src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_sse4_1_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse4_1_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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse4_1_double.h"
  50 #include "kernelutil_x86_sse4_1_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse4_1_double
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse4_1_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B;
  91     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B;
  93     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B;
  95     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B;
  97     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 116     __m128d          dummy_mask,cutoff_mask;
 117     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 118     __m128d          one     = _mm_set1_pd(1.0);
 119     __m128d          two     = _mm_set1_pd(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_pd(fr->epsfac);
 132     charge           = mdatoms->chargeA;
 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     /* Avoid stupid compiler warnings */
 161     jnrA = jnrB = 0;
 162     j_coord_offsetA = 0;
 163     j_coord_offsetB = 0;
 164
 165     outeriter        = 0;
 166     inneriter        = 0;
 167
 168     /* Start outer loop over neighborlists */
 169     for(iidx=0; iidx<nri; iidx++)
 170     {
 171         /* Load shift vector for this list */
 172         i_shift_offset   = DIM*shiftidx[iidx];
 173
 174         /* Load limits for loop over neighbors */
 175         j_index_start    = jindex[iidx];
 176         j_index_end      = jindex[iidx+1];
 177
 178         /* Get outer coordinate index */
 179         inr              = iinr[iidx];
 180         i_coord_offset   = DIM*inr;
 181
 182         /* Load i particle coords and add shift vector */
 183         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 184                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 185
 186         fix0             = _mm_setzero_pd();
 187         fiy0             = _mm_setzero_pd();
 188         fiz0             = _mm_setzero_pd();
 189         fix1             = _mm_setzero_pd();
 190         fiy1             = _mm_setzero_pd();
 191         fiz1             = _mm_setzero_pd();
 192         fix2             = _mm_setzero_pd();
 193         fiy2             = _mm_setzero_pd();
 194         fiz2             = _mm_setzero_pd();
 195         fix3             = _mm_setzero_pd();
 196         fiy3             = _mm_setzero_pd();
 197         fiz3             = _mm_setzero_pd();
 198
 199         /* Reset potential sums */
 200         velecsum         = _mm_setzero_pd();
 201         vvdwsum          = _mm_setzero_pd();
 202
 203         /* Start inner kernel loop */
 204         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 205         {
 206
 207             /* Get j neighbor index, and coordinate index */
 208             jnrA             = jjnr[jidx];
 209             jnrB             = jjnr[jidx+1];
 210             j_coord_offsetA  = DIM*jnrA;
 211             j_coord_offsetB  = DIM*jnrB;
 212
 213             /* load j atom coordinates */
 214             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 215                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 216                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 217
 218             /* Calculate displacement vector */
 219             dx00             = _mm_sub_pd(ix0,jx0);
 220             dy00             = _mm_sub_pd(iy0,jy0);
 221             dz00             = _mm_sub_pd(iz0,jz0);
 222             dx11             = _mm_sub_pd(ix1,jx1);
 223             dy11             = _mm_sub_pd(iy1,jy1);
 224             dz11             = _mm_sub_pd(iz1,jz1);
 225             dx12             = _mm_sub_pd(ix1,jx2);
 226             dy12             = _mm_sub_pd(iy1,jy2);
 227             dz12             = _mm_sub_pd(iz1,jz2);
 228             dx13             = _mm_sub_pd(ix1,jx3);
 229             dy13             = _mm_sub_pd(iy1,jy3);
 230             dz13             = _mm_sub_pd(iz1,jz3);
 231             dx21             = _mm_sub_pd(ix2,jx1);
 232             dy21             = _mm_sub_pd(iy2,jy1);
 233             dz21             = _mm_sub_pd(iz2,jz1);
 234             dx22             = _mm_sub_pd(ix2,jx2);
 235             dy22             = _mm_sub_pd(iy2,jy2);
 236             dz22             = _mm_sub_pd(iz2,jz2);
 237             dx23             = _mm_sub_pd(ix2,jx3);
 238             dy23             = _mm_sub_pd(iy2,jy3);
 239             dz23             = _mm_sub_pd(iz2,jz3);
 240             dx31             = _mm_sub_pd(ix3,jx1);
 241             dy31             = _mm_sub_pd(iy3,jy1);
 242             dz31             = _mm_sub_pd(iz3,jz1);
 243             dx32             = _mm_sub_pd(ix3,jx2);
 244             dy32             = _mm_sub_pd(iy3,jy2);
 245             dz32             = _mm_sub_pd(iz3,jz2);
 246             dx33             = _mm_sub_pd(ix3,jx3);
 247             dy33             = _mm_sub_pd(iy3,jy3);
 248             dz33             = _mm_sub_pd(iz3,jz3);
 249
 250             /* Calculate squared distance and things based on it */
 251             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 252             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 253             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 254             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 255             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 256             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 257             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 258             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 259             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 260             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 261
 262             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 263             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 264             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 265             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 266             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 267             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 268             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 269             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 270             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 271
 272             rinvsq00         = gmx_mm_inv_pd(rsq00);
 273             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 274             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 275             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 276             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 277             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 278             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 279             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 280             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 281             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 282
 283             fjx0             = _mm_setzero_pd();
 284             fjy0             = _mm_setzero_pd();
 285             fjz0             = _mm_setzero_pd();
 286             fjx1             = _mm_setzero_pd();
 287             fjy1             = _mm_setzero_pd();
 288             fjz1             = _mm_setzero_pd();
 289             fjx2             = _mm_setzero_pd();
 290             fjy2             = _mm_setzero_pd();
 291             fjz2             = _mm_setzero_pd();
 292             fjx3             = _mm_setzero_pd();
 293             fjy3             = _mm_setzero_pd();
 294             fjz3             = _mm_setzero_pd();
 295
 296             /**************************
 297              * CALCULATE INTERACTIONS *
 298              **************************/
 299
 300             /* LENNARD-JONES DISPERSION/REPULSION */
 301
 302             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 303             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 304             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 305             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 306             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 307
 308             /* Update potential sum for this i atom from the interaction with this j atom. */
 309             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 310
 311             fscal            = fvdw;
 312
 313             /* Calculate temporary vectorial force */
 314             tx               = _mm_mul_pd(fscal,dx00);
 315             ty               = _mm_mul_pd(fscal,dy00);
 316             tz               = _mm_mul_pd(fscal,dz00);
 317
 318             /* Update vectorial force */
 319             fix0             = _mm_add_pd(fix0,tx);
 320             fiy0             = _mm_add_pd(fiy0,ty);
 321             fiz0             = _mm_add_pd(fiz0,tz);
 322
 323             fjx0             = _mm_add_pd(fjx0,tx);
 324             fjy0             = _mm_add_pd(fjy0,ty);
 325             fjz0             = _mm_add_pd(fjz0,tz);
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             /* COULOMB ELECTROSTATICS */
 332             velec            = _mm_mul_pd(qq11,rinv11);
 333             felec            = _mm_mul_pd(velec,rinvsq11);
 334
 335             /* Update potential sum for this i atom from the interaction with this j atom. */
 336             velecsum         = _mm_add_pd(velecsum,velec);
 337
 338             fscal            = felec;
 339
 340             /* Calculate temporary vectorial force */
 341             tx               = _mm_mul_pd(fscal,dx11);
 342             ty               = _mm_mul_pd(fscal,dy11);
 343             tz               = _mm_mul_pd(fscal,dz11);
 344
 345             /* Update vectorial force */
 346             fix1             = _mm_add_pd(fix1,tx);
 347             fiy1             = _mm_add_pd(fiy1,ty);
 348             fiz1             = _mm_add_pd(fiz1,tz);
 349
 350             fjx1             = _mm_add_pd(fjx1,tx);
 351             fjy1             = _mm_add_pd(fjy1,ty);
 352             fjz1             = _mm_add_pd(fjz1,tz);
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             /* COULOMB ELECTROSTATICS */
 359             velec            = _mm_mul_pd(qq12,rinv12);
 360             felec            = _mm_mul_pd(velec,rinvsq12);
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velecsum         = _mm_add_pd(velecsum,velec);
 364
 365             fscal            = felec;
 366
 367             /* Calculate temporary vectorial force */
 368             tx               = _mm_mul_pd(fscal,dx12);
 369             ty               = _mm_mul_pd(fscal,dy12);
 370             tz               = _mm_mul_pd(fscal,dz12);
 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             fjx2             = _mm_add_pd(fjx2,tx);
 378             fjy2             = _mm_add_pd(fjy2,ty);
 379             fjz2             = _mm_add_pd(fjz2,tz);
 380
 381             /**************************
 382              * CALCULATE INTERACTIONS *
 383              **************************/
 384
 385             /* COULOMB ELECTROSTATICS */
 386             velec            = _mm_mul_pd(qq13,rinv13);
 387             felec            = _mm_mul_pd(velec,rinvsq13);
 388
 389             /* Update potential sum for this i atom from the interaction with this j atom. */
 390             velecsum         = _mm_add_pd(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             /* Calculate temporary vectorial force */
 395             tx               = _mm_mul_pd(fscal,dx13);
 396             ty               = _mm_mul_pd(fscal,dy13);
 397             tz               = _mm_mul_pd(fscal,dz13);
 398
 399             /* Update vectorial force */
 400             fix1             = _mm_add_pd(fix1,tx);
 401             fiy1             = _mm_add_pd(fiy1,ty);
 402             fiz1             = _mm_add_pd(fiz1,tz);
 403
 404             fjx3             = _mm_add_pd(fjx3,tx);
 405             fjy3             = _mm_add_pd(fjy3,ty);
 406             fjz3             = _mm_add_pd(fjz3,tz);
 407
 408             /**************************
 409              * CALCULATE INTERACTIONS *
 410              **************************/
 411
 412             /* COULOMB ELECTROSTATICS */
 413             velec            = _mm_mul_pd(qq21,rinv21);
 414             felec            = _mm_mul_pd(velec,rinvsq21);
 415
 416             /* Update potential sum for this i atom from the interaction with this j atom. */
 417             velecsum         = _mm_add_pd(velecsum,velec);
 418
 419             fscal            = felec;
 420
 421             /* Calculate temporary vectorial force */
 422             tx               = _mm_mul_pd(fscal,dx21);
 423             ty               = _mm_mul_pd(fscal,dy21);
 424             tz               = _mm_mul_pd(fscal,dz21);
 425
 426             /* Update vectorial force */
 427             fix2             = _mm_add_pd(fix2,tx);
 428             fiy2             = _mm_add_pd(fiy2,ty);
 429             fiz2             = _mm_add_pd(fiz2,tz);
 430
 431             fjx1             = _mm_add_pd(fjx1,tx);
 432             fjy1             = _mm_add_pd(fjy1,ty);
 433             fjz1             = _mm_add_pd(fjz1,tz);
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             /* COULOMB ELECTROSTATICS */
 440             velec            = _mm_mul_pd(qq22,rinv22);
 441             felec            = _mm_mul_pd(velec,rinvsq22);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velecsum         = _mm_add_pd(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448             /* Calculate temporary vectorial force */
 449             tx               = _mm_mul_pd(fscal,dx22);
 450             ty               = _mm_mul_pd(fscal,dy22);
 451             tz               = _mm_mul_pd(fscal,dz22);
 452
 453             /* Update vectorial force */
 454             fix2             = _mm_add_pd(fix2,tx);
 455             fiy2             = _mm_add_pd(fiy2,ty);
 456             fiz2             = _mm_add_pd(fiz2,tz);
 457
 458             fjx2             = _mm_add_pd(fjx2,tx);
 459             fjy2             = _mm_add_pd(fjy2,ty);
 460             fjz2             = _mm_add_pd(fjz2,tz);
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             /* COULOMB ELECTROSTATICS */
 467             velec            = _mm_mul_pd(qq23,rinv23);
 468             felec            = _mm_mul_pd(velec,rinvsq23);
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velecsum         = _mm_add_pd(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475             /* Calculate temporary vectorial force */
 476             tx               = _mm_mul_pd(fscal,dx23);
 477             ty               = _mm_mul_pd(fscal,dy23);
 478             tz               = _mm_mul_pd(fscal,dz23);
 479
 480             /* Update vectorial force */
 481             fix2             = _mm_add_pd(fix2,tx);
 482             fiy2             = _mm_add_pd(fiy2,ty);
 483             fiz2             = _mm_add_pd(fiz2,tz);
 484
 485             fjx3             = _mm_add_pd(fjx3,tx);
 486             fjy3             = _mm_add_pd(fjy3,ty);
 487             fjz3             = _mm_add_pd(fjz3,tz);
 488
 489             /**************************
 490              * CALCULATE INTERACTIONS *
 491              **************************/
 492
 493             /* COULOMB ELECTROSTATICS */
 494             velec            = _mm_mul_pd(qq31,rinv31);
 495             felec            = _mm_mul_pd(velec,rinvsq31);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velecsum         = _mm_add_pd(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502             /* Calculate temporary vectorial force */
 503             tx               = _mm_mul_pd(fscal,dx31);
 504             ty               = _mm_mul_pd(fscal,dy31);
 505             tz               = _mm_mul_pd(fscal,dz31);
 506
 507             /* Update vectorial force */
 508             fix3             = _mm_add_pd(fix3,tx);
 509             fiy3             = _mm_add_pd(fiy3,ty);
 510             fiz3             = _mm_add_pd(fiz3,tz);
 511
 512             fjx1             = _mm_add_pd(fjx1,tx);
 513             fjy1             = _mm_add_pd(fjy1,ty);
 514             fjz1             = _mm_add_pd(fjz1,tz);
 515
 516             /**************************
 517              * CALCULATE INTERACTIONS *
 518              **************************/
 519
 520             /* COULOMB ELECTROSTATICS */
 521             velec            = _mm_mul_pd(qq32,rinv32);
 522             felec            = _mm_mul_pd(velec,rinvsq32);
 523
 524             /* Update potential sum for this i atom from the interaction with this j atom. */
 525             velecsum         = _mm_add_pd(velecsum,velec);
 526
 527             fscal            = felec;
 528
 529             /* Calculate temporary vectorial force */
 530             tx               = _mm_mul_pd(fscal,dx32);
 531             ty               = _mm_mul_pd(fscal,dy32);
 532             tz               = _mm_mul_pd(fscal,dz32);
 533
 534             /* Update vectorial force */
 535             fix3             = _mm_add_pd(fix3,tx);
 536             fiy3             = _mm_add_pd(fiy3,ty);
 537             fiz3             = _mm_add_pd(fiz3,tz);
 538
 539             fjx2             = _mm_add_pd(fjx2,tx);
 540             fjy2             = _mm_add_pd(fjy2,ty);
 541             fjz2             = _mm_add_pd(fjz2,tz);
 542
 543             /**************************
 544              * CALCULATE INTERACTIONS *
 545              **************************/
 546
 547             /* COULOMB ELECTROSTATICS */
 548             velec            = _mm_mul_pd(qq33,rinv33);
 549             felec            = _mm_mul_pd(velec,rinvsq33);
 550
 551             /* Update potential sum for this i atom from the interaction with this j atom. */
 552             velecsum         = _mm_add_pd(velecsum,velec);
 553
 554             fscal            = felec;
 555
 556             /* Calculate temporary vectorial force */
 557             tx               = _mm_mul_pd(fscal,dx33);
 558             ty               = _mm_mul_pd(fscal,dy33);
 559             tz               = _mm_mul_pd(fscal,dz33);
 560
 561             /* Update vectorial force */
 562             fix3             = _mm_add_pd(fix3,tx);
 563             fiy3             = _mm_add_pd(fiy3,ty);
 564             fiz3             = _mm_add_pd(fiz3,tz);
 565
 566             fjx3             = _mm_add_pd(fjx3,tx);
 567             fjy3             = _mm_add_pd(fjy3,ty);
 568             fjz3             = _mm_add_pd(fjz3,tz);
 569
 570             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);
 571
 572             /* Inner loop uses 287 flops */
 573         }
 574
 575         if(jidx<j_index_end)
 576         {
 577
 578             jnrA             = jjnr[jidx];
 579             j_coord_offsetA  = DIM*jnrA;
 580
 581             /* load j atom coordinates */
 582             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 583                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 584                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 585
 586             /* Calculate displacement vector */
 587             dx00             = _mm_sub_pd(ix0,jx0);
 588             dy00             = _mm_sub_pd(iy0,jy0);
 589             dz00             = _mm_sub_pd(iz0,jz0);
 590             dx11             = _mm_sub_pd(ix1,jx1);
 591             dy11             = _mm_sub_pd(iy1,jy1);
 592             dz11             = _mm_sub_pd(iz1,jz1);
 593             dx12             = _mm_sub_pd(ix1,jx2);
 594             dy12             = _mm_sub_pd(iy1,jy2);
 595             dz12             = _mm_sub_pd(iz1,jz2);
 596             dx13             = _mm_sub_pd(ix1,jx3);
 597             dy13             = _mm_sub_pd(iy1,jy3);
 598             dz13             = _mm_sub_pd(iz1,jz3);
 599             dx21             = _mm_sub_pd(ix2,jx1);
 600             dy21             = _mm_sub_pd(iy2,jy1);
 601             dz21             = _mm_sub_pd(iz2,jz1);
 602             dx22             = _mm_sub_pd(ix2,jx2);
 603             dy22             = _mm_sub_pd(iy2,jy2);
 604             dz22             = _mm_sub_pd(iz2,jz2);
 605             dx23             = _mm_sub_pd(ix2,jx3);
 606             dy23             = _mm_sub_pd(iy2,jy3);
 607             dz23             = _mm_sub_pd(iz2,jz3);
 608             dx31             = _mm_sub_pd(ix3,jx1);
 609             dy31             = _mm_sub_pd(iy3,jy1);
 610             dz31             = _mm_sub_pd(iz3,jz1);
 611             dx32             = _mm_sub_pd(ix3,jx2);
 612             dy32             = _mm_sub_pd(iy3,jy2);
 613             dz32             = _mm_sub_pd(iz3,jz2);
 614             dx33             = _mm_sub_pd(ix3,jx3);
 615             dy33             = _mm_sub_pd(iy3,jy3);
 616             dz33             = _mm_sub_pd(iz3,jz3);
 617
 618             /* Calculate squared distance and things based on it */
 619             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 620             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 621             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 622             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 623             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 624             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 625             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 626             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 627             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 628             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 629
 630             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 631             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 632             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 633             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 634             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 635             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 636             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 637             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 638             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 639
 640             rinvsq00         = gmx_mm_inv_pd(rsq00);
 641             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 642             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 643             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 644             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 645             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 646             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 647             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 648             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 649             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 650
 651             fjx0             = _mm_setzero_pd();
 652             fjy0             = _mm_setzero_pd();
 653             fjz0             = _mm_setzero_pd();
 654             fjx1             = _mm_setzero_pd();
 655             fjy1             = _mm_setzero_pd();
 656             fjz1             = _mm_setzero_pd();
 657             fjx2             = _mm_setzero_pd();
 658             fjy2             = _mm_setzero_pd();
 659             fjz2             = _mm_setzero_pd();
 660             fjx3             = _mm_setzero_pd();
 661             fjy3             = _mm_setzero_pd();
 662             fjz3             = _mm_setzero_pd();
 663
 664             /**************************
 665              * CALCULATE INTERACTIONS *
 666              **************************/
 667
 668             /* LENNARD-JONES DISPERSION/REPULSION */
 669
 670             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 671             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 672             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 673             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 674             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 675
 676             /* Update potential sum for this i atom from the interaction with this j atom. */
 677             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 678             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 679
 680             fscal            = fvdw;
 681
 682             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 683
 684             /* Calculate temporary vectorial force */
 685             tx               = _mm_mul_pd(fscal,dx00);
 686             ty               = _mm_mul_pd(fscal,dy00);
 687             tz               = _mm_mul_pd(fscal,dz00);
 688
 689             /* Update vectorial force */
 690             fix0             = _mm_add_pd(fix0,tx);
 691             fiy0             = _mm_add_pd(fiy0,ty);
 692             fiz0             = _mm_add_pd(fiz0,tz);
 693
 694             fjx0             = _mm_add_pd(fjx0,tx);
 695             fjy0             = _mm_add_pd(fjy0,ty);
 696             fjz0             = _mm_add_pd(fjz0,tz);
 697
 698             /**************************
 699              * CALCULATE INTERACTIONS *
 700              **************************/
 701
 702             /* COULOMB ELECTROSTATICS */
 703             velec            = _mm_mul_pd(qq11,rinv11);
 704             felec            = _mm_mul_pd(velec,rinvsq11);
 705
 706             /* Update potential sum for this i atom from the interaction with this j atom. */
 707             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 708             velecsum         = _mm_add_pd(velecsum,velec);
 709
 710             fscal            = felec;
 711
 712             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 713
 714             /* Calculate temporary vectorial force */
 715             tx               = _mm_mul_pd(fscal,dx11);
 716             ty               = _mm_mul_pd(fscal,dy11);
 717             tz               = _mm_mul_pd(fscal,dz11);
 718
 719             /* Update vectorial force */
 720             fix1             = _mm_add_pd(fix1,tx);
 721             fiy1             = _mm_add_pd(fiy1,ty);
 722             fiz1             = _mm_add_pd(fiz1,tz);
 723
 724             fjx1             = _mm_add_pd(fjx1,tx);
 725             fjy1             = _mm_add_pd(fjy1,ty);
 726             fjz1             = _mm_add_pd(fjz1,tz);
 727
 728             /**************************
 729              * CALCULATE INTERACTIONS *
 730              **************************/
 731
 732             /* COULOMB ELECTROSTATICS */
 733             velec            = _mm_mul_pd(qq12,rinv12);
 734             felec            = _mm_mul_pd(velec,rinvsq12);
 735
 736             /* Update potential sum for this i atom from the interaction with this j atom. */
 737             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 738             velecsum         = _mm_add_pd(velecsum,velec);
 739
 740             fscal            = felec;
 741
 742             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 743
 744             /* Calculate temporary vectorial force */
 745             tx               = _mm_mul_pd(fscal,dx12);
 746             ty               = _mm_mul_pd(fscal,dy12);
 747             tz               = _mm_mul_pd(fscal,dz12);
 748
 749             /* Update vectorial force */
 750             fix1             = _mm_add_pd(fix1,tx);
 751             fiy1             = _mm_add_pd(fiy1,ty);
 752             fiz1             = _mm_add_pd(fiz1,tz);
 753
 754             fjx2             = _mm_add_pd(fjx2,tx);
 755             fjy2             = _mm_add_pd(fjy2,ty);
 756             fjz2             = _mm_add_pd(fjz2,tz);
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             /* COULOMB ELECTROSTATICS */
 763             velec            = _mm_mul_pd(qq13,rinv13);
 764             felec            = _mm_mul_pd(velec,rinvsq13);
 765
 766             /* Update potential sum for this i atom from the interaction with this j atom. */
 767             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 768             velecsum         = _mm_add_pd(velecsum,velec);
 769
 770             fscal            = felec;
 771
 772             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 773
 774             /* Calculate temporary vectorial force */
 775             tx               = _mm_mul_pd(fscal,dx13);
 776             ty               = _mm_mul_pd(fscal,dy13);
 777             tz               = _mm_mul_pd(fscal,dz13);
 778
 779             /* Update vectorial force */
 780             fix1             = _mm_add_pd(fix1,tx);
 781             fiy1             = _mm_add_pd(fiy1,ty);
 782             fiz1             = _mm_add_pd(fiz1,tz);
 783
 784             fjx3             = _mm_add_pd(fjx3,tx);
 785             fjy3             = _mm_add_pd(fjy3,ty);
 786             fjz3             = _mm_add_pd(fjz3,tz);
 787
 788             /**************************
 789              * CALCULATE INTERACTIONS *
 790              **************************/
 791
 792             /* COULOMB ELECTROSTATICS */
 793             velec            = _mm_mul_pd(qq21,rinv21);
 794             felec            = _mm_mul_pd(velec,rinvsq21);
 795
 796             /* Update potential sum for this i atom from the interaction with this j atom. */
 797             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 798             velecsum         = _mm_add_pd(velecsum,velec);
 799
 800             fscal            = felec;
 801
 802             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 803
 804             /* Calculate temporary vectorial force */
 805             tx               = _mm_mul_pd(fscal,dx21);
 806             ty               = _mm_mul_pd(fscal,dy21);
 807             tz               = _mm_mul_pd(fscal,dz21);
 808
 809             /* Update vectorial force */
 810             fix2             = _mm_add_pd(fix2,tx);
 811             fiy2             = _mm_add_pd(fiy2,ty);
 812             fiz2             = _mm_add_pd(fiz2,tz);
 813
 814             fjx1             = _mm_add_pd(fjx1,tx);
 815             fjy1             = _mm_add_pd(fjy1,ty);
 816             fjz1             = _mm_add_pd(fjz1,tz);
 817
 818             /**************************
 819              * CALCULATE INTERACTIONS *
 820              **************************/
 821
 822             /* COULOMB ELECTROSTATICS */
 823             velec            = _mm_mul_pd(qq22,rinv22);
 824             felec            = _mm_mul_pd(velec,rinvsq22);
 825
 826             /* Update potential sum for this i atom from the interaction with this j atom. */
 827             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 828             velecsum         = _mm_add_pd(velecsum,velec);
 829
 830             fscal            = felec;
 831
 832             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 833
 834             /* Calculate temporary vectorial force */
 835             tx               = _mm_mul_pd(fscal,dx22);
 836             ty               = _mm_mul_pd(fscal,dy22);
 837             tz               = _mm_mul_pd(fscal,dz22);
 838
 839             /* Update vectorial force */
 840             fix2             = _mm_add_pd(fix2,tx);
 841             fiy2             = _mm_add_pd(fiy2,ty);
 842             fiz2             = _mm_add_pd(fiz2,tz);
 843
 844             fjx2             = _mm_add_pd(fjx2,tx);
 845             fjy2             = _mm_add_pd(fjy2,ty);
 846             fjz2             = _mm_add_pd(fjz2,tz);
 847
 848             /**************************
 849              * CALCULATE INTERACTIONS *
 850              **************************/
 851
 852             /* COULOMB ELECTROSTATICS */
 853             velec            = _mm_mul_pd(qq23,rinv23);
 854             felec            = _mm_mul_pd(velec,rinvsq23);
 855
 856             /* Update potential sum for this i atom from the interaction with this j atom. */
 857             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 858             velecsum         = _mm_add_pd(velecsum,velec);
 859
 860             fscal            = felec;
 861
 862             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 863
 864             /* Calculate temporary vectorial force */
 865             tx               = _mm_mul_pd(fscal,dx23);
 866             ty               = _mm_mul_pd(fscal,dy23);
 867             tz               = _mm_mul_pd(fscal,dz23);
 868
 869             /* Update vectorial force */
 870             fix2             = _mm_add_pd(fix2,tx);
 871             fiy2             = _mm_add_pd(fiy2,ty);
 872             fiz2             = _mm_add_pd(fiz2,tz);
 873
 874             fjx3             = _mm_add_pd(fjx3,tx);
 875             fjy3             = _mm_add_pd(fjy3,ty);
 876             fjz3             = _mm_add_pd(fjz3,tz);
 877
 878             /**************************
 879              * CALCULATE INTERACTIONS *
 880              **************************/
 881
 882             /* COULOMB ELECTROSTATICS */
 883             velec            = _mm_mul_pd(qq31,rinv31);
 884             felec            = _mm_mul_pd(velec,rinvsq31);
 885
 886             /* Update potential sum for this i atom from the interaction with this j atom. */
 887             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 888             velecsum         = _mm_add_pd(velecsum,velec);
 889
 890             fscal            = felec;
 891
 892             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 893
 894             /* Calculate temporary vectorial force */
 895             tx               = _mm_mul_pd(fscal,dx31);
 896             ty               = _mm_mul_pd(fscal,dy31);
 897             tz               = _mm_mul_pd(fscal,dz31);
 898
 899             /* Update vectorial force */
 900             fix3             = _mm_add_pd(fix3,tx);
 901             fiy3             = _mm_add_pd(fiy3,ty);
 902             fiz3             = _mm_add_pd(fiz3,tz);
 903
 904             fjx1             = _mm_add_pd(fjx1,tx);
 905             fjy1             = _mm_add_pd(fjy1,ty);
 906             fjz1             = _mm_add_pd(fjz1,tz);
 907
 908             /**************************
 909              * CALCULATE INTERACTIONS *
 910              **************************/
 911
 912             /* COULOMB ELECTROSTATICS */
 913             velec            = _mm_mul_pd(qq32,rinv32);
 914             felec            = _mm_mul_pd(velec,rinvsq32);
 915
 916             /* Update potential sum for this i atom from the interaction with this j atom. */
 917             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 918             velecsum         = _mm_add_pd(velecsum,velec);
 919
 920             fscal            = felec;
 921
 922             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 923
 924             /* Calculate temporary vectorial force */
 925             tx               = _mm_mul_pd(fscal,dx32);
 926             ty               = _mm_mul_pd(fscal,dy32);
 927             tz               = _mm_mul_pd(fscal,dz32);
 928
 929             /* Update vectorial force */
 930             fix3             = _mm_add_pd(fix3,tx);
 931             fiy3             = _mm_add_pd(fiy3,ty);
 932             fiz3             = _mm_add_pd(fiz3,tz);
 933
 934             fjx2             = _mm_add_pd(fjx2,tx);
 935             fjy2             = _mm_add_pd(fjy2,ty);
 936             fjz2             = _mm_add_pd(fjz2,tz);
 937
 938             /**************************
 939              * CALCULATE INTERACTIONS *
 940              **************************/
 941
 942             /* COULOMB ELECTROSTATICS */
 943             velec            = _mm_mul_pd(qq33,rinv33);
 944             felec            = _mm_mul_pd(velec,rinvsq33);
 945
 946             /* Update potential sum for this i atom from the interaction with this j atom. */
 947             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 948             velecsum         = _mm_add_pd(velecsum,velec);
 949
 950             fscal            = felec;
 951
 952             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 953
 954             /* Calculate temporary vectorial force */
 955             tx               = _mm_mul_pd(fscal,dx33);
 956             ty               = _mm_mul_pd(fscal,dy33);
 957             tz               = _mm_mul_pd(fscal,dz33);
 958
 959             /* Update vectorial force */
 960             fix3             = _mm_add_pd(fix3,tx);
 961             fiy3             = _mm_add_pd(fiy3,ty);
 962             fiz3             = _mm_add_pd(fiz3,tz);
 963
 964             fjx3             = _mm_add_pd(fjx3,tx);
 965             fjy3             = _mm_add_pd(fjy3,ty);
 966             fjz3             = _mm_add_pd(fjz3,tz);
 967
 968             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 969
 970             /* Inner loop uses 287 flops */
 971         }
 972
 973         /* End of innermost loop */
 974
 975         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 976                                               f+i_coord_offset,fshift+i_shift_offset);
 977
 978         ggid                        = gid[iidx];
 979         /* Update potential energies */
 980         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 981         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 982
 983         /* Increment number of inner iterations */
 984         inneriter                  += j_index_end - j_index_start;
 985
 986         /* Outer loop uses 26 flops */
 987     }
 988
 989     /* Increment number of outer iterations */
 990     outeriter        += nri;
 991
 992     /* Update outer/inner flops */
 993
 994     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*287);
 995 }
 996 /*
 997  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_double
 998  * Electrostatics interaction: Coulomb
 999  * VdW interaction:            LennardJones
1000  * Geometry:                   Water4-Water4
1001  * Calculate force/pot:        Force
1002  */
1003 void
1004 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_double
1005                     (t_nblist                    * gmx_restrict       nlist,
1006                      rvec                        * gmx_restrict          xx,
1007                      rvec                        * gmx_restrict          ff,
1008                      t_forcerec                  * gmx_restrict          fr,
1009                      t_mdatoms                   * gmx_restrict     mdatoms,
1010                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1011                      t_nrnb                      * gmx_restrict        nrnb)
1012 {
1013     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1014      * just 0 for non-waters.
1015      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1016      * jnr indices corresponding to data put in the four positions in the SIMD register.
1017      */
1018     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1019     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1020     int              jnrA,jnrB;
1021     int              j_coord_offsetA,j_coord_offsetB;
1022     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1023     real             rcutoff_scalar;
1024     real             *shiftvec,*fshift,*x,*f;
1025     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1026     int              vdwioffset0;
1027     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1028     int              vdwioffset1;
1029     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1030     int              vdwioffset2;
1031     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1032     int              vdwioffset3;
1033     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1034     int              vdwjidx0A,vdwjidx0B;
1035     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1036     int              vdwjidx1A,vdwjidx1B;
1037     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1038     int              vdwjidx2A,vdwjidx2B;
1039     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1040     int              vdwjidx3A,vdwjidx3B;
1041     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1042     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1043     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1044     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1045     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1046     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1047     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1048     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1049     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1050     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1051     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1052     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1053     real             *charge;
1054     int              nvdwtype;
1055     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1056     int              *vdwtype;
1057     real             *vdwparam;
1058     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1059     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1060     __m128d          dummy_mask,cutoff_mask;
1061     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1062     __m128d          one     = _mm_set1_pd(1.0);
1063     __m128d          two     = _mm_set1_pd(2.0);
1064     x                = xx[0];
1065     f                = ff[0];
1066
1067     nri              = nlist->nri;
1068     iinr             = nlist->iinr;
1069     jindex           = nlist->jindex;
1070     jjnr             = nlist->jjnr;
1071     shiftidx         = nlist->shift;
1072     gid              = nlist->gid;
1073     shiftvec         = fr->shift_vec[0];
1074     fshift           = fr->fshift[0];
1075     facel            = _mm_set1_pd(fr->epsfac);
1076     charge           = mdatoms->chargeA;
1077     nvdwtype         = fr->ntype;
1078     vdwparam         = fr->nbfp;
1079     vdwtype          = mdatoms->typeA;
1080
1081     /* Setup water-specific parameters */
1082     inr              = nlist->iinr[0];
1083     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1084     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1085     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1086     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1087
1088     jq1              = _mm_set1_pd(charge[inr+1]);
1089     jq2              = _mm_set1_pd(charge[inr+2]);
1090     jq3              = _mm_set1_pd(charge[inr+3]);
1091     vdwjidx0A        = 2*vdwtype[inr+0];
1092     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1093     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1094     qq11             = _mm_mul_pd(iq1,jq1);
1095     qq12             = _mm_mul_pd(iq1,jq2);
1096     qq13             = _mm_mul_pd(iq1,jq3);
1097     qq21             = _mm_mul_pd(iq2,jq1);
1098     qq22             = _mm_mul_pd(iq2,jq2);
1099     qq23             = _mm_mul_pd(iq2,jq3);
1100     qq31             = _mm_mul_pd(iq3,jq1);
1101     qq32             = _mm_mul_pd(iq3,jq2);
1102     qq33             = _mm_mul_pd(iq3,jq3);
1103
1104     /* Avoid stupid compiler warnings */
1105     jnrA = jnrB = 0;
1106     j_coord_offsetA = 0;
1107     j_coord_offsetB = 0;
1108
1109     outeriter        = 0;
1110     inneriter        = 0;
1111
1112     /* Start outer loop over neighborlists */
1113     for(iidx=0; iidx<nri; iidx++)
1114     {
1115         /* Load shift vector for this list */
1116         i_shift_offset   = DIM*shiftidx[iidx];
1117
1118         /* Load limits for loop over neighbors */
1119         j_index_start    = jindex[iidx];
1120         j_index_end      = jindex[iidx+1];
1121
1122         /* Get outer coordinate index */
1123         inr              = iinr[iidx];
1124         i_coord_offset   = DIM*inr;
1125
1126         /* Load i particle coords and add shift vector */
1127         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1128                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1129
1130         fix0             = _mm_setzero_pd();
1131         fiy0             = _mm_setzero_pd();
1132         fiz0             = _mm_setzero_pd();
1133         fix1             = _mm_setzero_pd();
1134         fiy1             = _mm_setzero_pd();
1135         fiz1             = _mm_setzero_pd();
1136         fix2             = _mm_setzero_pd();
1137         fiy2             = _mm_setzero_pd();
1138         fiz2             = _mm_setzero_pd();
1139         fix3             = _mm_setzero_pd();
1140         fiy3             = _mm_setzero_pd();
1141         fiz3             = _mm_setzero_pd();
1142
1143         /* Start inner kernel loop */
1144         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1145         {
1146
1147             /* Get j neighbor index, and coordinate index */
1148             jnrA             = jjnr[jidx];
1149             jnrB             = jjnr[jidx+1];
1150             j_coord_offsetA  = DIM*jnrA;
1151             j_coord_offsetB  = DIM*jnrB;
1152
1153             /* load j atom coordinates */
1154             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1155                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1156                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1157
1158             /* Calculate displacement vector */
1159             dx00             = _mm_sub_pd(ix0,jx0);
1160             dy00             = _mm_sub_pd(iy0,jy0);
1161             dz00             = _mm_sub_pd(iz0,jz0);
1162             dx11             = _mm_sub_pd(ix1,jx1);
1163             dy11             = _mm_sub_pd(iy1,jy1);
1164             dz11             = _mm_sub_pd(iz1,jz1);
1165             dx12             = _mm_sub_pd(ix1,jx2);
1166             dy12             = _mm_sub_pd(iy1,jy2);
1167             dz12             = _mm_sub_pd(iz1,jz2);
1168             dx13             = _mm_sub_pd(ix1,jx3);
1169             dy13             = _mm_sub_pd(iy1,jy3);
1170             dz13             = _mm_sub_pd(iz1,jz3);
1171             dx21             = _mm_sub_pd(ix2,jx1);
1172             dy21             = _mm_sub_pd(iy2,jy1);
1173             dz21             = _mm_sub_pd(iz2,jz1);
1174             dx22             = _mm_sub_pd(ix2,jx2);
1175             dy22             = _mm_sub_pd(iy2,jy2);
1176             dz22             = _mm_sub_pd(iz2,jz2);
1177             dx23             = _mm_sub_pd(ix2,jx3);
1178             dy23             = _mm_sub_pd(iy2,jy3);
1179             dz23             = _mm_sub_pd(iz2,jz3);
1180             dx31             = _mm_sub_pd(ix3,jx1);
1181             dy31             = _mm_sub_pd(iy3,jy1);
1182             dz31             = _mm_sub_pd(iz3,jz1);
1183             dx32             = _mm_sub_pd(ix3,jx2);
1184             dy32             = _mm_sub_pd(iy3,jy2);
1185             dz32             = _mm_sub_pd(iz3,jz2);
1186             dx33             = _mm_sub_pd(ix3,jx3);
1187             dy33             = _mm_sub_pd(iy3,jy3);
1188             dz33             = _mm_sub_pd(iz3,jz3);
1189
1190             /* Calculate squared distance and things based on it */
1191             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1192             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1193             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1194             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1195             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1196             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1197             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1198             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1199             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1200             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1201
1202             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1203             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1204             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1205             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1206             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1207             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1208             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1209             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1210             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1211
1212             rinvsq00         = gmx_mm_inv_pd(rsq00);
1213             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1214             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1215             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1216             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1217             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1218             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1219             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1220             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1221             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1222
1223             fjx0             = _mm_setzero_pd();
1224             fjy0             = _mm_setzero_pd();
1225             fjz0             = _mm_setzero_pd();
1226             fjx1             = _mm_setzero_pd();
1227             fjy1             = _mm_setzero_pd();
1228             fjz1             = _mm_setzero_pd();
1229             fjx2             = _mm_setzero_pd();
1230             fjy2             = _mm_setzero_pd();
1231             fjz2             = _mm_setzero_pd();
1232             fjx3             = _mm_setzero_pd();
1233             fjy3             = _mm_setzero_pd();
1234             fjz3             = _mm_setzero_pd();
1235
1236             /**************************
1237              * CALCULATE INTERACTIONS *
1238              **************************/
1239
1240             /* LENNARD-JONES DISPERSION/REPULSION */
1241
1242             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1243             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1244
1245             fscal            = fvdw;
1246
1247             /* Calculate temporary vectorial force */
1248             tx               = _mm_mul_pd(fscal,dx00);
1249             ty               = _mm_mul_pd(fscal,dy00);
1250             tz               = _mm_mul_pd(fscal,dz00);
1251
1252             /* Update vectorial force */
1253             fix0             = _mm_add_pd(fix0,tx);
1254             fiy0             = _mm_add_pd(fiy0,ty);
1255             fiz0             = _mm_add_pd(fiz0,tz);
1256
1257             fjx0             = _mm_add_pd(fjx0,tx);
1258             fjy0             = _mm_add_pd(fjy0,ty);
1259             fjz0             = _mm_add_pd(fjz0,tz);
1260
1261             /**************************
1262              * CALCULATE INTERACTIONS *
1263              **************************/
1264
1265             /* COULOMB ELECTROSTATICS */
1266             velec            = _mm_mul_pd(qq11,rinv11);
1267             felec            = _mm_mul_pd(velec,rinvsq11);
1268
1269             fscal            = felec;
1270
1271             /* Calculate temporary vectorial force */
1272             tx               = _mm_mul_pd(fscal,dx11);
1273             ty               = _mm_mul_pd(fscal,dy11);
1274             tz               = _mm_mul_pd(fscal,dz11);
1275
1276             /* Update vectorial force */
1277             fix1             = _mm_add_pd(fix1,tx);
1278             fiy1             = _mm_add_pd(fiy1,ty);
1279             fiz1             = _mm_add_pd(fiz1,tz);
1280
1281             fjx1             = _mm_add_pd(fjx1,tx);
1282             fjy1             = _mm_add_pd(fjy1,ty);
1283             fjz1             = _mm_add_pd(fjz1,tz);
1284
1285             /**************************
1286              * CALCULATE INTERACTIONS *
1287              **************************/
1288
1289             /* COULOMB ELECTROSTATICS */
1290             velec            = _mm_mul_pd(qq12,rinv12);
1291             felec            = _mm_mul_pd(velec,rinvsq12);
1292
1293             fscal            = felec;
1294
1295             /* Calculate temporary vectorial force */
1296             tx               = _mm_mul_pd(fscal,dx12);
1297             ty               = _mm_mul_pd(fscal,dy12);
1298             tz               = _mm_mul_pd(fscal,dz12);
1299
1300             /* Update vectorial force */
1301             fix1             = _mm_add_pd(fix1,tx);
1302             fiy1             = _mm_add_pd(fiy1,ty);
1303             fiz1             = _mm_add_pd(fiz1,tz);
1304
1305             fjx2             = _mm_add_pd(fjx2,tx);
1306             fjy2             = _mm_add_pd(fjy2,ty);
1307             fjz2             = _mm_add_pd(fjz2,tz);
1308
1309             /**************************
1310              * CALCULATE INTERACTIONS *
1311              **************************/
1312
1313             /* COULOMB ELECTROSTATICS */
1314             velec            = _mm_mul_pd(qq13,rinv13);
1315             felec            = _mm_mul_pd(velec,rinvsq13);
1316
1317             fscal            = felec;
1318
1319             /* Calculate temporary vectorial force */
1320             tx               = _mm_mul_pd(fscal,dx13);
1321             ty               = _mm_mul_pd(fscal,dy13);
1322             tz               = _mm_mul_pd(fscal,dz13);
1323
1324             /* Update vectorial force */
1325             fix1             = _mm_add_pd(fix1,tx);
1326             fiy1             = _mm_add_pd(fiy1,ty);
1327             fiz1             = _mm_add_pd(fiz1,tz);
1328
1329             fjx3             = _mm_add_pd(fjx3,tx);
1330             fjy3             = _mm_add_pd(fjy3,ty);
1331             fjz3             = _mm_add_pd(fjz3,tz);
1332
1333             /**************************
1334              * CALCULATE INTERACTIONS *
1335              **************************/
1336
1337             /* COULOMB ELECTROSTATICS */
1338             velec            = _mm_mul_pd(qq21,rinv21);
1339             felec            = _mm_mul_pd(velec,rinvsq21);
1340
1341             fscal            = felec;
1342
1343             /* Calculate temporary vectorial force */
1344             tx               = _mm_mul_pd(fscal,dx21);
1345             ty               = _mm_mul_pd(fscal,dy21);
1346             tz               = _mm_mul_pd(fscal,dz21);
1347
1348             /* Update vectorial force */
1349             fix2             = _mm_add_pd(fix2,tx);
1350             fiy2             = _mm_add_pd(fiy2,ty);
1351             fiz2             = _mm_add_pd(fiz2,tz);
1352
1353             fjx1             = _mm_add_pd(fjx1,tx);
1354             fjy1             = _mm_add_pd(fjy1,ty);
1355             fjz1             = _mm_add_pd(fjz1,tz);
1356
1357             /**************************
1358              * CALCULATE INTERACTIONS *
1359              **************************/
1360
1361             /* COULOMB ELECTROSTATICS */
1362             velec            = _mm_mul_pd(qq22,rinv22);
1363             felec            = _mm_mul_pd(velec,rinvsq22);
1364
1365             fscal            = felec;
1366
1367             /* Calculate temporary vectorial force */
1368             tx               = _mm_mul_pd(fscal,dx22);
1369             ty               = _mm_mul_pd(fscal,dy22);
1370             tz               = _mm_mul_pd(fscal,dz22);
1371
1372             /* Update vectorial force */
1373             fix2             = _mm_add_pd(fix2,tx);
1374             fiy2             = _mm_add_pd(fiy2,ty);
1375             fiz2             = _mm_add_pd(fiz2,tz);
1376
1377             fjx2             = _mm_add_pd(fjx2,tx);
1378             fjy2             = _mm_add_pd(fjy2,ty);
1379             fjz2             = _mm_add_pd(fjz2,tz);
1380
1381             /**************************
1382              * CALCULATE INTERACTIONS *
1383              **************************/
1384
1385             /* COULOMB ELECTROSTATICS */
1386             velec            = _mm_mul_pd(qq23,rinv23);
1387             felec            = _mm_mul_pd(velec,rinvsq23);
1388
1389             fscal            = felec;
1390
1391             /* Calculate temporary vectorial force */
1392             tx               = _mm_mul_pd(fscal,dx23);
1393             ty               = _mm_mul_pd(fscal,dy23);
1394             tz               = _mm_mul_pd(fscal,dz23);
1395
1396             /* Update vectorial force */
1397             fix2             = _mm_add_pd(fix2,tx);
1398             fiy2             = _mm_add_pd(fiy2,ty);
1399             fiz2             = _mm_add_pd(fiz2,tz);
1400
1401             fjx3             = _mm_add_pd(fjx3,tx);
1402             fjy3             = _mm_add_pd(fjy3,ty);
1403             fjz3             = _mm_add_pd(fjz3,tz);
1404
1405             /**************************
1406              * CALCULATE INTERACTIONS *
1407              **************************/
1408
1409             /* COULOMB ELECTROSTATICS */
1410             velec            = _mm_mul_pd(qq31,rinv31);
1411             felec            = _mm_mul_pd(velec,rinvsq31);
1412
1413             fscal            = felec;
1414
1415             /* Calculate temporary vectorial force */
1416             tx               = _mm_mul_pd(fscal,dx31);
1417             ty               = _mm_mul_pd(fscal,dy31);
1418             tz               = _mm_mul_pd(fscal,dz31);
1419
1420             /* Update vectorial force */
1421             fix3             = _mm_add_pd(fix3,tx);
1422             fiy3             = _mm_add_pd(fiy3,ty);
1423             fiz3             = _mm_add_pd(fiz3,tz);
1424
1425             fjx1             = _mm_add_pd(fjx1,tx);
1426             fjy1             = _mm_add_pd(fjy1,ty);
1427             fjz1             = _mm_add_pd(fjz1,tz);
1428
1429             /**************************
1430              * CALCULATE INTERACTIONS *
1431              **************************/
1432
1433             /* COULOMB ELECTROSTATICS */
1434             velec            = _mm_mul_pd(qq32,rinv32);
1435             felec            = _mm_mul_pd(velec,rinvsq32);
1436
1437             fscal            = felec;
1438
1439             /* Calculate temporary vectorial force */
1440             tx               = _mm_mul_pd(fscal,dx32);
1441             ty               = _mm_mul_pd(fscal,dy32);
1442             tz               = _mm_mul_pd(fscal,dz32);
1443
1444             /* Update vectorial force */
1445             fix3             = _mm_add_pd(fix3,tx);
1446             fiy3             = _mm_add_pd(fiy3,ty);
1447             fiz3             = _mm_add_pd(fiz3,tz);
1448
1449             fjx2             = _mm_add_pd(fjx2,tx);
1450             fjy2             = _mm_add_pd(fjy2,ty);
1451             fjz2             = _mm_add_pd(fjz2,tz);
1452
1453             /**************************
1454              * CALCULATE INTERACTIONS *
1455              **************************/
1456
1457             /* COULOMB ELECTROSTATICS */
1458             velec            = _mm_mul_pd(qq33,rinv33);
1459             felec            = _mm_mul_pd(velec,rinvsq33);
1460
1461             fscal            = felec;
1462
1463             /* Calculate temporary vectorial force */
1464             tx               = _mm_mul_pd(fscal,dx33);
1465             ty               = _mm_mul_pd(fscal,dy33);
1466             tz               = _mm_mul_pd(fscal,dz33);
1467
1468             /* Update vectorial force */
1469             fix3             = _mm_add_pd(fix3,tx);
1470             fiy3             = _mm_add_pd(fiy3,ty);
1471             fiz3             = _mm_add_pd(fiz3,tz);
1472
1473             fjx3             = _mm_add_pd(fjx3,tx);
1474             fjy3             = _mm_add_pd(fjy3,ty);
1475             fjz3             = _mm_add_pd(fjz3,tz);
1476
1477             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);
1478
1479             /* Inner loop uses 273 flops */
1480         }
1481
1482         if(jidx<j_index_end)
1483         {
1484
1485             jnrA             = jjnr[jidx];
1486             j_coord_offsetA  = DIM*jnrA;
1487
1488             /* load j atom coordinates */
1489             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1490                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1491                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1492
1493             /* Calculate displacement vector */
1494             dx00             = _mm_sub_pd(ix0,jx0);
1495             dy00             = _mm_sub_pd(iy0,jy0);
1496             dz00             = _mm_sub_pd(iz0,jz0);
1497             dx11             = _mm_sub_pd(ix1,jx1);
1498             dy11             = _mm_sub_pd(iy1,jy1);
1499             dz11             = _mm_sub_pd(iz1,jz1);
1500             dx12             = _mm_sub_pd(ix1,jx2);
1501             dy12             = _mm_sub_pd(iy1,jy2);
1502             dz12             = _mm_sub_pd(iz1,jz2);
1503             dx13             = _mm_sub_pd(ix1,jx3);
1504             dy13             = _mm_sub_pd(iy1,jy3);
1505             dz13             = _mm_sub_pd(iz1,jz3);
1506             dx21             = _mm_sub_pd(ix2,jx1);
1507             dy21             = _mm_sub_pd(iy2,jy1);
1508             dz21             = _mm_sub_pd(iz2,jz1);
1509             dx22             = _mm_sub_pd(ix2,jx2);
1510             dy22             = _mm_sub_pd(iy2,jy2);
1511             dz22             = _mm_sub_pd(iz2,jz2);
1512             dx23             = _mm_sub_pd(ix2,jx3);
1513             dy23             = _mm_sub_pd(iy2,jy3);
1514             dz23             = _mm_sub_pd(iz2,jz3);
1515             dx31             = _mm_sub_pd(ix3,jx1);
1516             dy31             = _mm_sub_pd(iy3,jy1);
1517             dz31             = _mm_sub_pd(iz3,jz1);
1518             dx32             = _mm_sub_pd(ix3,jx2);
1519             dy32             = _mm_sub_pd(iy3,jy2);
1520             dz32             = _mm_sub_pd(iz3,jz2);
1521             dx33             = _mm_sub_pd(ix3,jx3);
1522             dy33             = _mm_sub_pd(iy3,jy3);
1523             dz33             = _mm_sub_pd(iz3,jz3);
1524
1525             /* Calculate squared distance and things based on it */
1526             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1527             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1528             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1529             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1530             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1531             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1532             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1533             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1534             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1535             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1536
1537             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1538             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1539             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1540             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1541             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1542             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1543             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1544             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1545             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1546
1547             rinvsq00         = gmx_mm_inv_pd(rsq00);
1548             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1549             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1550             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1551             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1552             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1553             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1554             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1555             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1556             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1557
1558             fjx0             = _mm_setzero_pd();
1559             fjy0             = _mm_setzero_pd();
1560             fjz0             = _mm_setzero_pd();
1561             fjx1             = _mm_setzero_pd();
1562             fjy1             = _mm_setzero_pd();
1563             fjz1             = _mm_setzero_pd();
1564             fjx2             = _mm_setzero_pd();
1565             fjy2             = _mm_setzero_pd();
1566             fjz2             = _mm_setzero_pd();
1567             fjx3             = _mm_setzero_pd();
1568             fjy3             = _mm_setzero_pd();
1569             fjz3             = _mm_setzero_pd();
1570
1571             /**************************
1572              * CALCULATE INTERACTIONS *
1573              **************************/
1574
1575             /* LENNARD-JONES DISPERSION/REPULSION */
1576
1577             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1578             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1579
1580             fscal            = fvdw;
1581
1582             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1583
1584             /* Calculate temporary vectorial force */
1585             tx               = _mm_mul_pd(fscal,dx00);
1586             ty               = _mm_mul_pd(fscal,dy00);
1587             tz               = _mm_mul_pd(fscal,dz00);
1588
1589             /* Update vectorial force */
1590             fix0             = _mm_add_pd(fix0,tx);
1591             fiy0             = _mm_add_pd(fiy0,ty);
1592             fiz0             = _mm_add_pd(fiz0,tz);
1593
1594             fjx0             = _mm_add_pd(fjx0,tx);
1595             fjy0             = _mm_add_pd(fjy0,ty);
1596             fjz0             = _mm_add_pd(fjz0,tz);
1597
1598             /**************************
1599              * CALCULATE INTERACTIONS *
1600              **************************/
1601
1602             /* COULOMB ELECTROSTATICS */
1603             velec            = _mm_mul_pd(qq11,rinv11);
1604             felec            = _mm_mul_pd(velec,rinvsq11);
1605
1606             fscal            = felec;
1607
1608             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1609
1610             /* Calculate temporary vectorial force */
1611             tx               = _mm_mul_pd(fscal,dx11);
1612             ty               = _mm_mul_pd(fscal,dy11);
1613             tz               = _mm_mul_pd(fscal,dz11);
1614
1615             /* Update vectorial force */
1616             fix1             = _mm_add_pd(fix1,tx);
1617             fiy1             = _mm_add_pd(fiy1,ty);
1618             fiz1             = _mm_add_pd(fiz1,tz);
1619
1620             fjx1             = _mm_add_pd(fjx1,tx);
1621             fjy1             = _mm_add_pd(fjy1,ty);
1622             fjz1             = _mm_add_pd(fjz1,tz);
1623
1624             /**************************
1625              * CALCULATE INTERACTIONS *
1626              **************************/
1627
1628             /* COULOMB ELECTROSTATICS */
1629             velec            = _mm_mul_pd(qq12,rinv12);
1630             felec            = _mm_mul_pd(velec,rinvsq12);
1631
1632             fscal            = felec;
1633
1634             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1635
1636             /* Calculate temporary vectorial force */
1637             tx               = _mm_mul_pd(fscal,dx12);
1638             ty               = _mm_mul_pd(fscal,dy12);
1639             tz               = _mm_mul_pd(fscal,dz12);
1640
1641             /* Update vectorial force */
1642             fix1             = _mm_add_pd(fix1,tx);
1643             fiy1             = _mm_add_pd(fiy1,ty);
1644             fiz1             = _mm_add_pd(fiz1,tz);
1645
1646             fjx2             = _mm_add_pd(fjx2,tx);
1647             fjy2             = _mm_add_pd(fjy2,ty);
1648             fjz2             = _mm_add_pd(fjz2,tz);
1649
1650             /**************************
1651              * CALCULATE INTERACTIONS *
1652              **************************/
1653
1654             /* COULOMB ELECTROSTATICS */
1655             velec            = _mm_mul_pd(qq13,rinv13);
1656             felec            = _mm_mul_pd(velec,rinvsq13);
1657
1658             fscal            = felec;
1659
1660             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1661
1662             /* Calculate temporary vectorial force */
1663             tx               = _mm_mul_pd(fscal,dx13);
1664             ty               = _mm_mul_pd(fscal,dy13);
1665             tz               = _mm_mul_pd(fscal,dz13);
1666
1667             /* Update vectorial force */
1668             fix1             = _mm_add_pd(fix1,tx);
1669             fiy1             = _mm_add_pd(fiy1,ty);
1670             fiz1             = _mm_add_pd(fiz1,tz);
1671
1672             fjx3             = _mm_add_pd(fjx3,tx);
1673             fjy3             = _mm_add_pd(fjy3,ty);
1674             fjz3             = _mm_add_pd(fjz3,tz);
1675
1676             /**************************
1677              * CALCULATE INTERACTIONS *
1678              **************************/
1679
1680             /* COULOMB ELECTROSTATICS */
1681             velec            = _mm_mul_pd(qq21,rinv21);
1682             felec            = _mm_mul_pd(velec,rinvsq21);
1683
1684             fscal            = felec;
1685
1686             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1687
1688             /* Calculate temporary vectorial force */
1689             tx               = _mm_mul_pd(fscal,dx21);
1690             ty               = _mm_mul_pd(fscal,dy21);
1691             tz               = _mm_mul_pd(fscal,dz21);
1692
1693             /* Update vectorial force */
1694             fix2             = _mm_add_pd(fix2,tx);
1695             fiy2             = _mm_add_pd(fiy2,ty);
1696             fiz2             = _mm_add_pd(fiz2,tz);
1697
1698             fjx1             = _mm_add_pd(fjx1,tx);
1699             fjy1             = _mm_add_pd(fjy1,ty);
1700             fjz1             = _mm_add_pd(fjz1,tz);
1701
1702             /**************************
1703              * CALCULATE INTERACTIONS *
1704              **************************/
1705
1706             /* COULOMB ELECTROSTATICS */
1707             velec            = _mm_mul_pd(qq22,rinv22);
1708             felec            = _mm_mul_pd(velec,rinvsq22);
1709
1710             fscal            = felec;
1711
1712             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1713
1714             /* Calculate temporary vectorial force */
1715             tx               = _mm_mul_pd(fscal,dx22);
1716             ty               = _mm_mul_pd(fscal,dy22);
1717             tz               = _mm_mul_pd(fscal,dz22);
1718
1719             /* Update vectorial force */
1720             fix2             = _mm_add_pd(fix2,tx);
1721             fiy2             = _mm_add_pd(fiy2,ty);
1722             fiz2             = _mm_add_pd(fiz2,tz);
1723
1724             fjx2             = _mm_add_pd(fjx2,tx);
1725             fjy2             = _mm_add_pd(fjy2,ty);
1726             fjz2             = _mm_add_pd(fjz2,tz);
1727
1728             /**************************
1729              * CALCULATE INTERACTIONS *
1730              **************************/
1731
1732             /* COULOMB ELECTROSTATICS */
1733             velec            = _mm_mul_pd(qq23,rinv23);
1734             felec            = _mm_mul_pd(velec,rinvsq23);
1735
1736             fscal            = felec;
1737
1738             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1739
1740             /* Calculate temporary vectorial force */
1741             tx               = _mm_mul_pd(fscal,dx23);
1742             ty               = _mm_mul_pd(fscal,dy23);
1743             tz               = _mm_mul_pd(fscal,dz23);
1744
1745             /* Update vectorial force */
1746             fix2             = _mm_add_pd(fix2,tx);
1747             fiy2             = _mm_add_pd(fiy2,ty);
1748             fiz2             = _mm_add_pd(fiz2,tz);
1749
1750             fjx3             = _mm_add_pd(fjx3,tx);
1751             fjy3             = _mm_add_pd(fjy3,ty);
1752             fjz3             = _mm_add_pd(fjz3,tz);
1753
1754             /**************************
1755              * CALCULATE INTERACTIONS *
1756              **************************/
1757
1758             /* COULOMB ELECTROSTATICS */
1759             velec            = _mm_mul_pd(qq31,rinv31);
1760             felec            = _mm_mul_pd(velec,rinvsq31);
1761
1762             fscal            = felec;
1763
1764             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1765
1766             /* Calculate temporary vectorial force */
1767             tx               = _mm_mul_pd(fscal,dx31);
1768             ty               = _mm_mul_pd(fscal,dy31);
1769             tz               = _mm_mul_pd(fscal,dz31);
1770
1771             /* Update vectorial force */
1772             fix3             = _mm_add_pd(fix3,tx);
1773             fiy3             = _mm_add_pd(fiy3,ty);
1774             fiz3             = _mm_add_pd(fiz3,tz);
1775
1776             fjx1             = _mm_add_pd(fjx1,tx);
1777             fjy1             = _mm_add_pd(fjy1,ty);
1778             fjz1             = _mm_add_pd(fjz1,tz);
1779
1780             /**************************
1781              * CALCULATE INTERACTIONS *
1782              **************************/
1783
1784             /* COULOMB ELECTROSTATICS */
1785             velec            = _mm_mul_pd(qq32,rinv32);
1786             felec            = _mm_mul_pd(velec,rinvsq32);
1787
1788             fscal            = felec;
1789
1790             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1791
1792             /* Calculate temporary vectorial force */
1793             tx               = _mm_mul_pd(fscal,dx32);
1794             ty               = _mm_mul_pd(fscal,dy32);
1795             tz               = _mm_mul_pd(fscal,dz32);
1796
1797             /* Update vectorial force */
1798             fix3             = _mm_add_pd(fix3,tx);
1799             fiy3             = _mm_add_pd(fiy3,ty);
1800             fiz3             = _mm_add_pd(fiz3,tz);
1801
1802             fjx2             = _mm_add_pd(fjx2,tx);
1803             fjy2             = _mm_add_pd(fjy2,ty);
1804             fjz2             = _mm_add_pd(fjz2,tz);
1805
1806             /**************************
1807              * CALCULATE INTERACTIONS *
1808              **************************/
1809
1810             /* COULOMB ELECTROSTATICS */
1811             velec            = _mm_mul_pd(qq33,rinv33);
1812             felec            = _mm_mul_pd(velec,rinvsq33);
1813
1814             fscal            = felec;
1815
1816             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1817
1818             /* Calculate temporary vectorial force */
1819             tx               = _mm_mul_pd(fscal,dx33);
1820             ty               = _mm_mul_pd(fscal,dy33);
1821             tz               = _mm_mul_pd(fscal,dz33);
1822
1823             /* Update vectorial force */
1824             fix3             = _mm_add_pd(fix3,tx);
1825             fiy3             = _mm_add_pd(fiy3,ty);
1826             fiz3             = _mm_add_pd(fiz3,tz);
1827
1828             fjx3             = _mm_add_pd(fjx3,tx);
1829             fjy3             = _mm_add_pd(fjy3,ty);
1830             fjz3             = _mm_add_pd(fjz3,tz);
1831
1832             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1833
1834             /* Inner loop uses 273 flops */
1835         }
1836
1837         /* End of innermost loop */
1838
1839         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1840                                               f+i_coord_offset,fshift+i_shift_offset);
1841
1842         /* Increment number of inner iterations */
1843         inneriter                  += j_index_end - j_index_start;
1844
1845         /* Outer loop uses 24 flops */
1846     }
1847
1848     /* Increment number of outer iterations */
1849     outeriter        += nri;
1850
1851     /* Update outer/inner flops */
1852
1853     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1854 }