src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_sse4_1_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_sse4_1_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwioffset3;
  76     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  77     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  78     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  79     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  80     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  81     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  82     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  83     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  84     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  85     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  86     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  87     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  88     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  89     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  90     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  91     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  92     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  93     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  94     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  95     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 102     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 103     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 104     real             rswitch_scalar,d_scalar;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_ps(fr->ic->k_rf);
 123     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_ps(fr->ic->c_rf);
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 132     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 133     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 134     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 135
 136     jq1              = _mm_set1_ps(charge[inr+1]);
 137     jq2              = _mm_set1_ps(charge[inr+2]);
 138     jq3              = _mm_set1_ps(charge[inr+3]);
 139     vdwjidx0A        = 2*vdwtype[inr+0];
 140     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 141     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 142     qq11             = _mm_mul_ps(iq1,jq1);
 143     qq12             = _mm_mul_ps(iq1,jq2);
 144     qq13             = _mm_mul_ps(iq1,jq3);
 145     qq21             = _mm_mul_ps(iq2,jq1);
 146     qq22             = _mm_mul_ps(iq2,jq2);
 147     qq23             = _mm_mul_ps(iq2,jq3);
 148     qq31             = _mm_mul_ps(iq3,jq1);
 149     qq32             = _mm_mul_ps(iq3,jq2);
 150     qq33             = _mm_mul_ps(iq3,jq3);
 151
 152     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 153     rcutoff_scalar   = fr->rcoulomb;
 154     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 155     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 156
 157     rswitch_scalar   = fr->rvdw_switch;
 158     rswitch          = _mm_set1_ps(rswitch_scalar);
 159     /* Setup switch parameters */
 160     d_scalar         = rcutoff_scalar-rswitch_scalar;
 161     d                = _mm_set1_ps(d_scalar);
 162     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 163     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 164     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 165     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 166     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 167     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 168
 169     /* Avoid stupid compiler warnings */
 170     jnrA = jnrB = jnrC = jnrD = 0;
 171     j_coord_offsetA = 0;
 172     j_coord_offsetB = 0;
 173     j_coord_offsetC = 0;
 174     j_coord_offsetD = 0;
 175
 176     outeriter        = 0;
 177     inneriter        = 0;
 178
 179     for(iidx=0;iidx<4*DIM;iidx++)
 180     {
 181         scratch[iidx] = 0.0;
 182     }
 183
 184     /* Start outer loop over neighborlists */
 185     for(iidx=0; iidx<nri; iidx++)
 186     {
 187         /* Load shift vector for this list */
 188         i_shift_offset   = DIM*shiftidx[iidx];
 189
 190         /* Load limits for loop over neighbors */
 191         j_index_start    = jindex[iidx];
 192         j_index_end      = jindex[iidx+1];
 193
 194         /* Get outer coordinate index */
 195         inr              = iinr[iidx];
 196         i_coord_offset   = DIM*inr;
 197
 198         /* Load i particle coords and add shift vector */
 199         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 200                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 201
 202         fix0             = _mm_setzero_ps();
 203         fiy0             = _mm_setzero_ps();
 204         fiz0             = _mm_setzero_ps();
 205         fix1             = _mm_setzero_ps();
 206         fiy1             = _mm_setzero_ps();
 207         fiz1             = _mm_setzero_ps();
 208         fix2             = _mm_setzero_ps();
 209         fiy2             = _mm_setzero_ps();
 210         fiz2             = _mm_setzero_ps();
 211         fix3             = _mm_setzero_ps();
 212         fiy3             = _mm_setzero_ps();
 213         fiz3             = _mm_setzero_ps();
 214
 215         /* Reset potential sums */
 216         velecsum         = _mm_setzero_ps();
 217         vvdwsum          = _mm_setzero_ps();
 218
 219         /* Start inner kernel loop */
 220         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 221         {
 222
 223             /* Get j neighbor index, and coordinate index */
 224             jnrA             = jjnr[jidx];
 225             jnrB             = jjnr[jidx+1];
 226             jnrC             = jjnr[jidx+2];
 227             jnrD             = jjnr[jidx+3];
 228             j_coord_offsetA  = DIM*jnrA;
 229             j_coord_offsetB  = DIM*jnrB;
 230             j_coord_offsetC  = DIM*jnrC;
 231             j_coord_offsetD  = DIM*jnrD;
 232
 233             /* load j atom coordinates */
 234             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 235                                               x+j_coord_offsetC,x+j_coord_offsetD,
 236                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 237                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 238
 239             /* Calculate displacement vector */
 240             dx00             = _mm_sub_ps(ix0,jx0);
 241             dy00             = _mm_sub_ps(iy0,jy0);
 242             dz00             = _mm_sub_ps(iz0,jz0);
 243             dx11             = _mm_sub_ps(ix1,jx1);
 244             dy11             = _mm_sub_ps(iy1,jy1);
 245             dz11             = _mm_sub_ps(iz1,jz1);
 246             dx12             = _mm_sub_ps(ix1,jx2);
 247             dy12             = _mm_sub_ps(iy1,jy2);
 248             dz12             = _mm_sub_ps(iz1,jz2);
 249             dx13             = _mm_sub_ps(ix1,jx3);
 250             dy13             = _mm_sub_ps(iy1,jy3);
 251             dz13             = _mm_sub_ps(iz1,jz3);
 252             dx21             = _mm_sub_ps(ix2,jx1);
 253             dy21             = _mm_sub_ps(iy2,jy1);
 254             dz21             = _mm_sub_ps(iz2,jz1);
 255             dx22             = _mm_sub_ps(ix2,jx2);
 256             dy22             = _mm_sub_ps(iy2,jy2);
 257             dz22             = _mm_sub_ps(iz2,jz2);
 258             dx23             = _mm_sub_ps(ix2,jx3);
 259             dy23             = _mm_sub_ps(iy2,jy3);
 260             dz23             = _mm_sub_ps(iz2,jz3);
 261             dx31             = _mm_sub_ps(ix3,jx1);
 262             dy31             = _mm_sub_ps(iy3,jy1);
 263             dz31             = _mm_sub_ps(iz3,jz1);
 264             dx32             = _mm_sub_ps(ix3,jx2);
 265             dy32             = _mm_sub_ps(iy3,jy2);
 266             dz32             = _mm_sub_ps(iz3,jz2);
 267             dx33             = _mm_sub_ps(ix3,jx3);
 268             dy33             = _mm_sub_ps(iy3,jy3);
 269             dz33             = _mm_sub_ps(iz3,jz3);
 270
 271             /* Calculate squared distance and things based on it */
 272             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 273             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 274             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 275             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 276             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 277             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 278             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 279             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 280             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 281             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 282
 283             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 284             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 285             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 286             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 287             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 288             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 289             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 290             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 291             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 292             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 293
 294             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 295             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 296             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 297             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 298             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 299             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 300             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 301             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 302             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 303             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 304
 305             fjx0             = _mm_setzero_ps();
 306             fjy0             = _mm_setzero_ps();
 307             fjz0             = _mm_setzero_ps();
 308             fjx1             = _mm_setzero_ps();
 309             fjy1             = _mm_setzero_ps();
 310             fjz1             = _mm_setzero_ps();
 311             fjx2             = _mm_setzero_ps();
 312             fjy2             = _mm_setzero_ps();
 313             fjz2             = _mm_setzero_ps();
 314             fjx3             = _mm_setzero_ps();
 315             fjy3             = _mm_setzero_ps();
 316             fjz3             = _mm_setzero_ps();
 317
 318             /**************************
 319              * CALCULATE INTERACTIONS *
 320              **************************/
 321
 322             if (gmx_mm_any_lt(rsq00,rcutoff2))
 323             {
 324
 325             r00              = _mm_mul_ps(rsq00,rinv00);
 326
 327             /* LENNARD-JONES DISPERSION/REPULSION */
 328
 329             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 330             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 331             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 332             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 333             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 334
 335             d                = _mm_sub_ps(r00,rswitch);
 336             d                = _mm_max_ps(d,_mm_setzero_ps());
 337             d2               = _mm_mul_ps(d,d);
 338             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 339
 340             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 341
 342             /* Evaluate switch function */
 343             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 344             fvdw             = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 345             vvdw             = _mm_mul_ps(vvdw,sw);
 346             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 347
 348             /* Update potential sum for this i atom from the interaction with this j atom. */
 349             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 350             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 351
 352             fscal            = fvdw;
 353
 354             fscal            = _mm_and_ps(fscal,cutoff_mask);
 355
 356             /* Calculate temporary vectorial force */
 357             tx               = _mm_mul_ps(fscal,dx00);
 358             ty               = _mm_mul_ps(fscal,dy00);
 359             tz               = _mm_mul_ps(fscal,dz00);
 360
 361             /* Update vectorial force */
 362             fix0             = _mm_add_ps(fix0,tx);
 363             fiy0             = _mm_add_ps(fiy0,ty);
 364             fiz0             = _mm_add_ps(fiz0,tz);
 365
 366             fjx0             = _mm_add_ps(fjx0,tx);
 367             fjy0             = _mm_add_ps(fjy0,ty);
 368             fjz0             = _mm_add_ps(fjz0,tz);
 369
 370             }
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             if (gmx_mm_any_lt(rsq11,rcutoff2))
 377             {
 378
 379             /* REACTION-FIELD ELECTROSTATICS */
 380             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 381             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 382
 383             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velec            = _mm_and_ps(velec,cutoff_mask);
 387             velecsum         = _mm_add_ps(velecsum,velec);
 388
 389             fscal            = felec;
 390
 391             fscal            = _mm_and_ps(fscal,cutoff_mask);
 392
 393             /* Calculate temporary vectorial force */
 394             tx               = _mm_mul_ps(fscal,dx11);
 395             ty               = _mm_mul_ps(fscal,dy11);
 396             tz               = _mm_mul_ps(fscal,dz11);
 397
 398             /* Update vectorial force */
 399             fix1             = _mm_add_ps(fix1,tx);
 400             fiy1             = _mm_add_ps(fiy1,ty);
 401             fiz1             = _mm_add_ps(fiz1,tz);
 402
 403             fjx1             = _mm_add_ps(fjx1,tx);
 404             fjy1             = _mm_add_ps(fjy1,ty);
 405             fjz1             = _mm_add_ps(fjz1,tz);
 406
 407             }
 408
 409             /**************************
 410              * CALCULATE INTERACTIONS *
 411              **************************/
 412
 413             if (gmx_mm_any_lt(rsq12,rcutoff2))
 414             {
 415
 416             /* REACTION-FIELD ELECTROSTATICS */
 417             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 418             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 419
 420             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 421
 422             /* Update potential sum for this i atom from the interaction with this j atom. */
 423             velec            = _mm_and_ps(velec,cutoff_mask);
 424             velecsum         = _mm_add_ps(velecsum,velec);
 425
 426             fscal            = felec;
 427
 428             fscal            = _mm_and_ps(fscal,cutoff_mask);
 429
 430             /* Calculate temporary vectorial force */
 431             tx               = _mm_mul_ps(fscal,dx12);
 432             ty               = _mm_mul_ps(fscal,dy12);
 433             tz               = _mm_mul_ps(fscal,dz12);
 434
 435             /* Update vectorial force */
 436             fix1             = _mm_add_ps(fix1,tx);
 437             fiy1             = _mm_add_ps(fiy1,ty);
 438             fiz1             = _mm_add_ps(fiz1,tz);
 439
 440             fjx2             = _mm_add_ps(fjx2,tx);
 441             fjy2             = _mm_add_ps(fjy2,ty);
 442             fjz2             = _mm_add_ps(fjz2,tz);
 443
 444             }
 445
 446             /**************************
 447              * CALCULATE INTERACTIONS *
 448              **************************/
 449
 450             if (gmx_mm_any_lt(rsq13,rcutoff2))
 451             {
 452
 453             /* REACTION-FIELD ELECTROSTATICS */
 454             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 455             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 456
 457             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 458
 459             /* Update potential sum for this i atom from the interaction with this j atom. */
 460             velec            = _mm_and_ps(velec,cutoff_mask);
 461             velecsum         = _mm_add_ps(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             fscal            = _mm_and_ps(fscal,cutoff_mask);
 466
 467             /* Calculate temporary vectorial force */
 468             tx               = _mm_mul_ps(fscal,dx13);
 469             ty               = _mm_mul_ps(fscal,dy13);
 470             tz               = _mm_mul_ps(fscal,dz13);
 471
 472             /* Update vectorial force */
 473             fix1             = _mm_add_ps(fix1,tx);
 474             fiy1             = _mm_add_ps(fiy1,ty);
 475             fiz1             = _mm_add_ps(fiz1,tz);
 476
 477             fjx3             = _mm_add_ps(fjx3,tx);
 478             fjy3             = _mm_add_ps(fjy3,ty);
 479             fjz3             = _mm_add_ps(fjz3,tz);
 480
 481             }
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             if (gmx_mm_any_lt(rsq21,rcutoff2))
 488             {
 489
 490             /* REACTION-FIELD ELECTROSTATICS */
 491             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 492             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 493
 494             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velec            = _mm_and_ps(velec,cutoff_mask);
 498             velecsum         = _mm_add_ps(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502             fscal            = _mm_and_ps(fscal,cutoff_mask);
 503
 504             /* Calculate temporary vectorial force */
 505             tx               = _mm_mul_ps(fscal,dx21);
 506             ty               = _mm_mul_ps(fscal,dy21);
 507             tz               = _mm_mul_ps(fscal,dz21);
 508
 509             /* Update vectorial force */
 510             fix2             = _mm_add_ps(fix2,tx);
 511             fiy2             = _mm_add_ps(fiy2,ty);
 512             fiz2             = _mm_add_ps(fiz2,tz);
 513
 514             fjx1             = _mm_add_ps(fjx1,tx);
 515             fjy1             = _mm_add_ps(fjy1,ty);
 516             fjz1             = _mm_add_ps(fjz1,tz);
 517
 518             }
 519
 520             /**************************
 521              * CALCULATE INTERACTIONS *
 522              **************************/
 523
 524             if (gmx_mm_any_lt(rsq22,rcutoff2))
 525             {
 526
 527             /* REACTION-FIELD ELECTROSTATICS */
 528             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 529             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 530
 531             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 532
 533             /* Update potential sum for this i atom from the interaction with this j atom. */
 534             velec            = _mm_and_ps(velec,cutoff_mask);
 535             velecsum         = _mm_add_ps(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539             fscal            = _mm_and_ps(fscal,cutoff_mask);
 540
 541             /* Calculate temporary vectorial force */
 542             tx               = _mm_mul_ps(fscal,dx22);
 543             ty               = _mm_mul_ps(fscal,dy22);
 544             tz               = _mm_mul_ps(fscal,dz22);
 545
 546             /* Update vectorial force */
 547             fix2             = _mm_add_ps(fix2,tx);
 548             fiy2             = _mm_add_ps(fiy2,ty);
 549             fiz2             = _mm_add_ps(fiz2,tz);
 550
 551             fjx2             = _mm_add_ps(fjx2,tx);
 552             fjy2             = _mm_add_ps(fjy2,ty);
 553             fjz2             = _mm_add_ps(fjz2,tz);
 554
 555             }
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             if (gmx_mm_any_lt(rsq23,rcutoff2))
 562             {
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 566             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 567
 568             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _mm_and_ps(velec,cutoff_mask);
 572             velecsum         = _mm_add_ps(velecsum,velec);
 573
 574             fscal            = felec;
 575
 576             fscal            = _mm_and_ps(fscal,cutoff_mask);
 577
 578             /* Calculate temporary vectorial force */
 579             tx               = _mm_mul_ps(fscal,dx23);
 580             ty               = _mm_mul_ps(fscal,dy23);
 581             tz               = _mm_mul_ps(fscal,dz23);
 582
 583             /* Update vectorial force */
 584             fix2             = _mm_add_ps(fix2,tx);
 585             fiy2             = _mm_add_ps(fiy2,ty);
 586             fiz2             = _mm_add_ps(fiz2,tz);
 587
 588             fjx3             = _mm_add_ps(fjx3,tx);
 589             fjy3             = _mm_add_ps(fjy3,ty);
 590             fjz3             = _mm_add_ps(fjz3,tz);
 591
 592             }
 593
 594             /**************************
 595              * CALCULATE INTERACTIONS *
 596              **************************/
 597
 598             if (gmx_mm_any_lt(rsq31,rcutoff2))
 599             {
 600
 601             /* REACTION-FIELD ELECTROSTATICS */
 602             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 603             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 604
 605             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 606
 607             /* Update potential sum for this i atom from the interaction with this j atom. */
 608             velec            = _mm_and_ps(velec,cutoff_mask);
 609             velecsum         = _mm_add_ps(velecsum,velec);
 610
 611             fscal            = felec;
 612
 613             fscal            = _mm_and_ps(fscal,cutoff_mask);
 614
 615             /* Calculate temporary vectorial force */
 616             tx               = _mm_mul_ps(fscal,dx31);
 617             ty               = _mm_mul_ps(fscal,dy31);
 618             tz               = _mm_mul_ps(fscal,dz31);
 619
 620             /* Update vectorial force */
 621             fix3             = _mm_add_ps(fix3,tx);
 622             fiy3             = _mm_add_ps(fiy3,ty);
 623             fiz3             = _mm_add_ps(fiz3,tz);
 624
 625             fjx1             = _mm_add_ps(fjx1,tx);
 626             fjy1             = _mm_add_ps(fjy1,ty);
 627             fjz1             = _mm_add_ps(fjz1,tz);
 628
 629             }
 630
 631             /**************************
 632              * CALCULATE INTERACTIONS *
 633              **************************/
 634
 635             if (gmx_mm_any_lt(rsq32,rcutoff2))
 636             {
 637
 638             /* REACTION-FIELD ELECTROSTATICS */
 639             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 640             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 641
 642             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 643
 644             /* Update potential sum for this i atom from the interaction with this j atom. */
 645             velec            = _mm_and_ps(velec,cutoff_mask);
 646             velecsum         = _mm_add_ps(velecsum,velec);
 647
 648             fscal            = felec;
 649
 650             fscal            = _mm_and_ps(fscal,cutoff_mask);
 651
 652             /* Calculate temporary vectorial force */
 653             tx               = _mm_mul_ps(fscal,dx32);
 654             ty               = _mm_mul_ps(fscal,dy32);
 655             tz               = _mm_mul_ps(fscal,dz32);
 656
 657             /* Update vectorial force */
 658             fix3             = _mm_add_ps(fix3,tx);
 659             fiy3             = _mm_add_ps(fiy3,ty);
 660             fiz3             = _mm_add_ps(fiz3,tz);
 661
 662             fjx2             = _mm_add_ps(fjx2,tx);
 663             fjy2             = _mm_add_ps(fjy2,ty);
 664             fjz2             = _mm_add_ps(fjz2,tz);
 665
 666             }
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             if (gmx_mm_any_lt(rsq33,rcutoff2))
 673             {
 674
 675             /* REACTION-FIELD ELECTROSTATICS */
 676             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 677             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 678
 679             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 680
 681             /* Update potential sum for this i atom from the interaction with this j atom. */
 682             velec            = _mm_and_ps(velec,cutoff_mask);
 683             velecsum         = _mm_add_ps(velecsum,velec);
 684
 685             fscal            = felec;
 686
 687             fscal            = _mm_and_ps(fscal,cutoff_mask);
 688
 689             /* Calculate temporary vectorial force */
 690             tx               = _mm_mul_ps(fscal,dx33);
 691             ty               = _mm_mul_ps(fscal,dy33);
 692             tz               = _mm_mul_ps(fscal,dz33);
 693
 694             /* Update vectorial force */
 695             fix3             = _mm_add_ps(fix3,tx);
 696             fiy3             = _mm_add_ps(fiy3,ty);
 697             fiz3             = _mm_add_ps(fiz3,tz);
 698
 699             fjx3             = _mm_add_ps(fjx3,tx);
 700             fjy3             = _mm_add_ps(fjy3,ty);
 701             fjz3             = _mm_add_ps(fjz3,tz);
 702
 703             }
 704
 705             fjptrA             = f+j_coord_offsetA;
 706             fjptrB             = f+j_coord_offsetB;
 707             fjptrC             = f+j_coord_offsetC;
 708             fjptrD             = f+j_coord_offsetD;
 709
 710             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 711                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 712                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 713
 714             /* Inner loop uses 386 flops */
 715         }
 716
 717         if(jidx<j_index_end)
 718         {
 719
 720             /* Get j neighbor index, and coordinate index */
 721             jnrlistA         = jjnr[jidx];
 722             jnrlistB         = jjnr[jidx+1];
 723             jnrlistC         = jjnr[jidx+2];
 724             jnrlistD         = jjnr[jidx+3];
 725             /* Sign of each element will be negative for non-real atoms.
 726              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 727              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 728              */
 729             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 730             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 731             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 732             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 733             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 734             j_coord_offsetA  = DIM*jnrA;
 735             j_coord_offsetB  = DIM*jnrB;
 736             j_coord_offsetC  = DIM*jnrC;
 737             j_coord_offsetD  = DIM*jnrD;
 738
 739             /* load j atom coordinates */
 740             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 741                                               x+j_coord_offsetC,x+j_coord_offsetD,
 742                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 743                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 744
 745             /* Calculate displacement vector */
 746             dx00             = _mm_sub_ps(ix0,jx0);
 747             dy00             = _mm_sub_ps(iy0,jy0);
 748             dz00             = _mm_sub_ps(iz0,jz0);
 749             dx11             = _mm_sub_ps(ix1,jx1);
 750             dy11             = _mm_sub_ps(iy1,jy1);
 751             dz11             = _mm_sub_ps(iz1,jz1);
 752             dx12             = _mm_sub_ps(ix1,jx2);
 753             dy12             = _mm_sub_ps(iy1,jy2);
 754             dz12             = _mm_sub_ps(iz1,jz2);
 755             dx13             = _mm_sub_ps(ix1,jx3);
 756             dy13             = _mm_sub_ps(iy1,jy3);
 757             dz13             = _mm_sub_ps(iz1,jz3);
 758             dx21             = _mm_sub_ps(ix2,jx1);
 759             dy21             = _mm_sub_ps(iy2,jy1);
 760             dz21             = _mm_sub_ps(iz2,jz1);
 761             dx22             = _mm_sub_ps(ix2,jx2);
 762             dy22             = _mm_sub_ps(iy2,jy2);
 763             dz22             = _mm_sub_ps(iz2,jz2);
 764             dx23             = _mm_sub_ps(ix2,jx3);
 765             dy23             = _mm_sub_ps(iy2,jy3);
 766             dz23             = _mm_sub_ps(iz2,jz3);
 767             dx31             = _mm_sub_ps(ix3,jx1);
 768             dy31             = _mm_sub_ps(iy3,jy1);
 769             dz31             = _mm_sub_ps(iz3,jz1);
 770             dx32             = _mm_sub_ps(ix3,jx2);
 771             dy32             = _mm_sub_ps(iy3,jy2);
 772             dz32             = _mm_sub_ps(iz3,jz2);
 773             dx33             = _mm_sub_ps(ix3,jx3);
 774             dy33             = _mm_sub_ps(iy3,jy3);
 775             dz33             = _mm_sub_ps(iz3,jz3);
 776
 777             /* Calculate squared distance and things based on it */
 778             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 779             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 780             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 781             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 782             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 783             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 784             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 785             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 786             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 787             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 788
 789             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 790             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 791             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 792             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 793             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 794             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 795             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 796             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 797             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 798             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 799
 800             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 801             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 802             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 803             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 804             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 805             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 806             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 807             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 808             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 809             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 810
 811             fjx0             = _mm_setzero_ps();
 812             fjy0             = _mm_setzero_ps();
 813             fjz0             = _mm_setzero_ps();
 814             fjx1             = _mm_setzero_ps();
 815             fjy1             = _mm_setzero_ps();
 816             fjz1             = _mm_setzero_ps();
 817             fjx2             = _mm_setzero_ps();
 818             fjy2             = _mm_setzero_ps();
 819             fjz2             = _mm_setzero_ps();
 820             fjx3             = _mm_setzero_ps();
 821             fjy3             = _mm_setzero_ps();
 822             fjz3             = _mm_setzero_ps();
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (gmx_mm_any_lt(rsq00,rcutoff2))
 829             {
 830
 831             r00              = _mm_mul_ps(rsq00,rinv00);
 832             r00              = _mm_andnot_ps(dummy_mask,r00);
 833
 834             /* LENNARD-JONES DISPERSION/REPULSION */
 835
 836             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 837             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 838             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 839             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 840             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 841
 842             d                = _mm_sub_ps(r00,rswitch);
 843             d                = _mm_max_ps(d,_mm_setzero_ps());
 844             d2               = _mm_mul_ps(d,d);
 845             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 846
 847             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 848
 849             /* Evaluate switch function */
 850             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 851             fvdw             = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 852             vvdw             = _mm_mul_ps(vvdw,sw);
 853             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 854
 855             /* Update potential sum for this i atom from the interaction with this j atom. */
 856             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 857             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 858             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 859
 860             fscal            = fvdw;
 861
 862             fscal            = _mm_and_ps(fscal,cutoff_mask);
 863
 864             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 865
 866             /* Calculate temporary vectorial force */
 867             tx               = _mm_mul_ps(fscal,dx00);
 868             ty               = _mm_mul_ps(fscal,dy00);
 869             tz               = _mm_mul_ps(fscal,dz00);
 870
 871             /* Update vectorial force */
 872             fix0             = _mm_add_ps(fix0,tx);
 873             fiy0             = _mm_add_ps(fiy0,ty);
 874             fiz0             = _mm_add_ps(fiz0,tz);
 875
 876             fjx0             = _mm_add_ps(fjx0,tx);
 877             fjy0             = _mm_add_ps(fjy0,ty);
 878             fjz0             = _mm_add_ps(fjz0,tz);
 879
 880             }
 881
 882             /**************************
 883              * CALCULATE INTERACTIONS *
 884              **************************/
 885
 886             if (gmx_mm_any_lt(rsq11,rcutoff2))
 887             {
 888
 889             /* REACTION-FIELD ELECTROSTATICS */
 890             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 891             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 892
 893             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 894
 895             /* Update potential sum for this i atom from the interaction with this j atom. */
 896             velec            = _mm_and_ps(velec,cutoff_mask);
 897             velec            = _mm_andnot_ps(dummy_mask,velec);
 898             velecsum         = _mm_add_ps(velecsum,velec);
 899
 900             fscal            = felec;
 901
 902             fscal            = _mm_and_ps(fscal,cutoff_mask);
 903
 904             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 905
 906             /* Calculate temporary vectorial force */
 907             tx               = _mm_mul_ps(fscal,dx11);
 908             ty               = _mm_mul_ps(fscal,dy11);
 909             tz               = _mm_mul_ps(fscal,dz11);
 910
 911             /* Update vectorial force */
 912             fix1             = _mm_add_ps(fix1,tx);
 913             fiy1             = _mm_add_ps(fiy1,ty);
 914             fiz1             = _mm_add_ps(fiz1,tz);
 915
 916             fjx1             = _mm_add_ps(fjx1,tx);
 917             fjy1             = _mm_add_ps(fjy1,ty);
 918             fjz1             = _mm_add_ps(fjz1,tz);
 919
 920             }
 921
 922             /**************************
 923              * CALCULATE INTERACTIONS *
 924              **************************/
 925
 926             if (gmx_mm_any_lt(rsq12,rcutoff2))
 927             {
 928
 929             /* REACTION-FIELD ELECTROSTATICS */
 930             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 931             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 932
 933             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 934
 935             /* Update potential sum for this i atom from the interaction with this j atom. */
 936             velec            = _mm_and_ps(velec,cutoff_mask);
 937             velec            = _mm_andnot_ps(dummy_mask,velec);
 938             velecsum         = _mm_add_ps(velecsum,velec);
 939
 940             fscal            = felec;
 941
 942             fscal            = _mm_and_ps(fscal,cutoff_mask);
 943
 944             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 945
 946             /* Calculate temporary vectorial force */
 947             tx               = _mm_mul_ps(fscal,dx12);
 948             ty               = _mm_mul_ps(fscal,dy12);
 949             tz               = _mm_mul_ps(fscal,dz12);
 950
 951             /* Update vectorial force */
 952             fix1             = _mm_add_ps(fix1,tx);
 953             fiy1             = _mm_add_ps(fiy1,ty);
 954             fiz1             = _mm_add_ps(fiz1,tz);
 955
 956             fjx2             = _mm_add_ps(fjx2,tx);
 957             fjy2             = _mm_add_ps(fjy2,ty);
 958             fjz2             = _mm_add_ps(fjz2,tz);
 959
 960             }
 961
 962             /**************************
 963              * CALCULATE INTERACTIONS *
 964              **************************/
 965
 966             if (gmx_mm_any_lt(rsq13,rcutoff2))
 967             {
 968
 969             /* REACTION-FIELD ELECTROSTATICS */
 970             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 971             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 972
 973             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 974
 975             /* Update potential sum for this i atom from the interaction with this j atom. */
 976             velec            = _mm_and_ps(velec,cutoff_mask);
 977             velec            = _mm_andnot_ps(dummy_mask,velec);
 978             velecsum         = _mm_add_ps(velecsum,velec);
 979
 980             fscal            = felec;
 981
 982             fscal            = _mm_and_ps(fscal,cutoff_mask);
 983
 984             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 985
 986             /* Calculate temporary vectorial force */
 987             tx               = _mm_mul_ps(fscal,dx13);
 988             ty               = _mm_mul_ps(fscal,dy13);
 989             tz               = _mm_mul_ps(fscal,dz13);
 990
 991             /* Update vectorial force */
 992             fix1             = _mm_add_ps(fix1,tx);
 993             fiy1             = _mm_add_ps(fiy1,ty);
 994             fiz1             = _mm_add_ps(fiz1,tz);
 995
 996             fjx3             = _mm_add_ps(fjx3,tx);
 997             fjy3             = _mm_add_ps(fjy3,ty);
 998             fjz3             = _mm_add_ps(fjz3,tz);
 999
1000             }
1001
1002             /**************************
1003              * CALCULATE INTERACTIONS *
1004              **************************/
1005
1006             if (gmx_mm_any_lt(rsq21,rcutoff2))
1007             {
1008
1009             /* REACTION-FIELD ELECTROSTATICS */
1010             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
1011             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1012
1013             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1014
1015             /* Update potential sum for this i atom from the interaction with this j atom. */
1016             velec            = _mm_and_ps(velec,cutoff_mask);
1017             velec            = _mm_andnot_ps(dummy_mask,velec);
1018             velecsum         = _mm_add_ps(velecsum,velec);
1019
1020             fscal            = felec;
1021
1022             fscal            = _mm_and_ps(fscal,cutoff_mask);
1023
1024             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1025
1026             /* Calculate temporary vectorial force */
1027             tx               = _mm_mul_ps(fscal,dx21);
1028             ty               = _mm_mul_ps(fscal,dy21);
1029             tz               = _mm_mul_ps(fscal,dz21);
1030
1031             /* Update vectorial force */
1032             fix2             = _mm_add_ps(fix2,tx);
1033             fiy2             = _mm_add_ps(fiy2,ty);
1034             fiz2             = _mm_add_ps(fiz2,tz);
1035
1036             fjx1             = _mm_add_ps(fjx1,tx);
1037             fjy1             = _mm_add_ps(fjy1,ty);
1038             fjz1             = _mm_add_ps(fjz1,tz);
1039
1040             }
1041
1042             /**************************
1043              * CALCULATE INTERACTIONS *
1044              **************************/
1045
1046             if (gmx_mm_any_lt(rsq22,rcutoff2))
1047             {
1048
1049             /* REACTION-FIELD ELECTROSTATICS */
1050             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1051             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1052
1053             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1054
1055             /* Update potential sum for this i atom from the interaction with this j atom. */
1056             velec            = _mm_and_ps(velec,cutoff_mask);
1057             velec            = _mm_andnot_ps(dummy_mask,velec);
1058             velecsum         = _mm_add_ps(velecsum,velec);
1059
1060             fscal            = felec;
1061
1062             fscal            = _mm_and_ps(fscal,cutoff_mask);
1063
1064             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1065
1066             /* Calculate temporary vectorial force */
1067             tx               = _mm_mul_ps(fscal,dx22);
1068             ty               = _mm_mul_ps(fscal,dy22);
1069             tz               = _mm_mul_ps(fscal,dz22);
1070
1071             /* Update vectorial force */
1072             fix2             = _mm_add_ps(fix2,tx);
1073             fiy2             = _mm_add_ps(fiy2,ty);
1074             fiz2             = _mm_add_ps(fiz2,tz);
1075
1076             fjx2             = _mm_add_ps(fjx2,tx);
1077             fjy2             = _mm_add_ps(fjy2,ty);
1078             fjz2             = _mm_add_ps(fjz2,tz);
1079
1080             }
1081
1082             /**************************
1083              * CALCULATE INTERACTIONS *
1084              **************************/
1085
1086             if (gmx_mm_any_lt(rsq23,rcutoff2))
1087             {
1088
1089             /* REACTION-FIELD ELECTROSTATICS */
1090             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
1091             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1092
1093             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1094
1095             /* Update potential sum for this i atom from the interaction with this j atom. */
1096             velec            = _mm_and_ps(velec,cutoff_mask);
1097             velec            = _mm_andnot_ps(dummy_mask,velec);
1098             velecsum         = _mm_add_ps(velecsum,velec);
1099
1100             fscal            = felec;
1101
1102             fscal            = _mm_and_ps(fscal,cutoff_mask);
1103
1104             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1105
1106             /* Calculate temporary vectorial force */
1107             tx               = _mm_mul_ps(fscal,dx23);
1108             ty               = _mm_mul_ps(fscal,dy23);
1109             tz               = _mm_mul_ps(fscal,dz23);
1110
1111             /* Update vectorial force */
1112             fix2             = _mm_add_ps(fix2,tx);
1113             fiy2             = _mm_add_ps(fiy2,ty);
1114             fiz2             = _mm_add_ps(fiz2,tz);
1115
1116             fjx3             = _mm_add_ps(fjx3,tx);
1117             fjy3             = _mm_add_ps(fjy3,ty);
1118             fjz3             = _mm_add_ps(fjz3,tz);
1119
1120             }
1121
1122             /**************************
1123              * CALCULATE INTERACTIONS *
1124              **************************/
1125
1126             if (gmx_mm_any_lt(rsq31,rcutoff2))
1127             {
1128
1129             /* REACTION-FIELD ELECTROSTATICS */
1130             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
1131             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1132
1133             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1134
1135             /* Update potential sum for this i atom from the interaction with this j atom. */
1136             velec            = _mm_and_ps(velec,cutoff_mask);
1137             velec            = _mm_andnot_ps(dummy_mask,velec);
1138             velecsum         = _mm_add_ps(velecsum,velec);
1139
1140             fscal            = felec;
1141
1142             fscal            = _mm_and_ps(fscal,cutoff_mask);
1143
1144             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1145
1146             /* Calculate temporary vectorial force */
1147             tx               = _mm_mul_ps(fscal,dx31);
1148             ty               = _mm_mul_ps(fscal,dy31);
1149             tz               = _mm_mul_ps(fscal,dz31);
1150
1151             /* Update vectorial force */
1152             fix3             = _mm_add_ps(fix3,tx);
1153             fiy3             = _mm_add_ps(fiy3,ty);
1154             fiz3             = _mm_add_ps(fiz3,tz);
1155
1156             fjx1             = _mm_add_ps(fjx1,tx);
1157             fjy1             = _mm_add_ps(fjy1,ty);
1158             fjz1             = _mm_add_ps(fjz1,tz);
1159
1160             }
1161
1162             /**************************
1163              * CALCULATE INTERACTIONS *
1164              **************************/
1165
1166             if (gmx_mm_any_lt(rsq32,rcutoff2))
1167             {
1168
1169             /* REACTION-FIELD ELECTROSTATICS */
1170             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1171             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1172
1173             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1174
1175             /* Update potential sum for this i atom from the interaction with this j atom. */
1176             velec            = _mm_and_ps(velec,cutoff_mask);
1177             velec            = _mm_andnot_ps(dummy_mask,velec);
1178             velecsum         = _mm_add_ps(velecsum,velec);
1179
1180             fscal            = felec;
1181
1182             fscal            = _mm_and_ps(fscal,cutoff_mask);
1183
1184             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1185
1186             /* Calculate temporary vectorial force */
1187             tx               = _mm_mul_ps(fscal,dx32);
1188             ty               = _mm_mul_ps(fscal,dy32);
1189             tz               = _mm_mul_ps(fscal,dz32);
1190
1191             /* Update vectorial force */
1192             fix3             = _mm_add_ps(fix3,tx);
1193             fiy3             = _mm_add_ps(fiy3,ty);
1194             fiz3             = _mm_add_ps(fiz3,tz);
1195
1196             fjx2             = _mm_add_ps(fjx2,tx);
1197             fjy2             = _mm_add_ps(fjy2,ty);
1198             fjz2             = _mm_add_ps(fjz2,tz);
1199
1200             }
1201
1202             /**************************
1203              * CALCULATE INTERACTIONS *
1204              **************************/
1205
1206             if (gmx_mm_any_lt(rsq33,rcutoff2))
1207             {
1208
1209             /* REACTION-FIELD ELECTROSTATICS */
1210             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1211             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1212
1213             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1214
1215             /* Update potential sum for this i atom from the interaction with this j atom. */
1216             velec            = _mm_and_ps(velec,cutoff_mask);
1217             velec            = _mm_andnot_ps(dummy_mask,velec);
1218             velecsum         = _mm_add_ps(velecsum,velec);
1219
1220             fscal            = felec;
1221
1222             fscal            = _mm_and_ps(fscal,cutoff_mask);
1223
1224             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1225
1226             /* Calculate temporary vectorial force */
1227             tx               = _mm_mul_ps(fscal,dx33);
1228             ty               = _mm_mul_ps(fscal,dy33);
1229             tz               = _mm_mul_ps(fscal,dz33);
1230
1231             /* Update vectorial force */
1232             fix3             = _mm_add_ps(fix3,tx);
1233             fiy3             = _mm_add_ps(fiy3,ty);
1234             fiz3             = _mm_add_ps(fiz3,tz);
1235
1236             fjx3             = _mm_add_ps(fjx3,tx);
1237             fjy3             = _mm_add_ps(fjy3,ty);
1238             fjz3             = _mm_add_ps(fjz3,tz);
1239
1240             }
1241
1242             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1243             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1244             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1245             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1246
1247             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1248                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1249                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1250
1251             /* Inner loop uses 387 flops */
1252         }
1253
1254         /* End of innermost loop */
1255
1256         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1257                                               f+i_coord_offset,fshift+i_shift_offset);
1258
1259         ggid                        = gid[iidx];
1260         /* Update potential energies */
1261         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1262         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1263
1264         /* Increment number of inner iterations */
1265         inneriter                  += j_index_end - j_index_start;
1266
1267         /* Outer loop uses 26 flops */
1268     }
1269
1270     /* Increment number of outer iterations */
1271     outeriter        += nri;
1272
1273     /* Update outer/inner flops */
1274
1275     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*387);
1276 }
1277 /*
1278  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sse4_1_single
1279  * Electrostatics interaction: ReactionField
1280  * VdW interaction:            LennardJones
1281  * Geometry:                   Water4-Water4
1282  * Calculate force/pot:        Force
1283  */
1284 void
1285 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sse4_1_single
1286                     (t_nblist * gmx_restrict                nlist,
1287                      rvec * gmx_restrict                    xx,
1288                      rvec * gmx_restrict                    ff,
1289                      t_forcerec * gmx_restrict              fr,
1290                      t_mdatoms * gmx_restrict               mdatoms,
1291                      nb_kernel_data_t * gmx_restrict        kernel_data,
1292                      t_nrnb * gmx_restrict                  nrnb)
1293 {
1294     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1295      * just 0 for non-waters.
1296      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1297      * jnr indices corresponding to data put in the four positions in the SIMD register.
1298      */
1299     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1300     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1301     int              jnrA,jnrB,jnrC,jnrD;
1302     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1303     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1304     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1305     real             rcutoff_scalar;
1306     real             *shiftvec,*fshift,*x,*f;
1307     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1308     real             scratch[4*DIM];
1309     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1310     int              vdwioffset0;
1311     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1312     int              vdwioffset1;
1313     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1314     int              vdwioffset2;
1315     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1316     int              vdwioffset3;
1317     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1318     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1319     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1320     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1321     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1322     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1323     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1324     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1325     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1326     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1327     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1328     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1329     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1330     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1331     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1332     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1333     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1334     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1335     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1336     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1337     real             *charge;
1338     int              nvdwtype;
1339     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1340     int              *vdwtype;
1341     real             *vdwparam;
1342     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1343     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1344     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1345     real             rswitch_scalar,d_scalar;
1346     __m128           dummy_mask,cutoff_mask;
1347     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1348     __m128           one     = _mm_set1_ps(1.0);
1349     __m128           two     = _mm_set1_ps(2.0);
1350     x                = xx[0];
1351     f                = ff[0];
1352
1353     nri              = nlist->nri;
1354     iinr             = nlist->iinr;
1355     jindex           = nlist->jindex;
1356     jjnr             = nlist->jjnr;
1357     shiftidx         = nlist->shift;
1358     gid              = nlist->gid;
1359     shiftvec         = fr->shift_vec[0];
1360     fshift           = fr->fshift[0];
1361     facel            = _mm_set1_ps(fr->epsfac);
1362     charge           = mdatoms->chargeA;
1363     krf              = _mm_set1_ps(fr->ic->k_rf);
1364     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1365     crf              = _mm_set1_ps(fr->ic->c_rf);
1366     nvdwtype         = fr->ntype;
1367     vdwparam         = fr->nbfp;
1368     vdwtype          = mdatoms->typeA;
1369
1370     /* Setup water-specific parameters */
1371     inr              = nlist->iinr[0];
1372     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1373     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1374     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1375     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1376
1377     jq1              = _mm_set1_ps(charge[inr+1]);
1378     jq2              = _mm_set1_ps(charge[inr+2]);
1379     jq3              = _mm_set1_ps(charge[inr+3]);
1380     vdwjidx0A        = 2*vdwtype[inr+0];
1381     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1382     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1383     qq11             = _mm_mul_ps(iq1,jq1);
1384     qq12             = _mm_mul_ps(iq1,jq2);
1385     qq13             = _mm_mul_ps(iq1,jq3);
1386     qq21             = _mm_mul_ps(iq2,jq1);
1387     qq22             = _mm_mul_ps(iq2,jq2);
1388     qq23             = _mm_mul_ps(iq2,jq3);
1389     qq31             = _mm_mul_ps(iq3,jq1);
1390     qq32             = _mm_mul_ps(iq3,jq2);
1391     qq33             = _mm_mul_ps(iq3,jq3);
1392
1393     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1394     rcutoff_scalar   = fr->rcoulomb;
1395     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1396     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1397
1398     rswitch_scalar   = fr->rvdw_switch;
1399     rswitch          = _mm_set1_ps(rswitch_scalar);
1400     /* Setup switch parameters */
1401     d_scalar         = rcutoff_scalar-rswitch_scalar;
1402     d                = _mm_set1_ps(d_scalar);
1403     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
1404     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1405     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1406     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
1407     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1408     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1409
1410     /* Avoid stupid compiler warnings */
1411     jnrA = jnrB = jnrC = jnrD = 0;
1412     j_coord_offsetA = 0;
1413     j_coord_offsetB = 0;
1414     j_coord_offsetC = 0;
1415     j_coord_offsetD = 0;
1416
1417     outeriter        = 0;
1418     inneriter        = 0;
1419
1420     for(iidx=0;iidx<4*DIM;iidx++)
1421     {
1422         scratch[iidx] = 0.0;
1423     }
1424
1425     /* Start outer loop over neighborlists */
1426     for(iidx=0; iidx<nri; iidx++)
1427     {
1428         /* Load shift vector for this list */
1429         i_shift_offset   = DIM*shiftidx[iidx];
1430
1431         /* Load limits for loop over neighbors */
1432         j_index_start    = jindex[iidx];
1433         j_index_end      = jindex[iidx+1];
1434
1435         /* Get outer coordinate index */
1436         inr              = iinr[iidx];
1437         i_coord_offset   = DIM*inr;
1438
1439         /* Load i particle coords and add shift vector */
1440         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1441                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1442
1443         fix0             = _mm_setzero_ps();
1444         fiy0             = _mm_setzero_ps();
1445         fiz0             = _mm_setzero_ps();
1446         fix1             = _mm_setzero_ps();
1447         fiy1             = _mm_setzero_ps();
1448         fiz1             = _mm_setzero_ps();
1449         fix2             = _mm_setzero_ps();
1450         fiy2             = _mm_setzero_ps();
1451         fiz2             = _mm_setzero_ps();
1452         fix3             = _mm_setzero_ps();
1453         fiy3             = _mm_setzero_ps();
1454         fiz3             = _mm_setzero_ps();
1455
1456         /* Start inner kernel loop */
1457         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1458         {
1459
1460             /* Get j neighbor index, and coordinate index */
1461             jnrA             = jjnr[jidx];
1462             jnrB             = jjnr[jidx+1];
1463             jnrC             = jjnr[jidx+2];
1464             jnrD             = jjnr[jidx+3];
1465             j_coord_offsetA  = DIM*jnrA;
1466             j_coord_offsetB  = DIM*jnrB;
1467             j_coord_offsetC  = DIM*jnrC;
1468             j_coord_offsetD  = DIM*jnrD;
1469
1470             /* load j atom coordinates */
1471             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1472                                               x+j_coord_offsetC,x+j_coord_offsetD,
1473                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1474                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1475
1476             /* Calculate displacement vector */
1477             dx00             = _mm_sub_ps(ix0,jx0);
1478             dy00             = _mm_sub_ps(iy0,jy0);
1479             dz00             = _mm_sub_ps(iz0,jz0);
1480             dx11             = _mm_sub_ps(ix1,jx1);
1481             dy11             = _mm_sub_ps(iy1,jy1);
1482             dz11             = _mm_sub_ps(iz1,jz1);
1483             dx12             = _mm_sub_ps(ix1,jx2);
1484             dy12             = _mm_sub_ps(iy1,jy2);
1485             dz12             = _mm_sub_ps(iz1,jz2);
1486             dx13             = _mm_sub_ps(ix1,jx3);
1487             dy13             = _mm_sub_ps(iy1,jy3);
1488             dz13             = _mm_sub_ps(iz1,jz3);
1489             dx21             = _mm_sub_ps(ix2,jx1);
1490             dy21             = _mm_sub_ps(iy2,jy1);
1491             dz21             = _mm_sub_ps(iz2,jz1);
1492             dx22             = _mm_sub_ps(ix2,jx2);
1493             dy22             = _mm_sub_ps(iy2,jy2);
1494             dz22             = _mm_sub_ps(iz2,jz2);
1495             dx23             = _mm_sub_ps(ix2,jx3);
1496             dy23             = _mm_sub_ps(iy2,jy3);
1497             dz23             = _mm_sub_ps(iz2,jz3);
1498             dx31             = _mm_sub_ps(ix3,jx1);
1499             dy31             = _mm_sub_ps(iy3,jy1);
1500             dz31             = _mm_sub_ps(iz3,jz1);
1501             dx32             = _mm_sub_ps(ix3,jx2);
1502             dy32             = _mm_sub_ps(iy3,jy2);
1503             dz32             = _mm_sub_ps(iz3,jz2);
1504             dx33             = _mm_sub_ps(ix3,jx3);
1505             dy33             = _mm_sub_ps(iy3,jy3);
1506             dz33             = _mm_sub_ps(iz3,jz3);
1507
1508             /* Calculate squared distance and things based on it */
1509             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1510             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1511             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1512             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1513             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1514             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1515             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1516             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1517             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1518             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1519
1520             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1521             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1522             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1523             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1524             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1525             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1526             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1527             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1528             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1529             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1530
1531             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1532             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1533             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1534             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1535             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1536             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1537             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1538             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1539             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1540             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1541
1542             fjx0             = _mm_setzero_ps();
1543             fjy0             = _mm_setzero_ps();
1544             fjz0             = _mm_setzero_ps();
1545             fjx1             = _mm_setzero_ps();
1546             fjy1             = _mm_setzero_ps();
1547             fjz1             = _mm_setzero_ps();
1548             fjx2             = _mm_setzero_ps();
1549             fjy2             = _mm_setzero_ps();
1550             fjz2             = _mm_setzero_ps();
1551             fjx3             = _mm_setzero_ps();
1552             fjy3             = _mm_setzero_ps();
1553             fjz3             = _mm_setzero_ps();
1554
1555             /**************************
1556              * CALCULATE INTERACTIONS *
1557              **************************/
1558
1559             if (gmx_mm_any_lt(rsq00,rcutoff2))
1560             {
1561
1562             r00              = _mm_mul_ps(rsq00,rinv00);
1563
1564             /* LENNARD-JONES DISPERSION/REPULSION */
1565
1566             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1567             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1568             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1569             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
1570             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1571
1572             d                = _mm_sub_ps(r00,rswitch);
1573             d                = _mm_max_ps(d,_mm_setzero_ps());
1574             d2               = _mm_mul_ps(d,d);
1575             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1576
1577             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1578
1579             /* Evaluate switch function */
1580             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1581             fvdw             = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1582             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1583
1584             fscal            = fvdw;
1585
1586             fscal            = _mm_and_ps(fscal,cutoff_mask);
1587
1588             /* Calculate temporary vectorial force */
1589             tx               = _mm_mul_ps(fscal,dx00);
1590             ty               = _mm_mul_ps(fscal,dy00);
1591             tz               = _mm_mul_ps(fscal,dz00);
1592
1593             /* Update vectorial force */
1594             fix0             = _mm_add_ps(fix0,tx);
1595             fiy0             = _mm_add_ps(fiy0,ty);
1596             fiz0             = _mm_add_ps(fiz0,tz);
1597
1598             fjx0             = _mm_add_ps(fjx0,tx);
1599             fjy0             = _mm_add_ps(fjy0,ty);
1600             fjz0             = _mm_add_ps(fjz0,tz);
1601
1602             }
1603
1604             /**************************
1605              * CALCULATE INTERACTIONS *
1606              **************************/
1607
1608             if (gmx_mm_any_lt(rsq11,rcutoff2))
1609             {
1610
1611             /* REACTION-FIELD ELECTROSTATICS */
1612             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1613
1614             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1615
1616             fscal            = felec;
1617
1618             fscal            = _mm_and_ps(fscal,cutoff_mask);
1619
1620             /* Calculate temporary vectorial force */
1621             tx               = _mm_mul_ps(fscal,dx11);
1622             ty               = _mm_mul_ps(fscal,dy11);
1623             tz               = _mm_mul_ps(fscal,dz11);
1624
1625             /* Update vectorial force */
1626             fix1             = _mm_add_ps(fix1,tx);
1627             fiy1             = _mm_add_ps(fiy1,ty);
1628             fiz1             = _mm_add_ps(fiz1,tz);
1629
1630             fjx1             = _mm_add_ps(fjx1,tx);
1631             fjy1             = _mm_add_ps(fjy1,ty);
1632             fjz1             = _mm_add_ps(fjz1,tz);
1633
1634             }
1635
1636             /**************************
1637              * CALCULATE INTERACTIONS *
1638              **************************/
1639
1640             if (gmx_mm_any_lt(rsq12,rcutoff2))
1641             {
1642
1643             /* REACTION-FIELD ELECTROSTATICS */
1644             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1645
1646             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1647
1648             fscal            = felec;
1649
1650             fscal            = _mm_and_ps(fscal,cutoff_mask);
1651
1652             /* Calculate temporary vectorial force */
1653             tx               = _mm_mul_ps(fscal,dx12);
1654             ty               = _mm_mul_ps(fscal,dy12);
1655             tz               = _mm_mul_ps(fscal,dz12);
1656
1657             /* Update vectorial force */
1658             fix1             = _mm_add_ps(fix1,tx);
1659             fiy1             = _mm_add_ps(fiy1,ty);
1660             fiz1             = _mm_add_ps(fiz1,tz);
1661
1662             fjx2             = _mm_add_ps(fjx2,tx);
1663             fjy2             = _mm_add_ps(fjy2,ty);
1664             fjz2             = _mm_add_ps(fjz2,tz);
1665
1666             }
1667
1668             /**************************
1669              * CALCULATE INTERACTIONS *
1670              **************************/
1671
1672             if (gmx_mm_any_lt(rsq13,rcutoff2))
1673             {
1674
1675             /* REACTION-FIELD ELECTROSTATICS */
1676             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1677
1678             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1679
1680             fscal            = felec;
1681
1682             fscal            = _mm_and_ps(fscal,cutoff_mask);
1683
1684             /* Calculate temporary vectorial force */
1685             tx               = _mm_mul_ps(fscal,dx13);
1686             ty               = _mm_mul_ps(fscal,dy13);
1687             tz               = _mm_mul_ps(fscal,dz13);
1688
1689             /* Update vectorial force */
1690             fix1             = _mm_add_ps(fix1,tx);
1691             fiy1             = _mm_add_ps(fiy1,ty);
1692             fiz1             = _mm_add_ps(fiz1,tz);
1693
1694             fjx3             = _mm_add_ps(fjx3,tx);
1695             fjy3             = _mm_add_ps(fjy3,ty);
1696             fjz3             = _mm_add_ps(fjz3,tz);
1697
1698             }
1699
1700             /**************************
1701              * CALCULATE INTERACTIONS *
1702              **************************/
1703
1704             if (gmx_mm_any_lt(rsq21,rcutoff2))
1705             {
1706
1707             /* REACTION-FIELD ELECTROSTATICS */
1708             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1709
1710             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1711
1712             fscal            = felec;
1713
1714             fscal            = _mm_and_ps(fscal,cutoff_mask);
1715
1716             /* Calculate temporary vectorial force */
1717             tx               = _mm_mul_ps(fscal,dx21);
1718             ty               = _mm_mul_ps(fscal,dy21);
1719             tz               = _mm_mul_ps(fscal,dz21);
1720
1721             /* Update vectorial force */
1722             fix2             = _mm_add_ps(fix2,tx);
1723             fiy2             = _mm_add_ps(fiy2,ty);
1724             fiz2             = _mm_add_ps(fiz2,tz);
1725
1726             fjx1             = _mm_add_ps(fjx1,tx);
1727             fjy1             = _mm_add_ps(fjy1,ty);
1728             fjz1             = _mm_add_ps(fjz1,tz);
1729
1730             }
1731
1732             /**************************
1733              * CALCULATE INTERACTIONS *
1734              **************************/
1735
1736             if (gmx_mm_any_lt(rsq22,rcutoff2))
1737             {
1738
1739             /* REACTION-FIELD ELECTROSTATICS */
1740             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1741
1742             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1743
1744             fscal            = felec;
1745
1746             fscal            = _mm_and_ps(fscal,cutoff_mask);
1747
1748             /* Calculate temporary vectorial force */
1749             tx               = _mm_mul_ps(fscal,dx22);
1750             ty               = _mm_mul_ps(fscal,dy22);
1751             tz               = _mm_mul_ps(fscal,dz22);
1752
1753             /* Update vectorial force */
1754             fix2             = _mm_add_ps(fix2,tx);
1755             fiy2             = _mm_add_ps(fiy2,ty);
1756             fiz2             = _mm_add_ps(fiz2,tz);
1757
1758             fjx2             = _mm_add_ps(fjx2,tx);
1759             fjy2             = _mm_add_ps(fjy2,ty);
1760             fjz2             = _mm_add_ps(fjz2,tz);
1761
1762             }
1763
1764             /**************************
1765              * CALCULATE INTERACTIONS *
1766              **************************/
1767
1768             if (gmx_mm_any_lt(rsq23,rcutoff2))
1769             {
1770
1771             /* REACTION-FIELD ELECTROSTATICS */
1772             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1773
1774             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1775
1776             fscal            = felec;
1777
1778             fscal            = _mm_and_ps(fscal,cutoff_mask);
1779
1780             /* Calculate temporary vectorial force */
1781             tx               = _mm_mul_ps(fscal,dx23);
1782             ty               = _mm_mul_ps(fscal,dy23);
1783             tz               = _mm_mul_ps(fscal,dz23);
1784
1785             /* Update vectorial force */
1786             fix2             = _mm_add_ps(fix2,tx);
1787             fiy2             = _mm_add_ps(fiy2,ty);
1788             fiz2             = _mm_add_ps(fiz2,tz);
1789
1790             fjx3             = _mm_add_ps(fjx3,tx);
1791             fjy3             = _mm_add_ps(fjy3,ty);
1792             fjz3             = _mm_add_ps(fjz3,tz);
1793
1794             }
1795
1796             /**************************
1797              * CALCULATE INTERACTIONS *
1798              **************************/
1799
1800             if (gmx_mm_any_lt(rsq31,rcutoff2))
1801             {
1802
1803             /* REACTION-FIELD ELECTROSTATICS */
1804             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1805
1806             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1807
1808             fscal            = felec;
1809
1810             fscal            = _mm_and_ps(fscal,cutoff_mask);
1811
1812             /* Calculate temporary vectorial force */
1813             tx               = _mm_mul_ps(fscal,dx31);
1814             ty               = _mm_mul_ps(fscal,dy31);
1815             tz               = _mm_mul_ps(fscal,dz31);
1816
1817             /* Update vectorial force */
1818             fix3             = _mm_add_ps(fix3,tx);
1819             fiy3             = _mm_add_ps(fiy3,ty);
1820             fiz3             = _mm_add_ps(fiz3,tz);
1821
1822             fjx1             = _mm_add_ps(fjx1,tx);
1823             fjy1             = _mm_add_ps(fjy1,ty);
1824             fjz1             = _mm_add_ps(fjz1,tz);
1825
1826             }
1827
1828             /**************************
1829              * CALCULATE INTERACTIONS *
1830              **************************/
1831
1832             if (gmx_mm_any_lt(rsq32,rcutoff2))
1833             {
1834
1835             /* REACTION-FIELD ELECTROSTATICS */
1836             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1837
1838             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1839
1840             fscal            = felec;
1841
1842             fscal            = _mm_and_ps(fscal,cutoff_mask);
1843
1844             /* Calculate temporary vectorial force */
1845             tx               = _mm_mul_ps(fscal,dx32);
1846             ty               = _mm_mul_ps(fscal,dy32);
1847             tz               = _mm_mul_ps(fscal,dz32);
1848
1849             /* Update vectorial force */
1850             fix3             = _mm_add_ps(fix3,tx);
1851             fiy3             = _mm_add_ps(fiy3,ty);
1852             fiz3             = _mm_add_ps(fiz3,tz);
1853
1854             fjx2             = _mm_add_ps(fjx2,tx);
1855             fjy2             = _mm_add_ps(fjy2,ty);
1856             fjz2             = _mm_add_ps(fjz2,tz);
1857
1858             }
1859
1860             /**************************
1861              * CALCULATE INTERACTIONS *
1862              **************************/
1863
1864             if (gmx_mm_any_lt(rsq33,rcutoff2))
1865             {
1866
1867             /* REACTION-FIELD ELECTROSTATICS */
1868             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1869
1870             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1871
1872             fscal            = felec;
1873
1874             fscal            = _mm_and_ps(fscal,cutoff_mask);
1875
1876             /* Calculate temporary vectorial force */
1877             tx               = _mm_mul_ps(fscal,dx33);
1878             ty               = _mm_mul_ps(fscal,dy33);
1879             tz               = _mm_mul_ps(fscal,dz33);
1880
1881             /* Update vectorial force */
1882             fix3             = _mm_add_ps(fix3,tx);
1883             fiy3             = _mm_add_ps(fiy3,ty);
1884             fiz3             = _mm_add_ps(fiz3,tz);
1885
1886             fjx3             = _mm_add_ps(fjx3,tx);
1887             fjy3             = _mm_add_ps(fjy3,ty);
1888             fjz3             = _mm_add_ps(fjz3,tz);
1889
1890             }
1891
1892             fjptrA             = f+j_coord_offsetA;
1893             fjptrB             = f+j_coord_offsetB;
1894             fjptrC             = f+j_coord_offsetC;
1895             fjptrD             = f+j_coord_offsetD;
1896
1897             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1898                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1899                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1900
1901             /* Inner loop uses 329 flops */
1902         }
1903
1904         if(jidx<j_index_end)
1905         {
1906
1907             /* Get j neighbor index, and coordinate index */
1908             jnrlistA         = jjnr[jidx];
1909             jnrlistB         = jjnr[jidx+1];
1910             jnrlistC         = jjnr[jidx+2];
1911             jnrlistD         = jjnr[jidx+3];
1912             /* Sign of each element will be negative for non-real atoms.
1913              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1914              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1915              */
1916             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1917             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1918             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1919             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1920             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1921             j_coord_offsetA  = DIM*jnrA;
1922             j_coord_offsetB  = DIM*jnrB;
1923             j_coord_offsetC  = DIM*jnrC;
1924             j_coord_offsetD  = DIM*jnrD;
1925
1926             /* load j atom coordinates */
1927             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1928                                               x+j_coord_offsetC,x+j_coord_offsetD,
1929                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1930                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1931
1932             /* Calculate displacement vector */
1933             dx00             = _mm_sub_ps(ix0,jx0);
1934             dy00             = _mm_sub_ps(iy0,jy0);
1935             dz00             = _mm_sub_ps(iz0,jz0);
1936             dx11             = _mm_sub_ps(ix1,jx1);
1937             dy11             = _mm_sub_ps(iy1,jy1);
1938             dz11             = _mm_sub_ps(iz1,jz1);
1939             dx12             = _mm_sub_ps(ix1,jx2);
1940             dy12             = _mm_sub_ps(iy1,jy2);
1941             dz12             = _mm_sub_ps(iz1,jz2);
1942             dx13             = _mm_sub_ps(ix1,jx3);
1943             dy13             = _mm_sub_ps(iy1,jy3);
1944             dz13             = _mm_sub_ps(iz1,jz3);
1945             dx21             = _mm_sub_ps(ix2,jx1);
1946             dy21             = _mm_sub_ps(iy2,jy1);
1947             dz21             = _mm_sub_ps(iz2,jz1);
1948             dx22             = _mm_sub_ps(ix2,jx2);
1949             dy22             = _mm_sub_ps(iy2,jy2);
1950             dz22             = _mm_sub_ps(iz2,jz2);
1951             dx23             = _mm_sub_ps(ix2,jx3);
1952             dy23             = _mm_sub_ps(iy2,jy3);
1953             dz23             = _mm_sub_ps(iz2,jz3);
1954             dx31             = _mm_sub_ps(ix3,jx1);
1955             dy31             = _mm_sub_ps(iy3,jy1);
1956             dz31             = _mm_sub_ps(iz3,jz1);
1957             dx32             = _mm_sub_ps(ix3,jx2);
1958             dy32             = _mm_sub_ps(iy3,jy2);
1959             dz32             = _mm_sub_ps(iz3,jz2);
1960             dx33             = _mm_sub_ps(ix3,jx3);
1961             dy33             = _mm_sub_ps(iy3,jy3);
1962             dz33             = _mm_sub_ps(iz3,jz3);
1963
1964             /* Calculate squared distance and things based on it */
1965             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1966             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1967             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1968             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1969             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1970             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1971             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1972             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1973             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1974             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1975
1976             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1977             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1978             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1979             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1980             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1981             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1982             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1983             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1984             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1985             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1986
1987             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1988             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1989             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1990             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1991             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1992             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1993             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1994             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1995             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1996             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1997
1998             fjx0             = _mm_setzero_ps();
1999             fjy0             = _mm_setzero_ps();
2000             fjz0             = _mm_setzero_ps();
2001             fjx1             = _mm_setzero_ps();
2002             fjy1             = _mm_setzero_ps();
2003             fjz1             = _mm_setzero_ps();
2004             fjx2             = _mm_setzero_ps();
2005             fjy2             = _mm_setzero_ps();
2006             fjz2             = _mm_setzero_ps();
2007             fjx3             = _mm_setzero_ps();
2008             fjy3             = _mm_setzero_ps();
2009             fjz3             = _mm_setzero_ps();
2010
2011             /**************************
2012              * CALCULATE INTERACTIONS *
2013              **************************/
2014
2015             if (gmx_mm_any_lt(rsq00,rcutoff2))
2016             {
2017
2018             r00              = _mm_mul_ps(rsq00,rinv00);
2019             r00              = _mm_andnot_ps(dummy_mask,r00);
2020
2021             /* LENNARD-JONES DISPERSION/REPULSION */
2022
2023             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
2024             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
2025             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
2026             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
2027             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
2028
2029             d                = _mm_sub_ps(r00,rswitch);
2030             d                = _mm_max_ps(d,_mm_setzero_ps());
2031             d2               = _mm_mul_ps(d,d);
2032             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
2033
2034             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
2035
2036             /* Evaluate switch function */
2037             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
2038             fvdw             = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
2039             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
2040
2041             fscal            = fvdw;
2042
2043             fscal            = _mm_and_ps(fscal,cutoff_mask);
2044
2045             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2046
2047             /* Calculate temporary vectorial force */
2048             tx               = _mm_mul_ps(fscal,dx00);
2049             ty               = _mm_mul_ps(fscal,dy00);
2050             tz               = _mm_mul_ps(fscal,dz00);
2051
2052             /* Update vectorial force */
2053             fix0             = _mm_add_ps(fix0,tx);
2054             fiy0             = _mm_add_ps(fiy0,ty);
2055             fiz0             = _mm_add_ps(fiz0,tz);
2056
2057             fjx0             = _mm_add_ps(fjx0,tx);
2058             fjy0             = _mm_add_ps(fjy0,ty);
2059             fjz0             = _mm_add_ps(fjz0,tz);
2060
2061             }
2062
2063             /**************************
2064              * CALCULATE INTERACTIONS *
2065              **************************/
2066
2067             if (gmx_mm_any_lt(rsq11,rcutoff2))
2068             {
2069
2070             /* REACTION-FIELD ELECTROSTATICS */
2071             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
2072
2073             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2074
2075             fscal            = felec;
2076
2077             fscal            = _mm_and_ps(fscal,cutoff_mask);
2078
2079             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2080
2081             /* Calculate temporary vectorial force */
2082             tx               = _mm_mul_ps(fscal,dx11);
2083             ty               = _mm_mul_ps(fscal,dy11);
2084             tz               = _mm_mul_ps(fscal,dz11);
2085
2086             /* Update vectorial force */
2087             fix1             = _mm_add_ps(fix1,tx);
2088             fiy1             = _mm_add_ps(fiy1,ty);
2089             fiz1             = _mm_add_ps(fiz1,tz);
2090
2091             fjx1             = _mm_add_ps(fjx1,tx);
2092             fjy1             = _mm_add_ps(fjy1,ty);
2093             fjz1             = _mm_add_ps(fjz1,tz);
2094
2095             }
2096
2097             /**************************
2098              * CALCULATE INTERACTIONS *
2099              **************************/
2100
2101             if (gmx_mm_any_lt(rsq12,rcutoff2))
2102             {
2103
2104             /* REACTION-FIELD ELECTROSTATICS */
2105             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
2106
2107             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2108
2109             fscal            = felec;
2110
2111             fscal            = _mm_and_ps(fscal,cutoff_mask);
2112
2113             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2114
2115             /* Calculate temporary vectorial force */
2116             tx               = _mm_mul_ps(fscal,dx12);
2117             ty               = _mm_mul_ps(fscal,dy12);
2118             tz               = _mm_mul_ps(fscal,dz12);
2119
2120             /* Update vectorial force */
2121             fix1             = _mm_add_ps(fix1,tx);
2122             fiy1             = _mm_add_ps(fiy1,ty);
2123             fiz1             = _mm_add_ps(fiz1,tz);
2124
2125             fjx2             = _mm_add_ps(fjx2,tx);
2126             fjy2             = _mm_add_ps(fjy2,ty);
2127             fjz2             = _mm_add_ps(fjz2,tz);
2128
2129             }
2130
2131             /**************************
2132              * CALCULATE INTERACTIONS *
2133              **************************/
2134
2135             if (gmx_mm_any_lt(rsq13,rcutoff2))
2136             {
2137
2138             /* REACTION-FIELD ELECTROSTATICS */
2139             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
2140
2141             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
2142
2143             fscal            = felec;
2144
2145             fscal            = _mm_and_ps(fscal,cutoff_mask);
2146
2147             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2148
2149             /* Calculate temporary vectorial force */
2150             tx               = _mm_mul_ps(fscal,dx13);
2151             ty               = _mm_mul_ps(fscal,dy13);
2152             tz               = _mm_mul_ps(fscal,dz13);
2153
2154             /* Update vectorial force */
2155             fix1             = _mm_add_ps(fix1,tx);
2156             fiy1             = _mm_add_ps(fiy1,ty);
2157             fiz1             = _mm_add_ps(fiz1,tz);
2158
2159             fjx3             = _mm_add_ps(fjx3,tx);
2160             fjy3             = _mm_add_ps(fjy3,ty);
2161             fjz3             = _mm_add_ps(fjz3,tz);
2162
2163             }
2164
2165             /**************************
2166              * CALCULATE INTERACTIONS *
2167              **************************/
2168
2169             if (gmx_mm_any_lt(rsq21,rcutoff2))
2170             {
2171
2172             /* REACTION-FIELD ELECTROSTATICS */
2173             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
2174
2175             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2176
2177             fscal            = felec;
2178
2179             fscal            = _mm_and_ps(fscal,cutoff_mask);
2180
2181             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2182
2183             /* Calculate temporary vectorial force */
2184             tx               = _mm_mul_ps(fscal,dx21);
2185             ty               = _mm_mul_ps(fscal,dy21);
2186             tz               = _mm_mul_ps(fscal,dz21);
2187
2188             /* Update vectorial force */
2189             fix2             = _mm_add_ps(fix2,tx);
2190             fiy2             = _mm_add_ps(fiy2,ty);
2191             fiz2             = _mm_add_ps(fiz2,tz);
2192
2193             fjx1             = _mm_add_ps(fjx1,tx);
2194             fjy1             = _mm_add_ps(fjy1,ty);
2195             fjz1             = _mm_add_ps(fjz1,tz);
2196
2197             }
2198
2199             /**************************
2200              * CALCULATE INTERACTIONS *
2201              **************************/
2202
2203             if (gmx_mm_any_lt(rsq22,rcutoff2))
2204             {
2205
2206             /* REACTION-FIELD ELECTROSTATICS */
2207             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2208
2209             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2210
2211             fscal            = felec;
2212
2213             fscal            = _mm_and_ps(fscal,cutoff_mask);
2214
2215             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2216
2217             /* Calculate temporary vectorial force */
2218             tx               = _mm_mul_ps(fscal,dx22);
2219             ty               = _mm_mul_ps(fscal,dy22);
2220             tz               = _mm_mul_ps(fscal,dz22);
2221
2222             /* Update vectorial force */
2223             fix2             = _mm_add_ps(fix2,tx);
2224             fiy2             = _mm_add_ps(fiy2,ty);
2225             fiz2             = _mm_add_ps(fiz2,tz);
2226
2227             fjx2             = _mm_add_ps(fjx2,tx);
2228             fjy2             = _mm_add_ps(fjy2,ty);
2229             fjz2             = _mm_add_ps(fjz2,tz);
2230
2231             }
2232
2233             /**************************
2234              * CALCULATE INTERACTIONS *
2235              **************************/
2236
2237             if (gmx_mm_any_lt(rsq23,rcutoff2))
2238             {
2239
2240             /* REACTION-FIELD ELECTROSTATICS */
2241             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
2242
2243             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2244
2245             fscal            = felec;
2246
2247             fscal            = _mm_and_ps(fscal,cutoff_mask);
2248
2249             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2250
2251             /* Calculate temporary vectorial force */
2252             tx               = _mm_mul_ps(fscal,dx23);
2253             ty               = _mm_mul_ps(fscal,dy23);
2254             tz               = _mm_mul_ps(fscal,dz23);
2255
2256             /* Update vectorial force */
2257             fix2             = _mm_add_ps(fix2,tx);
2258             fiy2             = _mm_add_ps(fiy2,ty);
2259             fiz2             = _mm_add_ps(fiz2,tz);
2260
2261             fjx3             = _mm_add_ps(fjx3,tx);
2262             fjy3             = _mm_add_ps(fjy3,ty);
2263             fjz3             = _mm_add_ps(fjz3,tz);
2264
2265             }
2266
2267             /**************************
2268              * CALCULATE INTERACTIONS *
2269              **************************/
2270
2271             if (gmx_mm_any_lt(rsq31,rcutoff2))
2272             {
2273
2274             /* REACTION-FIELD ELECTROSTATICS */
2275             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
2276
2277             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2278
2279             fscal            = felec;
2280
2281             fscal            = _mm_and_ps(fscal,cutoff_mask);
2282
2283             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2284
2285             /* Calculate temporary vectorial force */
2286             tx               = _mm_mul_ps(fscal,dx31);
2287             ty               = _mm_mul_ps(fscal,dy31);
2288             tz               = _mm_mul_ps(fscal,dz31);
2289
2290             /* Update vectorial force */
2291             fix3             = _mm_add_ps(fix3,tx);
2292             fiy3             = _mm_add_ps(fiy3,ty);
2293             fiz3             = _mm_add_ps(fiz3,tz);
2294
2295             fjx1             = _mm_add_ps(fjx1,tx);
2296             fjy1             = _mm_add_ps(fjy1,ty);
2297             fjz1             = _mm_add_ps(fjz1,tz);
2298
2299             }
2300
2301             /**************************
2302              * CALCULATE INTERACTIONS *
2303              **************************/
2304
2305             if (gmx_mm_any_lt(rsq32,rcutoff2))
2306             {
2307
2308             /* REACTION-FIELD ELECTROSTATICS */
2309             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
2310
2311             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2312
2313             fscal            = felec;
2314
2315             fscal            = _mm_and_ps(fscal,cutoff_mask);
2316
2317             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2318
2319             /* Calculate temporary vectorial force */
2320             tx               = _mm_mul_ps(fscal,dx32);
2321             ty               = _mm_mul_ps(fscal,dy32);
2322             tz               = _mm_mul_ps(fscal,dz32);
2323
2324             /* Update vectorial force */
2325             fix3             = _mm_add_ps(fix3,tx);
2326             fiy3             = _mm_add_ps(fiy3,ty);
2327             fiz3             = _mm_add_ps(fiz3,tz);
2328
2329             fjx2             = _mm_add_ps(fjx2,tx);
2330             fjy2             = _mm_add_ps(fjy2,ty);
2331             fjz2             = _mm_add_ps(fjz2,tz);
2332
2333             }
2334
2335             /**************************
2336              * CALCULATE INTERACTIONS *
2337              **************************/
2338
2339             if (gmx_mm_any_lt(rsq33,rcutoff2))
2340             {
2341
2342             /* REACTION-FIELD ELECTROSTATICS */
2343             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2344
2345             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2346
2347             fscal            = felec;
2348
2349             fscal            = _mm_and_ps(fscal,cutoff_mask);
2350
2351             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2352
2353             /* Calculate temporary vectorial force */
2354             tx               = _mm_mul_ps(fscal,dx33);
2355             ty               = _mm_mul_ps(fscal,dy33);
2356             tz               = _mm_mul_ps(fscal,dz33);
2357
2358             /* Update vectorial force */
2359             fix3             = _mm_add_ps(fix3,tx);
2360             fiy3             = _mm_add_ps(fiy3,ty);
2361             fiz3             = _mm_add_ps(fiz3,tz);
2362
2363             fjx3             = _mm_add_ps(fjx3,tx);
2364             fjy3             = _mm_add_ps(fjy3,ty);
2365             fjz3             = _mm_add_ps(fjz3,tz);
2366
2367             }
2368
2369             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2370             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2371             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2372             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2373
2374             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2375                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2376                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2377
2378             /* Inner loop uses 330 flops */
2379         }
2380
2381         /* End of innermost loop */
2382
2383         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2384                                               f+i_coord_offset,fshift+i_shift_offset);
2385
2386         /* Increment number of inner iterations */
2387         inneriter                  += j_index_end - j_index_start;
2388
2389         /* Outer loop uses 24 flops */
2390     }
2391
2392     /* Increment number of outer iterations */
2393     outeriter        += nri;
2394
2395     /* Update outer/inner flops */
2396
2397     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*330);
2398 }