src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_sse2_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse2_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse2_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_VF_sse2_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_VF_sse2_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwioffset3;
  86     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  87     int              vdwjidx0A,vdwjidx0B;
  88     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     int              vdwjidx3A,vdwjidx3B;
  94     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  95     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 101     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 102     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 103     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 104     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 105     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 106     real             *charge;
 107     int              nvdwtype;
 108     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 109     int              *vdwtype;
 110     real             *vdwparam;
 111     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 112     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 113     __m128d          dummy_mask,cutoff_mask;
 114     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 115     __m128d          one     = _mm_set1_pd(1.0);
 116     __m128d          two     = _mm_set1_pd(2.0);
 117     x                = xx[0];
 118     f                = ff[0];
 119
 120     nri              = nlist->nri;
 121     iinr             = nlist->iinr;
 122     jindex           = nlist->jindex;
 123     jjnr             = nlist->jjnr;
 124     shiftidx         = nlist->shift;
 125     gid              = nlist->gid;
 126     shiftvec         = fr->shift_vec[0];
 127     fshift           = fr->fshift[0];
 128     facel            = _mm_set1_pd(fr->ic->epsfac);
 129     charge           = mdatoms->chargeA;
 130     krf              = _mm_set1_pd(fr->ic->k_rf);
 131     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 132     crf              = _mm_set1_pd(fr->ic->c_rf);
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     /* Setup water-specific parameters */
 138     inr              = nlist->iinr[0];
 139     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 140     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 141     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 142     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 143
 144     jq1              = _mm_set1_pd(charge[inr+1]);
 145     jq2              = _mm_set1_pd(charge[inr+2]);
 146     jq3              = _mm_set1_pd(charge[inr+3]);
 147     vdwjidx0A        = 2*vdwtype[inr+0];
 148     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 149     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 150     qq11             = _mm_mul_pd(iq1,jq1);
 151     qq12             = _mm_mul_pd(iq1,jq2);
 152     qq13             = _mm_mul_pd(iq1,jq3);
 153     qq21             = _mm_mul_pd(iq2,jq1);
 154     qq22             = _mm_mul_pd(iq2,jq2);
 155     qq23             = _mm_mul_pd(iq2,jq3);
 156     qq31             = _mm_mul_pd(iq3,jq1);
 157     qq32             = _mm_mul_pd(iq3,jq2);
 158     qq33             = _mm_mul_pd(iq3,jq3);
 159
 160     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 161     rcutoff_scalar   = fr->ic->rcoulomb;
 162     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 163     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 164
 165     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 166     rvdw             = _mm_set1_pd(fr->ic->rvdw);
 167
 168     /* Avoid stupid compiler warnings */
 169     jnrA = jnrB = 0;
 170     j_coord_offsetA = 0;
 171     j_coord_offsetB = 0;
 172
 173     outeriter        = 0;
 174     inneriter        = 0;
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 193
 194         fix0             = _mm_setzero_pd();
 195         fiy0             = _mm_setzero_pd();
 196         fiz0             = _mm_setzero_pd();
 197         fix1             = _mm_setzero_pd();
 198         fiy1             = _mm_setzero_pd();
 199         fiz1             = _mm_setzero_pd();
 200         fix2             = _mm_setzero_pd();
 201         fiy2             = _mm_setzero_pd();
 202         fiz2             = _mm_setzero_pd();
 203         fix3             = _mm_setzero_pd();
 204         fiy3             = _mm_setzero_pd();
 205         fiz3             = _mm_setzero_pd();
 206
 207         /* Reset potential sums */
 208         velecsum         = _mm_setzero_pd();
 209         vvdwsum          = _mm_setzero_pd();
 210
 211         /* Start inner kernel loop */
 212         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 213         {
 214
 215             /* Get j neighbor index, and coordinate index */
 216             jnrA             = jjnr[jidx];
 217             jnrB             = jjnr[jidx+1];
 218             j_coord_offsetA  = DIM*jnrA;
 219             j_coord_offsetB  = DIM*jnrB;
 220
 221             /* load j atom coordinates */
 222             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 223                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 224                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 225
 226             /* Calculate displacement vector */
 227             dx00             = _mm_sub_pd(ix0,jx0);
 228             dy00             = _mm_sub_pd(iy0,jy0);
 229             dz00             = _mm_sub_pd(iz0,jz0);
 230             dx11             = _mm_sub_pd(ix1,jx1);
 231             dy11             = _mm_sub_pd(iy1,jy1);
 232             dz11             = _mm_sub_pd(iz1,jz1);
 233             dx12             = _mm_sub_pd(ix1,jx2);
 234             dy12             = _mm_sub_pd(iy1,jy2);
 235             dz12             = _mm_sub_pd(iz1,jz2);
 236             dx13             = _mm_sub_pd(ix1,jx3);
 237             dy13             = _mm_sub_pd(iy1,jy3);
 238             dz13             = _mm_sub_pd(iz1,jz3);
 239             dx21             = _mm_sub_pd(ix2,jx1);
 240             dy21             = _mm_sub_pd(iy2,jy1);
 241             dz21             = _mm_sub_pd(iz2,jz1);
 242             dx22             = _mm_sub_pd(ix2,jx2);
 243             dy22             = _mm_sub_pd(iy2,jy2);
 244             dz22             = _mm_sub_pd(iz2,jz2);
 245             dx23             = _mm_sub_pd(ix2,jx3);
 246             dy23             = _mm_sub_pd(iy2,jy3);
 247             dz23             = _mm_sub_pd(iz2,jz3);
 248             dx31             = _mm_sub_pd(ix3,jx1);
 249             dy31             = _mm_sub_pd(iy3,jy1);
 250             dz31             = _mm_sub_pd(iz3,jz1);
 251             dx32             = _mm_sub_pd(ix3,jx2);
 252             dy32             = _mm_sub_pd(iy3,jy2);
 253             dz32             = _mm_sub_pd(iz3,jz2);
 254             dx33             = _mm_sub_pd(ix3,jx3);
 255             dy33             = _mm_sub_pd(iy3,jy3);
 256             dz33             = _mm_sub_pd(iz3,jz3);
 257
 258             /* Calculate squared distance and things based on it */
 259             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 260             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 261             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 262             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 263             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 264             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 265             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 266             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 267             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 268             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 269
 270             rinv11           = sse2_invsqrt_d(rsq11);
 271             rinv12           = sse2_invsqrt_d(rsq12);
 272             rinv13           = sse2_invsqrt_d(rsq13);
 273             rinv21           = sse2_invsqrt_d(rsq21);
 274             rinv22           = sse2_invsqrt_d(rsq22);
 275             rinv23           = sse2_invsqrt_d(rsq23);
 276             rinv31           = sse2_invsqrt_d(rsq31);
 277             rinv32           = sse2_invsqrt_d(rsq32);
 278             rinv33           = sse2_invsqrt_d(rsq33);
 279
 280             rinvsq00         = sse2_inv_d(rsq00);
 281             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 282             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 283             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 284             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 285             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 286             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 287             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 288             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 289             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 290
 291             fjx0             = _mm_setzero_pd();
 292             fjy0             = _mm_setzero_pd();
 293             fjz0             = _mm_setzero_pd();
 294             fjx1             = _mm_setzero_pd();
 295             fjy1             = _mm_setzero_pd();
 296             fjz1             = _mm_setzero_pd();
 297             fjx2             = _mm_setzero_pd();
 298             fjy2             = _mm_setzero_pd();
 299             fjz2             = _mm_setzero_pd();
 300             fjx3             = _mm_setzero_pd();
 301             fjy3             = _mm_setzero_pd();
 302             fjz3             = _mm_setzero_pd();
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             if (gmx_mm_any_lt(rsq00,rcutoff2))
 309             {
 310
 311             /* LENNARD-JONES DISPERSION/REPULSION */
 312
 313             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 314             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 315             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 316             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 317                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 318             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 319
 320             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 321
 322             /* Update potential sum for this i atom from the interaction with this j atom. */
 323             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 324             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 325
 326             fscal            = fvdw;
 327
 328             fscal            = _mm_and_pd(fscal,cutoff_mask);
 329
 330             /* Calculate temporary vectorial force */
 331             tx               = _mm_mul_pd(fscal,dx00);
 332             ty               = _mm_mul_pd(fscal,dy00);
 333             tz               = _mm_mul_pd(fscal,dz00);
 334
 335             /* Update vectorial force */
 336             fix0             = _mm_add_pd(fix0,tx);
 337             fiy0             = _mm_add_pd(fiy0,ty);
 338             fiz0             = _mm_add_pd(fiz0,tz);
 339
 340             fjx0             = _mm_add_pd(fjx0,tx);
 341             fjy0             = _mm_add_pd(fjy0,ty);
 342             fjz0             = _mm_add_pd(fjz0,tz);
 343
 344             }
 345
 346             /**************************
 347              * CALCULATE INTERACTIONS *
 348              **************************/
 349
 350             if (gmx_mm_any_lt(rsq11,rcutoff2))
 351             {
 352
 353             /* REACTION-FIELD ELECTROSTATICS */
 354             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 355             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 356
 357             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velec            = _mm_and_pd(velec,cutoff_mask);
 361             velecsum         = _mm_add_pd(velecsum,velec);
 362
 363             fscal            = felec;
 364
 365             fscal            = _mm_and_pd(fscal,cutoff_mask);
 366
 367             /* Calculate temporary vectorial force */
 368             tx               = _mm_mul_pd(fscal,dx11);
 369             ty               = _mm_mul_pd(fscal,dy11);
 370             tz               = _mm_mul_pd(fscal,dz11);
 371
 372             /* Update vectorial force */
 373             fix1             = _mm_add_pd(fix1,tx);
 374             fiy1             = _mm_add_pd(fiy1,ty);
 375             fiz1             = _mm_add_pd(fiz1,tz);
 376
 377             fjx1             = _mm_add_pd(fjx1,tx);
 378             fjy1             = _mm_add_pd(fjy1,ty);
 379             fjz1             = _mm_add_pd(fjz1,tz);
 380
 381             }
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             if (gmx_mm_any_lt(rsq12,rcutoff2))
 388             {
 389
 390             /* REACTION-FIELD ELECTROSTATICS */
 391             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 392             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 393
 394             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velec            = _mm_and_pd(velec,cutoff_mask);
 398             velecsum         = _mm_add_pd(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402             fscal            = _mm_and_pd(fscal,cutoff_mask);
 403
 404             /* Calculate temporary vectorial force */
 405             tx               = _mm_mul_pd(fscal,dx12);
 406             ty               = _mm_mul_pd(fscal,dy12);
 407             tz               = _mm_mul_pd(fscal,dz12);
 408
 409             /* Update vectorial force */
 410             fix1             = _mm_add_pd(fix1,tx);
 411             fiy1             = _mm_add_pd(fiy1,ty);
 412             fiz1             = _mm_add_pd(fiz1,tz);
 413
 414             fjx2             = _mm_add_pd(fjx2,tx);
 415             fjy2             = _mm_add_pd(fjy2,ty);
 416             fjz2             = _mm_add_pd(fjz2,tz);
 417
 418             }
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             if (gmx_mm_any_lt(rsq13,rcutoff2))
 425             {
 426
 427             /* REACTION-FIELD ELECTROSTATICS */
 428             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 429             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 430
 431             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velec            = _mm_and_pd(velec,cutoff_mask);
 435             velecsum         = _mm_add_pd(velecsum,velec);
 436
 437             fscal            = felec;
 438
 439             fscal            = _mm_and_pd(fscal,cutoff_mask);
 440
 441             /* Calculate temporary vectorial force */
 442             tx               = _mm_mul_pd(fscal,dx13);
 443             ty               = _mm_mul_pd(fscal,dy13);
 444             tz               = _mm_mul_pd(fscal,dz13);
 445
 446             /* Update vectorial force */
 447             fix1             = _mm_add_pd(fix1,tx);
 448             fiy1             = _mm_add_pd(fiy1,ty);
 449             fiz1             = _mm_add_pd(fiz1,tz);
 450
 451             fjx3             = _mm_add_pd(fjx3,tx);
 452             fjy3             = _mm_add_pd(fjy3,ty);
 453             fjz3             = _mm_add_pd(fjz3,tz);
 454
 455             }
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             if (gmx_mm_any_lt(rsq21,rcutoff2))
 462             {
 463
 464             /* REACTION-FIELD ELECTROSTATICS */
 465             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 466             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 467
 468             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velec            = _mm_and_pd(velec,cutoff_mask);
 472             velecsum         = _mm_add_pd(velecsum,velec);
 473
 474             fscal            = felec;
 475
 476             fscal            = _mm_and_pd(fscal,cutoff_mask);
 477
 478             /* Calculate temporary vectorial force */
 479             tx               = _mm_mul_pd(fscal,dx21);
 480             ty               = _mm_mul_pd(fscal,dy21);
 481             tz               = _mm_mul_pd(fscal,dz21);
 482
 483             /* Update vectorial force */
 484             fix2             = _mm_add_pd(fix2,tx);
 485             fiy2             = _mm_add_pd(fiy2,ty);
 486             fiz2             = _mm_add_pd(fiz2,tz);
 487
 488             fjx1             = _mm_add_pd(fjx1,tx);
 489             fjy1             = _mm_add_pd(fjy1,ty);
 490             fjz1             = _mm_add_pd(fjz1,tz);
 491
 492             }
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             if (gmx_mm_any_lt(rsq22,rcutoff2))
 499             {
 500
 501             /* REACTION-FIELD ELECTROSTATICS */
 502             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 503             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 504
 505             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velec            = _mm_and_pd(velec,cutoff_mask);
 509             velecsum         = _mm_add_pd(velecsum,velec);
 510
 511             fscal            = felec;
 512
 513             fscal            = _mm_and_pd(fscal,cutoff_mask);
 514
 515             /* Calculate temporary vectorial force */
 516             tx               = _mm_mul_pd(fscal,dx22);
 517             ty               = _mm_mul_pd(fscal,dy22);
 518             tz               = _mm_mul_pd(fscal,dz22);
 519
 520             /* Update vectorial force */
 521             fix2             = _mm_add_pd(fix2,tx);
 522             fiy2             = _mm_add_pd(fiy2,ty);
 523             fiz2             = _mm_add_pd(fiz2,tz);
 524
 525             fjx2             = _mm_add_pd(fjx2,tx);
 526             fjy2             = _mm_add_pd(fjy2,ty);
 527             fjz2             = _mm_add_pd(fjz2,tz);
 528
 529             }
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             if (gmx_mm_any_lt(rsq23,rcutoff2))
 536             {
 537
 538             /* REACTION-FIELD ELECTROSTATICS */
 539             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 540             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 541
 542             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 543
 544             /* Update potential sum for this i atom from the interaction with this j atom. */
 545             velec            = _mm_and_pd(velec,cutoff_mask);
 546             velecsum         = _mm_add_pd(velecsum,velec);
 547
 548             fscal            = felec;
 549
 550             fscal            = _mm_and_pd(fscal,cutoff_mask);
 551
 552             /* Calculate temporary vectorial force */
 553             tx               = _mm_mul_pd(fscal,dx23);
 554             ty               = _mm_mul_pd(fscal,dy23);
 555             tz               = _mm_mul_pd(fscal,dz23);
 556
 557             /* Update vectorial force */
 558             fix2             = _mm_add_pd(fix2,tx);
 559             fiy2             = _mm_add_pd(fiy2,ty);
 560             fiz2             = _mm_add_pd(fiz2,tz);
 561
 562             fjx3             = _mm_add_pd(fjx3,tx);
 563             fjy3             = _mm_add_pd(fjy3,ty);
 564             fjz3             = _mm_add_pd(fjz3,tz);
 565
 566             }
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             if (gmx_mm_any_lt(rsq31,rcutoff2))
 573             {
 574
 575             /* REACTION-FIELD ELECTROSTATICS */
 576             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 577             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 578
 579             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 580
 581             /* Update potential sum for this i atom from the interaction with this j atom. */
 582             velec            = _mm_and_pd(velec,cutoff_mask);
 583             velecsum         = _mm_add_pd(velecsum,velec);
 584
 585             fscal            = felec;
 586
 587             fscal            = _mm_and_pd(fscal,cutoff_mask);
 588
 589             /* Calculate temporary vectorial force */
 590             tx               = _mm_mul_pd(fscal,dx31);
 591             ty               = _mm_mul_pd(fscal,dy31);
 592             tz               = _mm_mul_pd(fscal,dz31);
 593
 594             /* Update vectorial force */
 595             fix3             = _mm_add_pd(fix3,tx);
 596             fiy3             = _mm_add_pd(fiy3,ty);
 597             fiz3             = _mm_add_pd(fiz3,tz);
 598
 599             fjx1             = _mm_add_pd(fjx1,tx);
 600             fjy1             = _mm_add_pd(fjy1,ty);
 601             fjz1             = _mm_add_pd(fjz1,tz);
 602
 603             }
 604
 605             /**************************
 606              * CALCULATE INTERACTIONS *
 607              **************************/
 608
 609             if (gmx_mm_any_lt(rsq32,rcutoff2))
 610             {
 611
 612             /* REACTION-FIELD ELECTROSTATICS */
 613             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 614             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 615
 616             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 617
 618             /* Update potential sum for this i atom from the interaction with this j atom. */
 619             velec            = _mm_and_pd(velec,cutoff_mask);
 620             velecsum         = _mm_add_pd(velecsum,velec);
 621
 622             fscal            = felec;
 623
 624             fscal            = _mm_and_pd(fscal,cutoff_mask);
 625
 626             /* Calculate temporary vectorial force */
 627             tx               = _mm_mul_pd(fscal,dx32);
 628             ty               = _mm_mul_pd(fscal,dy32);
 629             tz               = _mm_mul_pd(fscal,dz32);
 630
 631             /* Update vectorial force */
 632             fix3             = _mm_add_pd(fix3,tx);
 633             fiy3             = _mm_add_pd(fiy3,ty);
 634             fiz3             = _mm_add_pd(fiz3,tz);
 635
 636             fjx2             = _mm_add_pd(fjx2,tx);
 637             fjy2             = _mm_add_pd(fjy2,ty);
 638             fjz2             = _mm_add_pd(fjz2,tz);
 639
 640             }
 641
 642             /**************************
 643              * CALCULATE INTERACTIONS *
 644              **************************/
 645
 646             if (gmx_mm_any_lt(rsq33,rcutoff2))
 647             {
 648
 649             /* REACTION-FIELD ELECTROSTATICS */
 650             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 651             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 652
 653             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 654
 655             /* Update potential sum for this i atom from the interaction with this j atom. */
 656             velec            = _mm_and_pd(velec,cutoff_mask);
 657             velecsum         = _mm_add_pd(velecsum,velec);
 658
 659             fscal            = felec;
 660
 661             fscal            = _mm_and_pd(fscal,cutoff_mask);
 662
 663             /* Calculate temporary vectorial force */
 664             tx               = _mm_mul_pd(fscal,dx33);
 665             ty               = _mm_mul_pd(fscal,dy33);
 666             tz               = _mm_mul_pd(fscal,dz33);
 667
 668             /* Update vectorial force */
 669             fix3             = _mm_add_pd(fix3,tx);
 670             fiy3             = _mm_add_pd(fiy3,ty);
 671             fiz3             = _mm_add_pd(fiz3,tz);
 672
 673             fjx3             = _mm_add_pd(fjx3,tx);
 674             fjy3             = _mm_add_pd(fjy3,ty);
 675             fjz3             = _mm_add_pd(fjz3,tz);
 676
 677             }
 678
 679             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 680
 681             /* Inner loop uses 368 flops */
 682         }
 683
 684         if(jidx<j_index_end)
 685         {
 686
 687             jnrA             = jjnr[jidx];
 688             j_coord_offsetA  = DIM*jnrA;
 689
 690             /* load j atom coordinates */
 691             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 692                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 693                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 694
 695             /* Calculate displacement vector */
 696             dx00             = _mm_sub_pd(ix0,jx0);
 697             dy00             = _mm_sub_pd(iy0,jy0);
 698             dz00             = _mm_sub_pd(iz0,jz0);
 699             dx11             = _mm_sub_pd(ix1,jx1);
 700             dy11             = _mm_sub_pd(iy1,jy1);
 701             dz11             = _mm_sub_pd(iz1,jz1);
 702             dx12             = _mm_sub_pd(ix1,jx2);
 703             dy12             = _mm_sub_pd(iy1,jy2);
 704             dz12             = _mm_sub_pd(iz1,jz2);
 705             dx13             = _mm_sub_pd(ix1,jx3);
 706             dy13             = _mm_sub_pd(iy1,jy3);
 707             dz13             = _mm_sub_pd(iz1,jz3);
 708             dx21             = _mm_sub_pd(ix2,jx1);
 709             dy21             = _mm_sub_pd(iy2,jy1);
 710             dz21             = _mm_sub_pd(iz2,jz1);
 711             dx22             = _mm_sub_pd(ix2,jx2);
 712             dy22             = _mm_sub_pd(iy2,jy2);
 713             dz22             = _mm_sub_pd(iz2,jz2);
 714             dx23             = _mm_sub_pd(ix2,jx3);
 715             dy23             = _mm_sub_pd(iy2,jy3);
 716             dz23             = _mm_sub_pd(iz2,jz3);
 717             dx31             = _mm_sub_pd(ix3,jx1);
 718             dy31             = _mm_sub_pd(iy3,jy1);
 719             dz31             = _mm_sub_pd(iz3,jz1);
 720             dx32             = _mm_sub_pd(ix3,jx2);
 721             dy32             = _mm_sub_pd(iy3,jy2);
 722             dz32             = _mm_sub_pd(iz3,jz2);
 723             dx33             = _mm_sub_pd(ix3,jx3);
 724             dy33             = _mm_sub_pd(iy3,jy3);
 725             dz33             = _mm_sub_pd(iz3,jz3);
 726
 727             /* Calculate squared distance and things based on it */
 728             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 729             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 730             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 731             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 732             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 733             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 734             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 735             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 736             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 737             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 738
 739             rinv11           = sse2_invsqrt_d(rsq11);
 740             rinv12           = sse2_invsqrt_d(rsq12);
 741             rinv13           = sse2_invsqrt_d(rsq13);
 742             rinv21           = sse2_invsqrt_d(rsq21);
 743             rinv22           = sse2_invsqrt_d(rsq22);
 744             rinv23           = sse2_invsqrt_d(rsq23);
 745             rinv31           = sse2_invsqrt_d(rsq31);
 746             rinv32           = sse2_invsqrt_d(rsq32);
 747             rinv33           = sse2_invsqrt_d(rsq33);
 748
 749             rinvsq00         = sse2_inv_d(rsq00);
 750             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 751             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 752             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 753             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 754             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 755             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 756             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 757             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 758             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 759
 760             fjx0             = _mm_setzero_pd();
 761             fjy0             = _mm_setzero_pd();
 762             fjz0             = _mm_setzero_pd();
 763             fjx1             = _mm_setzero_pd();
 764             fjy1             = _mm_setzero_pd();
 765             fjz1             = _mm_setzero_pd();
 766             fjx2             = _mm_setzero_pd();
 767             fjy2             = _mm_setzero_pd();
 768             fjz2             = _mm_setzero_pd();
 769             fjx3             = _mm_setzero_pd();
 770             fjy3             = _mm_setzero_pd();
 771             fjz3             = _mm_setzero_pd();
 772
 773             /**************************
 774              * CALCULATE INTERACTIONS *
 775              **************************/
 776
 777             if (gmx_mm_any_lt(rsq00,rcutoff2))
 778             {
 779
 780             /* LENNARD-JONES DISPERSION/REPULSION */
 781
 782             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 783             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 784             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 785             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 786                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 787             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 788
 789             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 790
 791             /* Update potential sum for this i atom from the interaction with this j atom. */
 792             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 793             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 794             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 795
 796             fscal            = fvdw;
 797
 798             fscal            = _mm_and_pd(fscal,cutoff_mask);
 799
 800             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 801
 802             /* Calculate temporary vectorial force */
 803             tx               = _mm_mul_pd(fscal,dx00);
 804             ty               = _mm_mul_pd(fscal,dy00);
 805             tz               = _mm_mul_pd(fscal,dz00);
 806
 807             /* Update vectorial force */
 808             fix0             = _mm_add_pd(fix0,tx);
 809             fiy0             = _mm_add_pd(fiy0,ty);
 810             fiz0             = _mm_add_pd(fiz0,tz);
 811
 812             fjx0             = _mm_add_pd(fjx0,tx);
 813             fjy0             = _mm_add_pd(fjy0,ty);
 814             fjz0             = _mm_add_pd(fjz0,tz);
 815
 816             }
 817
 818             /**************************
 819              * CALCULATE INTERACTIONS *
 820              **************************/
 821
 822             if (gmx_mm_any_lt(rsq11,rcutoff2))
 823             {
 824
 825             /* REACTION-FIELD ELECTROSTATICS */
 826             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 827             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 828
 829             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 830
 831             /* Update potential sum for this i atom from the interaction with this j atom. */
 832             velec            = _mm_and_pd(velec,cutoff_mask);
 833             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 834             velecsum         = _mm_add_pd(velecsum,velec);
 835
 836             fscal            = felec;
 837
 838             fscal            = _mm_and_pd(fscal,cutoff_mask);
 839
 840             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 841
 842             /* Calculate temporary vectorial force */
 843             tx               = _mm_mul_pd(fscal,dx11);
 844             ty               = _mm_mul_pd(fscal,dy11);
 845             tz               = _mm_mul_pd(fscal,dz11);
 846
 847             /* Update vectorial force */
 848             fix1             = _mm_add_pd(fix1,tx);
 849             fiy1             = _mm_add_pd(fiy1,ty);
 850             fiz1             = _mm_add_pd(fiz1,tz);
 851
 852             fjx1             = _mm_add_pd(fjx1,tx);
 853             fjy1             = _mm_add_pd(fjy1,ty);
 854             fjz1             = _mm_add_pd(fjz1,tz);
 855
 856             }
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             if (gmx_mm_any_lt(rsq12,rcutoff2))
 863             {
 864
 865             /* REACTION-FIELD ELECTROSTATICS */
 866             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 867             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 868
 869             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 870
 871             /* Update potential sum for this i atom from the interaction with this j atom. */
 872             velec            = _mm_and_pd(velec,cutoff_mask);
 873             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 874             velecsum         = _mm_add_pd(velecsum,velec);
 875
 876             fscal            = felec;
 877
 878             fscal            = _mm_and_pd(fscal,cutoff_mask);
 879
 880             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 881
 882             /* Calculate temporary vectorial force */
 883             tx               = _mm_mul_pd(fscal,dx12);
 884             ty               = _mm_mul_pd(fscal,dy12);
 885             tz               = _mm_mul_pd(fscal,dz12);
 886
 887             /* Update vectorial force */
 888             fix1             = _mm_add_pd(fix1,tx);
 889             fiy1             = _mm_add_pd(fiy1,ty);
 890             fiz1             = _mm_add_pd(fiz1,tz);
 891
 892             fjx2             = _mm_add_pd(fjx2,tx);
 893             fjy2             = _mm_add_pd(fjy2,ty);
 894             fjz2             = _mm_add_pd(fjz2,tz);
 895
 896             }
 897
 898             /**************************
 899              * CALCULATE INTERACTIONS *
 900              **************************/
 901
 902             if (gmx_mm_any_lt(rsq13,rcutoff2))
 903             {
 904
 905             /* REACTION-FIELD ELECTROSTATICS */
 906             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 907             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 908
 909             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 910
 911             /* Update potential sum for this i atom from the interaction with this j atom. */
 912             velec            = _mm_and_pd(velec,cutoff_mask);
 913             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 914             velecsum         = _mm_add_pd(velecsum,velec);
 915
 916             fscal            = felec;
 917
 918             fscal            = _mm_and_pd(fscal,cutoff_mask);
 919
 920             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 921
 922             /* Calculate temporary vectorial force */
 923             tx               = _mm_mul_pd(fscal,dx13);
 924             ty               = _mm_mul_pd(fscal,dy13);
 925             tz               = _mm_mul_pd(fscal,dz13);
 926
 927             /* Update vectorial force */
 928             fix1             = _mm_add_pd(fix1,tx);
 929             fiy1             = _mm_add_pd(fiy1,ty);
 930             fiz1             = _mm_add_pd(fiz1,tz);
 931
 932             fjx3             = _mm_add_pd(fjx3,tx);
 933             fjy3             = _mm_add_pd(fjy3,ty);
 934             fjz3             = _mm_add_pd(fjz3,tz);
 935
 936             }
 937
 938             /**************************
 939              * CALCULATE INTERACTIONS *
 940              **************************/
 941
 942             if (gmx_mm_any_lt(rsq21,rcutoff2))
 943             {
 944
 945             /* REACTION-FIELD ELECTROSTATICS */
 946             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 947             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 948
 949             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 950
 951             /* Update potential sum for this i atom from the interaction with this j atom. */
 952             velec            = _mm_and_pd(velec,cutoff_mask);
 953             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 954             velecsum         = _mm_add_pd(velecsum,velec);
 955
 956             fscal            = felec;
 957
 958             fscal            = _mm_and_pd(fscal,cutoff_mask);
 959
 960             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 961
 962             /* Calculate temporary vectorial force */
 963             tx               = _mm_mul_pd(fscal,dx21);
 964             ty               = _mm_mul_pd(fscal,dy21);
 965             tz               = _mm_mul_pd(fscal,dz21);
 966
 967             /* Update vectorial force */
 968             fix2             = _mm_add_pd(fix2,tx);
 969             fiy2             = _mm_add_pd(fiy2,ty);
 970             fiz2             = _mm_add_pd(fiz2,tz);
 971
 972             fjx1             = _mm_add_pd(fjx1,tx);
 973             fjy1             = _mm_add_pd(fjy1,ty);
 974             fjz1             = _mm_add_pd(fjz1,tz);
 975
 976             }
 977
 978             /**************************
 979              * CALCULATE INTERACTIONS *
 980              **************************/
 981
 982             if (gmx_mm_any_lt(rsq22,rcutoff2))
 983             {
 984
 985             /* REACTION-FIELD ELECTROSTATICS */
 986             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 987             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 988
 989             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 990
 991             /* Update potential sum for this i atom from the interaction with this j atom. */
 992             velec            = _mm_and_pd(velec,cutoff_mask);
 993             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 994             velecsum         = _mm_add_pd(velecsum,velec);
 995
 996             fscal            = felec;
 997
 998             fscal            = _mm_and_pd(fscal,cutoff_mask);
 999
1000             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1001
1002             /* Calculate temporary vectorial force */
1003             tx               = _mm_mul_pd(fscal,dx22);
1004             ty               = _mm_mul_pd(fscal,dy22);
1005             tz               = _mm_mul_pd(fscal,dz22);
1006
1007             /* Update vectorial force */
1008             fix2             = _mm_add_pd(fix2,tx);
1009             fiy2             = _mm_add_pd(fiy2,ty);
1010             fiz2             = _mm_add_pd(fiz2,tz);
1011
1012             fjx2             = _mm_add_pd(fjx2,tx);
1013             fjy2             = _mm_add_pd(fjy2,ty);
1014             fjz2             = _mm_add_pd(fjz2,tz);
1015
1016             }
1017
1018             /**************************
1019              * CALCULATE INTERACTIONS *
1020              **************************/
1021
1022             if (gmx_mm_any_lt(rsq23,rcutoff2))
1023             {
1024
1025             /* REACTION-FIELD ELECTROSTATICS */
1026             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
1027             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1028
1029             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1030
1031             /* Update potential sum for this i atom from the interaction with this j atom. */
1032             velec            = _mm_and_pd(velec,cutoff_mask);
1033             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1034             velecsum         = _mm_add_pd(velecsum,velec);
1035
1036             fscal            = felec;
1037
1038             fscal            = _mm_and_pd(fscal,cutoff_mask);
1039
1040             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1041
1042             /* Calculate temporary vectorial force */
1043             tx               = _mm_mul_pd(fscal,dx23);
1044             ty               = _mm_mul_pd(fscal,dy23);
1045             tz               = _mm_mul_pd(fscal,dz23);
1046
1047             /* Update vectorial force */
1048             fix2             = _mm_add_pd(fix2,tx);
1049             fiy2             = _mm_add_pd(fiy2,ty);
1050             fiz2             = _mm_add_pd(fiz2,tz);
1051
1052             fjx3             = _mm_add_pd(fjx3,tx);
1053             fjy3             = _mm_add_pd(fjy3,ty);
1054             fjz3             = _mm_add_pd(fjz3,tz);
1055
1056             }
1057
1058             /**************************
1059              * CALCULATE INTERACTIONS *
1060              **************************/
1061
1062             if (gmx_mm_any_lt(rsq31,rcutoff2))
1063             {
1064
1065             /* REACTION-FIELD ELECTROSTATICS */
1066             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
1067             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1068
1069             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1070
1071             /* Update potential sum for this i atom from the interaction with this j atom. */
1072             velec            = _mm_and_pd(velec,cutoff_mask);
1073             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1074             velecsum         = _mm_add_pd(velecsum,velec);
1075
1076             fscal            = felec;
1077
1078             fscal            = _mm_and_pd(fscal,cutoff_mask);
1079
1080             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1081
1082             /* Calculate temporary vectorial force */
1083             tx               = _mm_mul_pd(fscal,dx31);
1084             ty               = _mm_mul_pd(fscal,dy31);
1085             tz               = _mm_mul_pd(fscal,dz31);
1086
1087             /* Update vectorial force */
1088             fix3             = _mm_add_pd(fix3,tx);
1089             fiy3             = _mm_add_pd(fiy3,ty);
1090             fiz3             = _mm_add_pd(fiz3,tz);
1091
1092             fjx1             = _mm_add_pd(fjx1,tx);
1093             fjy1             = _mm_add_pd(fjy1,ty);
1094             fjz1             = _mm_add_pd(fjz1,tz);
1095
1096             }
1097
1098             /**************************
1099              * CALCULATE INTERACTIONS *
1100              **************************/
1101
1102             if (gmx_mm_any_lt(rsq32,rcutoff2))
1103             {
1104
1105             /* REACTION-FIELD ELECTROSTATICS */
1106             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
1107             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1108
1109             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1110
1111             /* Update potential sum for this i atom from the interaction with this j atom. */
1112             velec            = _mm_and_pd(velec,cutoff_mask);
1113             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1114             velecsum         = _mm_add_pd(velecsum,velec);
1115
1116             fscal            = felec;
1117
1118             fscal            = _mm_and_pd(fscal,cutoff_mask);
1119
1120             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1121
1122             /* Calculate temporary vectorial force */
1123             tx               = _mm_mul_pd(fscal,dx32);
1124             ty               = _mm_mul_pd(fscal,dy32);
1125             tz               = _mm_mul_pd(fscal,dz32);
1126
1127             /* Update vectorial force */
1128             fix3             = _mm_add_pd(fix3,tx);
1129             fiy3             = _mm_add_pd(fiy3,ty);
1130             fiz3             = _mm_add_pd(fiz3,tz);
1131
1132             fjx2             = _mm_add_pd(fjx2,tx);
1133             fjy2             = _mm_add_pd(fjy2,ty);
1134             fjz2             = _mm_add_pd(fjz2,tz);
1135
1136             }
1137
1138             /**************************
1139              * CALCULATE INTERACTIONS *
1140              **************************/
1141
1142             if (gmx_mm_any_lt(rsq33,rcutoff2))
1143             {
1144
1145             /* REACTION-FIELD ELECTROSTATICS */
1146             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1147             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1148
1149             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1150
1151             /* Update potential sum for this i atom from the interaction with this j atom. */
1152             velec            = _mm_and_pd(velec,cutoff_mask);
1153             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1154             velecsum         = _mm_add_pd(velecsum,velec);
1155
1156             fscal            = felec;
1157
1158             fscal            = _mm_and_pd(fscal,cutoff_mask);
1159
1160             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1161
1162             /* Calculate temporary vectorial force */
1163             tx               = _mm_mul_pd(fscal,dx33);
1164             ty               = _mm_mul_pd(fscal,dy33);
1165             tz               = _mm_mul_pd(fscal,dz33);
1166
1167             /* Update vectorial force */
1168             fix3             = _mm_add_pd(fix3,tx);
1169             fiy3             = _mm_add_pd(fiy3,ty);
1170             fiz3             = _mm_add_pd(fiz3,tz);
1171
1172             fjx3             = _mm_add_pd(fjx3,tx);
1173             fjy3             = _mm_add_pd(fjy3,ty);
1174             fjz3             = _mm_add_pd(fjz3,tz);
1175
1176             }
1177
1178             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1179
1180             /* Inner loop uses 368 flops */
1181         }
1182
1183         /* End of innermost loop */
1184
1185         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1186                                               f+i_coord_offset,fshift+i_shift_offset);
1187
1188         ggid                        = gid[iidx];
1189         /* Update potential energies */
1190         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1191         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1192
1193         /* Increment number of inner iterations */
1194         inneriter                  += j_index_end - j_index_start;
1195
1196         /* Outer loop uses 26 flops */
1197     }
1198
1199     /* Increment number of outer iterations */
1200     outeriter        += nri;
1201
1202     /* Update outer/inner flops */
1203
1204     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*368);
1205 }
1206 /*
1207  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_sse2_double
1208  * Electrostatics interaction: ReactionField
1209  * VdW interaction:            LennardJones
1210  * Geometry:                   Water4-Water4
1211  * Calculate force/pot:        Force
1212  */
1213 void
1214 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_sse2_double
1215                     (t_nblist                    * gmx_restrict       nlist,
1216                      rvec                        * gmx_restrict          xx,
1217                      rvec                        * gmx_restrict          ff,
1218                      struct t_forcerec           * gmx_restrict          fr,
1219                      t_mdatoms                   * gmx_restrict     mdatoms,
1220                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1221                      t_nrnb                      * gmx_restrict        nrnb)
1222 {
1223     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1224      * just 0 for non-waters.
1225      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1226      * jnr indices corresponding to data put in the four positions in the SIMD register.
1227      */
1228     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1229     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1230     int              jnrA,jnrB;
1231     int              j_coord_offsetA,j_coord_offsetB;
1232     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1233     real             rcutoff_scalar;
1234     real             *shiftvec,*fshift,*x,*f;
1235     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1236     int              vdwioffset0;
1237     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1238     int              vdwioffset1;
1239     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1240     int              vdwioffset2;
1241     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1242     int              vdwioffset3;
1243     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1244     int              vdwjidx0A,vdwjidx0B;
1245     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1246     int              vdwjidx1A,vdwjidx1B;
1247     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1248     int              vdwjidx2A,vdwjidx2B;
1249     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1250     int              vdwjidx3A,vdwjidx3B;
1251     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1252     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1253     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1254     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1255     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1256     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1257     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1258     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1259     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1260     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1261     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1262     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1263     real             *charge;
1264     int              nvdwtype;
1265     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1266     int              *vdwtype;
1267     real             *vdwparam;
1268     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1269     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1270     __m128d          dummy_mask,cutoff_mask;
1271     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1272     __m128d          one     = _mm_set1_pd(1.0);
1273     __m128d          two     = _mm_set1_pd(2.0);
1274     x                = xx[0];
1275     f                = ff[0];
1276
1277     nri              = nlist->nri;
1278     iinr             = nlist->iinr;
1279     jindex           = nlist->jindex;
1280     jjnr             = nlist->jjnr;
1281     shiftidx         = nlist->shift;
1282     gid              = nlist->gid;
1283     shiftvec         = fr->shift_vec[0];
1284     fshift           = fr->fshift[0];
1285     facel            = _mm_set1_pd(fr->ic->epsfac);
1286     charge           = mdatoms->chargeA;
1287     krf              = _mm_set1_pd(fr->ic->k_rf);
1288     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1289     crf              = _mm_set1_pd(fr->ic->c_rf);
1290     nvdwtype         = fr->ntype;
1291     vdwparam         = fr->nbfp;
1292     vdwtype          = mdatoms->typeA;
1293
1294     /* Setup water-specific parameters */
1295     inr              = nlist->iinr[0];
1296     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1297     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1298     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1299     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1300
1301     jq1              = _mm_set1_pd(charge[inr+1]);
1302     jq2              = _mm_set1_pd(charge[inr+2]);
1303     jq3              = _mm_set1_pd(charge[inr+3]);
1304     vdwjidx0A        = 2*vdwtype[inr+0];
1305     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1306     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1307     qq11             = _mm_mul_pd(iq1,jq1);
1308     qq12             = _mm_mul_pd(iq1,jq2);
1309     qq13             = _mm_mul_pd(iq1,jq3);
1310     qq21             = _mm_mul_pd(iq2,jq1);
1311     qq22             = _mm_mul_pd(iq2,jq2);
1312     qq23             = _mm_mul_pd(iq2,jq3);
1313     qq31             = _mm_mul_pd(iq3,jq1);
1314     qq32             = _mm_mul_pd(iq3,jq2);
1315     qq33             = _mm_mul_pd(iq3,jq3);
1316
1317     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1318     rcutoff_scalar   = fr->ic->rcoulomb;
1319     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1320     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1321
1322     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
1323     rvdw             = _mm_set1_pd(fr->ic->rvdw);
1324
1325     /* Avoid stupid compiler warnings */
1326     jnrA = jnrB = 0;
1327     j_coord_offsetA = 0;
1328     j_coord_offsetB = 0;
1329
1330     outeriter        = 0;
1331     inneriter        = 0;
1332
1333     /* Start outer loop over neighborlists */
1334     for(iidx=0; iidx<nri; iidx++)
1335     {
1336         /* Load shift vector for this list */
1337         i_shift_offset   = DIM*shiftidx[iidx];
1338
1339         /* Load limits for loop over neighbors */
1340         j_index_start    = jindex[iidx];
1341         j_index_end      = jindex[iidx+1];
1342
1343         /* Get outer coordinate index */
1344         inr              = iinr[iidx];
1345         i_coord_offset   = DIM*inr;
1346
1347         /* Load i particle coords and add shift vector */
1348         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1349                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1350
1351         fix0             = _mm_setzero_pd();
1352         fiy0             = _mm_setzero_pd();
1353         fiz0             = _mm_setzero_pd();
1354         fix1             = _mm_setzero_pd();
1355         fiy1             = _mm_setzero_pd();
1356         fiz1             = _mm_setzero_pd();
1357         fix2             = _mm_setzero_pd();
1358         fiy2             = _mm_setzero_pd();
1359         fiz2             = _mm_setzero_pd();
1360         fix3             = _mm_setzero_pd();
1361         fiy3             = _mm_setzero_pd();
1362         fiz3             = _mm_setzero_pd();
1363
1364         /* Start inner kernel loop */
1365         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1366         {
1367
1368             /* Get j neighbor index, and coordinate index */
1369             jnrA             = jjnr[jidx];
1370             jnrB             = jjnr[jidx+1];
1371             j_coord_offsetA  = DIM*jnrA;
1372             j_coord_offsetB  = DIM*jnrB;
1373
1374             /* load j atom coordinates */
1375             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1376                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1377                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1378
1379             /* Calculate displacement vector */
1380             dx00             = _mm_sub_pd(ix0,jx0);
1381             dy00             = _mm_sub_pd(iy0,jy0);
1382             dz00             = _mm_sub_pd(iz0,jz0);
1383             dx11             = _mm_sub_pd(ix1,jx1);
1384             dy11             = _mm_sub_pd(iy1,jy1);
1385             dz11             = _mm_sub_pd(iz1,jz1);
1386             dx12             = _mm_sub_pd(ix1,jx2);
1387             dy12             = _mm_sub_pd(iy1,jy2);
1388             dz12             = _mm_sub_pd(iz1,jz2);
1389             dx13             = _mm_sub_pd(ix1,jx3);
1390             dy13             = _mm_sub_pd(iy1,jy3);
1391             dz13             = _mm_sub_pd(iz1,jz3);
1392             dx21             = _mm_sub_pd(ix2,jx1);
1393             dy21             = _mm_sub_pd(iy2,jy1);
1394             dz21             = _mm_sub_pd(iz2,jz1);
1395             dx22             = _mm_sub_pd(ix2,jx2);
1396             dy22             = _mm_sub_pd(iy2,jy2);
1397             dz22             = _mm_sub_pd(iz2,jz2);
1398             dx23             = _mm_sub_pd(ix2,jx3);
1399             dy23             = _mm_sub_pd(iy2,jy3);
1400             dz23             = _mm_sub_pd(iz2,jz3);
1401             dx31             = _mm_sub_pd(ix3,jx1);
1402             dy31             = _mm_sub_pd(iy3,jy1);
1403             dz31             = _mm_sub_pd(iz3,jz1);
1404             dx32             = _mm_sub_pd(ix3,jx2);
1405             dy32             = _mm_sub_pd(iy3,jy2);
1406             dz32             = _mm_sub_pd(iz3,jz2);
1407             dx33             = _mm_sub_pd(ix3,jx3);
1408             dy33             = _mm_sub_pd(iy3,jy3);
1409             dz33             = _mm_sub_pd(iz3,jz3);
1410
1411             /* Calculate squared distance and things based on it */
1412             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1413             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1414             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1415             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1416             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1417             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1418             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1419             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1420             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1421             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1422
1423             rinv11           = sse2_invsqrt_d(rsq11);
1424             rinv12           = sse2_invsqrt_d(rsq12);
1425             rinv13           = sse2_invsqrt_d(rsq13);
1426             rinv21           = sse2_invsqrt_d(rsq21);
1427             rinv22           = sse2_invsqrt_d(rsq22);
1428             rinv23           = sse2_invsqrt_d(rsq23);
1429             rinv31           = sse2_invsqrt_d(rsq31);
1430             rinv32           = sse2_invsqrt_d(rsq32);
1431             rinv33           = sse2_invsqrt_d(rsq33);
1432
1433             rinvsq00         = sse2_inv_d(rsq00);
1434             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1435             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1436             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1437             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1438             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1439             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1440             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1441             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1442             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1443
1444             fjx0             = _mm_setzero_pd();
1445             fjy0             = _mm_setzero_pd();
1446             fjz0             = _mm_setzero_pd();
1447             fjx1             = _mm_setzero_pd();
1448             fjy1             = _mm_setzero_pd();
1449             fjz1             = _mm_setzero_pd();
1450             fjx2             = _mm_setzero_pd();
1451             fjy2             = _mm_setzero_pd();
1452             fjz2             = _mm_setzero_pd();
1453             fjx3             = _mm_setzero_pd();
1454             fjy3             = _mm_setzero_pd();
1455             fjz3             = _mm_setzero_pd();
1456
1457             /**************************
1458              * CALCULATE INTERACTIONS *
1459              **************************/
1460
1461             if (gmx_mm_any_lt(rsq00,rcutoff2))
1462             {
1463
1464             /* LENNARD-JONES DISPERSION/REPULSION */
1465
1466             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1467             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1468
1469             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1470
1471             fscal            = fvdw;
1472
1473             fscal            = _mm_and_pd(fscal,cutoff_mask);
1474
1475             /* Calculate temporary vectorial force */
1476             tx               = _mm_mul_pd(fscal,dx00);
1477             ty               = _mm_mul_pd(fscal,dy00);
1478             tz               = _mm_mul_pd(fscal,dz00);
1479
1480             /* Update vectorial force */
1481             fix0             = _mm_add_pd(fix0,tx);
1482             fiy0             = _mm_add_pd(fiy0,ty);
1483             fiz0             = _mm_add_pd(fiz0,tz);
1484
1485             fjx0             = _mm_add_pd(fjx0,tx);
1486             fjy0             = _mm_add_pd(fjy0,ty);
1487             fjz0             = _mm_add_pd(fjz0,tz);
1488
1489             }
1490
1491             /**************************
1492              * CALCULATE INTERACTIONS *
1493              **************************/
1494
1495             if (gmx_mm_any_lt(rsq11,rcutoff2))
1496             {
1497
1498             /* REACTION-FIELD ELECTROSTATICS */
1499             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1500
1501             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1502
1503             fscal            = felec;
1504
1505             fscal            = _mm_and_pd(fscal,cutoff_mask);
1506
1507             /* Calculate temporary vectorial force */
1508             tx               = _mm_mul_pd(fscal,dx11);
1509             ty               = _mm_mul_pd(fscal,dy11);
1510             tz               = _mm_mul_pd(fscal,dz11);
1511
1512             /* Update vectorial force */
1513             fix1             = _mm_add_pd(fix1,tx);
1514             fiy1             = _mm_add_pd(fiy1,ty);
1515             fiz1             = _mm_add_pd(fiz1,tz);
1516
1517             fjx1             = _mm_add_pd(fjx1,tx);
1518             fjy1             = _mm_add_pd(fjy1,ty);
1519             fjz1             = _mm_add_pd(fjz1,tz);
1520
1521             }
1522
1523             /**************************
1524              * CALCULATE INTERACTIONS *
1525              **************************/
1526
1527             if (gmx_mm_any_lt(rsq12,rcutoff2))
1528             {
1529
1530             /* REACTION-FIELD ELECTROSTATICS */
1531             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1532
1533             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1534
1535             fscal            = felec;
1536
1537             fscal            = _mm_and_pd(fscal,cutoff_mask);
1538
1539             /* Calculate temporary vectorial force */
1540             tx               = _mm_mul_pd(fscal,dx12);
1541             ty               = _mm_mul_pd(fscal,dy12);
1542             tz               = _mm_mul_pd(fscal,dz12);
1543
1544             /* Update vectorial force */
1545             fix1             = _mm_add_pd(fix1,tx);
1546             fiy1             = _mm_add_pd(fiy1,ty);
1547             fiz1             = _mm_add_pd(fiz1,tz);
1548
1549             fjx2             = _mm_add_pd(fjx2,tx);
1550             fjy2             = _mm_add_pd(fjy2,ty);
1551             fjz2             = _mm_add_pd(fjz2,tz);
1552
1553             }
1554
1555             /**************************
1556              * CALCULATE INTERACTIONS *
1557              **************************/
1558
1559             if (gmx_mm_any_lt(rsq13,rcutoff2))
1560             {
1561
1562             /* REACTION-FIELD ELECTROSTATICS */
1563             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1564
1565             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1566
1567             fscal            = felec;
1568
1569             fscal            = _mm_and_pd(fscal,cutoff_mask);
1570
1571             /* Calculate temporary vectorial force */
1572             tx               = _mm_mul_pd(fscal,dx13);
1573             ty               = _mm_mul_pd(fscal,dy13);
1574             tz               = _mm_mul_pd(fscal,dz13);
1575
1576             /* Update vectorial force */
1577             fix1             = _mm_add_pd(fix1,tx);
1578             fiy1             = _mm_add_pd(fiy1,ty);
1579             fiz1             = _mm_add_pd(fiz1,tz);
1580
1581             fjx3             = _mm_add_pd(fjx3,tx);
1582             fjy3             = _mm_add_pd(fjy3,ty);
1583             fjz3             = _mm_add_pd(fjz3,tz);
1584
1585             }
1586
1587             /**************************
1588              * CALCULATE INTERACTIONS *
1589              **************************/
1590
1591             if (gmx_mm_any_lt(rsq21,rcutoff2))
1592             {
1593
1594             /* REACTION-FIELD ELECTROSTATICS */
1595             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1596
1597             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1598
1599             fscal            = felec;
1600
1601             fscal            = _mm_and_pd(fscal,cutoff_mask);
1602
1603             /* Calculate temporary vectorial force */
1604             tx               = _mm_mul_pd(fscal,dx21);
1605             ty               = _mm_mul_pd(fscal,dy21);
1606             tz               = _mm_mul_pd(fscal,dz21);
1607
1608             /* Update vectorial force */
1609             fix2             = _mm_add_pd(fix2,tx);
1610             fiy2             = _mm_add_pd(fiy2,ty);
1611             fiz2             = _mm_add_pd(fiz2,tz);
1612
1613             fjx1             = _mm_add_pd(fjx1,tx);
1614             fjy1             = _mm_add_pd(fjy1,ty);
1615             fjz1             = _mm_add_pd(fjz1,tz);
1616
1617             }
1618
1619             /**************************
1620              * CALCULATE INTERACTIONS *
1621              **************************/
1622
1623             if (gmx_mm_any_lt(rsq22,rcutoff2))
1624             {
1625
1626             /* REACTION-FIELD ELECTROSTATICS */
1627             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1628
1629             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1630
1631             fscal            = felec;
1632
1633             fscal            = _mm_and_pd(fscal,cutoff_mask);
1634
1635             /* Calculate temporary vectorial force */
1636             tx               = _mm_mul_pd(fscal,dx22);
1637             ty               = _mm_mul_pd(fscal,dy22);
1638             tz               = _mm_mul_pd(fscal,dz22);
1639
1640             /* Update vectorial force */
1641             fix2             = _mm_add_pd(fix2,tx);
1642             fiy2             = _mm_add_pd(fiy2,ty);
1643             fiz2             = _mm_add_pd(fiz2,tz);
1644
1645             fjx2             = _mm_add_pd(fjx2,tx);
1646             fjy2             = _mm_add_pd(fjy2,ty);
1647             fjz2             = _mm_add_pd(fjz2,tz);
1648
1649             }
1650
1651             /**************************
1652              * CALCULATE INTERACTIONS *
1653              **************************/
1654
1655             if (gmx_mm_any_lt(rsq23,rcutoff2))
1656             {
1657
1658             /* REACTION-FIELD ELECTROSTATICS */
1659             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1660
1661             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1662
1663             fscal            = felec;
1664
1665             fscal            = _mm_and_pd(fscal,cutoff_mask);
1666
1667             /* Calculate temporary vectorial force */
1668             tx               = _mm_mul_pd(fscal,dx23);
1669             ty               = _mm_mul_pd(fscal,dy23);
1670             tz               = _mm_mul_pd(fscal,dz23);
1671
1672             /* Update vectorial force */
1673             fix2             = _mm_add_pd(fix2,tx);
1674             fiy2             = _mm_add_pd(fiy2,ty);
1675             fiz2             = _mm_add_pd(fiz2,tz);
1676
1677             fjx3             = _mm_add_pd(fjx3,tx);
1678             fjy3             = _mm_add_pd(fjy3,ty);
1679             fjz3             = _mm_add_pd(fjz3,tz);
1680
1681             }
1682
1683             /**************************
1684              * CALCULATE INTERACTIONS *
1685              **************************/
1686
1687             if (gmx_mm_any_lt(rsq31,rcutoff2))
1688             {
1689
1690             /* REACTION-FIELD ELECTROSTATICS */
1691             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1692
1693             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1694
1695             fscal            = felec;
1696
1697             fscal            = _mm_and_pd(fscal,cutoff_mask);
1698
1699             /* Calculate temporary vectorial force */
1700             tx               = _mm_mul_pd(fscal,dx31);
1701             ty               = _mm_mul_pd(fscal,dy31);
1702             tz               = _mm_mul_pd(fscal,dz31);
1703
1704             /* Update vectorial force */
1705             fix3             = _mm_add_pd(fix3,tx);
1706             fiy3             = _mm_add_pd(fiy3,ty);
1707             fiz3             = _mm_add_pd(fiz3,tz);
1708
1709             fjx1             = _mm_add_pd(fjx1,tx);
1710             fjy1             = _mm_add_pd(fjy1,ty);
1711             fjz1             = _mm_add_pd(fjz1,tz);
1712
1713             }
1714
1715             /**************************
1716              * CALCULATE INTERACTIONS *
1717              **************************/
1718
1719             if (gmx_mm_any_lt(rsq32,rcutoff2))
1720             {
1721
1722             /* REACTION-FIELD ELECTROSTATICS */
1723             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1724
1725             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_and_pd(fscal,cutoff_mask);
1730
1731             /* Calculate temporary vectorial force */
1732             tx               = _mm_mul_pd(fscal,dx32);
1733             ty               = _mm_mul_pd(fscal,dy32);
1734             tz               = _mm_mul_pd(fscal,dz32);
1735
1736             /* Update vectorial force */
1737             fix3             = _mm_add_pd(fix3,tx);
1738             fiy3             = _mm_add_pd(fiy3,ty);
1739             fiz3             = _mm_add_pd(fiz3,tz);
1740
1741             fjx2             = _mm_add_pd(fjx2,tx);
1742             fjy2             = _mm_add_pd(fjy2,ty);
1743             fjz2             = _mm_add_pd(fjz2,tz);
1744
1745             }
1746
1747             /**************************
1748              * CALCULATE INTERACTIONS *
1749              **************************/
1750
1751             if (gmx_mm_any_lt(rsq33,rcutoff2))
1752             {
1753
1754             /* REACTION-FIELD ELECTROSTATICS */
1755             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1756
1757             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1758
1759             fscal            = felec;
1760
1761             fscal            = _mm_and_pd(fscal,cutoff_mask);
1762
1763             /* Calculate temporary vectorial force */
1764             tx               = _mm_mul_pd(fscal,dx33);
1765             ty               = _mm_mul_pd(fscal,dy33);
1766             tz               = _mm_mul_pd(fscal,dz33);
1767
1768             /* Update vectorial force */
1769             fix3             = _mm_add_pd(fix3,tx);
1770             fiy3             = _mm_add_pd(fiy3,ty);
1771             fiz3             = _mm_add_pd(fiz3,tz);
1772
1773             fjx3             = _mm_add_pd(fjx3,tx);
1774             fjy3             = _mm_add_pd(fjy3,ty);
1775             fjz3             = _mm_add_pd(fjz3,tz);
1776
1777             }
1778
1779             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1780
1781             /* Inner loop uses 303 flops */
1782         }
1783
1784         if(jidx<j_index_end)
1785         {
1786
1787             jnrA             = jjnr[jidx];
1788             j_coord_offsetA  = DIM*jnrA;
1789
1790             /* load j atom coordinates */
1791             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1792                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1793                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1794
1795             /* Calculate displacement vector */
1796             dx00             = _mm_sub_pd(ix0,jx0);
1797             dy00             = _mm_sub_pd(iy0,jy0);
1798             dz00             = _mm_sub_pd(iz0,jz0);
1799             dx11             = _mm_sub_pd(ix1,jx1);
1800             dy11             = _mm_sub_pd(iy1,jy1);
1801             dz11             = _mm_sub_pd(iz1,jz1);
1802             dx12             = _mm_sub_pd(ix1,jx2);
1803             dy12             = _mm_sub_pd(iy1,jy2);
1804             dz12             = _mm_sub_pd(iz1,jz2);
1805             dx13             = _mm_sub_pd(ix1,jx3);
1806             dy13             = _mm_sub_pd(iy1,jy3);
1807             dz13             = _mm_sub_pd(iz1,jz3);
1808             dx21             = _mm_sub_pd(ix2,jx1);
1809             dy21             = _mm_sub_pd(iy2,jy1);
1810             dz21             = _mm_sub_pd(iz2,jz1);
1811             dx22             = _mm_sub_pd(ix2,jx2);
1812             dy22             = _mm_sub_pd(iy2,jy2);
1813             dz22             = _mm_sub_pd(iz2,jz2);
1814             dx23             = _mm_sub_pd(ix2,jx3);
1815             dy23             = _mm_sub_pd(iy2,jy3);
1816             dz23             = _mm_sub_pd(iz2,jz3);
1817             dx31             = _mm_sub_pd(ix3,jx1);
1818             dy31             = _mm_sub_pd(iy3,jy1);
1819             dz31             = _mm_sub_pd(iz3,jz1);
1820             dx32             = _mm_sub_pd(ix3,jx2);
1821             dy32             = _mm_sub_pd(iy3,jy2);
1822             dz32             = _mm_sub_pd(iz3,jz2);
1823             dx33             = _mm_sub_pd(ix3,jx3);
1824             dy33             = _mm_sub_pd(iy3,jy3);
1825             dz33             = _mm_sub_pd(iz3,jz3);
1826
1827             /* Calculate squared distance and things based on it */
1828             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1829             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1830             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1831             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1832             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1833             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1834             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1835             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1836             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1837             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1838
1839             rinv11           = sse2_invsqrt_d(rsq11);
1840             rinv12           = sse2_invsqrt_d(rsq12);
1841             rinv13           = sse2_invsqrt_d(rsq13);
1842             rinv21           = sse2_invsqrt_d(rsq21);
1843             rinv22           = sse2_invsqrt_d(rsq22);
1844             rinv23           = sse2_invsqrt_d(rsq23);
1845             rinv31           = sse2_invsqrt_d(rsq31);
1846             rinv32           = sse2_invsqrt_d(rsq32);
1847             rinv33           = sse2_invsqrt_d(rsq33);
1848
1849             rinvsq00         = sse2_inv_d(rsq00);
1850             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1851             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1852             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1853             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1854             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1855             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1856             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1857             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1858             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1859
1860             fjx0             = _mm_setzero_pd();
1861             fjy0             = _mm_setzero_pd();
1862             fjz0             = _mm_setzero_pd();
1863             fjx1             = _mm_setzero_pd();
1864             fjy1             = _mm_setzero_pd();
1865             fjz1             = _mm_setzero_pd();
1866             fjx2             = _mm_setzero_pd();
1867             fjy2             = _mm_setzero_pd();
1868             fjz2             = _mm_setzero_pd();
1869             fjx3             = _mm_setzero_pd();
1870             fjy3             = _mm_setzero_pd();
1871             fjz3             = _mm_setzero_pd();
1872
1873             /**************************
1874              * CALCULATE INTERACTIONS *
1875              **************************/
1876
1877             if (gmx_mm_any_lt(rsq00,rcutoff2))
1878             {
1879
1880             /* LENNARD-JONES DISPERSION/REPULSION */
1881
1882             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1883             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1884
1885             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1886
1887             fscal            = fvdw;
1888
1889             fscal            = _mm_and_pd(fscal,cutoff_mask);
1890
1891             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1892
1893             /* Calculate temporary vectorial force */
1894             tx               = _mm_mul_pd(fscal,dx00);
1895             ty               = _mm_mul_pd(fscal,dy00);
1896             tz               = _mm_mul_pd(fscal,dz00);
1897
1898             /* Update vectorial force */
1899             fix0             = _mm_add_pd(fix0,tx);
1900             fiy0             = _mm_add_pd(fiy0,ty);
1901             fiz0             = _mm_add_pd(fiz0,tz);
1902
1903             fjx0             = _mm_add_pd(fjx0,tx);
1904             fjy0             = _mm_add_pd(fjy0,ty);
1905             fjz0             = _mm_add_pd(fjz0,tz);
1906
1907             }
1908
1909             /**************************
1910              * CALCULATE INTERACTIONS *
1911              **************************/
1912
1913             if (gmx_mm_any_lt(rsq11,rcutoff2))
1914             {
1915
1916             /* REACTION-FIELD ELECTROSTATICS */
1917             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1918
1919             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1920
1921             fscal            = felec;
1922
1923             fscal            = _mm_and_pd(fscal,cutoff_mask);
1924
1925             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1926
1927             /* Calculate temporary vectorial force */
1928             tx               = _mm_mul_pd(fscal,dx11);
1929             ty               = _mm_mul_pd(fscal,dy11);
1930             tz               = _mm_mul_pd(fscal,dz11);
1931
1932             /* Update vectorial force */
1933             fix1             = _mm_add_pd(fix1,tx);
1934             fiy1             = _mm_add_pd(fiy1,ty);
1935             fiz1             = _mm_add_pd(fiz1,tz);
1936
1937             fjx1             = _mm_add_pd(fjx1,tx);
1938             fjy1             = _mm_add_pd(fjy1,ty);
1939             fjz1             = _mm_add_pd(fjz1,tz);
1940
1941             }
1942
1943             /**************************
1944              * CALCULATE INTERACTIONS *
1945              **************************/
1946
1947             if (gmx_mm_any_lt(rsq12,rcutoff2))
1948             {
1949
1950             /* REACTION-FIELD ELECTROSTATICS */
1951             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1952
1953             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1954
1955             fscal            = felec;
1956
1957             fscal            = _mm_and_pd(fscal,cutoff_mask);
1958
1959             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1960
1961             /* Calculate temporary vectorial force */
1962             tx               = _mm_mul_pd(fscal,dx12);
1963             ty               = _mm_mul_pd(fscal,dy12);
1964             tz               = _mm_mul_pd(fscal,dz12);
1965
1966             /* Update vectorial force */
1967             fix1             = _mm_add_pd(fix1,tx);
1968             fiy1             = _mm_add_pd(fiy1,ty);
1969             fiz1             = _mm_add_pd(fiz1,tz);
1970
1971             fjx2             = _mm_add_pd(fjx2,tx);
1972             fjy2             = _mm_add_pd(fjy2,ty);
1973             fjz2             = _mm_add_pd(fjz2,tz);
1974
1975             }
1976
1977             /**************************
1978              * CALCULATE INTERACTIONS *
1979              **************************/
1980
1981             if (gmx_mm_any_lt(rsq13,rcutoff2))
1982             {
1983
1984             /* REACTION-FIELD ELECTROSTATICS */
1985             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1986
1987             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1988
1989             fscal            = felec;
1990
1991             fscal            = _mm_and_pd(fscal,cutoff_mask);
1992
1993             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1994
1995             /* Calculate temporary vectorial force */
1996             tx               = _mm_mul_pd(fscal,dx13);
1997             ty               = _mm_mul_pd(fscal,dy13);
1998             tz               = _mm_mul_pd(fscal,dz13);
1999
2000             /* Update vectorial force */
2001             fix1             = _mm_add_pd(fix1,tx);
2002             fiy1             = _mm_add_pd(fiy1,ty);
2003             fiz1             = _mm_add_pd(fiz1,tz);
2004
2005             fjx3             = _mm_add_pd(fjx3,tx);
2006             fjy3             = _mm_add_pd(fjy3,ty);
2007             fjz3             = _mm_add_pd(fjz3,tz);
2008
2009             }
2010
2011             /**************************
2012              * CALCULATE INTERACTIONS *
2013              **************************/
2014
2015             if (gmx_mm_any_lt(rsq21,rcutoff2))
2016             {
2017
2018             /* REACTION-FIELD ELECTROSTATICS */
2019             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2020
2021             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2022
2023             fscal            = felec;
2024
2025             fscal            = _mm_and_pd(fscal,cutoff_mask);
2026
2027             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2028
2029             /* Calculate temporary vectorial force */
2030             tx               = _mm_mul_pd(fscal,dx21);
2031             ty               = _mm_mul_pd(fscal,dy21);
2032             tz               = _mm_mul_pd(fscal,dz21);
2033
2034             /* Update vectorial force */
2035             fix2             = _mm_add_pd(fix2,tx);
2036             fiy2             = _mm_add_pd(fiy2,ty);
2037             fiz2             = _mm_add_pd(fiz2,tz);
2038
2039             fjx1             = _mm_add_pd(fjx1,tx);
2040             fjy1             = _mm_add_pd(fjy1,ty);
2041             fjz1             = _mm_add_pd(fjz1,tz);
2042
2043             }
2044
2045             /**************************
2046              * CALCULATE INTERACTIONS *
2047              **************************/
2048
2049             if (gmx_mm_any_lt(rsq22,rcutoff2))
2050             {
2051
2052             /* REACTION-FIELD ELECTROSTATICS */
2053             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2054
2055             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2056
2057             fscal            = felec;
2058
2059             fscal            = _mm_and_pd(fscal,cutoff_mask);
2060
2061             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2062
2063             /* Calculate temporary vectorial force */
2064             tx               = _mm_mul_pd(fscal,dx22);
2065             ty               = _mm_mul_pd(fscal,dy22);
2066             tz               = _mm_mul_pd(fscal,dz22);
2067
2068             /* Update vectorial force */
2069             fix2             = _mm_add_pd(fix2,tx);
2070             fiy2             = _mm_add_pd(fiy2,ty);
2071             fiz2             = _mm_add_pd(fiz2,tz);
2072
2073             fjx2             = _mm_add_pd(fjx2,tx);
2074             fjy2             = _mm_add_pd(fjy2,ty);
2075             fjz2             = _mm_add_pd(fjz2,tz);
2076
2077             }
2078
2079             /**************************
2080              * CALCULATE INTERACTIONS *
2081              **************************/
2082
2083             if (gmx_mm_any_lt(rsq23,rcutoff2))
2084             {
2085
2086             /* REACTION-FIELD ELECTROSTATICS */
2087             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
2088
2089             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
2090
2091             fscal            = felec;
2092
2093             fscal            = _mm_and_pd(fscal,cutoff_mask);
2094
2095             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2096
2097             /* Calculate temporary vectorial force */
2098             tx               = _mm_mul_pd(fscal,dx23);
2099             ty               = _mm_mul_pd(fscal,dy23);
2100             tz               = _mm_mul_pd(fscal,dz23);
2101
2102             /* Update vectorial force */
2103             fix2             = _mm_add_pd(fix2,tx);
2104             fiy2             = _mm_add_pd(fiy2,ty);
2105             fiz2             = _mm_add_pd(fiz2,tz);
2106
2107             fjx3             = _mm_add_pd(fjx3,tx);
2108             fjy3             = _mm_add_pd(fjy3,ty);
2109             fjz3             = _mm_add_pd(fjz3,tz);
2110
2111             }
2112
2113             /**************************
2114              * CALCULATE INTERACTIONS *
2115              **************************/
2116
2117             if (gmx_mm_any_lt(rsq31,rcutoff2))
2118             {
2119
2120             /* REACTION-FIELD ELECTROSTATICS */
2121             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
2122
2123             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2124
2125             fscal            = felec;
2126
2127             fscal            = _mm_and_pd(fscal,cutoff_mask);
2128
2129             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2130
2131             /* Calculate temporary vectorial force */
2132             tx               = _mm_mul_pd(fscal,dx31);
2133             ty               = _mm_mul_pd(fscal,dy31);
2134             tz               = _mm_mul_pd(fscal,dz31);
2135
2136             /* Update vectorial force */
2137             fix3             = _mm_add_pd(fix3,tx);
2138             fiy3             = _mm_add_pd(fiy3,ty);
2139             fiz3             = _mm_add_pd(fiz3,tz);
2140
2141             fjx1             = _mm_add_pd(fjx1,tx);
2142             fjy1             = _mm_add_pd(fjy1,ty);
2143             fjz1             = _mm_add_pd(fjz1,tz);
2144
2145             }
2146
2147             /**************************
2148              * CALCULATE INTERACTIONS *
2149              **************************/
2150
2151             if (gmx_mm_any_lt(rsq32,rcutoff2))
2152             {
2153
2154             /* REACTION-FIELD ELECTROSTATICS */
2155             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
2156
2157             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2158
2159             fscal            = felec;
2160
2161             fscal            = _mm_and_pd(fscal,cutoff_mask);
2162
2163             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2164
2165             /* Calculate temporary vectorial force */
2166             tx               = _mm_mul_pd(fscal,dx32);
2167             ty               = _mm_mul_pd(fscal,dy32);
2168             tz               = _mm_mul_pd(fscal,dz32);
2169
2170             /* Update vectorial force */
2171             fix3             = _mm_add_pd(fix3,tx);
2172             fiy3             = _mm_add_pd(fiy3,ty);
2173             fiz3             = _mm_add_pd(fiz3,tz);
2174
2175             fjx2             = _mm_add_pd(fjx2,tx);
2176             fjy2             = _mm_add_pd(fjy2,ty);
2177             fjz2             = _mm_add_pd(fjz2,tz);
2178
2179             }
2180
2181             /**************************
2182              * CALCULATE INTERACTIONS *
2183              **************************/
2184
2185             if (gmx_mm_any_lt(rsq33,rcutoff2))
2186             {
2187
2188             /* REACTION-FIELD ELECTROSTATICS */
2189             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
2190
2191             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2192
2193             fscal            = felec;
2194
2195             fscal            = _mm_and_pd(fscal,cutoff_mask);
2196
2197             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2198
2199             /* Calculate temporary vectorial force */
2200             tx               = _mm_mul_pd(fscal,dx33);
2201             ty               = _mm_mul_pd(fscal,dy33);
2202             tz               = _mm_mul_pd(fscal,dz33);
2203
2204             /* Update vectorial force */
2205             fix3             = _mm_add_pd(fix3,tx);
2206             fiy3             = _mm_add_pd(fiy3,ty);
2207             fiz3             = _mm_add_pd(fiz3,tz);
2208
2209             fjx3             = _mm_add_pd(fjx3,tx);
2210             fjy3             = _mm_add_pd(fjy3,ty);
2211             fjz3             = _mm_add_pd(fjz3,tz);
2212
2213             }
2214
2215             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2216
2217             /* Inner loop uses 303 flops */
2218         }
2219
2220         /* End of innermost loop */
2221
2222         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2223                                               f+i_coord_offset,fshift+i_shift_offset);
2224
2225         /* Increment number of inner iterations */
2226         inneriter                  += j_index_end - j_index_start;
2227
2228         /* Outer loop uses 24 flops */
2229     }
2230
2231     /* Increment number of outer iterations */
2232     outeriter        += nri;
2233
2234     /* Update outer/inner flops */
2235
2236     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
2237 }