src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRFCut_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_ElecRFCut_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_ElecRFCut_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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 172     rcutoff_scalar   = fr->rcoulomb;
 173     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 174     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 175
 176     /* Avoid stupid compiler warnings */
 177     jnrA = jnrB = jnrC = jnrD = 0;
 178     j_coord_offsetA = 0;
 179     j_coord_offsetB = 0;
 180     j_coord_offsetC = 0;
 181     j_coord_offsetD = 0;
 182
 183     outeriter        = 0;
 184     inneriter        = 0;
 185
 186     for(iidx=0;iidx<4*DIM;iidx++)
 187     {
 188         scratch[iidx] = 0.0;
 189     }
 190
 191     /* Start outer loop over neighborlists */
 192     for(iidx=0; iidx<nri; iidx++)
 193     {
 194         /* Load shift vector for this list */
 195         i_shift_offset   = DIM*shiftidx[iidx];
 196
 197         /* Load limits for loop over neighbors */
 198         j_index_start    = jindex[iidx];
 199         j_index_end      = jindex[iidx+1];
 200
 201         /* Get outer coordinate index */
 202         inr              = iinr[iidx];
 203         i_coord_offset   = DIM*inr;
 204
 205         /* Load i particle coords and add shift vector */
 206         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 207                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 208
 209         fix0             = _mm_setzero_ps();
 210         fiy0             = _mm_setzero_ps();
 211         fiz0             = _mm_setzero_ps();
 212         fix1             = _mm_setzero_ps();
 213         fiy1             = _mm_setzero_ps();
 214         fiz1             = _mm_setzero_ps();
 215         fix2             = _mm_setzero_ps();
 216         fiy2             = _mm_setzero_ps();
 217         fiz2             = _mm_setzero_ps();
 218         fix3             = _mm_setzero_ps();
 219         fiy3             = _mm_setzero_ps();
 220         fiz3             = _mm_setzero_ps();
 221
 222         /* Reset potential sums */
 223         velecsum         = _mm_setzero_ps();
 224         vvdwsum          = _mm_setzero_ps();
 225
 226         /* Start inner kernel loop */
 227         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 228         {
 229
 230             /* Get j neighbor index, and coordinate index */
 231             jnrA             = jjnr[jidx];
 232             jnrB             = jjnr[jidx+1];
 233             jnrC             = jjnr[jidx+2];
 234             jnrD             = jjnr[jidx+3];
 235             j_coord_offsetA  = DIM*jnrA;
 236             j_coord_offsetB  = DIM*jnrB;
 237             j_coord_offsetC  = DIM*jnrC;
 238             j_coord_offsetD  = DIM*jnrD;
 239
 240             /* load j atom coordinates */
 241             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 242                                               x+j_coord_offsetC,x+j_coord_offsetD,
 243                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 244                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 245
 246             /* Calculate displacement vector */
 247             dx00             = _mm_sub_ps(ix0,jx0);
 248             dy00             = _mm_sub_ps(iy0,jy0);
 249             dz00             = _mm_sub_ps(iz0,jz0);
 250             dx11             = _mm_sub_ps(ix1,jx1);
 251             dy11             = _mm_sub_ps(iy1,jy1);
 252             dz11             = _mm_sub_ps(iz1,jz1);
 253             dx12             = _mm_sub_ps(ix1,jx2);
 254             dy12             = _mm_sub_ps(iy1,jy2);
 255             dz12             = _mm_sub_ps(iz1,jz2);
 256             dx13             = _mm_sub_ps(ix1,jx3);
 257             dy13             = _mm_sub_ps(iy1,jy3);
 258             dz13             = _mm_sub_ps(iz1,jz3);
 259             dx21             = _mm_sub_ps(ix2,jx1);
 260             dy21             = _mm_sub_ps(iy2,jy1);
 261             dz21             = _mm_sub_ps(iz2,jz1);
 262             dx22             = _mm_sub_ps(ix2,jx2);
 263             dy22             = _mm_sub_ps(iy2,jy2);
 264             dz22             = _mm_sub_ps(iz2,jz2);
 265             dx23             = _mm_sub_ps(ix2,jx3);
 266             dy23             = _mm_sub_ps(iy2,jy3);
 267             dz23             = _mm_sub_ps(iz2,jz3);
 268             dx31             = _mm_sub_ps(ix3,jx1);
 269             dy31             = _mm_sub_ps(iy3,jy1);
 270             dz31             = _mm_sub_ps(iz3,jz1);
 271             dx32             = _mm_sub_ps(ix3,jx2);
 272             dy32             = _mm_sub_ps(iy3,jy2);
 273             dz32             = _mm_sub_ps(iz3,jz2);
 274             dx33             = _mm_sub_ps(ix3,jx3);
 275             dy33             = _mm_sub_ps(iy3,jy3);
 276             dz33             = _mm_sub_ps(iz3,jz3);
 277
 278             /* Calculate squared distance and things based on it */
 279             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 280             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 281             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 282             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 283             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 284             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 285             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 286             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 287             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 288             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 289
 290             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 291             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 292             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 293             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 294             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 295             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 296             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 297             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 298             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 299             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 300
 301             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 302             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 303             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 304             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 305             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 306             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 307             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 308             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 309             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 310
 311             fjx0             = _mm_setzero_ps();
 312             fjy0             = _mm_setzero_ps();
 313             fjz0             = _mm_setzero_ps();
 314             fjx1             = _mm_setzero_ps();
 315             fjy1             = _mm_setzero_ps();
 316             fjz1             = _mm_setzero_ps();
 317             fjx2             = _mm_setzero_ps();
 318             fjy2             = _mm_setzero_ps();
 319             fjz2             = _mm_setzero_ps();
 320             fjx3             = _mm_setzero_ps();
 321             fjy3             = _mm_setzero_ps();
 322             fjz3             = _mm_setzero_ps();
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             if (gmx_mm_any_lt(rsq00,rcutoff2))
 329             {
 330
 331             r00              = _mm_mul_ps(rsq00,rinv00);
 332
 333             /* Calculate table index by multiplying r with table scale and truncate to integer */
 334             rt               = _mm_mul_ps(r00,vftabscale);
 335             vfitab           = _mm_cvttps_epi32(rt);
 336             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 337             vfitab           = _mm_slli_epi32(vfitab,3);
 338
 339             /* CUBIC SPLINE TABLE DISPERSION */
 340             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 341             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 342             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 343             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 344             _MM_TRANSPOSE4_PS(Y,F,G,H);
 345             Heps             = _mm_mul_ps(vfeps,H);
 346             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 347             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 348             vvdw6            = _mm_mul_ps(c6_00,VV);
 349             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 350             fvdw6            = _mm_mul_ps(c6_00,FF);
 351
 352             /* CUBIC SPLINE TABLE REPULSION */
 353             vfitab           = _mm_add_epi32(vfitab,ifour);
 354             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 355             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 356             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 357             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 358             _MM_TRANSPOSE4_PS(Y,F,G,H);
 359             Heps             = _mm_mul_ps(vfeps,H);
 360             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 361             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 362             vvdw12           = _mm_mul_ps(c12_00,VV);
 363             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 364             fvdw12           = _mm_mul_ps(c12_00,FF);
 365             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 366             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 367
 368             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 372             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 373
 374             fscal            = fvdw;
 375
 376             fscal            = _mm_and_ps(fscal,cutoff_mask);
 377
 378             /* Calculate temporary vectorial force */
 379             tx               = _mm_mul_ps(fscal,dx00);
 380             ty               = _mm_mul_ps(fscal,dy00);
 381             tz               = _mm_mul_ps(fscal,dz00);
 382
 383             /* Update vectorial force */
 384             fix0             = _mm_add_ps(fix0,tx);
 385             fiy0             = _mm_add_ps(fiy0,ty);
 386             fiz0             = _mm_add_ps(fiz0,tz);
 387
 388             fjx0             = _mm_add_ps(fjx0,tx);
 389             fjy0             = _mm_add_ps(fjy0,ty);
 390             fjz0             = _mm_add_ps(fjz0,tz);
 391
 392             }
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             if (gmx_mm_any_lt(rsq11,rcutoff2))
 399             {
 400
 401             /* REACTION-FIELD ELECTROSTATICS */
 402             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 403             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 404
 405             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velec            = _mm_and_ps(velec,cutoff_mask);
 409             velecsum         = _mm_add_ps(velecsum,velec);
 410
 411             fscal            = felec;
 412
 413             fscal            = _mm_and_ps(fscal,cutoff_mask);
 414
 415             /* Calculate temporary vectorial force */
 416             tx               = _mm_mul_ps(fscal,dx11);
 417             ty               = _mm_mul_ps(fscal,dy11);
 418             tz               = _mm_mul_ps(fscal,dz11);
 419
 420             /* Update vectorial force */
 421             fix1             = _mm_add_ps(fix1,tx);
 422             fiy1             = _mm_add_ps(fiy1,ty);
 423             fiz1             = _mm_add_ps(fiz1,tz);
 424
 425             fjx1             = _mm_add_ps(fjx1,tx);
 426             fjy1             = _mm_add_ps(fjy1,ty);
 427             fjz1             = _mm_add_ps(fjz1,tz);
 428
 429             }
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             if (gmx_mm_any_lt(rsq12,rcutoff2))
 436             {
 437
 438             /* REACTION-FIELD ELECTROSTATICS */
 439             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 440             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 441
 442             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 443
 444             /* Update potential sum for this i atom from the interaction with this j atom. */
 445             velec            = _mm_and_ps(velec,cutoff_mask);
 446             velecsum         = _mm_add_ps(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             fscal            = _mm_and_ps(fscal,cutoff_mask);
 451
 452             /* Calculate temporary vectorial force */
 453             tx               = _mm_mul_ps(fscal,dx12);
 454             ty               = _mm_mul_ps(fscal,dy12);
 455             tz               = _mm_mul_ps(fscal,dz12);
 456
 457             /* Update vectorial force */
 458             fix1             = _mm_add_ps(fix1,tx);
 459             fiy1             = _mm_add_ps(fiy1,ty);
 460             fiz1             = _mm_add_ps(fiz1,tz);
 461
 462             fjx2             = _mm_add_ps(fjx2,tx);
 463             fjy2             = _mm_add_ps(fjy2,ty);
 464             fjz2             = _mm_add_ps(fjz2,tz);
 465
 466             }
 467
 468             /**************************
 469              * CALCULATE INTERACTIONS *
 470              **************************/
 471
 472             if (gmx_mm_any_lt(rsq13,rcutoff2))
 473             {
 474
 475             /* REACTION-FIELD ELECTROSTATICS */
 476             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 477             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 478
 479             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 480
 481             /* Update potential sum for this i atom from the interaction with this j atom. */
 482             velec            = _mm_and_ps(velec,cutoff_mask);
 483             velecsum         = _mm_add_ps(velecsum,velec);
 484
 485             fscal            = felec;
 486
 487             fscal            = _mm_and_ps(fscal,cutoff_mask);
 488
 489             /* Calculate temporary vectorial force */
 490             tx               = _mm_mul_ps(fscal,dx13);
 491             ty               = _mm_mul_ps(fscal,dy13);
 492             tz               = _mm_mul_ps(fscal,dz13);
 493
 494             /* Update vectorial force */
 495             fix1             = _mm_add_ps(fix1,tx);
 496             fiy1             = _mm_add_ps(fiy1,ty);
 497             fiz1             = _mm_add_ps(fiz1,tz);
 498
 499             fjx3             = _mm_add_ps(fjx3,tx);
 500             fjy3             = _mm_add_ps(fjy3,ty);
 501             fjz3             = _mm_add_ps(fjz3,tz);
 502
 503             }
 504
 505             /**************************
 506              * CALCULATE INTERACTIONS *
 507              **************************/
 508
 509             if (gmx_mm_any_lt(rsq21,rcutoff2))
 510             {
 511
 512             /* REACTION-FIELD ELECTROSTATICS */
 513             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 514             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 515
 516             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 517
 518             /* Update potential sum for this i atom from the interaction with this j atom. */
 519             velec            = _mm_and_ps(velec,cutoff_mask);
 520             velecsum         = _mm_add_ps(velecsum,velec);
 521
 522             fscal            = felec;
 523
 524             fscal            = _mm_and_ps(fscal,cutoff_mask);
 525
 526             /* Calculate temporary vectorial force */
 527             tx               = _mm_mul_ps(fscal,dx21);
 528             ty               = _mm_mul_ps(fscal,dy21);
 529             tz               = _mm_mul_ps(fscal,dz21);
 530
 531             /* Update vectorial force */
 532             fix2             = _mm_add_ps(fix2,tx);
 533             fiy2             = _mm_add_ps(fiy2,ty);
 534             fiz2             = _mm_add_ps(fiz2,tz);
 535
 536             fjx1             = _mm_add_ps(fjx1,tx);
 537             fjy1             = _mm_add_ps(fjy1,ty);
 538             fjz1             = _mm_add_ps(fjz1,tz);
 539
 540             }
 541
 542             /**************************
 543              * CALCULATE INTERACTIONS *
 544              **************************/
 545
 546             if (gmx_mm_any_lt(rsq22,rcutoff2))
 547             {
 548
 549             /* REACTION-FIELD ELECTROSTATICS */
 550             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 551             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 552
 553             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 554
 555             /* Update potential sum for this i atom from the interaction with this j atom. */
 556             velec            = _mm_and_ps(velec,cutoff_mask);
 557             velecsum         = _mm_add_ps(velecsum,velec);
 558
 559             fscal            = felec;
 560
 561             fscal            = _mm_and_ps(fscal,cutoff_mask);
 562
 563             /* Calculate temporary vectorial force */
 564             tx               = _mm_mul_ps(fscal,dx22);
 565             ty               = _mm_mul_ps(fscal,dy22);
 566             tz               = _mm_mul_ps(fscal,dz22);
 567
 568             /* Update vectorial force */
 569             fix2             = _mm_add_ps(fix2,tx);
 570             fiy2             = _mm_add_ps(fiy2,ty);
 571             fiz2             = _mm_add_ps(fiz2,tz);
 572
 573             fjx2             = _mm_add_ps(fjx2,tx);
 574             fjy2             = _mm_add_ps(fjy2,ty);
 575             fjz2             = _mm_add_ps(fjz2,tz);
 576
 577             }
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             if (gmx_mm_any_lt(rsq23,rcutoff2))
 584             {
 585
 586             /* REACTION-FIELD ELECTROSTATICS */
 587             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 588             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 589
 590             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 591
 592             /* Update potential sum for this i atom from the interaction with this j atom. */
 593             velec            = _mm_and_ps(velec,cutoff_mask);
 594             velecsum         = _mm_add_ps(velecsum,velec);
 595
 596             fscal            = felec;
 597
 598             fscal            = _mm_and_ps(fscal,cutoff_mask);
 599
 600             /* Calculate temporary vectorial force */
 601             tx               = _mm_mul_ps(fscal,dx23);
 602             ty               = _mm_mul_ps(fscal,dy23);
 603             tz               = _mm_mul_ps(fscal,dz23);
 604
 605             /* Update vectorial force */
 606             fix2             = _mm_add_ps(fix2,tx);
 607             fiy2             = _mm_add_ps(fiy2,ty);
 608             fiz2             = _mm_add_ps(fiz2,tz);
 609
 610             fjx3             = _mm_add_ps(fjx3,tx);
 611             fjy3             = _mm_add_ps(fjy3,ty);
 612             fjz3             = _mm_add_ps(fjz3,tz);
 613
 614             }
 615
 616             /**************************
 617              * CALCULATE INTERACTIONS *
 618              **************************/
 619
 620             if (gmx_mm_any_lt(rsq31,rcutoff2))
 621             {
 622
 623             /* REACTION-FIELD ELECTROSTATICS */
 624             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 625             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 626
 627             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 628
 629             /* Update potential sum for this i atom from the interaction with this j atom. */
 630             velec            = _mm_and_ps(velec,cutoff_mask);
 631             velecsum         = _mm_add_ps(velecsum,velec);
 632
 633             fscal            = felec;
 634
 635             fscal            = _mm_and_ps(fscal,cutoff_mask);
 636
 637             /* Calculate temporary vectorial force */
 638             tx               = _mm_mul_ps(fscal,dx31);
 639             ty               = _mm_mul_ps(fscal,dy31);
 640             tz               = _mm_mul_ps(fscal,dz31);
 641
 642             /* Update vectorial force */
 643             fix3             = _mm_add_ps(fix3,tx);
 644             fiy3             = _mm_add_ps(fiy3,ty);
 645             fiz3             = _mm_add_ps(fiz3,tz);
 646
 647             fjx1             = _mm_add_ps(fjx1,tx);
 648             fjy1             = _mm_add_ps(fjy1,ty);
 649             fjz1             = _mm_add_ps(fjz1,tz);
 650
 651             }
 652
 653             /**************************
 654              * CALCULATE INTERACTIONS *
 655              **************************/
 656
 657             if (gmx_mm_any_lt(rsq32,rcutoff2))
 658             {
 659
 660             /* REACTION-FIELD ELECTROSTATICS */
 661             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 662             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 663
 664             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 665
 666             /* Update potential sum for this i atom from the interaction with this j atom. */
 667             velec            = _mm_and_ps(velec,cutoff_mask);
 668             velecsum         = _mm_add_ps(velecsum,velec);
 669
 670             fscal            = felec;
 671
 672             fscal            = _mm_and_ps(fscal,cutoff_mask);
 673
 674             /* Calculate temporary vectorial force */
 675             tx               = _mm_mul_ps(fscal,dx32);
 676             ty               = _mm_mul_ps(fscal,dy32);
 677             tz               = _mm_mul_ps(fscal,dz32);
 678
 679             /* Update vectorial force */
 680             fix3             = _mm_add_ps(fix3,tx);
 681             fiy3             = _mm_add_ps(fiy3,ty);
 682             fiz3             = _mm_add_ps(fiz3,tz);
 683
 684             fjx2             = _mm_add_ps(fjx2,tx);
 685             fjy2             = _mm_add_ps(fjy2,ty);
 686             fjz2             = _mm_add_ps(fjz2,tz);
 687
 688             }
 689
 690             /**************************
 691              * CALCULATE INTERACTIONS *
 692              **************************/
 693
 694             if (gmx_mm_any_lt(rsq33,rcutoff2))
 695             {
 696
 697             /* REACTION-FIELD ELECTROSTATICS */
 698             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 699             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 700
 701             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 702
 703             /* Update potential sum for this i atom from the interaction with this j atom. */
 704             velec            = _mm_and_ps(velec,cutoff_mask);
 705             velecsum         = _mm_add_ps(velecsum,velec);
 706
 707             fscal            = felec;
 708
 709             fscal            = _mm_and_ps(fscal,cutoff_mask);
 710
 711             /* Calculate temporary vectorial force */
 712             tx               = _mm_mul_ps(fscal,dx33);
 713             ty               = _mm_mul_ps(fscal,dy33);
 714             tz               = _mm_mul_ps(fscal,dz33);
 715
 716             /* Update vectorial force */
 717             fix3             = _mm_add_ps(fix3,tx);
 718             fiy3             = _mm_add_ps(fiy3,ty);
 719             fiz3             = _mm_add_ps(fiz3,tz);
 720
 721             fjx3             = _mm_add_ps(fjx3,tx);
 722             fjy3             = _mm_add_ps(fjy3,ty);
 723             fjz3             = _mm_add_ps(fjz3,tz);
 724
 725             }
 726
 727             fjptrA             = f+j_coord_offsetA;
 728             fjptrB             = f+j_coord_offsetB;
 729             fjptrC             = f+j_coord_offsetC;
 730             fjptrD             = f+j_coord_offsetD;
 731
 732             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 733                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 734                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 735
 736             /* Inner loop uses 387 flops */
 737         }
 738
 739         if(jidx<j_index_end)
 740         {
 741
 742             /* Get j neighbor index, and coordinate index */
 743             jnrlistA         = jjnr[jidx];
 744             jnrlistB         = jjnr[jidx+1];
 745             jnrlistC         = jjnr[jidx+2];
 746             jnrlistD         = jjnr[jidx+3];
 747             /* Sign of each element will be negative for non-real atoms.
 748              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 749              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 750              */
 751             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 752             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 753             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 754             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 755             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 756             j_coord_offsetA  = DIM*jnrA;
 757             j_coord_offsetB  = DIM*jnrB;
 758             j_coord_offsetC  = DIM*jnrC;
 759             j_coord_offsetD  = DIM*jnrD;
 760
 761             /* load j atom coordinates */
 762             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 763                                               x+j_coord_offsetC,x+j_coord_offsetD,
 764                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 765                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 766
 767             /* Calculate displacement vector */
 768             dx00             = _mm_sub_ps(ix0,jx0);
 769             dy00             = _mm_sub_ps(iy0,jy0);
 770             dz00             = _mm_sub_ps(iz0,jz0);
 771             dx11             = _mm_sub_ps(ix1,jx1);
 772             dy11             = _mm_sub_ps(iy1,jy1);
 773             dz11             = _mm_sub_ps(iz1,jz1);
 774             dx12             = _mm_sub_ps(ix1,jx2);
 775             dy12             = _mm_sub_ps(iy1,jy2);
 776             dz12             = _mm_sub_ps(iz1,jz2);
 777             dx13             = _mm_sub_ps(ix1,jx3);
 778             dy13             = _mm_sub_ps(iy1,jy3);
 779             dz13             = _mm_sub_ps(iz1,jz3);
 780             dx21             = _mm_sub_ps(ix2,jx1);
 781             dy21             = _mm_sub_ps(iy2,jy1);
 782             dz21             = _mm_sub_ps(iz2,jz1);
 783             dx22             = _mm_sub_ps(ix2,jx2);
 784             dy22             = _mm_sub_ps(iy2,jy2);
 785             dz22             = _mm_sub_ps(iz2,jz2);
 786             dx23             = _mm_sub_ps(ix2,jx3);
 787             dy23             = _mm_sub_ps(iy2,jy3);
 788             dz23             = _mm_sub_ps(iz2,jz3);
 789             dx31             = _mm_sub_ps(ix3,jx1);
 790             dy31             = _mm_sub_ps(iy3,jy1);
 791             dz31             = _mm_sub_ps(iz3,jz1);
 792             dx32             = _mm_sub_ps(ix3,jx2);
 793             dy32             = _mm_sub_ps(iy3,jy2);
 794             dz32             = _mm_sub_ps(iz3,jz2);
 795             dx33             = _mm_sub_ps(ix3,jx3);
 796             dy33             = _mm_sub_ps(iy3,jy3);
 797             dz33             = _mm_sub_ps(iz3,jz3);
 798
 799             /* Calculate squared distance and things based on it */
 800             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 801             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 802             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 803             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 804             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 805             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 806             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 807             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 808             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 809             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 810
 811             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 812             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 813             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 814             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 815             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 816             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 817             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 818             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 819             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 820             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 821
 822             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 823             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 824             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 825             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 826             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 827             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 828             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 829             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 830             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 831
 832             fjx0             = _mm_setzero_ps();
 833             fjy0             = _mm_setzero_ps();
 834             fjz0             = _mm_setzero_ps();
 835             fjx1             = _mm_setzero_ps();
 836             fjy1             = _mm_setzero_ps();
 837             fjz1             = _mm_setzero_ps();
 838             fjx2             = _mm_setzero_ps();
 839             fjy2             = _mm_setzero_ps();
 840             fjz2             = _mm_setzero_ps();
 841             fjx3             = _mm_setzero_ps();
 842             fjy3             = _mm_setzero_ps();
 843             fjz3             = _mm_setzero_ps();
 844
 845             /**************************
 846              * CALCULATE INTERACTIONS *
 847              **************************/
 848
 849             if (gmx_mm_any_lt(rsq00,rcutoff2))
 850             {
 851
 852             r00              = _mm_mul_ps(rsq00,rinv00);
 853             r00              = _mm_andnot_ps(dummy_mask,r00);
 854
 855             /* Calculate table index by multiplying r with table scale and truncate to integer */
 856             rt               = _mm_mul_ps(r00,vftabscale);
 857             vfitab           = _mm_cvttps_epi32(rt);
 858             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 859             vfitab           = _mm_slli_epi32(vfitab,3);
 860
 861             /* CUBIC SPLINE TABLE DISPERSION */
 862             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 863             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 864             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 865             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 866             _MM_TRANSPOSE4_PS(Y,F,G,H);
 867             Heps             = _mm_mul_ps(vfeps,H);
 868             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 869             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 870             vvdw6            = _mm_mul_ps(c6_00,VV);
 871             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 872             fvdw6            = _mm_mul_ps(c6_00,FF);
 873
 874             /* CUBIC SPLINE TABLE REPULSION */
 875             vfitab           = _mm_add_epi32(vfitab,ifour);
 876             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 877             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 878             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 879             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 880             _MM_TRANSPOSE4_PS(Y,F,G,H);
 881             Heps             = _mm_mul_ps(vfeps,H);
 882             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 883             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 884             vvdw12           = _mm_mul_ps(c12_00,VV);
 885             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 886             fvdw12           = _mm_mul_ps(c12_00,FF);
 887             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 888             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 889
 890             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 891
 892             /* Update potential sum for this i atom from the interaction with this j atom. */
 893             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 894             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 895             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 896
 897             fscal            = fvdw;
 898
 899             fscal            = _mm_and_ps(fscal,cutoff_mask);
 900
 901             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 902
 903             /* Calculate temporary vectorial force */
 904             tx               = _mm_mul_ps(fscal,dx00);
 905             ty               = _mm_mul_ps(fscal,dy00);
 906             tz               = _mm_mul_ps(fscal,dz00);
 907
 908             /* Update vectorial force */
 909             fix0             = _mm_add_ps(fix0,tx);
 910             fiy0             = _mm_add_ps(fiy0,ty);
 911             fiz0             = _mm_add_ps(fiz0,tz);
 912
 913             fjx0             = _mm_add_ps(fjx0,tx);
 914             fjy0             = _mm_add_ps(fjy0,ty);
 915             fjz0             = _mm_add_ps(fjz0,tz);
 916
 917             }
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             if (gmx_mm_any_lt(rsq11,rcutoff2))
 924             {
 925
 926             /* REACTION-FIELD ELECTROSTATICS */
 927             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 928             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 929
 930             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 931
 932             /* Update potential sum for this i atom from the interaction with this j atom. */
 933             velec            = _mm_and_ps(velec,cutoff_mask);
 934             velec            = _mm_andnot_ps(dummy_mask,velec);
 935             velecsum         = _mm_add_ps(velecsum,velec);
 936
 937             fscal            = felec;
 938
 939             fscal            = _mm_and_ps(fscal,cutoff_mask);
 940
 941             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 942
 943             /* Calculate temporary vectorial force */
 944             tx               = _mm_mul_ps(fscal,dx11);
 945             ty               = _mm_mul_ps(fscal,dy11);
 946             tz               = _mm_mul_ps(fscal,dz11);
 947
 948             /* Update vectorial force */
 949             fix1             = _mm_add_ps(fix1,tx);
 950             fiy1             = _mm_add_ps(fiy1,ty);
 951             fiz1             = _mm_add_ps(fiz1,tz);
 952
 953             fjx1             = _mm_add_ps(fjx1,tx);
 954             fjy1             = _mm_add_ps(fjy1,ty);
 955             fjz1             = _mm_add_ps(fjz1,tz);
 956
 957             }
 958
 959             /**************************
 960              * CALCULATE INTERACTIONS *
 961              **************************/
 962
 963             if (gmx_mm_any_lt(rsq12,rcutoff2))
 964             {
 965
 966             /* REACTION-FIELD ELECTROSTATICS */
 967             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 968             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 969
 970             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 971
 972             /* Update potential sum for this i atom from the interaction with this j atom. */
 973             velec            = _mm_and_ps(velec,cutoff_mask);
 974             velec            = _mm_andnot_ps(dummy_mask,velec);
 975             velecsum         = _mm_add_ps(velecsum,velec);
 976
 977             fscal            = felec;
 978
 979             fscal            = _mm_and_ps(fscal,cutoff_mask);
 980
 981             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 982
 983             /* Calculate temporary vectorial force */
 984             tx               = _mm_mul_ps(fscal,dx12);
 985             ty               = _mm_mul_ps(fscal,dy12);
 986             tz               = _mm_mul_ps(fscal,dz12);
 987
 988             /* Update vectorial force */
 989             fix1             = _mm_add_ps(fix1,tx);
 990             fiy1             = _mm_add_ps(fiy1,ty);
 991             fiz1             = _mm_add_ps(fiz1,tz);
 992
 993             fjx2             = _mm_add_ps(fjx2,tx);
 994             fjy2             = _mm_add_ps(fjy2,ty);
 995             fjz2             = _mm_add_ps(fjz2,tz);
 996
 997             }
 998
 999             /**************************
1000              * CALCULATE INTERACTIONS *
1001              **************************/
1002
1003             if (gmx_mm_any_lt(rsq13,rcutoff2))
1004             {
1005
1006             /* REACTION-FIELD ELECTROSTATICS */
1007             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
1008             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1009
1010             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1011
1012             /* Update potential sum for this i atom from the interaction with this j atom. */
1013             velec            = _mm_and_ps(velec,cutoff_mask);
1014             velec            = _mm_andnot_ps(dummy_mask,velec);
1015             velecsum         = _mm_add_ps(velecsum,velec);
1016
1017             fscal            = felec;
1018
1019             fscal            = _mm_and_ps(fscal,cutoff_mask);
1020
1021             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1022
1023             /* Calculate temporary vectorial force */
1024             tx               = _mm_mul_ps(fscal,dx13);
1025             ty               = _mm_mul_ps(fscal,dy13);
1026             tz               = _mm_mul_ps(fscal,dz13);
1027
1028             /* Update vectorial force */
1029             fix1             = _mm_add_ps(fix1,tx);
1030             fiy1             = _mm_add_ps(fiy1,ty);
1031             fiz1             = _mm_add_ps(fiz1,tz);
1032
1033             fjx3             = _mm_add_ps(fjx3,tx);
1034             fjy3             = _mm_add_ps(fjy3,ty);
1035             fjz3             = _mm_add_ps(fjz3,tz);
1036
1037             }
1038
1039             /**************************
1040              * CALCULATE INTERACTIONS *
1041              **************************/
1042
1043             if (gmx_mm_any_lt(rsq21,rcutoff2))
1044             {
1045
1046             /* REACTION-FIELD ELECTROSTATICS */
1047             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
1048             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1049
1050             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1051
1052             /* Update potential sum for this i atom from the interaction with this j atom. */
1053             velec            = _mm_and_ps(velec,cutoff_mask);
1054             velec            = _mm_andnot_ps(dummy_mask,velec);
1055             velecsum         = _mm_add_ps(velecsum,velec);
1056
1057             fscal            = felec;
1058
1059             fscal            = _mm_and_ps(fscal,cutoff_mask);
1060
1061             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1062
1063             /* Calculate temporary vectorial force */
1064             tx               = _mm_mul_ps(fscal,dx21);
1065             ty               = _mm_mul_ps(fscal,dy21);
1066             tz               = _mm_mul_ps(fscal,dz21);
1067
1068             /* Update vectorial force */
1069             fix2             = _mm_add_ps(fix2,tx);
1070             fiy2             = _mm_add_ps(fiy2,ty);
1071             fiz2             = _mm_add_ps(fiz2,tz);
1072
1073             fjx1             = _mm_add_ps(fjx1,tx);
1074             fjy1             = _mm_add_ps(fjy1,ty);
1075             fjz1             = _mm_add_ps(fjz1,tz);
1076
1077             }
1078
1079             /**************************
1080              * CALCULATE INTERACTIONS *
1081              **************************/
1082
1083             if (gmx_mm_any_lt(rsq22,rcutoff2))
1084             {
1085
1086             /* REACTION-FIELD ELECTROSTATICS */
1087             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1088             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1089
1090             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1091
1092             /* Update potential sum for this i atom from the interaction with this j atom. */
1093             velec            = _mm_and_ps(velec,cutoff_mask);
1094             velec            = _mm_andnot_ps(dummy_mask,velec);
1095             velecsum         = _mm_add_ps(velecsum,velec);
1096
1097             fscal            = felec;
1098
1099             fscal            = _mm_and_ps(fscal,cutoff_mask);
1100
1101             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1102
1103             /* Calculate temporary vectorial force */
1104             tx               = _mm_mul_ps(fscal,dx22);
1105             ty               = _mm_mul_ps(fscal,dy22);
1106             tz               = _mm_mul_ps(fscal,dz22);
1107
1108             /* Update vectorial force */
1109             fix2             = _mm_add_ps(fix2,tx);
1110             fiy2             = _mm_add_ps(fiy2,ty);
1111             fiz2             = _mm_add_ps(fiz2,tz);
1112
1113             fjx2             = _mm_add_ps(fjx2,tx);
1114             fjy2             = _mm_add_ps(fjy2,ty);
1115             fjz2             = _mm_add_ps(fjz2,tz);
1116
1117             }
1118
1119             /**************************
1120              * CALCULATE INTERACTIONS *
1121              **************************/
1122
1123             if (gmx_mm_any_lt(rsq23,rcutoff2))
1124             {
1125
1126             /* REACTION-FIELD ELECTROSTATICS */
1127             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
1128             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1129
1130             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1131
1132             /* Update potential sum for this i atom from the interaction with this j atom. */
1133             velec            = _mm_and_ps(velec,cutoff_mask);
1134             velec            = _mm_andnot_ps(dummy_mask,velec);
1135             velecsum         = _mm_add_ps(velecsum,velec);
1136
1137             fscal            = felec;
1138
1139             fscal            = _mm_and_ps(fscal,cutoff_mask);
1140
1141             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1142
1143             /* Calculate temporary vectorial force */
1144             tx               = _mm_mul_ps(fscal,dx23);
1145             ty               = _mm_mul_ps(fscal,dy23);
1146             tz               = _mm_mul_ps(fscal,dz23);
1147
1148             /* Update vectorial force */
1149             fix2             = _mm_add_ps(fix2,tx);
1150             fiy2             = _mm_add_ps(fiy2,ty);
1151             fiz2             = _mm_add_ps(fiz2,tz);
1152
1153             fjx3             = _mm_add_ps(fjx3,tx);
1154             fjy3             = _mm_add_ps(fjy3,ty);
1155             fjz3             = _mm_add_ps(fjz3,tz);
1156
1157             }
1158
1159             /**************************
1160              * CALCULATE INTERACTIONS *
1161              **************************/
1162
1163             if (gmx_mm_any_lt(rsq31,rcutoff2))
1164             {
1165
1166             /* REACTION-FIELD ELECTROSTATICS */
1167             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
1168             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1169
1170             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1171
1172             /* Update potential sum for this i atom from the interaction with this j atom. */
1173             velec            = _mm_and_ps(velec,cutoff_mask);
1174             velec            = _mm_andnot_ps(dummy_mask,velec);
1175             velecsum         = _mm_add_ps(velecsum,velec);
1176
1177             fscal            = felec;
1178
1179             fscal            = _mm_and_ps(fscal,cutoff_mask);
1180
1181             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1182
1183             /* Calculate temporary vectorial force */
1184             tx               = _mm_mul_ps(fscal,dx31);
1185             ty               = _mm_mul_ps(fscal,dy31);
1186             tz               = _mm_mul_ps(fscal,dz31);
1187
1188             /* Update vectorial force */
1189             fix3             = _mm_add_ps(fix3,tx);
1190             fiy3             = _mm_add_ps(fiy3,ty);
1191             fiz3             = _mm_add_ps(fiz3,tz);
1192
1193             fjx1             = _mm_add_ps(fjx1,tx);
1194             fjy1             = _mm_add_ps(fjy1,ty);
1195             fjz1             = _mm_add_ps(fjz1,tz);
1196
1197             }
1198
1199             /**************************
1200              * CALCULATE INTERACTIONS *
1201              **************************/
1202
1203             if (gmx_mm_any_lt(rsq32,rcutoff2))
1204             {
1205
1206             /* REACTION-FIELD ELECTROSTATICS */
1207             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1208             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1209
1210             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1211
1212             /* Update potential sum for this i atom from the interaction with this j atom. */
1213             velec            = _mm_and_ps(velec,cutoff_mask);
1214             velec            = _mm_andnot_ps(dummy_mask,velec);
1215             velecsum         = _mm_add_ps(velecsum,velec);
1216
1217             fscal            = felec;
1218
1219             fscal            = _mm_and_ps(fscal,cutoff_mask);
1220
1221             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1222
1223             /* Calculate temporary vectorial force */
1224             tx               = _mm_mul_ps(fscal,dx32);
1225             ty               = _mm_mul_ps(fscal,dy32);
1226             tz               = _mm_mul_ps(fscal,dz32);
1227
1228             /* Update vectorial force */
1229             fix3             = _mm_add_ps(fix3,tx);
1230             fiy3             = _mm_add_ps(fiy3,ty);
1231             fiz3             = _mm_add_ps(fiz3,tz);
1232
1233             fjx2             = _mm_add_ps(fjx2,tx);
1234             fjy2             = _mm_add_ps(fjy2,ty);
1235             fjz2             = _mm_add_ps(fjz2,tz);
1236
1237             }
1238
1239             /**************************
1240              * CALCULATE INTERACTIONS *
1241              **************************/
1242
1243             if (gmx_mm_any_lt(rsq33,rcutoff2))
1244             {
1245
1246             /* REACTION-FIELD ELECTROSTATICS */
1247             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1248             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1249
1250             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1251
1252             /* Update potential sum for this i atom from the interaction with this j atom. */
1253             velec            = _mm_and_ps(velec,cutoff_mask);
1254             velec            = _mm_andnot_ps(dummy_mask,velec);
1255             velecsum         = _mm_add_ps(velecsum,velec);
1256
1257             fscal            = felec;
1258
1259             fscal            = _mm_and_ps(fscal,cutoff_mask);
1260
1261             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1262
1263             /* Calculate temporary vectorial force */
1264             tx               = _mm_mul_ps(fscal,dx33);
1265             ty               = _mm_mul_ps(fscal,dy33);
1266             tz               = _mm_mul_ps(fscal,dz33);
1267
1268             /* Update vectorial force */
1269             fix3             = _mm_add_ps(fix3,tx);
1270             fiy3             = _mm_add_ps(fiy3,ty);
1271             fiz3             = _mm_add_ps(fiz3,tz);
1272
1273             fjx3             = _mm_add_ps(fjx3,tx);
1274             fjy3             = _mm_add_ps(fjy3,ty);
1275             fjz3             = _mm_add_ps(fjz3,tz);
1276
1277             }
1278
1279             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1280             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1281             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1282             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1283
1284             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1285                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1286                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1287
1288             /* Inner loop uses 388 flops */
1289         }
1290
1291         /* End of innermost loop */
1292
1293         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1294                                               f+i_coord_offset,fshift+i_shift_offset);
1295
1296         ggid                        = gid[iidx];
1297         /* Update potential energies */
1298         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1299         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1300
1301         /* Increment number of inner iterations */
1302         inneriter                  += j_index_end - j_index_start;
1303
1304         /* Outer loop uses 26 flops */
1305     }
1306
1307     /* Increment number of outer iterations */
1308     outeriter        += nri;
1309
1310     /* Update outer/inner flops */
1311
1312     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*388);
1313 }
1314 /*
1315  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_sse4_1_single
1316  * Electrostatics interaction: ReactionField
1317  * VdW interaction:            CubicSplineTable
1318  * Geometry:                   Water4-Water4
1319  * Calculate force/pot:        Force
1320  */
1321 void
1322 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_sse4_1_single
1323                     (t_nblist                    * gmx_restrict       nlist,
1324                      rvec                        * gmx_restrict          xx,
1325                      rvec                        * gmx_restrict          ff,
1326                      t_forcerec                  * gmx_restrict          fr,
1327                      t_mdatoms                   * gmx_restrict     mdatoms,
1328                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1329                      t_nrnb                      * gmx_restrict        nrnb)
1330 {
1331     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1332      * just 0 for non-waters.
1333      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1334      * jnr indices corresponding to data put in the four positions in the SIMD register.
1335      */
1336     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1337     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1338     int              jnrA,jnrB,jnrC,jnrD;
1339     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1340     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1341     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1342     real             rcutoff_scalar;
1343     real             *shiftvec,*fshift,*x,*f;
1344     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1345     real             scratch[4*DIM];
1346     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1347     int              vdwioffset0;
1348     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1349     int              vdwioffset1;
1350     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1351     int              vdwioffset2;
1352     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1353     int              vdwioffset3;
1354     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1355     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1356     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1357     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1358     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1359     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1360     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1361     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1362     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1363     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1364     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1365     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1366     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1367     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1368     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1369     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1370     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1371     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1372     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1373     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1374     real             *charge;
1375     int              nvdwtype;
1376     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1377     int              *vdwtype;
1378     real             *vdwparam;
1379     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1380     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1381     __m128i          vfitab;
1382     __m128i          ifour       = _mm_set1_epi32(4);
1383     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1384     real             *vftab;
1385     __m128           dummy_mask,cutoff_mask;
1386     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1387     __m128           one     = _mm_set1_ps(1.0);
1388     __m128           two     = _mm_set1_ps(2.0);
1389     x                = xx[0];
1390     f                = ff[0];
1391
1392     nri              = nlist->nri;
1393     iinr             = nlist->iinr;
1394     jindex           = nlist->jindex;
1395     jjnr             = nlist->jjnr;
1396     shiftidx         = nlist->shift;
1397     gid              = nlist->gid;
1398     shiftvec         = fr->shift_vec[0];
1399     fshift           = fr->fshift[0];
1400     facel            = _mm_set1_ps(fr->epsfac);
1401     charge           = mdatoms->chargeA;
1402     krf              = _mm_set1_ps(fr->ic->k_rf);
1403     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1404     crf              = _mm_set1_ps(fr->ic->c_rf);
1405     nvdwtype         = fr->ntype;
1406     vdwparam         = fr->nbfp;
1407     vdwtype          = mdatoms->typeA;
1408
1409     vftab            = kernel_data->table_vdw->data;
1410     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1411
1412     /* Setup water-specific parameters */
1413     inr              = nlist->iinr[0];
1414     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1415     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1416     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1417     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1418
1419     jq1              = _mm_set1_ps(charge[inr+1]);
1420     jq2              = _mm_set1_ps(charge[inr+2]);
1421     jq3              = _mm_set1_ps(charge[inr+3]);
1422     vdwjidx0A        = 2*vdwtype[inr+0];
1423     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1424     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1425     qq11             = _mm_mul_ps(iq1,jq1);
1426     qq12             = _mm_mul_ps(iq1,jq2);
1427     qq13             = _mm_mul_ps(iq1,jq3);
1428     qq21             = _mm_mul_ps(iq2,jq1);
1429     qq22             = _mm_mul_ps(iq2,jq2);
1430     qq23             = _mm_mul_ps(iq2,jq3);
1431     qq31             = _mm_mul_ps(iq3,jq1);
1432     qq32             = _mm_mul_ps(iq3,jq2);
1433     qq33             = _mm_mul_ps(iq3,jq3);
1434
1435     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1436     rcutoff_scalar   = fr->rcoulomb;
1437     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1438     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1439
1440     /* Avoid stupid compiler warnings */
1441     jnrA = jnrB = jnrC = jnrD = 0;
1442     j_coord_offsetA = 0;
1443     j_coord_offsetB = 0;
1444     j_coord_offsetC = 0;
1445     j_coord_offsetD = 0;
1446
1447     outeriter        = 0;
1448     inneriter        = 0;
1449
1450     for(iidx=0;iidx<4*DIM;iidx++)
1451     {
1452         scratch[iidx] = 0.0;
1453     }
1454
1455     /* Start outer loop over neighborlists */
1456     for(iidx=0; iidx<nri; iidx++)
1457     {
1458         /* Load shift vector for this list */
1459         i_shift_offset   = DIM*shiftidx[iidx];
1460
1461         /* Load limits for loop over neighbors */
1462         j_index_start    = jindex[iidx];
1463         j_index_end      = jindex[iidx+1];
1464
1465         /* Get outer coordinate index */
1466         inr              = iinr[iidx];
1467         i_coord_offset   = DIM*inr;
1468
1469         /* Load i particle coords and add shift vector */
1470         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1471                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1472
1473         fix0             = _mm_setzero_ps();
1474         fiy0             = _mm_setzero_ps();
1475         fiz0             = _mm_setzero_ps();
1476         fix1             = _mm_setzero_ps();
1477         fiy1             = _mm_setzero_ps();
1478         fiz1             = _mm_setzero_ps();
1479         fix2             = _mm_setzero_ps();
1480         fiy2             = _mm_setzero_ps();
1481         fiz2             = _mm_setzero_ps();
1482         fix3             = _mm_setzero_ps();
1483         fiy3             = _mm_setzero_ps();
1484         fiz3             = _mm_setzero_ps();
1485
1486         /* Start inner kernel loop */
1487         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1488         {
1489
1490             /* Get j neighbor index, and coordinate index */
1491             jnrA             = jjnr[jidx];
1492             jnrB             = jjnr[jidx+1];
1493             jnrC             = jjnr[jidx+2];
1494             jnrD             = jjnr[jidx+3];
1495             j_coord_offsetA  = DIM*jnrA;
1496             j_coord_offsetB  = DIM*jnrB;
1497             j_coord_offsetC  = DIM*jnrC;
1498             j_coord_offsetD  = DIM*jnrD;
1499
1500             /* load j atom coordinates */
1501             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1502                                               x+j_coord_offsetC,x+j_coord_offsetD,
1503                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1504                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1505
1506             /* Calculate displacement vector */
1507             dx00             = _mm_sub_ps(ix0,jx0);
1508             dy00             = _mm_sub_ps(iy0,jy0);
1509             dz00             = _mm_sub_ps(iz0,jz0);
1510             dx11             = _mm_sub_ps(ix1,jx1);
1511             dy11             = _mm_sub_ps(iy1,jy1);
1512             dz11             = _mm_sub_ps(iz1,jz1);
1513             dx12             = _mm_sub_ps(ix1,jx2);
1514             dy12             = _mm_sub_ps(iy1,jy2);
1515             dz12             = _mm_sub_ps(iz1,jz2);
1516             dx13             = _mm_sub_ps(ix1,jx3);
1517             dy13             = _mm_sub_ps(iy1,jy3);
1518             dz13             = _mm_sub_ps(iz1,jz3);
1519             dx21             = _mm_sub_ps(ix2,jx1);
1520             dy21             = _mm_sub_ps(iy2,jy1);
1521             dz21             = _mm_sub_ps(iz2,jz1);
1522             dx22             = _mm_sub_ps(ix2,jx2);
1523             dy22             = _mm_sub_ps(iy2,jy2);
1524             dz22             = _mm_sub_ps(iz2,jz2);
1525             dx23             = _mm_sub_ps(ix2,jx3);
1526             dy23             = _mm_sub_ps(iy2,jy3);
1527             dz23             = _mm_sub_ps(iz2,jz3);
1528             dx31             = _mm_sub_ps(ix3,jx1);
1529             dy31             = _mm_sub_ps(iy3,jy1);
1530             dz31             = _mm_sub_ps(iz3,jz1);
1531             dx32             = _mm_sub_ps(ix3,jx2);
1532             dy32             = _mm_sub_ps(iy3,jy2);
1533             dz32             = _mm_sub_ps(iz3,jz2);
1534             dx33             = _mm_sub_ps(ix3,jx3);
1535             dy33             = _mm_sub_ps(iy3,jy3);
1536             dz33             = _mm_sub_ps(iz3,jz3);
1537
1538             /* Calculate squared distance and things based on it */
1539             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1540             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1541             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1542             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1543             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1544             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1545             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1546             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1547             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1548             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1549
1550             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1551             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1552             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1553             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1554             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1555             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1556             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1557             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1558             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1559             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1560
1561             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1562             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1563             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1564             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1565             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1566             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1567             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1568             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1569             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1570
1571             fjx0             = _mm_setzero_ps();
1572             fjy0             = _mm_setzero_ps();
1573             fjz0             = _mm_setzero_ps();
1574             fjx1             = _mm_setzero_ps();
1575             fjy1             = _mm_setzero_ps();
1576             fjz1             = _mm_setzero_ps();
1577             fjx2             = _mm_setzero_ps();
1578             fjy2             = _mm_setzero_ps();
1579             fjz2             = _mm_setzero_ps();
1580             fjx3             = _mm_setzero_ps();
1581             fjy3             = _mm_setzero_ps();
1582             fjz3             = _mm_setzero_ps();
1583
1584             /**************************
1585              * CALCULATE INTERACTIONS *
1586              **************************/
1587
1588             if (gmx_mm_any_lt(rsq00,rcutoff2))
1589             {
1590
1591             r00              = _mm_mul_ps(rsq00,rinv00);
1592
1593             /* Calculate table index by multiplying r with table scale and truncate to integer */
1594             rt               = _mm_mul_ps(r00,vftabscale);
1595             vfitab           = _mm_cvttps_epi32(rt);
1596             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1597             vfitab           = _mm_slli_epi32(vfitab,3);
1598
1599             /* CUBIC SPLINE TABLE DISPERSION */
1600             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1601             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1602             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1603             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1604             _MM_TRANSPOSE4_PS(Y,F,G,H);
1605             Heps             = _mm_mul_ps(vfeps,H);
1606             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1607             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1608             fvdw6            = _mm_mul_ps(c6_00,FF);
1609
1610             /* CUBIC SPLINE TABLE REPULSION */
1611             vfitab           = _mm_add_epi32(vfitab,ifour);
1612             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1613             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1614             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1615             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1616             _MM_TRANSPOSE4_PS(Y,F,G,H);
1617             Heps             = _mm_mul_ps(vfeps,H);
1618             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1619             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1620             fvdw12           = _mm_mul_ps(c12_00,FF);
1621             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1622
1623             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1624
1625             fscal            = fvdw;
1626
1627             fscal            = _mm_and_ps(fscal,cutoff_mask);
1628
1629             /* Calculate temporary vectorial force */
1630             tx               = _mm_mul_ps(fscal,dx00);
1631             ty               = _mm_mul_ps(fscal,dy00);
1632             tz               = _mm_mul_ps(fscal,dz00);
1633
1634             /* Update vectorial force */
1635             fix0             = _mm_add_ps(fix0,tx);
1636             fiy0             = _mm_add_ps(fiy0,ty);
1637             fiz0             = _mm_add_ps(fiz0,tz);
1638
1639             fjx0             = _mm_add_ps(fjx0,tx);
1640             fjy0             = _mm_add_ps(fjy0,ty);
1641             fjz0             = _mm_add_ps(fjz0,tz);
1642
1643             }
1644
1645             /**************************
1646              * CALCULATE INTERACTIONS *
1647              **************************/
1648
1649             if (gmx_mm_any_lt(rsq11,rcutoff2))
1650             {
1651
1652             /* REACTION-FIELD ELECTROSTATICS */
1653             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1654
1655             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1656
1657             fscal            = felec;
1658
1659             fscal            = _mm_and_ps(fscal,cutoff_mask);
1660
1661             /* Calculate temporary vectorial force */
1662             tx               = _mm_mul_ps(fscal,dx11);
1663             ty               = _mm_mul_ps(fscal,dy11);
1664             tz               = _mm_mul_ps(fscal,dz11);
1665
1666             /* Update vectorial force */
1667             fix1             = _mm_add_ps(fix1,tx);
1668             fiy1             = _mm_add_ps(fiy1,ty);
1669             fiz1             = _mm_add_ps(fiz1,tz);
1670
1671             fjx1             = _mm_add_ps(fjx1,tx);
1672             fjy1             = _mm_add_ps(fjy1,ty);
1673             fjz1             = _mm_add_ps(fjz1,tz);
1674
1675             }
1676
1677             /**************************
1678              * CALCULATE INTERACTIONS *
1679              **************************/
1680
1681             if (gmx_mm_any_lt(rsq12,rcutoff2))
1682             {
1683
1684             /* REACTION-FIELD ELECTROSTATICS */
1685             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1686
1687             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1688
1689             fscal            = felec;
1690
1691             fscal            = _mm_and_ps(fscal,cutoff_mask);
1692
1693             /* Calculate temporary vectorial force */
1694             tx               = _mm_mul_ps(fscal,dx12);
1695             ty               = _mm_mul_ps(fscal,dy12);
1696             tz               = _mm_mul_ps(fscal,dz12);
1697
1698             /* Update vectorial force */
1699             fix1             = _mm_add_ps(fix1,tx);
1700             fiy1             = _mm_add_ps(fiy1,ty);
1701             fiz1             = _mm_add_ps(fiz1,tz);
1702
1703             fjx2             = _mm_add_ps(fjx2,tx);
1704             fjy2             = _mm_add_ps(fjy2,ty);
1705             fjz2             = _mm_add_ps(fjz2,tz);
1706
1707             }
1708
1709             /**************************
1710              * CALCULATE INTERACTIONS *
1711              **************************/
1712
1713             if (gmx_mm_any_lt(rsq13,rcutoff2))
1714             {
1715
1716             /* REACTION-FIELD ELECTROSTATICS */
1717             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1718
1719             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1720
1721             fscal            = felec;
1722
1723             fscal            = _mm_and_ps(fscal,cutoff_mask);
1724
1725             /* Calculate temporary vectorial force */
1726             tx               = _mm_mul_ps(fscal,dx13);
1727             ty               = _mm_mul_ps(fscal,dy13);
1728             tz               = _mm_mul_ps(fscal,dz13);
1729
1730             /* Update vectorial force */
1731             fix1             = _mm_add_ps(fix1,tx);
1732             fiy1             = _mm_add_ps(fiy1,ty);
1733             fiz1             = _mm_add_ps(fiz1,tz);
1734
1735             fjx3             = _mm_add_ps(fjx3,tx);
1736             fjy3             = _mm_add_ps(fjy3,ty);
1737             fjz3             = _mm_add_ps(fjz3,tz);
1738
1739             }
1740
1741             /**************************
1742              * CALCULATE INTERACTIONS *
1743              **************************/
1744
1745             if (gmx_mm_any_lt(rsq21,rcutoff2))
1746             {
1747
1748             /* REACTION-FIELD ELECTROSTATICS */
1749             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1750
1751             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1752
1753             fscal            = felec;
1754
1755             fscal            = _mm_and_ps(fscal,cutoff_mask);
1756
1757             /* Calculate temporary vectorial force */
1758             tx               = _mm_mul_ps(fscal,dx21);
1759             ty               = _mm_mul_ps(fscal,dy21);
1760             tz               = _mm_mul_ps(fscal,dz21);
1761
1762             /* Update vectorial force */
1763             fix2             = _mm_add_ps(fix2,tx);
1764             fiy2             = _mm_add_ps(fiy2,ty);
1765             fiz2             = _mm_add_ps(fiz2,tz);
1766
1767             fjx1             = _mm_add_ps(fjx1,tx);
1768             fjy1             = _mm_add_ps(fjy1,ty);
1769             fjz1             = _mm_add_ps(fjz1,tz);
1770
1771             }
1772
1773             /**************************
1774              * CALCULATE INTERACTIONS *
1775              **************************/
1776
1777             if (gmx_mm_any_lt(rsq22,rcutoff2))
1778             {
1779
1780             /* REACTION-FIELD ELECTROSTATICS */
1781             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1782
1783             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1784
1785             fscal            = felec;
1786
1787             fscal            = _mm_and_ps(fscal,cutoff_mask);
1788
1789             /* Calculate temporary vectorial force */
1790             tx               = _mm_mul_ps(fscal,dx22);
1791             ty               = _mm_mul_ps(fscal,dy22);
1792             tz               = _mm_mul_ps(fscal,dz22);
1793
1794             /* Update vectorial force */
1795             fix2             = _mm_add_ps(fix2,tx);
1796             fiy2             = _mm_add_ps(fiy2,ty);
1797             fiz2             = _mm_add_ps(fiz2,tz);
1798
1799             fjx2             = _mm_add_ps(fjx2,tx);
1800             fjy2             = _mm_add_ps(fjy2,ty);
1801             fjz2             = _mm_add_ps(fjz2,tz);
1802
1803             }
1804
1805             /**************************
1806              * CALCULATE INTERACTIONS *
1807              **************************/
1808
1809             if (gmx_mm_any_lt(rsq23,rcutoff2))
1810             {
1811
1812             /* REACTION-FIELD ELECTROSTATICS */
1813             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1814
1815             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1816
1817             fscal            = felec;
1818
1819             fscal            = _mm_and_ps(fscal,cutoff_mask);
1820
1821             /* Calculate temporary vectorial force */
1822             tx               = _mm_mul_ps(fscal,dx23);
1823             ty               = _mm_mul_ps(fscal,dy23);
1824             tz               = _mm_mul_ps(fscal,dz23);
1825
1826             /* Update vectorial force */
1827             fix2             = _mm_add_ps(fix2,tx);
1828             fiy2             = _mm_add_ps(fiy2,ty);
1829             fiz2             = _mm_add_ps(fiz2,tz);
1830
1831             fjx3             = _mm_add_ps(fjx3,tx);
1832             fjy3             = _mm_add_ps(fjy3,ty);
1833             fjz3             = _mm_add_ps(fjz3,tz);
1834
1835             }
1836
1837             /**************************
1838              * CALCULATE INTERACTIONS *
1839              **************************/
1840
1841             if (gmx_mm_any_lt(rsq31,rcutoff2))
1842             {
1843
1844             /* REACTION-FIELD ELECTROSTATICS */
1845             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1846
1847             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1848
1849             fscal            = felec;
1850
1851             fscal            = _mm_and_ps(fscal,cutoff_mask);
1852
1853             /* Calculate temporary vectorial force */
1854             tx               = _mm_mul_ps(fscal,dx31);
1855             ty               = _mm_mul_ps(fscal,dy31);
1856             tz               = _mm_mul_ps(fscal,dz31);
1857
1858             /* Update vectorial force */
1859             fix3             = _mm_add_ps(fix3,tx);
1860             fiy3             = _mm_add_ps(fiy3,ty);
1861             fiz3             = _mm_add_ps(fiz3,tz);
1862
1863             fjx1             = _mm_add_ps(fjx1,tx);
1864             fjy1             = _mm_add_ps(fjy1,ty);
1865             fjz1             = _mm_add_ps(fjz1,tz);
1866
1867             }
1868
1869             /**************************
1870              * CALCULATE INTERACTIONS *
1871              **************************/
1872
1873             if (gmx_mm_any_lt(rsq32,rcutoff2))
1874             {
1875
1876             /* REACTION-FIELD ELECTROSTATICS */
1877             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1878
1879             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1880
1881             fscal            = felec;
1882
1883             fscal            = _mm_and_ps(fscal,cutoff_mask);
1884
1885             /* Calculate temporary vectorial force */
1886             tx               = _mm_mul_ps(fscal,dx32);
1887             ty               = _mm_mul_ps(fscal,dy32);
1888             tz               = _mm_mul_ps(fscal,dz32);
1889
1890             /* Update vectorial force */
1891             fix3             = _mm_add_ps(fix3,tx);
1892             fiy3             = _mm_add_ps(fiy3,ty);
1893             fiz3             = _mm_add_ps(fiz3,tz);
1894
1895             fjx2             = _mm_add_ps(fjx2,tx);
1896             fjy2             = _mm_add_ps(fjy2,ty);
1897             fjz2             = _mm_add_ps(fjz2,tz);
1898
1899             }
1900
1901             /**************************
1902              * CALCULATE INTERACTIONS *
1903              **************************/
1904
1905             if (gmx_mm_any_lt(rsq33,rcutoff2))
1906             {
1907
1908             /* REACTION-FIELD ELECTROSTATICS */
1909             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1910
1911             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1912
1913             fscal            = felec;
1914
1915             fscal            = _mm_and_ps(fscal,cutoff_mask);
1916
1917             /* Calculate temporary vectorial force */
1918             tx               = _mm_mul_ps(fscal,dx33);
1919             ty               = _mm_mul_ps(fscal,dy33);
1920             tz               = _mm_mul_ps(fscal,dz33);
1921
1922             /* Update vectorial force */
1923             fix3             = _mm_add_ps(fix3,tx);
1924             fiy3             = _mm_add_ps(fiy3,ty);
1925             fiz3             = _mm_add_ps(fiz3,tz);
1926
1927             fjx3             = _mm_add_ps(fjx3,tx);
1928             fjy3             = _mm_add_ps(fjy3,ty);
1929             fjz3             = _mm_add_ps(fjz3,tz);
1930
1931             }
1932
1933             fjptrA             = f+j_coord_offsetA;
1934             fjptrB             = f+j_coord_offsetB;
1935             fjptrC             = f+j_coord_offsetC;
1936             fjptrD             = f+j_coord_offsetD;
1937
1938             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1939                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1940                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1941
1942             /* Inner loop uses 324 flops */
1943         }
1944
1945         if(jidx<j_index_end)
1946         {
1947
1948             /* Get j neighbor index, and coordinate index */
1949             jnrlistA         = jjnr[jidx];
1950             jnrlistB         = jjnr[jidx+1];
1951             jnrlistC         = jjnr[jidx+2];
1952             jnrlistD         = jjnr[jidx+3];
1953             /* Sign of each element will be negative for non-real atoms.
1954              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1955              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1956              */
1957             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1958             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1959             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1960             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1961             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1962             j_coord_offsetA  = DIM*jnrA;
1963             j_coord_offsetB  = DIM*jnrB;
1964             j_coord_offsetC  = DIM*jnrC;
1965             j_coord_offsetD  = DIM*jnrD;
1966
1967             /* load j atom coordinates */
1968             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1969                                               x+j_coord_offsetC,x+j_coord_offsetD,
1970                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1971                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1972
1973             /* Calculate displacement vector */
1974             dx00             = _mm_sub_ps(ix0,jx0);
1975             dy00             = _mm_sub_ps(iy0,jy0);
1976             dz00             = _mm_sub_ps(iz0,jz0);
1977             dx11             = _mm_sub_ps(ix1,jx1);
1978             dy11             = _mm_sub_ps(iy1,jy1);
1979             dz11             = _mm_sub_ps(iz1,jz1);
1980             dx12             = _mm_sub_ps(ix1,jx2);
1981             dy12             = _mm_sub_ps(iy1,jy2);
1982             dz12             = _mm_sub_ps(iz1,jz2);
1983             dx13             = _mm_sub_ps(ix1,jx3);
1984             dy13             = _mm_sub_ps(iy1,jy3);
1985             dz13             = _mm_sub_ps(iz1,jz3);
1986             dx21             = _mm_sub_ps(ix2,jx1);
1987             dy21             = _mm_sub_ps(iy2,jy1);
1988             dz21             = _mm_sub_ps(iz2,jz1);
1989             dx22             = _mm_sub_ps(ix2,jx2);
1990             dy22             = _mm_sub_ps(iy2,jy2);
1991             dz22             = _mm_sub_ps(iz2,jz2);
1992             dx23             = _mm_sub_ps(ix2,jx3);
1993             dy23             = _mm_sub_ps(iy2,jy3);
1994             dz23             = _mm_sub_ps(iz2,jz3);
1995             dx31             = _mm_sub_ps(ix3,jx1);
1996             dy31             = _mm_sub_ps(iy3,jy1);
1997             dz31             = _mm_sub_ps(iz3,jz1);
1998             dx32             = _mm_sub_ps(ix3,jx2);
1999             dy32             = _mm_sub_ps(iy3,jy2);
2000             dz32             = _mm_sub_ps(iz3,jz2);
2001             dx33             = _mm_sub_ps(ix3,jx3);
2002             dy33             = _mm_sub_ps(iy3,jy3);
2003             dz33             = _mm_sub_ps(iz3,jz3);
2004
2005             /* Calculate squared distance and things based on it */
2006             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
2007             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
2008             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
2009             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
2010             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
2011             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
2012             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
2013             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
2014             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
2015             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
2016
2017             rinv00           = gmx_mm_invsqrt_ps(rsq00);
2018             rinv11           = gmx_mm_invsqrt_ps(rsq11);
2019             rinv12           = gmx_mm_invsqrt_ps(rsq12);
2020             rinv13           = gmx_mm_invsqrt_ps(rsq13);
2021             rinv21           = gmx_mm_invsqrt_ps(rsq21);
2022             rinv22           = gmx_mm_invsqrt_ps(rsq22);
2023             rinv23           = gmx_mm_invsqrt_ps(rsq23);
2024             rinv31           = gmx_mm_invsqrt_ps(rsq31);
2025             rinv32           = gmx_mm_invsqrt_ps(rsq32);
2026             rinv33           = gmx_mm_invsqrt_ps(rsq33);
2027
2028             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
2029             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
2030             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
2031             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
2032             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
2033             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
2034             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
2035             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
2036             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
2037
2038             fjx0             = _mm_setzero_ps();
2039             fjy0             = _mm_setzero_ps();
2040             fjz0             = _mm_setzero_ps();
2041             fjx1             = _mm_setzero_ps();
2042             fjy1             = _mm_setzero_ps();
2043             fjz1             = _mm_setzero_ps();
2044             fjx2             = _mm_setzero_ps();
2045             fjy2             = _mm_setzero_ps();
2046             fjz2             = _mm_setzero_ps();
2047             fjx3             = _mm_setzero_ps();
2048             fjy3             = _mm_setzero_ps();
2049             fjz3             = _mm_setzero_ps();
2050
2051             /**************************
2052              * CALCULATE INTERACTIONS *
2053              **************************/
2054
2055             if (gmx_mm_any_lt(rsq00,rcutoff2))
2056             {
2057
2058             r00              = _mm_mul_ps(rsq00,rinv00);
2059             r00              = _mm_andnot_ps(dummy_mask,r00);
2060
2061             /* Calculate table index by multiplying r with table scale and truncate to integer */
2062             rt               = _mm_mul_ps(r00,vftabscale);
2063             vfitab           = _mm_cvttps_epi32(rt);
2064             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
2065             vfitab           = _mm_slli_epi32(vfitab,3);
2066
2067             /* CUBIC SPLINE TABLE DISPERSION */
2068             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
2069             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
2070             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
2071             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
2072             _MM_TRANSPOSE4_PS(Y,F,G,H);
2073             Heps             = _mm_mul_ps(vfeps,H);
2074             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
2075             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
2076             fvdw6            = _mm_mul_ps(c6_00,FF);
2077
2078             /* CUBIC SPLINE TABLE REPULSION */
2079             vfitab           = _mm_add_epi32(vfitab,ifour);
2080             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
2081             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
2082             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
2083             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
2084             _MM_TRANSPOSE4_PS(Y,F,G,H);
2085             Heps             = _mm_mul_ps(vfeps,H);
2086             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
2087             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
2088             fvdw12           = _mm_mul_ps(c12_00,FF);
2089             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
2090
2091             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
2092
2093             fscal            = fvdw;
2094
2095             fscal            = _mm_and_ps(fscal,cutoff_mask);
2096
2097             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2098
2099             /* Calculate temporary vectorial force */
2100             tx               = _mm_mul_ps(fscal,dx00);
2101             ty               = _mm_mul_ps(fscal,dy00);
2102             tz               = _mm_mul_ps(fscal,dz00);
2103
2104             /* Update vectorial force */
2105             fix0             = _mm_add_ps(fix0,tx);
2106             fiy0             = _mm_add_ps(fiy0,ty);
2107             fiz0             = _mm_add_ps(fiz0,tz);
2108
2109             fjx0             = _mm_add_ps(fjx0,tx);
2110             fjy0             = _mm_add_ps(fjy0,ty);
2111             fjz0             = _mm_add_ps(fjz0,tz);
2112
2113             }
2114
2115             /**************************
2116              * CALCULATE INTERACTIONS *
2117              **************************/
2118
2119             if (gmx_mm_any_lt(rsq11,rcutoff2))
2120             {
2121
2122             /* REACTION-FIELD ELECTROSTATICS */
2123             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
2124
2125             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2126
2127             fscal            = felec;
2128
2129             fscal            = _mm_and_ps(fscal,cutoff_mask);
2130
2131             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2132
2133             /* Calculate temporary vectorial force */
2134             tx               = _mm_mul_ps(fscal,dx11);
2135             ty               = _mm_mul_ps(fscal,dy11);
2136             tz               = _mm_mul_ps(fscal,dz11);
2137
2138             /* Update vectorial force */
2139             fix1             = _mm_add_ps(fix1,tx);
2140             fiy1             = _mm_add_ps(fiy1,ty);
2141             fiz1             = _mm_add_ps(fiz1,tz);
2142
2143             fjx1             = _mm_add_ps(fjx1,tx);
2144             fjy1             = _mm_add_ps(fjy1,ty);
2145             fjz1             = _mm_add_ps(fjz1,tz);
2146
2147             }
2148
2149             /**************************
2150              * CALCULATE INTERACTIONS *
2151              **************************/
2152
2153             if (gmx_mm_any_lt(rsq12,rcutoff2))
2154             {
2155
2156             /* REACTION-FIELD ELECTROSTATICS */
2157             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
2158
2159             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2160
2161             fscal            = felec;
2162
2163             fscal            = _mm_and_ps(fscal,cutoff_mask);
2164
2165             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2166
2167             /* Calculate temporary vectorial force */
2168             tx               = _mm_mul_ps(fscal,dx12);
2169             ty               = _mm_mul_ps(fscal,dy12);
2170             tz               = _mm_mul_ps(fscal,dz12);
2171
2172             /* Update vectorial force */
2173             fix1             = _mm_add_ps(fix1,tx);
2174             fiy1             = _mm_add_ps(fiy1,ty);
2175             fiz1             = _mm_add_ps(fiz1,tz);
2176
2177             fjx2             = _mm_add_ps(fjx2,tx);
2178             fjy2             = _mm_add_ps(fjy2,ty);
2179             fjz2             = _mm_add_ps(fjz2,tz);
2180
2181             }
2182
2183             /**************************
2184              * CALCULATE INTERACTIONS *
2185              **************************/
2186
2187             if (gmx_mm_any_lt(rsq13,rcutoff2))
2188             {
2189
2190             /* REACTION-FIELD ELECTROSTATICS */
2191             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
2192
2193             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
2194
2195             fscal            = felec;
2196
2197             fscal            = _mm_and_ps(fscal,cutoff_mask);
2198
2199             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2200
2201             /* Calculate temporary vectorial force */
2202             tx               = _mm_mul_ps(fscal,dx13);
2203             ty               = _mm_mul_ps(fscal,dy13);
2204             tz               = _mm_mul_ps(fscal,dz13);
2205
2206             /* Update vectorial force */
2207             fix1             = _mm_add_ps(fix1,tx);
2208             fiy1             = _mm_add_ps(fiy1,ty);
2209             fiz1             = _mm_add_ps(fiz1,tz);
2210
2211             fjx3             = _mm_add_ps(fjx3,tx);
2212             fjy3             = _mm_add_ps(fjy3,ty);
2213             fjz3             = _mm_add_ps(fjz3,tz);
2214
2215             }
2216
2217             /**************************
2218              * CALCULATE INTERACTIONS *
2219              **************************/
2220
2221             if (gmx_mm_any_lt(rsq21,rcutoff2))
2222             {
2223
2224             /* REACTION-FIELD ELECTROSTATICS */
2225             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
2226
2227             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2228
2229             fscal            = felec;
2230
2231             fscal            = _mm_and_ps(fscal,cutoff_mask);
2232
2233             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2234
2235             /* Calculate temporary vectorial force */
2236             tx               = _mm_mul_ps(fscal,dx21);
2237             ty               = _mm_mul_ps(fscal,dy21);
2238             tz               = _mm_mul_ps(fscal,dz21);
2239
2240             /* Update vectorial force */
2241             fix2             = _mm_add_ps(fix2,tx);
2242             fiy2             = _mm_add_ps(fiy2,ty);
2243             fiz2             = _mm_add_ps(fiz2,tz);
2244
2245             fjx1             = _mm_add_ps(fjx1,tx);
2246             fjy1             = _mm_add_ps(fjy1,ty);
2247             fjz1             = _mm_add_ps(fjz1,tz);
2248
2249             }
2250
2251             /**************************
2252              * CALCULATE INTERACTIONS *
2253              **************************/
2254
2255             if (gmx_mm_any_lt(rsq22,rcutoff2))
2256             {
2257
2258             /* REACTION-FIELD ELECTROSTATICS */
2259             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2260
2261             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2262
2263             fscal            = felec;
2264
2265             fscal            = _mm_and_ps(fscal,cutoff_mask);
2266
2267             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2268
2269             /* Calculate temporary vectorial force */
2270             tx               = _mm_mul_ps(fscal,dx22);
2271             ty               = _mm_mul_ps(fscal,dy22);
2272             tz               = _mm_mul_ps(fscal,dz22);
2273
2274             /* Update vectorial force */
2275             fix2             = _mm_add_ps(fix2,tx);
2276             fiy2             = _mm_add_ps(fiy2,ty);
2277             fiz2             = _mm_add_ps(fiz2,tz);
2278
2279             fjx2             = _mm_add_ps(fjx2,tx);
2280             fjy2             = _mm_add_ps(fjy2,ty);
2281             fjz2             = _mm_add_ps(fjz2,tz);
2282
2283             }
2284
2285             /**************************
2286              * CALCULATE INTERACTIONS *
2287              **************************/
2288
2289             if (gmx_mm_any_lt(rsq23,rcutoff2))
2290             {
2291
2292             /* REACTION-FIELD ELECTROSTATICS */
2293             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
2294
2295             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2296
2297             fscal            = felec;
2298
2299             fscal            = _mm_and_ps(fscal,cutoff_mask);
2300
2301             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2302
2303             /* Calculate temporary vectorial force */
2304             tx               = _mm_mul_ps(fscal,dx23);
2305             ty               = _mm_mul_ps(fscal,dy23);
2306             tz               = _mm_mul_ps(fscal,dz23);
2307
2308             /* Update vectorial force */
2309             fix2             = _mm_add_ps(fix2,tx);
2310             fiy2             = _mm_add_ps(fiy2,ty);
2311             fiz2             = _mm_add_ps(fiz2,tz);
2312
2313             fjx3             = _mm_add_ps(fjx3,tx);
2314             fjy3             = _mm_add_ps(fjy3,ty);
2315             fjz3             = _mm_add_ps(fjz3,tz);
2316
2317             }
2318
2319             /**************************
2320              * CALCULATE INTERACTIONS *
2321              **************************/
2322
2323             if (gmx_mm_any_lt(rsq31,rcutoff2))
2324             {
2325
2326             /* REACTION-FIELD ELECTROSTATICS */
2327             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
2328
2329             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2330
2331             fscal            = felec;
2332
2333             fscal            = _mm_and_ps(fscal,cutoff_mask);
2334
2335             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2336
2337             /* Calculate temporary vectorial force */
2338             tx               = _mm_mul_ps(fscal,dx31);
2339             ty               = _mm_mul_ps(fscal,dy31);
2340             tz               = _mm_mul_ps(fscal,dz31);
2341
2342             /* Update vectorial force */
2343             fix3             = _mm_add_ps(fix3,tx);
2344             fiy3             = _mm_add_ps(fiy3,ty);
2345             fiz3             = _mm_add_ps(fiz3,tz);
2346
2347             fjx1             = _mm_add_ps(fjx1,tx);
2348             fjy1             = _mm_add_ps(fjy1,ty);
2349             fjz1             = _mm_add_ps(fjz1,tz);
2350
2351             }
2352
2353             /**************************
2354              * CALCULATE INTERACTIONS *
2355              **************************/
2356
2357             if (gmx_mm_any_lt(rsq32,rcutoff2))
2358             {
2359
2360             /* REACTION-FIELD ELECTROSTATICS */
2361             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
2362
2363             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2364
2365             fscal            = felec;
2366
2367             fscal            = _mm_and_ps(fscal,cutoff_mask);
2368
2369             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2370
2371             /* Calculate temporary vectorial force */
2372             tx               = _mm_mul_ps(fscal,dx32);
2373             ty               = _mm_mul_ps(fscal,dy32);
2374             tz               = _mm_mul_ps(fscal,dz32);
2375
2376             /* Update vectorial force */
2377             fix3             = _mm_add_ps(fix3,tx);
2378             fiy3             = _mm_add_ps(fiy3,ty);
2379             fiz3             = _mm_add_ps(fiz3,tz);
2380
2381             fjx2             = _mm_add_ps(fjx2,tx);
2382             fjy2             = _mm_add_ps(fjy2,ty);
2383             fjz2             = _mm_add_ps(fjz2,tz);
2384
2385             }
2386
2387             /**************************
2388              * CALCULATE INTERACTIONS *
2389              **************************/
2390
2391             if (gmx_mm_any_lt(rsq33,rcutoff2))
2392             {
2393
2394             /* REACTION-FIELD ELECTROSTATICS */
2395             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2396
2397             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2398
2399             fscal            = felec;
2400
2401             fscal            = _mm_and_ps(fscal,cutoff_mask);
2402
2403             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2404
2405             /* Calculate temporary vectorial force */
2406             tx               = _mm_mul_ps(fscal,dx33);
2407             ty               = _mm_mul_ps(fscal,dy33);
2408             tz               = _mm_mul_ps(fscal,dz33);
2409
2410             /* Update vectorial force */
2411             fix3             = _mm_add_ps(fix3,tx);
2412             fiy3             = _mm_add_ps(fiy3,ty);
2413             fiz3             = _mm_add_ps(fiz3,tz);
2414
2415             fjx3             = _mm_add_ps(fjx3,tx);
2416             fjy3             = _mm_add_ps(fjy3,ty);
2417             fjz3             = _mm_add_ps(fjz3,tz);
2418
2419             }
2420
2421             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2422             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2423             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2424             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2425
2426             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2427                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2428                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2429
2430             /* Inner loop uses 325 flops */
2431         }
2432
2433         /* End of innermost loop */
2434
2435         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2436                                               f+i_coord_offset,fshift+i_shift_offset);
2437
2438         /* Increment number of inner iterations */
2439         inneriter                  += j_index_end - j_index_start;
2440
2441         /* Outer loop uses 24 flops */
2442     }
2443
2444     /* Increment number of outer iterations */
2445     outeriter        += nri;
2446
2447     /* Update outer/inner flops */
2448
2449     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*325);
2450 }