mm_mul_ps(c6_00,VV);
 341             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 342             fvdw6            = _mm_mul_ps(c6_00,FF);
 343
 344             /* CUBIC SPLINE TABLE REPULSION */
 345             vfitab           = _mm_add_epi32(vfitab,ifour);
 346             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 347             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 348             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 349             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 350             _MM_TRANSPOSE4_PS(Y,F,G,H);
 351             Heps             = _mm_mul_ps(vfeps,H);
 352             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 353             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 354             vvdw12           = _mm_mul_ps(c12_00,VV);
 355             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 356             fvdw12           = _mm_mul_ps(c12_00,FF);
 357             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 358             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 359
 360             /* Update potential sum for this i atom from the interaction with this j atom. */
 361             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 362
 363             fscal            = fvdw;
 364
 365             /* Calculate temporary vectorial force */
 366             tx               = _mm_mul_ps(fscal,dx00);
 367             ty               = _mm_mul_ps(fscal,dy00);
 368             tz               = _mm_mul_ps(fscal,dz00);
 369
 370             /* Update vectorial force */
 371             fix0             = _mm_add_ps(fix0,tx);
 372             fiy0             = _mm_add_ps(fiy0,ty);
 373             fiz0             = _mm_add_ps(fiz0,tz);
 374
 375             fjx0             = _mm_add_ps(fjx0,tx);
 376             fjy0             = _mm_add_ps(fjy0,ty);
 377             fjz0             = _mm_add_ps(fjz0,tz);
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* REACTION-FIELD ELECTROSTATICS */
 384             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 385             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velecsum         = _mm_add_ps(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392             /* Calculate temporary vectorial force */
 393             tx               = _mm_mul_ps(fscal,dx11);
 394             ty               = _mm_mul_ps(fscal,dy11);
 395             tz               = _mm_mul_ps(fscal,dz11);
 396
 397             /* Update vectorial force */
 398             fix1             = _mm_add_ps(fix1,tx);
 399             fiy1             = _mm_add_ps(fiy1,ty);
 400             fiz1             = _mm_add_ps(fiz1,tz);
 401
 402             fjx1             = _mm_add_ps(fjx1,tx);
 403             fjy1             = _mm_add_ps(fjy1,ty);
 404             fjz1             = _mm_add_ps(fjz1,tz);
 405
 406             /**************************
 407              * CALCULATE INTERACTIONS *
 408              **************************/
 409
 410             /* REACTION-FIELD ELECTROSTATICS */
 411             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 412             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velecsum         = _mm_add_ps(velecsum,velec);
 416
 417             fscal            = felec;
 418
 419             /* Calculate temporary vectorial force */
 420             tx               = _mm_mul_ps(fscal,dx12);
 421             ty               = _mm_mul_ps(fscal,dy12);
 422             tz               = _mm_mul_ps(fscal,dz12);
 423
 424             /* Update vectorial force */
 425             fix1             = _mm_add_ps(fix1,tx);
 426             fiy1             = _mm_add_ps(fiy1,ty);
 427             fiz1             = _mm_add_ps(fiz1,tz);
 428
 429             fjx2             = _mm_add_ps(fjx2,tx);
 430             fjy2             = _mm_add_ps(fjy2,ty);
 431             fjz2             = _mm_add_ps(fjz2,tz);
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 439             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 440
 441             /* Update potential sum for this i atom from the interaction with this j atom. */
 442             velecsum         = _mm_add_ps(velecsum,velec);
 443
 444             fscal            = felec;
 445
 446             /* Calculate temporary vectorial force */
 447             tx               = _mm_mul_ps(fscal,dx13);
 448             ty               = _mm_mul_ps(fscal,dy13);
 449             tz               = _mm_mul_ps(fscal,dz13);
 450
 451             /* Update vectorial force */
 452             fix1             = _mm_add_ps(fix1,tx);
 453             fiy1             = _mm_add_ps(fiy1,ty);
 454             fiz1             = _mm_add_ps(fiz1,tz);
 455
 456             fjx3             = _mm_add_ps(fjx3,tx);
 457             fjy3             = _mm_add_ps(fjy3,ty);
 458             fjz3             = _mm_add_ps(fjz3,tz);
 459
 460             /**************************
 461              * CALCULATE INTERACTIONS *
 462              **************************/
 463
 464             /* REACTION-FIELD ELECTROSTATICS */
 465             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 466             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 467
 468             /* Update potential sum for this i atom from the interaction with this j atom. */
 469             velecsum         = _mm_add_ps(velecsum,velec);
 470
 471             fscal            = felec;
 472
 473             /* Calculate temporary vectorial force */
 474             tx               = _mm_mul_ps(fscal,dx21);
 475             ty               = _mm_mul_ps(fscal,dy21);
 476             tz               = _mm_mul_ps(fscal,dz21);
 477
 478             /* Update vectorial force */
 479             fix2             = _mm_add_ps(fix2,tx);
 480             fiy2             = _mm_add_ps(fiy2,ty);
 481             fiz2             = _mm_add_ps(fiz2,tz);
 482
 483             fjx1             = _mm_add_ps(fjx1,tx);
 484             fjy1             = _mm_add_ps(fjy1,ty);
 485             fjz1             = _mm_add_ps(fjz1,tz);
 486
 487             /**************************
 488              * CALCULATE INTERACTIONS *
 489              **************************/
 490
 491             /* REACTION-FIELD ELECTROSTATICS */
 492             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 493             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 494
 495             /* Update potential sum for this i atom from the interaction with this j atom. */
 496             velecsum         = _mm_add_ps(velecsum,velec);
 497
 498             fscal            = felec;
 499
 500             /* Calculate temporary vectorial force */
 501             tx               = _mm_mul_ps(fscal,dx22);
 502             ty               = _mm_mul_ps(fscal,dy22);
 503             tz               = _mm_mul_ps(fscal,dz22);
 504
 505             /* Update vectorial force */
 506             fix2             = _mm_add_ps(fix2,tx);
 507             fiy2             = _mm_add_ps(fiy2,ty);
 508             fiz2             = _mm_add_ps(fiz2,tz);
 509
 510             fjx2             = _mm_add_ps(fjx2,tx);
 511             fjy2             = _mm_add_ps(fjy2,ty);
 512             fjz2             = _mm_add_ps(fjz2,tz);
 513
 514             /**************************
 515              * CALCULATE INTERACTIONS *
 516              **************************/
 517
 518             /* REACTION-FIELD ELECTROSTATICS */
 519             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 520             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 521
 522             /* Update potential sum for this i atom from the interaction with this j atom. */
 523             velecsum         = _mm_add_ps(velecsum,velec);
 524
 525             fscal            = felec;
 526
 527             /* Calculate temporary vectorial force */
 528             tx               = _mm_mul_ps(fscal,dx23);
 529             ty               = _mm_mul_ps(fscal,dy23);
 530             tz               = _mm_mul_ps(fscal,dz23);
 531
 532             /* Update vectorial force */
 533             fix2             = _mm_add_ps(fix2,tx);
 534             fiy2             = _mm_add_ps(fiy2,ty);
 535             fiz2             = _mm_add_ps(fiz2,tz);
 536
 537             fjx3             = _mm_add_ps(fjx3,tx);
 538             fjy3             = _mm_add_ps(fjy3,ty);
 539             fjz3             = _mm_add_ps(fjz3,tz);
 540
 541             /**************************
 542              * CALCULATE INTERACTIONS *
 543              **************************/
 544
 545             /* REACTION-FIELD ELECTROSTATICS */
 546             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 547             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 548
 549             /* Update potential sum for this i atom from the interaction with this j atom. */
 550             velecsum         = _mm_add_ps(velecsum,velec);
 551
 552             fscal            = felec;
 553
 554             /* Calculate temporary vectorial force */
 555             tx               = _mm_mul_ps(fscal,dx31);
 556             ty               = _mm_mul_ps(fscal,dy31);
 557             tz               = _mm_mul_ps(fscal,dz31);
 558
 559             /* Update vectorial force */
 560             fix3             = _mm_add_ps(fix3,tx);
 561             fiy3             = _mm_add_ps(fiy3,ty);
 562             fiz3             = _mm_add_ps(fiz3,tz);
 563
 564             fjx1             = _mm_add_ps(fjx1,tx);
 565             fjy1             = _mm_add_ps(fjy1,ty);
 566             fjz1             = _mm_add_ps(fjz1,tz);
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             /* REACTION-FIELD ELECTROSTATICS */
 573             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 574             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 575
 576             /* Update potential sum for this i atom from the interaction with this j atom. */
 577             velecsum         = _mm_add_ps(velecsum,velec);
 578
 579             fscal            = felec;
 580
 581             /* Calculate temporary vectorial force */
 582             tx               = _mm_mul_ps(fscal,dx32);
 583             ty               = _mm_mul_ps(fscal,dy32);
 584             tz               = _mm_mul_ps(fscal,dz32);
 585
 586             /* Update vectorial force */
 587             fix3             = _mm_add_ps(fix3,tx);
 588             fiy3             = _mm_add_ps(fiy3,ty);
 589             fiz3             = _mm_add_ps(fiz3,tz);
 590
 591             fjx2             = _mm_add_ps(fjx2,tx);
 592             fjy2             = _mm_add_ps(fjy2,ty);
 593             fjz2             = _mm_add_ps(fjz2,tz);
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             /* REACTION-FIELD ELECTROSTATICS */
 600             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 601             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 602
 603             /* Update potential sum for this i atom from the interaction with this j atom. */
 604             velecsum         = _mm_add_ps(velecsum,velec);
 605
 606             fscal            = felec;
 607
 608             /* Calculate temporary vectorial force */
 609             tx               = _mm_mul_ps(fscal,dx33);
 610             ty               = _mm_mul_ps(fscal,dy33);
 611             tz               = _mm_mul_ps(fscal,dz33);
 612
 613             /* Update vectorial force */
 614             fix3             = _mm_add_ps(fix3,tx);
 615             fiy3             = _mm_add_ps(fiy3,ty);
 616             fiz3             = _mm_add_ps(fiz3,tz);
 617
 618             fjx3             = _mm_add_ps(fjx3,tx);
 619             fjy3             = _mm_add_ps(fjy3,ty);
 620             fjz3             = _mm_add_ps(fjz3,tz);
 621
 622             fjptrA             = f+j_coord_offsetA;
 623             fjptrB             = f+j_coord_offsetB;
 624             fjptrC             = f+j_coord_offsetC;
 625             fjptrD             = f+j_coord_offsetD;
 626
 627             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 628                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 629                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 630
 631             /* Inner loop uses 347 flops */
 632         }
 633
 634         if(jidx<j_index_end)
 635         {
 636
 637             /* Get j neighbor index, and coordinate index */
 638             jnrlistA         = jjnr[jidx];
 639             jnrlistB         = jjnr[jidx+1];
 640             jnrlistC         = jjnr[jidx+2];
 641             jnrlistD         = jjnr[jidx+3];
 642             /* Sign of each element will be negative for non-real atoms.
 643              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 644              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 645              */
 646             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 647             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 648             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 649             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 650             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 651             j_coord_offsetA  = DIM*jnrA;
 652             j_coord_offsetB  = DIM*jnrB;
 653             j_coord_offsetC  = DIM*jnrC;
 654             j_coord_offsetD  = DIM*jnrD;
 655
 656             /* load j atom coordinates */
 657             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 658                                               x+j_coord_offsetC,x+j_coord_offsetD,
 659                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 660                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 661
 662             /* Calculate displacement vector */
 663             dx00             = _mm_sub_ps(ix0,jx0);
 664             dy00             = _mm_sub_ps(iy0,jy0);
 665             dz00             = _mm_sub_ps(iz0,jz0);
 666             dx11             = _mm_sub_ps(ix1,jx1);
 667             dy11             = _mm_sub_ps(iy1,jy1);
 668             dz11             = _mm_sub_ps(iz1,jz1);
 669             dx12             = _mm_sub_ps(ix1,jx2);
 670             dy12             = _mm_sub_ps(iy1,jy2);
 671             dz12             = _mm_sub_ps(iz1,jz2);
 672             dx13             = _mm_sub_ps(ix1,jx3);
 673             dy13             = _mm_sub_ps(iy1,jy3);
 674             dz13             = _mm_sub_ps(iz1,jz3);
 675             dx21             = _mm_sub_ps(ix2,jx1);
 676             dy21             = _mm_sub_ps(iy2,jy1);
 677             dz21             = _mm_sub_ps(iz2,jz1);
 678             dx22             = _mm_sub_ps(ix2,jx2);
 679             dy22             = _mm_sub_ps(iy2,jy2);
 680             dz22             = _mm_sub_ps(iz2,jz2);
 681             dx23             = _mm_sub_ps(ix2,jx3);
 682             dy23             = _mm_sub_ps(iy2,jy3);
 683             dz23             = _mm_sub_ps(iz2,jz3);
 684             dx31             = _mm_sub_ps(ix3,jx1);
 685             dy31             = _mm_sub_ps(iy3,jy1);
 686             dz31             = _mm_sub_ps(iz3,jz1);
 687             dx32             = _mm_sub_ps(ix3,jx2);
 688             dy32             = _mm_sub_ps(iy3,jy2);
 689             dz32             = _mm_sub_ps(iz3,jz2);
 690             dx33             = _mm_sub_ps(ix3,jx3);
 691             dy33             = _mm_sub_ps(iy3,jy3);
 692             dz33             = _mm_sub_ps(iz3,jz3);
 693
 694             /* Calculate squared distance and things based on it */
 695             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 696             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 697             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 698             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 699             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 700             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 701             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 702             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 703             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 704             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 705
 706             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 707             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 708             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 709             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 710             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 711             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 712             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 713             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 714             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 715             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 716
 717             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 718             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 719             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 720             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 721             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 722             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 723             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 724             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 725             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 726
 727             fjx0             = _mm_setzero_ps();
 728             fjy0             = _mm_setzero_ps();
 729             fjz0             = _mm_setzero_ps();
 730             fjx1             = _mm_setzero_ps();
 731             fjy1             = _mm_setzero_ps();
 732             fjz1             = _mm_setzero_ps();
 733             fjx2             = _mm_setzero_ps();
 734             fjy2             = _mm_setzero_ps();
 735             fjz2             = _mm_setzero_ps();
 736             fjx3             = _mm_setzero_ps();
 737             fjy3             = _mm_setzero_ps();
 738             fjz3             = _mm_setzero_ps();
 739
 740             /**************************
 741              * CALCULATE INTERACTIONS *
 742              **************************/
 743
 744             r00              = _mm_mul_ps(rsq00,rinv00);
 745             r00              = _mm_andnot_ps(dummy_mask,r00);
 746
 747             /* Calculate table index by multiplying r with table scale and truncate to integer */
 748             rt               = _mm_mul_ps(r00,vftabscale);
 749             vfitab           = _mm_cvttps_epi32(rt);
 750             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 751             vfitab           = _mm_slli_epi32(vfitab,3);
 752
 753             /* CUBIC SPLINE TABLE DISPERSION */
 754             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 755             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 756             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 757             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 758             _MM_TRANSPOSE4_PS(Y,F,G,H);
 759             Heps             = _mm_mul_ps(vfeps,H);
 760             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 761             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 762             vvdw6            = _mm_mul_ps(c6_00,VV);
 763             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 764             fvdw6            = _mm_mul_ps(c6_00,FF);
 765
 766             /* CUBIC SPLINE TABLE REPULSION */
 767             vfitab           = _mm_add_epi32(vfitab,ifour);
 768             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 769             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 770             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 771             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 772             _MM_TRANSPOSE4_PS(Y,F,G,H);
 773             Heps             = _mm_mul_ps(vfeps,H);
 774             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 775             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 776             vvdw12           = _mm_mul_ps(c12_00,VV);
 777             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 778             fvdw12           = _mm_mul_ps(c12_00,FF);
 779             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 780             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 781
 782             /* Update potential sum for this i atom from the interaction with this j atom. */
 783             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 784             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 785
 786             fscal            = fvdw;
 787
 788             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 789
 790             /* Calculate temporary vectorial force */
 791             tx               = _mm_mul_ps(fscal,dx00);
 792             ty               = _mm_mul_ps(fscal,dy00);
 793             tz               = _mm_mul_ps(fscal,dz00);
 794
 795             /* Update vectorial force */
 796             fix0             = _mm_add_ps(fix0,tx);
 797             fiy0             = _mm_add_ps(fiy0,ty);
 798             fiz0             = _mm_add_ps(fiz0,tz);
 799
 800             fjx0             = _mm_add_ps(fjx0,tx);
 801             fjy0             = _mm_add_ps(fjy0,ty);
 802             fjz0             = _mm_add_ps(fjz0,tz);
 803
 804             /**************************
 805              * CALCULATE INTERACTIONS *
 806              **************************/
 807
 808             /* REACTION-FIELD ELECTROSTATICS */
 809             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 810             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 811
 812             /* Update potential sum for this i atom from the interaction with this j atom. */
 813             velec            = _mm_andnot_ps(dummy_mask,velec);
 814             velecsum         = _mm_add_ps(velecsum,velec);
 815
 816             fscal            = felec;
 817
 818             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 819
 820             /* Calculate temporary vectorial force */
 821             tx               = _mm_mul_ps(fscal,dx11);
 822             ty               = _mm_mul_ps(fscal,dy11);
 823             tz               = _mm_mul_ps(fscal,dz11);
 824
 825             /* Update vectorial force */
 826             fix1             = _mm_add_ps(fix1,tx);
 827             fiy1             = _mm_add_ps(fiy1,ty);
 828             fiz1             = _mm_add_ps(fiz1,tz);
 829
 830             fjx1             = _mm_add_ps(fjx1,tx);
 831             fjy1             = _mm_add_ps(fjy1,ty);
 832             fjz1             = _mm_add_ps(fjz1,tz);
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             /* REACTION-FIELD ELECTROSTATICS */
 839             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 840             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 841
 842             /* Update potential sum for this i atom from the interaction with this j atom. */
 843             velec            = _mm_andnot_ps(dummy_mask,velec);
 844             velecsum         = _mm_add_ps(velecsum,velec);
 845
 846             fscal            = felec;
 847
 848             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 849
 850             /* Calculate temporary vectorial force */
 851             tx               = _mm_mul_ps(fscal,dx12);
 852             ty               = _mm_mul_ps(fscal,dy12);
 853             tz               = _mm_mul_ps(fscal,dz12);
 854
 855             /* Update vectorial force */
 856             fix1             = _mm_add_ps(fix1,tx);
 857             fiy1             = _mm_add_ps(fiy1,ty);
 858             fiz1             = _mm_add_ps(fiz1,tz);
 859
 860             fjx2             = _mm_add_ps(fjx2,tx);
 861             fjy2             = _mm_add_ps(fjy2,ty);
 862             fjz2             = _mm_add_ps(fjz2,tz);
 863
 864             /**************************
 865              * CALCULATE INTERACTIONS *
 866              **************************/
 867
 868             /* REACTION-FIELD ELECTROSTATICS */
 869             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 870             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 871
 872             /* Update potential sum for this i atom from the interaction with this j atom. */
 873             velec            = _mm_andnot_ps(dummy_mask,velec);
 874             velecsum         = _mm_add_ps(velecsum,velec);
 875
 876             fscal            = felec;
 877
 878             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 879
 880             /* Calculate temporary vectorial force */
 881             tx               = _mm_mul_ps(fscal,dx13);
 882             ty               = _mm_mul_ps(fscal,dy13);
 883             tz               = _mm_mul_ps(fscal,dz13);
 884
 885             /* Update vectorial force */
 886             fix1             = _mm_add_ps(fix1,tx);
 887             fiy1             = _mm_add_ps(fiy1,ty);
 888             fiz1             = _mm_add_ps(fiz1,tz);
 889
 890             fjx3             = _mm_add_ps(fjx3,tx);
 891             fjy3             = _mm_add_ps(fjy3,ty);
 892             fjz3             = _mm_add_ps(fjz3,tz);
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             /* REACTION-FIELD ELECTROSTATICS */
 899             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 900             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 901
 902             /* Update potential sum for this i atom from the interaction with this j atom. */
 903             velec            = _mm_andnot_ps(dummy_mask,velec);
 904             velecsum         = _mm_add_ps(velecsum,velec);
 905
 906             fscal            = felec;
 907
 908             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 909
 910             /* Calculate temporary vectorial force */
 911             tx               = _mm_mul_ps(fscal,dx21);
 912             ty               = _mm_mul_ps(fscal,dy21);
 913             tz               = _mm_mul_ps(fscal,dz21);
 914
 915             /* Update vectorial force */
 916             fix2             = _mm_add_ps(fix2,tx);
 917             fiy2             = _mm_add_ps(fiy2,ty);
 918             fiz2             = _mm_add_ps(fiz2,tz);
 919
 920             fjx1             = _mm_add_ps(fjx1,tx);
 921             fjy1             = _mm_add_ps(fjy1,ty);
 922             fjz1             = _mm_add_ps(fjz1,tz);
 923
 924             /**************************
 925              * CALCULATE INTERACTIONS *
 926              **************************/
 927
 928             /* REACTION-FIELD ELECTROSTATICS */
 929             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 930             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 931
 932             /* Update potential sum for this i atom from the interaction with this j atom. */
 933             velec            = _mm_andnot_ps(dummy_mask,velec);
 934             velecsum         = _mm_add_ps(velecsum,velec);
 935
 936             fscal            = felec;
 937
 938             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 939
 940             /* Calculate temporary vectorial force */
 941             tx               = _mm_mul_ps(fscal,dx22);
 942             ty               = _mm_mul_ps(fscal,dy22);
 943             tz               = _mm_mul_ps(fscal,dz22);
 944
 945             /* Update vectorial force */
 946             fix2             = _mm_add_ps(fix2,tx);
 947             fiy2             = _mm_add_ps(fiy2,ty);
 948             fiz2             = _mm_add_ps(fiz2,tz);
 949
 950             fjx2             = _mm_add_ps(fjx2,tx);
 951             fjy2             = _mm_add_ps(fjy2,ty);
 952             fjz2             = _mm_add_ps(fjz2,tz);
 953
 954             /**************************
 955              * CALCULATE INTERACTIONS *
 956              **************************/
 957
 958             /* REACTION-FIELD ELECTROSTATICS */
 959             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 960             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 961
 962             /* Update potential sum for this i atom from the interaction with this j atom. */
 963             velec            = _mm_andnot_ps(dummy_mask,velec);
 964             velecsum         = _mm_add_ps(velecsum,velec);
 965
 966             fscal            = felec;
 967
 968             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 969
 970             /* Calculate temporary vectorial force */
 971             tx               = _mm_mul_ps(fscal,dx23);
 972             ty               = _mm_mul_ps(fscal,dy23);
 973             tz               = _mm_mul_ps(fscal,dz23);
 974
 975             /* Update vectorial force */
 976             fix2             = _mm_add_ps(fix2,tx);
 977             fiy2             = _mm_add_ps(fiy2,ty);
 978             fiz2             = _mm_add_ps(fiz2,tz);
 979
 980             fjx3             = _mm_add_ps(fjx3,tx);
 981             fjy3             = _mm_add_ps(fjy3,ty);
 982             fjz3             = _mm_add_ps(fjz3,tz);
 983
 984             /**************************
 985              * CALCULATE INTERACTIONS *
 986              **************************/
 987
 988             /* REACTION-FIELD ELECTROSTATICS */
 989             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 990             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 991
 992             /* Update potential sum for this i atom from the interaction with this j atom. */
 993             velec            = _mm_andnot_ps(dummy_mask,velec);
 994             velecsum         = _mm_add_ps(velecsum,velec);
 995
 996             fscal            = felec;
 997
 998             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 999
1000             /* Calculate temporary vectorial force */
1001             tx               = _mm_mul_ps(fscal,dx31);
1002             ty               = _mm_mul_ps(fscal,dy31);
1003             tz               = _mm_mul_ps(fscal,dz31);
1004
1005             /* Update vectorial force */
1006             fix3             = _mm_add_ps(fix3,tx);
1007             fiy3             = _mm_add_ps(fiy3,ty);
1008             fiz3             = _mm_add_ps(fiz3,tz);
1009
1010             fjx1             = _mm_add_ps(fjx1,tx);
1011             fjy1             = _mm_add_ps(fjy1,ty);
1012             fjz1             = _mm_add_ps(fjz1,tz);
1013
1014             /**************************
1015              * CALCULATE INTERACTIONS *
1016              **************************/
1017
1018             /* REACTION-FIELD ELECTROSTATICS */
1019             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1020             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1021
1022             /* Update potential sum for this i atom from the interaction with this j atom. */
1023             velec            = _mm_andnot_ps(dummy_mask,velec);
1024             velecsum         = _mm_add_ps(velecsum,velec);
1025
1026             fscal            = felec;
1027
1028             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1029
1030             /* Calculate temporary vectorial force */
1031             tx               = _mm_mul_ps(fscal,dx32);
1032             ty               = _mm_mul_ps(fscal,dy32);
1033             tz               = _mm_mul_ps(fscal,dz32);
1034
1035             /* Update vectorial force */
1036             fix3             = _mm_add_ps(fix3,tx);
1037             fiy3             = _mm_add_ps(fiy3,ty);
1038             fiz3             = _mm_add_ps(fiz3,tz);
1039
1040             fjx2             = _mm_add_ps(fjx2,tx);
1041             fjy2             = _mm_add_ps(fjy2,ty);
1042             fjz2             = _mm_add_ps(fjz2,tz);
1043
1044             /**************************
1045              * CALCULATE INTERACTIONS *
1046              **************************/
1047
1048             /* REACTION-FIELD ELECTROSTATICS */
1049             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1050             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1051
1052             /* Update potential sum for this i atom from the interaction with this j atom. */
1053             velec            = _mm_andnot_ps(dummy_mask,velec);
1054             velecsum         = _mm_add_ps(velecsum,velec);
1055
1056             fscal            = felec;
1057
1058             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1059
1060             /* Calculate temporary vectorial force */
1061             tx               = _mm_mul_ps(fscal,dx33);
1062             ty               = _mm_mul_ps(fscal,dy33);
1063             tz               = _mm_mul_ps(fscal,dz33);
1064
1065             /* Update vectorial force */
1066             fix3             = _mm_add_ps(fix3,tx);
1067             fiy3             = _mm_add_ps(fiy3,ty);
1068             fiz3             = _mm_add_ps(fiz3,tz);
1069
1070             fjx3             = _mm_add_ps(fjx3,tx);
1071             fjy3             = _mm_add_ps(fjy3,ty);
1072             fjz3             = _mm_add_ps(fjz3,tz);
1073
1074             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1075             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1076             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1077             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1078
1079             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1080                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1081                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1082
1083             /* Inner loop uses 348 flops */
1084         }
1085
1086         /* End of innermost loop */
1087
1088         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1089                                               f+i_coord_offset,fshift+i_shift_offset);
1090
1091         ggid                        = gid[iidx];
1092         /* Update potential energies */
1093         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1094         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1095
1096         /* Increment number of inner iterations */
1097         inneriter                  += j_index_end - j_index_start;
1098
1099         /* Outer loop uses 26 flops */
1100     }
1101
1102     /* Increment number of outer iterations */
1103     outeriter        += nri;
1104
1105     /* Update outer/inner flops */
1106
1107     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*348);
1108 }
1109 /*
1110  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse4_1_single
1111  * Electrostatics interaction: ReactionField
1112  * VdW interaction:            CubicSplineTable
1113  * Geometry:                   Water4-Water4
1114  * Calculate force/pot:        Force
1115  */
1116 void
1117 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse4_1_single
1118                     (t_nblist                    * gmx_restrict       nlist,
1119                      rvec                        * gmx_restrict          xx,
1120                      rvec                        * gmx_restrict          ff,
1121                      t_forcerec                  * gmx_restrict          fr,
1122                      t_mdatoms                   * gmx_restrict     mdatoms,
1123                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1124                      t_nrnb                      * gmx_restrict        nrnb)
1125 {
1126     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1127      * just 0 for non-waters.
1128      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1129      * jnr indices corresponding to data put in the four positions in the SIMD register.
1130      */
1131     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1132     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1133     int              jnrA,jnrB,jnrC,jnrD;
1134     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1135     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1136     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1137     real             rcutoff_scalar;
1138     real             *shiftvec,*fshift,*x,*f;
1139     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1140     real             scratch[4*DIM];
1141     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1142     int              vdwioffset0;
1143     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1144     int              vdwioffset1;
1145     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1146     int              vdwioffset2;
1147     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1148     int              vdwioffset3;
1149     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1150     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1151     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1152     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1153     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1154     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1155     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1156     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1157     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1158     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1159     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1160     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1161     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1162     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1163     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1164     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1165     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1166     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1167     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1168     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1169     real             *charge;
1170     int              nvdwtype;
1171     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1172     int              *vdwtype;
1173     real             *vdwparam;
1174     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1175     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1176     __m128i          vfitab;
1177     __m128i          ifour       = _mm_set1_epi32(4);
1178     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1179     real             *vftab;
1180     __m128           dummy_mask,cutoff_mask;
1181     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1182     __m128           one     = _mm_set1_ps(1.0);
1183     __m128           two     = _mm_set1_ps(2.0);
1184     x                = xx[0];
1185     f                = ff[0];
1186
1187     nri              = nlist->nri;
1188     iinr             = nlist->iinr;
1189     jindex           = nlist->jindex;
1190     jjnr             = nlist->jjnr;
1191     shiftidx         = nlist->shift;
1192     gid              = nlist->gid;
1193     shiftvec         = fr->shift_vec[0];
1194     fshift           = fr->fshift[0];
1195     facel            = _mm_set1_ps(fr->epsfac);
1196     charge           = mdatoms->chargeA;
1197     krf              = _mm_set1_ps(fr->ic->k_rf);
1198     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1199     crf              = _mm_set1_ps(fr->ic->c_rf);
1200     nvdwtype         = fr->ntype;
1201     vdwparam         = fr->nbfp;
1202     vdwtype          = mdatoms->typeA;
1203
1204     vftab            = kernel_data->table_vdw->data;
1205     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1206
1207     /* Setup water-specific parameters */
1208     inr              = nlist->iinr[0];
1209     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1210     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1211     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1212     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1213
1214     jq1              = _mm_set1_ps(charge[inr+1]);
1215     jq2              = _mm_set1_ps(charge[inr+2]);
1216     jq3              = _mm_set1_ps(charge[inr+3]);
1217     vdwjidx0A        = 2*vdwtype[inr+0];
1218     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1219     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1220     qq11             = _mm_mul_ps(iq1,jq1);
1221     qq12             = _mm_mul_ps(iq1,jq2);
1222     qq13             = _mm_mul_ps(iq1,jq3);
1223     qq21             = _mm_mul_ps(iq2,jq1);
1224     qq22             = _mm_mul_ps(iq2,jq2);
1225     qq23             = _mm_mul_ps(iq2,jq3);
1226     qq31             = _mm_mul_ps(iq3,jq1);
1227     qq32             = _mm_mul_ps(iq3,jq2);
1228     qq33             = _mm_mul_ps(iq3,jq3);
1229
1230     /* Avoid stupid compiler warnings */
1231     jnrA = jnrB = jnrC = jnrD = 0;
1232     j_coord_offsetA = 0;
1233     j_coord_offsetB = 0;
1234     j_coord_offsetC = 0;
1235     j_coord_offsetD = 0;
1236
1237     outeriter        = 0;
1238     inneriter        = 0;
1239
1240     for(iidx=0;iidx<4*DIM;iidx++)
1241     {
1242         scratch[iidx] = 0.0;
1243     }
1244
1245     /* Start outer loop over neighborlists */
1246     for(iidx=0; iidx<nri; iidx++)
1247     {
1248         /* Load shift vector for this list */
1249         i_shift_offset   = DIM*shiftidx[iidx];
1250
1251         /* Load limits for loop over neighbors */
1252         j_index_start    = jindex[iidx];
1253         j_index_end      = jindex[iidx+1];
1254
1255         /* Get outer coordinate index */
1256         inr              = iinr[iidx];
1257         i_coord_offset   = DIM*inr;
1258
1259         /* Load i particle coords and add shift vector */
1260         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1261                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1262
1263         fix0             = _mm_setzero_ps();
1264         fiy0             = _mm_setzero_ps();
1265         fiz0             = _mm_setzero_ps();
1266         fix1             = _mm_setzero_ps();
1267         fiy1             = _mm_setzero_ps();
1268         fiz1             = _mm_setzero_ps();
1269         fix2             = _mm_setzero_ps();
1270         fiy2             = _mm_setzero_ps();
1271         fiz2             = _mm_setzero_ps();
1272         fix3             = _mm_setzero_ps();
1273         fiy3             = _mm_setzero_ps();
1274         fiz3             = _mm_setzero_ps();
1275
1276         /* Start inner kernel loop */
1277         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1278         {
1279
1280             /* Get j neighbor index, and coordinate index */
1281             jnrA             = jjnr[jidx];
1282             jnrB             = jjnr[jidx+1];
1283             jnrC             = jjnr[jidx+2];
1284             jnrD             = jjnr[jidx+3];
1285             j_coord_offsetA  = DIM*jnrA;
1286             j_coord_offsetB  = DIM*jnrB;
1287             j_coord_offsetC  = DIM*jnrC;
1288             j_coord_offsetD  = DIM*jnrD;
1289
1290             /* load j atom coordinates */
1291             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1292                                               x+j_coord_offsetC,x+j_coord_offsetD,
1293                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1294                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1295
1296             /* Calculate displacement vector */
1297             dx00             = _mm_sub_ps(ix0,jx0);
1298             dy00             = _mm_sub_ps(iy0,jy0);
1299             dz00             = _mm_sub_ps(iz0,jz0);
1300             dx11             = _mm_sub_ps(ix1,jx1);
1301             dy11             = _mm_sub_ps(iy1,jy1);
1302             dz11             = _mm_sub_ps(iz1,jz1);
1303             dx12             = _mm_sub_ps(ix1,jx2);
1304             dy12             = _mm_sub_ps(iy1,jy2);
1305             dz12             = _mm_sub_ps(iz1,jz2);
1306             dx13             = _mm_sub_ps(ix1,jx3);
1307             dy13             = _mm_sub_ps(iy1,jy3);
1308             dz13             = _mm_sub_ps(iz1,jz3);
1309             dx21             = _mm_sub_ps(ix2,jx1);
1310             dy21             = _mm_sub_ps(iy2,jy1);
1311             dz21             = _mm_sub_ps(iz2,jz1);
1312             dx22             = _mm_sub_ps(ix2,jx2);
1313             dy22             = _mm_sub_ps(iy2,jy2);
1314             dz22             = _mm_sub_ps(iz2,jz2);
1315             dx23             = _mm_sub_ps(ix2,jx3);
1316             dy23             = _mm_sub_ps(iy2,jy3);
1317             dz23             = _mm_sub_ps(iz2,jz3);
1318             dx31             = _mm_sub_ps(ix3,jx1);
1319             dy31             = _mm_sub_ps(iy3,jy1);
1320             dz31             = _mm_sub_ps(iz3,jz1);
1321             dx32             = _mm_sub_ps(ix3,jx2);
1322             dy32             = _mm_sub_ps(iy3,jy2);
1323             dz32             = _mm_sub_ps(iz3,jz2);
1324             dx33             = _mm_sub_ps(ix3,jx3);
1325             dy33             = _mm_sub_ps(iy3,jy3);
1326             dz33             = _mm_sub_ps(iz3,jz3);
1327
1328             /* Calculate squared distance and things based on it */
1329             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1330             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1331             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1332             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1333             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1334             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1335             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1336             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1337             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1338             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1339
1340             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1341             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1342             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1343             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1344             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1345             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1346             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1347             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1348             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1349             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1350
1351             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1352             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1353             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1354             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1355             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1356             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1357             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1358             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1359             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1360
1361             fjx0             = _mm_setzero_ps();
1362             fjy0             = _mm_setzero_ps();
1363             fjz0             = _mm_setzero_ps();
1364             fjx1             = _mm_setzero_ps();
1365             fjy1             = _mm_setzero_ps();
1366             fjz1             = _mm_setzero_ps();
1367             fjx2             = _mm_setzero_ps();
1368             fjy2             = _mm_setzero_ps();
1369             fjz2             = _mm_setzero_ps();
1370             fjx3             = _mm_setzero_ps();
1371             fjy3             = _mm_setzero_ps();
1372             fjz3             = _mm_setzero_ps();
1373
1374             /**************************
1375              * CALCULATE INTERACTIONS *
1376              **************************/
1377
1378             r00              = _mm_mul_ps(rsq00,rinv00);
1379
1380             /* Calculate table index by multiplying r with table scale and truncate to integer */
1381             rt               = _mm_mul_ps(r00,vftabscale);
1382             vfitab           = _mm_cvttps_epi32(rt);
1383             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1384             vfitab           = _mm_slli_epi32(vfitab,3);
1385
1386             /* CUBIC SPLINE TABLE DISPERSION */
1387             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1388             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1389             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1390             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1391             _MM_TRANSPOSE4_PS(Y,F,G,H);
1392             Heps             = _mm_mul_ps(vfeps,H);
1393             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1394             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1395             fvdw6            = _mm_mul_ps(c6_00,FF);
1396
1397             /* CUBIC SPLINE TABLE REPULSION */
1398             vfitab           = _mm_add_epi32(vfitab,ifour);
1399             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1400             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1401             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1402             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1403             _MM_TRANSPOSE4_PS(Y,F,G,H);
1404             Heps             = _mm_mul_ps(vfeps,H);
1405             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1406             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1407             fvdw12           = _mm_mul_ps(c12_00,FF);
1408             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1409
1410             fscal            = fvdw;
1411
1412             /* Calculate temporary vectorial force */
1413             tx               = _mm_mul_ps(fscal,dx00);
1414             ty               = _mm_mul_ps(fscal,dy00);
1415             tz               = _mm_mul_ps(fscal,dz00);
1416
1417             /* Update vectorial force */
1418             fix0             = _mm_add_ps(fix0,tx);
1419             fiy0             = _mm_add_ps(fiy0,ty);
1420             fiz0             = _mm_add_ps(fiz0,tz);
1421
1422             fjx0             = _mm_add_ps(fjx0,tx);
1423             fjy0             = _mm_add_ps(fjy0,ty);
1424             fjz0             = _mm_add_ps(fjz0,tz);
1425
1426             /**************************
1427              * CALCULATE INTERACTIONS *
1428              **************************/
1429
1430             /* REACTION-FIELD ELECTROSTATICS */
1431             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1432
1433             fscal            = felec;
1434
1435             /* Calculate temporary vectorial force */
1436             tx               = _mm_mul_ps(fscal,dx11);
1437             ty               = _mm_mul_ps(fscal,dy11);
1438             tz               = _mm_mul_ps(fscal,dz11);
1439
1440             /* Update vectorial force */
1441             fix1             = _mm_add_ps(fix1,tx);
1442             fiy1             = _mm_add_ps(fiy1,ty);
1443             fiz1             = _mm_add_ps(fiz1,tz);
1444
1445             fjx1             = _mm_add_ps(fjx1,tx);
1446             fjy1             = _mm_add_ps(fjy1,ty);
1447             fjz1             = _mm_add_ps(fjz1,tz);
1448
1449             /**************************
1450              * CALCULATE INTERACTIONS *
1451              **************************/
1452
1453             /* REACTION-FIELD ELECTROSTATICS */
1454             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1455
1456             fscal            = felec;
1457
1458             /* Calculate temporary vectorial force */
1459             tx               = _mm_mul_ps(fscal,dx12);
1460             ty               = _mm_mul_ps(fscal,dy12);
1461             tz               = _mm_mul_ps(fscal,dz12);
1462
1463             /* Update vectorial force */
1464             fix1             = _mm_add_ps(fix1,tx);
1465             fiy1             = _mm_add_ps(fiy1,ty);
1466             fiz1             = _mm_add_ps(fiz1,tz);
1467
1468             fjx2             = _mm_add_ps(fjx2,tx);
1469             fjy2             = _mm_add_ps(fjy2,ty);
1470             fjz2             = _mm_add_ps(fjz2,tz);
1471
1472             /**************************
1473              * CALCULATE INTERACTIONS *
1474              **************************/
1475
1476             /* REACTION-FIELD ELECTROSTATICS */
1477             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1478
1479             fscal            = felec;
1480
1481             /* Calculate temporary vectorial force */
1482             tx               = _mm_mul_ps(fscal,dx13);
1483             ty               = _mm_mul_ps(fscal,dy13);
1484             tz               = _mm_mul_ps(fscal,dz13);
1485
1486             /* Update vectorial force */
1487             fix1             = _mm_add_ps(fix1,tx);
1488             fiy1             = _mm_add_ps(fiy1,ty);
1489             fiz1             = _mm_add_ps(fiz1,tz);
1490
1491             fjx3             = _mm_add_ps(fjx3,tx);
1492             fjy3             = _mm_add_ps(fjy3,ty);
1493             fjz3             = _mm_add_ps(fjz3,tz);
1494
1495             /**************************
1496              * CALCULATE INTERACTIONS *
1497              **************************/
1498
1499             /* REACTION-FIELD ELECTROSTATICS */
1500             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1501
1502             fscal            = felec;
1503
1504             /* Calculate temporary vectorial force */
1505             tx               = _mm_mul_ps(fscal,dx21);
1506             ty               = _mm_mul_ps(fscal,dy21);
1507             tz               = _mm_mul_ps(fscal,dz21);
1508
1509             /* Update vectorial force */
1510             fix2             = _mm_add_ps(fix2,tx);
1511             fiy2             = _mm_add_ps(fiy2,ty);
1512             fiz2             = _mm_add_ps(fiz2,tz);
1513
1514             fjx1             = _mm_add_ps(fjx1,tx);
1515             fjy1             = _mm_add_ps(fjy1,ty);
1516             fjz1             = _mm_add_ps(fjz1,tz);
1517
1518             /**************************
1519              * CALCULATE INTERACTIONS *
1520              **************************/
1521
1522             /* REACTION-FIELD ELECTROSTATICS */
1523             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1524
1525             fscal            = felec;
1526
1527             /* Calculate temporary vectorial force */
1528             tx               = _mm_mul_ps(fscal,dx22);
1529             ty               = _mm_mul_ps(fscal,dy22);
1530             tz               = _mm_mul_ps(fscal,dz22);
1531
1532             /* Update vectorial force */
1533             fix2             = _mm_add_ps(fix2,tx);
1534             fiy2             = _mm_add_ps(fiy2,ty);
1535             fiz2             = _mm_add_ps(fiz2,tz);
1536
1537             fjx2             = _mm_add_ps(fjx2,tx);
1538             fjy2             = _mm_add_ps(fjy2,ty);
1539             fjz2             = _mm_add_ps(fjz2,tz);
1540
1541             /**************************
1542              * CALCULATE INTERACTIONS *
1543              **************************/
1544
1545             /* REACTION-FIELD ELECTROSTATICS */
1546             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1547
1548             fscal            = felec;
1549
1550             /* Calculate temporary vectorial force */
1551             tx               = _mm_mul_ps(fscal,dx23);
1552             ty               = _mm_mul_ps(fscal,dy23);
1553             tz               = _mm_mul_ps(fscal,dz23);
1554
1555             /* Update vectorial force */
1556             fix2             = _mm_add_ps(fix2,tx);
1557             fiy2             = _mm_add_ps(fiy2,ty);
1558             fiz2             = _mm_add_ps(fiz2,tz);
1559
1560             fjx3             = _mm_add_ps(fjx3,tx);
1561             fjy3             = _mm_add_ps(fjy3,ty);
1562             fjz3             = _mm_add_ps(fjz3,tz);
1563
1564             /**************************
1565              * CALCULATE INTERACTIONS *
1566              **************************/
1567
1568             /* REACTION-FIELD ELECTROSTATICS */
1569             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1570
1571             fscal            = felec;
1572
1573             /* Calculate temporary vectorial force */
1574             tx               = _mm_mul_ps(fscal,dx31);
1575             ty               = _mm_mul_ps(fscal,dy31);
1576             tz               = _mm_mul_ps(fscal,dz31);
1577
1578             /* Update vectorial force */
1579             fix3             = _mm_add_ps(fix3,tx);
1580             fiy3             = _mm_add_ps(fiy3,ty);
1581             fiz3             = _mm_add_ps(fiz3,tz);
1582
1583             fjx1             = _mm_add_ps(fjx1,tx);
1584             fjy1             = _mm_add_ps(fjy1,ty);
1585             fjz1             = _mm_add_ps(fjz1,tz);
1586
1587             /**************************
1588              * CALCULATE INTERACTIONS *
1589              **************************/
1590
1591             /* REACTION-FIELD ELECTROSTATICS */
1592             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1593
1594             fscal            = felec;
1595
1596             /* Calculate temporary vectorial force */
1597             tx               = _mm_mul_ps(fscal,dx32);
1598             ty               = _mm_mul_ps(fscal,dy32);
1599             tz               = _mm_mul_ps(fscal,dz32);
1600
1601             /* Update vectorial force */
1602             fix3             = _mm_add_ps(fix3,tx);
1603             fiy3             = _mm_add_ps(fiy3,ty);
1604             fiz3             = _mm_add_ps(fiz3,tz);
1605
1606             fjx2             = _mm_add_ps(fjx2,tx);
1607             fjy2             = _mm_add_ps(fjy2,ty);
1608             fjz2             = _mm_add_ps(fjz2,tz);
1609
1610             /**************************
1611              * CALCULATE INTERACTIONS *
1612              **************************/
1613
1614             /* REACTION-FIELD ELECTROSTATICS */
1615             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1616
1617             fscal            = felec;
1618
1619             /* Calculate temporary vectorial force */
1620             tx               = _mm_mul_ps(fscal,dx33);
1621             ty               = _mm_mul_ps(fscal,dy33);
1622             tz               = _mm_mul_ps(fscal,dz33);
1623
1624             /* Update vectorial force */
1625             fix3             = _mm_add_ps(fix3,tx);
1626             fiy3             = _mm_add_ps(fiy3,ty);
1627             fiz3             = _mm_add_ps(fiz3,tz);
1628
1629             fjx3             = _mm_add_ps(fjx3,tx);
1630             fjy3             = _mm_add_ps(fjy3,ty);
1631             fjz3             = _mm_add_ps(fjz3,tz);
1632
1633             fjptrA             = f+j_coord_offsetA;
1634             fjptrB             = f+j_coord_offsetB;
1635             fjptrC             = f+j_coord_offsetC;
1636             fjptrD             = f+j_coord_offsetD;
1637
1638             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1639                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1640                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1641
1642             /* Inner loop uses 294 flops */
1643         }
1644
1645         if(jidx<j_index_end)
1646         {
1647
1648             /* Get j neighbor index, and coordinate index */
1649             jnrlistA         = jjnr[jidx];
1650             jnrlistB         = jjnr[jidx+1];
1651             jnrlistC         = jjnr[jidx+2];
1652             jnrlistD         = jjnr[jidx+3];
1653             /* Sign of each element will be negative for non-real atoms.
1654              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1655              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1656              */
1657             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1658             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1659             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1660             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1661             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1662             j_coord_offsetA  = DIM*jnrA;
1663             j_coord_offsetB  = DIM*jnrB;
1664             j_coord_offsetC  = DIM*jnrC;
1665             j_coord_offsetD  = DIM*jnrD;
1666
1667             /* load j atom coordinates */
1668             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1669                                               x+j_coord_offsetC,x+j_coord_offsetD,
1670                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1671                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1672
1673             /* Calculate displacement vector */
1674             dx00             = _mm_sub_ps(ix0,jx0);
1675             dy00             = _mm_sub_ps(iy0,jy0);
1676             dz00             = _mm_sub_ps(iz0,jz0);
1677             dx11             = _mm_sub_ps(ix1,jx1);
1678             dy11             = _mm_sub_ps(iy1,jy1);
1679             dz11             = _mm_sub_ps(iz1,jz1);
1680             dx12             = _mm_sub_ps(ix1,jx2);
1681             dy12             = _mm_sub_ps(iy1,jy2);
1682             dz12             = _mm_sub_ps(iz1,jz2);
1683             dx13             = _mm_sub_ps(ix1,jx3);
1684             dy13             = _mm_sub_ps(iy1,jy3);
1685             dz13             = _mm_sub_ps(iz1,jz3);
1686             dx21             = _mm_sub_ps(ix2,jx1);
1687             dy21             = _mm_sub_ps(iy2,jy1);
1688             dz21             = _mm_sub_ps(iz2,jz1);
1689             dx22             = _mm_sub_ps(ix2,jx2);
1690             dy22             = _mm_sub_ps(iy2,jy2);
1691             dz22             = _mm_sub_ps(iz2,jz2);
1692             dx23             = _mm_sub_ps(ix2,jx3);
1693             dy23             = _mm_sub_ps(iy2,jy3);
1694             dz23             = _mm_sub_ps(iz2,jz3);
1695             dx31             = _mm_sub_ps(ix3,jx1);
1696             dy31             = _mm_sub_ps(iy3,jy1);
1697             dz31             = _mm_sub_ps(iz3,jz1);
1698             dx32             = _mm_sub_ps(ix3,jx2);
1699             dy32             = _mm_sub_ps(iy3,jy2);
1700             dz32             = _mm_sub_ps(iz3,jz2);
1701             dx33             = _mm_sub_ps(ix3,jx3);
1702             dy33             = _mm_sub_ps(iy3,jy3);
1703             dz33             = _mm_sub_ps(iz3,jz3);
1704
1705             /* Calculate squared distance and things based on it */
1706             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1707             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1708             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1709             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1710             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1711             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1712             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1713             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1714             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1715             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1716
1717             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1718             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1719             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1720             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1721             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1722             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1723             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1724             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1725             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1726             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1727
1728             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1729             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1730             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1731             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1732             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1733             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1734             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1735             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1736             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1737
1738             fjx0             = _mm_setzero_ps();
1739             fjy0             = _mm_setzero_ps();
1740             fjz0             = _mm_setzero_ps();
1741             fjx1             = _mm_setzero_ps();
1742             fjy1             = _mm_setzero_ps();
1743             fjz1             = _mm_setzero_ps();
1744             fjx2             = _mm_setzero_ps();
1745             fjy2             = _mm_setzero_ps();
1746             fjz2             = _mm_setzero_ps();
1747             fjx3             = _mm_setzero_ps();
1748             fjy3             = _mm_setzero_ps();
1749             fjz3             = _mm_setzero_ps();
1750
1751             /**************************
1752              * CALCULATE INTERACTIONS *
1753              **************************/
1754
1755             r00              = _mm_mul_ps(rsq00,rinv00);
1756             r00              = _mm_andnot_ps(dummy_mask,r00);
1757
1758             /* Calculate table index by multiplying r with table scale and truncate to integer */
1759             rt               = _mm_mul_ps(r00,vftabscale);
1760             vfitab           = _mm_cvttps_epi32(rt);
1761             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1762             vfitab           = _mm_slli_epi32(vfitab,3);
1763
1764             /* CUBIC SPLINE TABLE DISPERSION */
1765             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1766             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1767             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1768             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1769             _MM_TRANSPOSE4_PS(Y,F,G,H);
1770             Heps             = _mm_mul_ps(vfeps,H);
1771             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1772             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1773             fvdw6            = _mm_mul_ps(c6_00,FF);
1774
1775             /* CUBIC SPLINE TABLE REPULSION */
1776             vfitab           = _mm_add_epi32(vfitab,ifour);
1777             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1778             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1779             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1780             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1781             _MM_TRANSPOSE4_PS(Y,F,G,H);
1782             Heps             = _mm_mul_ps(vfeps,H);
1783             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1784             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1785             fvdw12           = _mm_mul_ps(c12_00,FF);
1786             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1787
1788             fscal            = fvdw;
1789
1790             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1791
1792             /* Calculate temporary vectorial force */
1793             tx               = _mm_mul_ps(fscal,dx00);
1794             ty               = _mm_mul_ps(fscal,dy00);
1795             tz               = _mm_mul_ps(fscal,dz00);
1796
1797             /* Update vectorial force */
1798             fix0             = _mm_add_ps(fix0,tx);
1799             fiy0             = _mm_add_ps(fiy0,ty);
1800             fiz0             = _mm_add_ps(fiz0,tz);
1801
1802             fjx0             = _mm_add_ps(fjx0,tx);
1803             fjy0             = _mm_add_ps(fjy0,ty);
1804             fjz0             = _mm_add_ps(fjz0,tz);
1805
1806             /**************************
1807              * CALCULATE INTERACTIONS *
1808              **************************/
1809
1810             /* REACTION-FIELD ELECTROSTATICS */
1811             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1812
1813             fscal            = felec;
1814
1815             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1816
1817             /* Calculate temporary vectorial force */
1818             tx               = _mm_mul_ps(fscal,dx11);
1819             ty               = _mm_mul_ps(fscal,dy11);
1820             tz               = _mm_mul_ps(fscal,dz11);
1821
1822             /* Update vectorial force */
1823             fix1             = _mm_add_ps(fix1,tx);
1824             fiy1             = _mm_add_ps(fiy1,ty);
1825             fiz1             = _mm_add_ps(fiz1,tz);
1826
1827             fjx1             = _mm_add_ps(fjx1,tx);
1828             fjy1             = _mm_add_ps(fjy1,ty);
1829             fjz1             = _mm_add_ps(fjz1,tz);
1830
1831             /**************************
1832              * CALCULATE INTERACTIONS *
1833              **************************/
1834
1835             /* REACTION-FIELD ELECTROSTATICS */
1836             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1837
1838             fscal            = felec;
1839
1840             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1841
1842             /* Calculate temporary vectorial force */
1843             tx               = _mm_mul_ps(fscal,dx12);
1844             ty               = _mm_mul_ps(fscal,dy12);
1845             tz               = _mm_mul_ps(fscal,dz12);
1846
1847             /* Update vectorial force */
1848             fix1             = _mm_add_ps(fix1,tx);
1849             fiy1             = _mm_add_ps(fiy1,ty);
1850             fiz1             = _mm_add_ps(fiz1,tz);
1851
1852             fjx2             = _mm_add_ps(fjx2,tx);
1853             fjy2             = _mm_add_ps(fjy2,ty);
1854             fjz2             = _mm_add_ps(fjz2,tz);
1855
1856             /**************************
1857              * CALCULATE INTERACTIONS *
1858              **************************/
1859
1860             /* REACTION-FIELD ELECTROSTATICS */
1861             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1862
1863             fscal            = felec;
1864
1865             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1866
1867             /* Calculate temporary vectorial force */
1868             tx               = _mm_mul_ps(fscal,dx13);
1869             ty               = _mm_mul_ps(fscal,dy13);
1870             tz               = _mm_mul_ps(fscal,dz13);
1871
1872             /* Update vectorial force */
1873             fix1             = _mm_add_ps(fix1,tx);
1874             fiy1             = _mm_add_ps(fiy1,ty);
1875             fiz1             = _mm_add_ps(fiz1,tz);
1876
1877             fjx3             = _mm_add_ps(fjx3,tx);
1878             fjy3             = _mm_add_ps(fjy3,ty);
1879             fjz3             = _mm_add_ps(fjz3,tz);
1880
1881             /**************************
1882              * CALCULATE INTERACTIONS *
1883              **************************/
1884
1885             /* REACTION-FIELD ELECTROSTATICS */
1886             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1887
1888             fscal            = felec;
1889
1890             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1891
1892             /* Calculate temporary vectorial force */
1893             tx               = _mm_mul_ps(fscal,dx21);
1894             ty               = _mm_mul_ps(fscal,dy21);
1895             tz               = _mm_mul_ps(fscal,dz21);
1896
1897             /* Update vectorial force */
1898             fix2             = _mm_add_ps(fix2,tx);
1899             fiy2             = _mm_add_ps(fiy2,ty);
1900             fiz2             = _mm_add_ps(fiz2,tz);
1901
1902             fjx1             = _mm_add_ps(fjx1,tx);
1903             fjy1             = _mm_add_ps(fjy1,ty);
1904             fjz1             = _mm_add_ps(fjz1,tz);
1905
1906             /**************************
1907              * CALCULATE INTERACTIONS *
1908              **************************/
1909
1910             /* REACTION-FIELD ELECTROSTATICS */
1911             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1912
1913             fscal            = felec;
1914
1915             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1916
1917             /* Calculate temporary vectorial force */
1918             tx               = _mm_mul_ps(fscal,dx22);
1919             ty               = _mm_mul_ps(fscal,dy22);
1920             tz               = _mm_mul_ps(fscal,dz22);
1921
1922             /* Update vectorial force */
1923             fix2             = _mm_add_ps(fix2,tx);
1924             fiy2             = _mm_add_ps(fiy2,ty);
1925             fiz2             = _mm_add_ps(fiz2,tz);
1926
1927             fjx2             = _mm_add_ps(fjx2,tx);
1928             fjy2             = _mm_add_ps(fjy2,ty);
1929             fjz2             = _mm_add_ps(fjz2,tz);
1930
1931             /**************************
1932              * CALCULATE INTERACTIONS *
1933              **************************/
1934
1935             /* REACTION-FIELD ELECTROSTATICS */
1936             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1937
1938             fscal            = felec;
1939
1940             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1941
1942             /* Calculate temporary vectorial force */
1943             tx               = _mm_mul_ps(fscal,dx23);
1944             ty               = _mm_mul_ps(fscal,dy23);
1945             tz               = _mm_mul_ps(fscal,dz23);
1946
1947             /* Update vectorial force */
1948             fix2             = _mm_add_ps(fix2,tx);
1949             fiy2             = _mm_add_ps(fiy2,ty);
1950             fiz2             = _mm_add_ps(fiz2,tz);
1951
1952             fjx3             = _mm_add_ps(fjx3,tx);
1953             fjy3             = _mm_add_ps(fjy3,ty);
1954             fjz3             = _mm_add_ps(fjz3,tz);
1955
1956             /**************************
1957              * CALCULATE INTERACTIONS *
1958              **************************/
1959
1960             /* REACTION-FIELD ELECTROSTATICS */
1961             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1962
1963             fscal            = felec;
1964
1965             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1966
1967             /* Calculate temporary vectorial force */
1968             tx               = _mm_mul_ps(fscal,dx31);
1969             ty               = _mm_mul_ps(fscal,dy31);
1970             tz               = _mm_mul_ps(fscal,dz31);
1971
1972             /* Update vectorial force */
1973             fix3             = _mm_add_ps(fix3,tx);
1974             fiy3             = _mm_add_ps(fiy3,ty);
1975             fiz3             = _mm_add_ps(fiz3,tz);
1976
1977             fjx1             = _mm_add_ps(fjx1,tx);
1978             fjy1             = _mm_add_ps(fjy1,ty);
1979             fjz1             = _mm_add_ps(fjz1,tz);
1980
1981             /**************************
1982              * CALCULATE INTERACTIONS *
1983              **************************/
1984
1985             /* REACTION-FIELD ELECTROSTATICS */
1986             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1987
1988             fscal            = felec;
1989
1990             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1991
1992             /* Calculate temporary vectorial force */
1993             tx               = _mm_mul_ps(fscal,dx32);
1994             ty               = _mm_mul_ps(fscal,dy32);
1995             tz               = _mm_mul_ps(fscal,dz32);
1996
1997             /* Update vectorial force */
1998             fix3             = _mm_add_ps(fix3,tx);
1999             fiy3             = _mm_add_ps(fiy3,ty);
2000             fiz3             = _mm_add_ps(fiz3,tz);
2001
2002             fjx2             = _mm_add_ps(fjx2,tx);
2003             fjy2             = _mm_add_ps(fjy2,ty);
2004             fjz2             = _mm_add_ps(fjz2,tz);
2005
2006             /**************************
2007              * CALCULATE INTERACTIONS *
2008              **************************/
2009
2010             /* REACTION-FIELD ELECTROSTATICS */
2011             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2012
2013             fscal            = felec;
2014
2015             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2016
2017             /* Calculate temporary vectorial force */
2018             tx               = _mm_mul_ps(fscal,dx33);
2019             ty               = _mm_mul_ps(fscal,dy33);
2020             tz               = _mm_mul_ps(fscal,dz33);
2021
2022             /* Update vectorial force */
2023             fix3             = _mm_add_ps(fix3,tx);
2024             fiy3             = _mm_add_ps(fiy3,ty);
2025             fiz3             = _mm_add_ps(fiz3,tz);
2026
2027             fjx3             = _mm_add_ps(fjx3,tx);
2028             fjy3             = _mm_add_ps(fjy3,ty);
2029             fjz3             = _mm_add_ps(fjz3,tz);
2030
2031             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2032             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2033             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2034             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2035
2036             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2037                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2038                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2039
2040             /* Inner loop uses 295 flops */
2041         }
2042
2043         /* End of innermost loop */
2044
2045         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2046                                               f+i_coord_offset,fshift+i_shift_offset);
2047
2048         /* Increment number of inner iterations */
2049         inneriter                  += j_index_end - j_index_start;
2050
2051         /* Outer loop uses 24 flops */
2052     }
2053
2054     /* Increment number of outer iterations */
2055     outeriter        += nri;
2056
2057     /* Update outer/inner flops */
2058
2059     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*295);
2060 }