src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 avx_128_fma_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_avx_128_fma_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_avx_128_fma_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_msub_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             /* Update vectorial force */
 314             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 315             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 316             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 317
 318             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 319             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 320             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             /* COULOMB ELECTROSTATICS */
 327             velec            = _mm_mul_pd(qq11,rinv11);
 328             felec            = _mm_mul_pd(velec,rinvsq11);
 329
 330             /* Update potential sum for this i atom from the interaction with this j atom. */
 331             velecsum         = _mm_add_pd(velecsum,velec);
 332
 333             fscal            = felec;
 334
 335             /* Update vectorial force */
 336             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 337             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 338             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 339
 340             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 341             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 342             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 343
 344             /**************************
 345              * CALCULATE INTERACTIONS *
 346              **************************/
 347
 348             /* COULOMB ELECTROSTATICS */
 349             velec            = _mm_mul_pd(qq12,rinv12);
 350             felec            = _mm_mul_pd(velec,rinvsq12);
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             velecsum         = _mm_add_pd(velecsum,velec);
 354
 355             fscal            = felec;
 356
 357             /* Update vectorial force */
 358             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 359             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 360             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 361
 362             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 363             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 364             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             /* COULOMB ELECTROSTATICS */
 371             velec            = _mm_mul_pd(qq13,rinv13);
 372             felec            = _mm_mul_pd(velec,rinvsq13);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _mm_add_pd(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379             /* Update vectorial force */
 380             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 381             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 382             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 383
 384             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 385             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 386             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             /* COULOMB ELECTROSTATICS */
 393             velec            = _mm_mul_pd(qq21,rinv21);
 394             felec            = _mm_mul_pd(velec,rinvsq21);
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velecsum         = _mm_add_pd(velecsum,velec);
 398
 399             fscal            = felec;
 400
 401             /* Update vectorial force */
 402             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 403             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 404             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 405
 406             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 407             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 408             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             /* COULOMB ELECTROSTATICS */
 415             velec            = _mm_mul_pd(qq22,rinv22);
 416             felec            = _mm_mul_pd(velec,rinvsq22);
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velecsum         = _mm_add_pd(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423             /* Update vectorial force */
 424             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 425             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 426             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 427
 428             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 429             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 430             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             /* COULOMB ELECTROSTATICS */
 437             velec            = _mm_mul_pd(qq23,rinv23);
 438             felec            = _mm_mul_pd(velec,rinvsq23);
 439
 440             /* Update potential sum for this i atom from the interaction with this j atom. */
 441             velecsum         = _mm_add_pd(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445             /* Update vectorial force */
 446             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 447             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 448             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 449
 450             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 451             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 452             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 453
 454             /**************************
 455              * CALCULATE INTERACTIONS *
 456              **************************/
 457
 458             /* COULOMB ELECTROSTATICS */
 459             velec            = _mm_mul_pd(qq31,rinv31);
 460             felec            = _mm_mul_pd(velec,rinvsq31);
 461
 462             /* Update potential sum for this i atom from the interaction with this j atom. */
 463             velecsum         = _mm_add_pd(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467             /* Update vectorial force */
 468             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 469             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 470             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 471
 472             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 473             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 474             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 475
 476             /**************************
 477              * CALCULATE INTERACTIONS *
 478              **************************/
 479
 480             /* COULOMB ELECTROSTATICS */
 481             velec            = _mm_mul_pd(qq32,rinv32);
 482             felec            = _mm_mul_pd(velec,rinvsq32);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velecsum         = _mm_add_pd(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             /* Update vectorial force */
 490             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 491             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 492             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 493
 494             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 495             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 496             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             /* COULOMB ELECTROSTATICS */
 503             velec            = _mm_mul_pd(qq33,rinv33);
 504             felec            = _mm_mul_pd(velec,rinvsq33);
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velecsum         = _mm_add_pd(velecsum,velec);
 508
 509             fscal            = felec;
 510
 511             /* Update vectorial force */
 512             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 513             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 514             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 515
 516             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 517             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 518             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 519
 520             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);
 521
 522             /* Inner loop uses 317 flops */
 523         }
 524
 525         if(jidx<j_index_end)
 526         {
 527
 528             jnrA             = jjnr[jidx];
 529             j_coord_offsetA  = DIM*jnrA;
 530
 531             /* load j atom coordinates */
 532             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 533                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 534                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 535
 536             /* Calculate displacement vector */
 537             dx00             = _mm_sub_pd(ix0,jx0);
 538             dy00             = _mm_sub_pd(iy0,jy0);
 539             dz00             = _mm_sub_pd(iz0,jz0);
 540             dx11             = _mm_sub_pd(ix1,jx1);
 541             dy11             = _mm_sub_pd(iy1,jy1);
 542             dz11             = _mm_sub_pd(iz1,jz1);
 543             dx12             = _mm_sub_pd(ix1,jx2);
 544             dy12             = _mm_sub_pd(iy1,jy2);
 545             dz12             = _mm_sub_pd(iz1,jz2);
 546             dx13             = _mm_sub_pd(ix1,jx3);
 547             dy13             = _mm_sub_pd(iy1,jy3);
 548             dz13             = _mm_sub_pd(iz1,jz3);
 549             dx21             = _mm_sub_pd(ix2,jx1);
 550             dy21             = _mm_sub_pd(iy2,jy1);
 551             dz21             = _mm_sub_pd(iz2,jz1);
 552             dx22             = _mm_sub_pd(ix2,jx2);
 553             dy22             = _mm_sub_pd(iy2,jy2);
 554             dz22             = _mm_sub_pd(iz2,jz2);
 555             dx23             = _mm_sub_pd(ix2,jx3);
 556             dy23             = _mm_sub_pd(iy2,jy3);
 557             dz23             = _mm_sub_pd(iz2,jz3);
 558             dx31             = _mm_sub_pd(ix3,jx1);
 559             dy31             = _mm_sub_pd(iy3,jy1);
 560             dz31             = _mm_sub_pd(iz3,jz1);
 561             dx32             = _mm_sub_pd(ix3,jx2);
 562             dy32             = _mm_sub_pd(iy3,jy2);
 563             dz32             = _mm_sub_pd(iz3,jz2);
 564             dx33             = _mm_sub_pd(ix3,jx3);
 565             dy33             = _mm_sub_pd(iy3,jy3);
 566             dz33             = _mm_sub_pd(iz3,jz3);
 567
 568             /* Calculate squared distance and things based on it */
 569             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 570             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 571             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 572             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 573             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 574             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 575             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 576             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 577             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 578             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 579
 580             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 581             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 582             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 583             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 584             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 585             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 586             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 587             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 588             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 589
 590             rinvsq00         = gmx_mm_inv_pd(rsq00);
 591             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 592             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 593             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 594             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 595             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 596             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 597             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 598             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 599             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 600
 601             fjx0             = _mm_setzero_pd();
 602             fjy0             = _mm_setzero_pd();
 603             fjz0             = _mm_setzero_pd();
 604             fjx1             = _mm_setzero_pd();
 605             fjy1             = _mm_setzero_pd();
 606             fjz1             = _mm_setzero_pd();
 607             fjx2             = _mm_setzero_pd();
 608             fjy2             = _mm_setzero_pd();
 609             fjz2             = _mm_setzero_pd();
 610             fjx3             = _mm_setzero_pd();
 611             fjy3             = _mm_setzero_pd();
 612             fjz3             = _mm_setzero_pd();
 613
 614             /**************************
 615              * CALCULATE INTERACTIONS *
 616              **************************/
 617
 618             /* LENNARD-JONES DISPERSION/REPULSION */
 619
 620             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 621             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 622             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 623             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 624             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 625
 626             /* Update potential sum for this i atom from the interaction with this j atom. */
 627             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 628             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 629
 630             fscal            = fvdw;
 631
 632             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 633
 634             /* Update vectorial force */
 635             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 636             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 637             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 638
 639             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 640             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 641             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 642
 643             /**************************
 644              * CALCULATE INTERACTIONS *
 645              **************************/
 646
 647             /* COULOMB ELECTROSTATICS */
 648             velec            = _mm_mul_pd(qq11,rinv11);
 649             felec            = _mm_mul_pd(velec,rinvsq11);
 650
 651             /* Update potential sum for this i atom from the interaction with this j atom. */
 652             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 653             velecsum         = _mm_add_pd(velecsum,velec);
 654
 655             fscal            = felec;
 656
 657             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 658
 659             /* Update vectorial force */
 660             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 661             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 662             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 663
 664             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 665             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 666             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             /* COULOMB ELECTROSTATICS */
 673             velec            = _mm_mul_pd(qq12,rinv12);
 674             felec            = _mm_mul_pd(velec,rinvsq12);
 675
 676             /* Update potential sum for this i atom from the interaction with this j atom. */
 677             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 678             velecsum         = _mm_add_pd(velecsum,velec);
 679
 680             fscal            = felec;
 681
 682             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 683
 684             /* Update vectorial force */
 685             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 686             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 687             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 688
 689             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 690             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 691             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 692
 693             /**************************
 694              * CALCULATE INTERACTIONS *
 695              **************************/
 696
 697             /* COULOMB ELECTROSTATICS */
 698             velec            = _mm_mul_pd(qq13,rinv13);
 699             felec            = _mm_mul_pd(velec,rinvsq13);
 700
 701             /* Update potential sum for this i atom from the interaction with this j atom. */
 702             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 703             velecsum         = _mm_add_pd(velecsum,velec);
 704
 705             fscal            = felec;
 706
 707             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 708
 709             /* Update vectorial force */
 710             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 711             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 712             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 713
 714             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 715             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 716             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 717
 718             /**************************
 719              * CALCULATE INTERACTIONS *
 720              **************************/
 721
 722             /* COULOMB ELECTROSTATICS */
 723             velec            = _mm_mul_pd(qq21,rinv21);
 724             felec            = _mm_mul_pd(velec,rinvsq21);
 725
 726             /* Update potential sum for this i atom from the interaction with this j atom. */
 727             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 728             velecsum         = _mm_add_pd(velecsum,velec);
 729
 730             fscal            = felec;
 731
 732             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 733
 734             /* Update vectorial force */
 735             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 736             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 737             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 738
 739             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 740             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 741             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 742
 743             /**************************
 744              * CALCULATE INTERACTIONS *
 745              **************************/
 746
 747             /* COULOMB ELECTROSTATICS */
 748             velec            = _mm_mul_pd(qq22,rinv22);
 749             felec            = _mm_mul_pd(velec,rinvsq22);
 750
 751             /* Update potential sum for this i atom from the interaction with this j atom. */
 752             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 753             velecsum         = _mm_add_pd(velecsum,velec);
 754
 755             fscal            = felec;
 756
 757             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 758
 759             /* Update vectorial force */
 760             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 761             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 762             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 763
 764             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 765             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 766             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             /* COULOMB ELECTROSTATICS */
 773             velec            = _mm_mul_pd(qq23,rinv23);
 774             felec            = _mm_mul_pd(velec,rinvsq23);
 775
 776             /* Update potential sum for this i atom from the interaction with this j atom. */
 777             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 778             velecsum         = _mm_add_pd(velecsum,velec);
 779
 780             fscal            = felec;
 781
 782             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 783
 784             /* Update vectorial force */
 785             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 786             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 787             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 788
 789             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 790             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 791             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 792
 793             /**************************
 794              * CALCULATE INTERACTIONS *
 795              **************************/
 796
 797             /* COULOMB ELECTROSTATICS */
 798             velec            = _mm_mul_pd(qq31,rinv31);
 799             felec            = _mm_mul_pd(velec,rinvsq31);
 800
 801             /* Update potential sum for this i atom from the interaction with this j atom. */
 802             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 803             velecsum         = _mm_add_pd(velecsum,velec);
 804
 805             fscal            = felec;
 806
 807             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 808
 809             /* Update vectorial force */
 810             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 811             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 812             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 813
 814             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 815             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 816             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 817
 818             /**************************
 819              * CALCULATE INTERACTIONS *
 820              **************************/
 821
 822             /* COULOMB ELECTROSTATICS */
 823             velec            = _mm_mul_pd(qq32,rinv32);
 824             felec            = _mm_mul_pd(velec,rinvsq32);
 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             /* Update vectorial force */
 835             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 836             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 837             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 838
 839             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 840             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 841             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 842
 843             /**************************
 844              * CALCULATE INTERACTIONS *
 845              **************************/
 846
 847             /* COULOMB ELECTROSTATICS */
 848             velec            = _mm_mul_pd(qq33,rinv33);
 849             felec            = _mm_mul_pd(velec,rinvsq33);
 850
 851             /* Update potential sum for this i atom from the interaction with this j atom. */
 852             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 853             velecsum         = _mm_add_pd(velecsum,velec);
 854
 855             fscal            = felec;
 856
 857             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 858
 859             /* Update vectorial force */
 860             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 861             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 862             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 863
 864             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 865             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 866             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 867
 868             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 869
 870             /* Inner loop uses 317 flops */
 871         }
 872
 873         /* End of innermost loop */
 874
 875         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 876                                               f+i_coord_offset,fshift+i_shift_offset);
 877
 878         ggid                        = gid[iidx];
 879         /* Update potential energies */
 880         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 881         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 882
 883         /* Increment number of inner iterations */
 884         inneriter                  += j_index_end - j_index_start;
 885
 886         /* Outer loop uses 26 flops */
 887     }
 888
 889     /* Increment number of outer iterations */
 890     outeriter        += nri;
 891
 892     /* Update outer/inner flops */
 893
 894     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*317);
 895 }
 896 /*
 897  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_double
 898  * Electrostatics interaction: Coulomb
 899  * VdW interaction:            LennardJones
 900  * Geometry:                   Water4-Water4
 901  * Calculate force/pot:        Force
 902  */
 903 void
 904 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_double
 905                     (t_nblist                    * gmx_restrict       nlist,
 906                      rvec                        * gmx_restrict          xx,
 907                      rvec                        * gmx_restrict          ff,
 908                      t_forcerec                  * gmx_restrict          fr,
 909                      t_mdatoms                   * gmx_restrict     mdatoms,
 910                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 911                      t_nrnb                      * gmx_restrict        nrnb)
 912 {
 913     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 914      * just 0 for non-waters.
 915      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 916      * jnr indices corresponding to data put in the four positions in the SIMD register.
 917      */
 918     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 919     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 920     int              jnrA,jnrB;
 921     int              j_coord_offsetA,j_coord_offsetB;
 922     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 923     real             rcutoff_scalar;
 924     real             *shiftvec,*fshift,*x,*f;
 925     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 926     int              vdwioffset0;
 927     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 928     int              vdwioffset1;
 929     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 930     int              vdwioffset2;
 931     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 932     int              vdwioffset3;
 933     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 934     int              vdwjidx0A,vdwjidx0B;
 935     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 936     int              vdwjidx1A,vdwjidx1B;
 937     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 938     int              vdwjidx2A,vdwjidx2B;
 939     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 940     int              vdwjidx3A,vdwjidx3B;
 941     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 942     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 943     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 944     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 945     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 946     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 947     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 948     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 949     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 950     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 951     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 952     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 953     real             *charge;
 954     int              nvdwtype;
 955     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 956     int              *vdwtype;
 957     real             *vdwparam;
 958     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 959     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 960     __m128d          dummy_mask,cutoff_mask;
 961     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 962     __m128d          one     = _mm_set1_pd(1.0);
 963     __m128d          two     = _mm_set1_pd(2.0);
 964     x                = xx[0];
 965     f                = ff[0];
 966
 967     nri              = nlist->nri;
 968     iinr             = nlist->iinr;
 969     jindex           = nlist->jindex;
 970     jjnr             = nlist->jjnr;
 971     shiftidx         = nlist->shift;
 972     gid              = nlist->gid;
 973     shiftvec         = fr->shift_vec[0];
 974     fshift           = fr->fshift[0];
 975     facel            = _mm_set1_pd(fr->epsfac);
 976     charge           = mdatoms->chargeA;
 977     nvdwtype         = fr->ntype;
 978     vdwparam         = fr->nbfp;
 979     vdwtype          = mdatoms->typeA;
 980
 981     /* Setup water-specific parameters */
 982     inr              = nlist->iinr[0];
 983     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 984     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 985     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 986     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 987
 988     jq1              = _mm_set1_pd(charge[inr+1]);
 989     jq2              = _mm_set1_pd(charge[inr+2]);
 990     jq3              = _mm_set1_pd(charge[inr+3]);
 991     vdwjidx0A        = 2*vdwtype[inr+0];
 992     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 993     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 994     qq11             = _mm_mul_pd(iq1,jq1);
 995     qq12             = _mm_mul_pd(iq1,jq2);
 996     qq13             = _mm_mul_pd(iq1,jq3);
 997     qq21             = _mm_mul_pd(iq2,jq1);
 998     qq22             = _mm_mul_pd(iq2,jq2);
 999     qq23             = _mm_mul_pd(iq2,jq3);
1000     qq31             = _mm_mul_pd(iq3,jq1);
1001     qq32             = _mm_mul_pd(iq3,jq2);
1002     qq33             = _mm_mul_pd(iq3,jq3);
1003
1004     /* Avoid stupid compiler warnings */
1005     jnrA = jnrB = 0;
1006     j_coord_offsetA = 0;
1007     j_coord_offsetB = 0;
1008
1009     outeriter        = 0;
1010     inneriter        = 0;
1011
1012     /* Start outer loop over neighborlists */
1013     for(iidx=0; iidx<nri; iidx++)
1014     {
1015         /* Load shift vector for this list */
1016         i_shift_offset   = DIM*shiftidx[iidx];
1017
1018         /* Load limits for loop over neighbors */
1019         j_index_start    = jindex[iidx];
1020         j_index_end      = jindex[iidx+1];
1021
1022         /* Get outer coordinate index */
1023         inr              = iinr[iidx];
1024         i_coord_offset   = DIM*inr;
1025
1026         /* Load i particle coords and add shift vector */
1027         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1028                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1029
1030         fix0             = _mm_setzero_pd();
1031         fiy0             = _mm_setzero_pd();
1032         fiz0             = _mm_setzero_pd();
1033         fix1             = _mm_setzero_pd();
1034         fiy1             = _mm_setzero_pd();
1035         fiz1             = _mm_setzero_pd();
1036         fix2             = _mm_setzero_pd();
1037         fiy2             = _mm_setzero_pd();
1038         fiz2             = _mm_setzero_pd();
1039         fix3             = _mm_setzero_pd();
1040         fiy3             = _mm_setzero_pd();
1041         fiz3             = _mm_setzero_pd();
1042
1043         /* Start inner kernel loop */
1044         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1045         {
1046
1047             /* Get j neighbor index, and coordinate index */
1048             jnrA             = jjnr[jidx];
1049             jnrB             = jjnr[jidx+1];
1050             j_coord_offsetA  = DIM*jnrA;
1051             j_coord_offsetB  = DIM*jnrB;
1052
1053             /* load j atom coordinates */
1054             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1055                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1056                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1057
1058             /* Calculate displacement vector */
1059             dx00             = _mm_sub_pd(ix0,jx0);
1060             dy00             = _mm_sub_pd(iy0,jy0);
1061             dz00             = _mm_sub_pd(iz0,jz0);
1062             dx11             = _mm_sub_pd(ix1,jx1);
1063             dy11             = _mm_sub_pd(iy1,jy1);
1064             dz11             = _mm_sub_pd(iz1,jz1);
1065             dx12             = _mm_sub_pd(ix1,jx2);
1066             dy12             = _mm_sub_pd(iy1,jy2);
1067             dz12             = _mm_sub_pd(iz1,jz2);
1068             dx13             = _mm_sub_pd(ix1,jx3);
1069             dy13             = _mm_sub_pd(iy1,jy3);
1070             dz13             = _mm_sub_pd(iz1,jz3);
1071             dx21             = _mm_sub_pd(ix2,jx1);
1072             dy21             = _mm_sub_pd(iy2,jy1);
1073             dz21             = _mm_sub_pd(iz2,jz1);
1074             dx22             = _mm_sub_pd(ix2,jx2);
1075             dy22             = _mm_sub_pd(iy2,jy2);
1076             dz22             = _mm_sub_pd(iz2,jz2);
1077             dx23             = _mm_sub_pd(ix2,jx3);
1078             dy23             = _mm_sub_pd(iy2,jy3);
1079             dz23             = _mm_sub_pd(iz2,jz3);
1080             dx31             = _mm_sub_pd(ix3,jx1);
1081             dy31             = _mm_sub_pd(iy3,jy1);
1082             dz31             = _mm_sub_pd(iz3,jz1);
1083             dx32             = _mm_sub_pd(ix3,jx2);
1084             dy32             = _mm_sub_pd(iy3,jy2);
1085             dz32             = _mm_sub_pd(iz3,jz2);
1086             dx33             = _mm_sub_pd(ix3,jx3);
1087             dy33             = _mm_sub_pd(iy3,jy3);
1088             dz33             = _mm_sub_pd(iz3,jz3);
1089
1090             /* Calculate squared distance and things based on it */
1091             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1092             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1093             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1094             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1095             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1096             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1097             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1098             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1099             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1100             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1101
1102             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1103             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1104             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1105             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1106             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1107             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1108             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1109             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1110             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1111
1112             rinvsq00         = gmx_mm_inv_pd(rsq00);
1113             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1114             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1115             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1116             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1117             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1118             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1119             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1120             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1121             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1122
1123             fjx0             = _mm_setzero_pd();
1124             fjy0             = _mm_setzero_pd();
1125             fjz0             = _mm_setzero_pd();
1126             fjx1             = _mm_setzero_pd();
1127             fjy1             = _mm_setzero_pd();
1128             fjz1             = _mm_setzero_pd();
1129             fjx2             = _mm_setzero_pd();
1130             fjy2             = _mm_setzero_pd();
1131             fjz2             = _mm_setzero_pd();
1132             fjx3             = _mm_setzero_pd();
1133             fjy3             = _mm_setzero_pd();
1134             fjz3             = _mm_setzero_pd();
1135
1136             /**************************
1137              * CALCULATE INTERACTIONS *
1138              **************************/
1139
1140             /* LENNARD-JONES DISPERSION/REPULSION */
1141
1142             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1143             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1144
1145             fscal            = fvdw;
1146
1147             /* Update vectorial force */
1148             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1149             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1150             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1151
1152             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1153             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1154             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1155
1156             /**************************
1157              * CALCULATE INTERACTIONS *
1158              **************************/
1159
1160             /* COULOMB ELECTROSTATICS */
1161             velec            = _mm_mul_pd(qq11,rinv11);
1162             felec            = _mm_mul_pd(velec,rinvsq11);
1163
1164             fscal            = felec;
1165
1166             /* Update vectorial force */
1167             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1168             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1169             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1170
1171             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1172             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1173             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1174
1175             /**************************
1176              * CALCULATE INTERACTIONS *
1177              **************************/
1178
1179             /* COULOMB ELECTROSTATICS */
1180             velec            = _mm_mul_pd(qq12,rinv12);
1181             felec            = _mm_mul_pd(velec,rinvsq12);
1182
1183             fscal            = felec;
1184
1185             /* Update vectorial force */
1186             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1187             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1188             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1189
1190             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1191             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1192             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1193
1194             /**************************
1195              * CALCULATE INTERACTIONS *
1196              **************************/
1197
1198             /* COULOMB ELECTROSTATICS */
1199             velec            = _mm_mul_pd(qq13,rinv13);
1200             felec            = _mm_mul_pd(velec,rinvsq13);
1201
1202             fscal            = felec;
1203
1204             /* Update vectorial force */
1205             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1206             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1207             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1208
1209             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1210             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1211             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1212
1213             /**************************
1214              * CALCULATE INTERACTIONS *
1215              **************************/
1216
1217             /* COULOMB ELECTROSTATICS */
1218             velec            = _mm_mul_pd(qq21,rinv21);
1219             felec            = _mm_mul_pd(velec,rinvsq21);
1220
1221             fscal            = felec;
1222
1223             /* Update vectorial force */
1224             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1225             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1226             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1227
1228             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1229             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1230             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1231
1232             /**************************
1233              * CALCULATE INTERACTIONS *
1234              **************************/
1235
1236             /* COULOMB ELECTROSTATICS */
1237             velec            = _mm_mul_pd(qq22,rinv22);
1238             felec            = _mm_mul_pd(velec,rinvsq22);
1239
1240             fscal            = felec;
1241
1242             /* Update vectorial force */
1243             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1244             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1245             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1246
1247             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1248             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1249             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1250
1251             /**************************
1252              * CALCULATE INTERACTIONS *
1253              **************************/
1254
1255             /* COULOMB ELECTROSTATICS */
1256             velec            = _mm_mul_pd(qq23,rinv23);
1257             felec            = _mm_mul_pd(velec,rinvsq23);
1258
1259             fscal            = felec;
1260
1261             /* Update vectorial force */
1262             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1263             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1264             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1265
1266             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1267             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1268             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1269
1270             /**************************
1271              * CALCULATE INTERACTIONS *
1272              **************************/
1273
1274             /* COULOMB ELECTROSTATICS */
1275             velec            = _mm_mul_pd(qq31,rinv31);
1276             felec            = _mm_mul_pd(velec,rinvsq31);
1277
1278             fscal            = felec;
1279
1280             /* Update vectorial force */
1281             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1282             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1283             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1284
1285             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1286             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1287             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1288
1289             /**************************
1290              * CALCULATE INTERACTIONS *
1291              **************************/
1292
1293             /* COULOMB ELECTROSTATICS */
1294             velec            = _mm_mul_pd(qq32,rinv32);
1295             felec            = _mm_mul_pd(velec,rinvsq32);
1296
1297             fscal            = felec;
1298
1299             /* Update vectorial force */
1300             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1301             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1302             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1303
1304             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1305             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1306             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1307
1308             /**************************
1309              * CALCULATE INTERACTIONS *
1310              **************************/
1311
1312             /* COULOMB ELECTROSTATICS */
1313             velec            = _mm_mul_pd(qq33,rinv33);
1314             felec            = _mm_mul_pd(velec,rinvsq33);
1315
1316             fscal            = felec;
1317
1318             /* Update vectorial force */
1319             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1320             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1321             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1322
1323             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1324             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1325             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1326
1327             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);
1328
1329             /* Inner loop uses 303 flops */
1330         }
1331
1332         if(jidx<j_index_end)
1333         {
1334
1335             jnrA             = jjnr[jidx];
1336             j_coord_offsetA  = DIM*jnrA;
1337
1338             /* load j atom coordinates */
1339             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1340                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1341                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1342
1343             /* Calculate displacement vector */
1344             dx00             = _mm_sub_pd(ix0,jx0);
1345             dy00             = _mm_sub_pd(iy0,jy0);
1346             dz00             = _mm_sub_pd(iz0,jz0);
1347             dx11             = _mm_sub_pd(ix1,jx1);
1348             dy11             = _mm_sub_pd(iy1,jy1);
1349             dz11             = _mm_sub_pd(iz1,jz1);
1350             dx12             = _mm_sub_pd(ix1,jx2);
1351             dy12             = _mm_sub_pd(iy1,jy2);
1352             dz12             = _mm_sub_pd(iz1,jz2);
1353             dx13             = _mm_sub_pd(ix1,jx3);
1354             dy13             = _mm_sub_pd(iy1,jy3);
1355             dz13             = _mm_sub_pd(iz1,jz3);
1356             dx21             = _mm_sub_pd(ix2,jx1);
1357             dy21             = _mm_sub_pd(iy2,jy1);
1358             dz21             = _mm_sub_pd(iz2,jz1);
1359             dx22             = _mm_sub_pd(ix2,jx2);
1360             dy22             = _mm_sub_pd(iy2,jy2);
1361             dz22             = _mm_sub_pd(iz2,jz2);
1362             dx23             = _mm_sub_pd(ix2,jx3);
1363             dy23             = _mm_sub_pd(iy2,jy3);
1364             dz23             = _mm_sub_pd(iz2,jz3);
1365             dx31             = _mm_sub_pd(ix3,jx1);
1366             dy31             = _mm_sub_pd(iy3,jy1);
1367             dz31             = _mm_sub_pd(iz3,jz1);
1368             dx32             = _mm_sub_pd(ix3,jx2);
1369             dy32             = _mm_sub_pd(iy3,jy2);
1370             dz32             = _mm_sub_pd(iz3,jz2);
1371             dx33             = _mm_sub_pd(ix3,jx3);
1372             dy33             = _mm_sub_pd(iy3,jy3);
1373             dz33             = _mm_sub_pd(iz3,jz3);
1374
1375             /* Calculate squared distance and things based on it */
1376             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1377             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1378             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1379             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1380             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1381             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1382             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1383             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1384             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1385             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1386
1387             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1388             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1389             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1390             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1391             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1392             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1393             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1394             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1395             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1396
1397             rinvsq00         = gmx_mm_inv_pd(rsq00);
1398             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1399             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1400             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1401             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1402             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1403             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1404             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1405             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1406             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1407
1408             fjx0             = _mm_setzero_pd();
1409             fjy0             = _mm_setzero_pd();
1410             fjz0             = _mm_setzero_pd();
1411             fjx1             = _mm_setzero_pd();
1412             fjy1             = _mm_setzero_pd();
1413             fjz1             = _mm_setzero_pd();
1414             fjx2             = _mm_setzero_pd();
1415             fjy2             = _mm_setzero_pd();
1416             fjz2             = _mm_setzero_pd();
1417             fjx3             = _mm_setzero_pd();
1418             fjy3             = _mm_setzero_pd();
1419             fjz3             = _mm_setzero_pd();
1420
1421             /**************************
1422              * CALCULATE INTERACTIONS *
1423              **************************/
1424
1425             /* LENNARD-JONES DISPERSION/REPULSION */
1426
1427             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1428             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1429
1430             fscal            = fvdw;
1431
1432             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1433
1434             /* Update vectorial force */
1435             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1436             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1437             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1438
1439             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1440             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1441             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1442
1443             /**************************
1444              * CALCULATE INTERACTIONS *
1445              **************************/
1446
1447             /* COULOMB ELECTROSTATICS */
1448             velec            = _mm_mul_pd(qq11,rinv11);
1449             felec            = _mm_mul_pd(velec,rinvsq11);
1450
1451             fscal            = felec;
1452
1453             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1454
1455             /* Update vectorial force */
1456             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1457             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1458             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1459
1460             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1461             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1462             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             /* COULOMB ELECTROSTATICS */
1469             velec            = _mm_mul_pd(qq12,rinv12);
1470             felec            = _mm_mul_pd(velec,rinvsq12);
1471
1472             fscal            = felec;
1473
1474             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1475
1476             /* Update vectorial force */
1477             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1478             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1479             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1480
1481             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1482             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1483             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1484
1485             /**************************
1486              * CALCULATE INTERACTIONS *
1487              **************************/
1488
1489             /* COULOMB ELECTROSTATICS */
1490             velec            = _mm_mul_pd(qq13,rinv13);
1491             felec            = _mm_mul_pd(velec,rinvsq13);
1492
1493             fscal            = felec;
1494
1495             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1496
1497             /* Update vectorial force */
1498             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1499             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1500             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1501
1502             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1503             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1504             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1505
1506             /**************************
1507              * CALCULATE INTERACTIONS *
1508              **************************/
1509
1510             /* COULOMB ELECTROSTATICS */
1511             velec            = _mm_mul_pd(qq21,rinv21);
1512             felec            = _mm_mul_pd(velec,rinvsq21);
1513
1514             fscal            = felec;
1515
1516             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1517
1518             /* Update vectorial force */
1519             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1520             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1521             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1522
1523             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1524             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1525             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1526
1527             /**************************
1528              * CALCULATE INTERACTIONS *
1529              **************************/
1530
1531             /* COULOMB ELECTROSTATICS */
1532             velec            = _mm_mul_pd(qq22,rinv22);
1533             felec            = _mm_mul_pd(velec,rinvsq22);
1534
1535             fscal            = felec;
1536
1537             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1538
1539             /* Update vectorial force */
1540             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1541             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1542             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1543
1544             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1545             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1546             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1547
1548             /**************************
1549              * CALCULATE INTERACTIONS *
1550              **************************/
1551
1552             /* COULOMB ELECTROSTATICS */
1553             velec            = _mm_mul_pd(qq23,rinv23);
1554             felec            = _mm_mul_pd(velec,rinvsq23);
1555
1556             fscal            = felec;
1557
1558             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1559
1560             /* Update vectorial force */
1561             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1562             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1563             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1564
1565             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1566             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1567             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1568
1569             /**************************
1570              * CALCULATE INTERACTIONS *
1571              **************************/
1572
1573             /* COULOMB ELECTROSTATICS */
1574             velec            = _mm_mul_pd(qq31,rinv31);
1575             felec            = _mm_mul_pd(velec,rinvsq31);
1576
1577             fscal            = felec;
1578
1579             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1580
1581             /* Update vectorial force */
1582             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1583             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1584             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1585
1586             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1587             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1588             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             /* COULOMB ELECTROSTATICS */
1595             velec            = _mm_mul_pd(qq32,rinv32);
1596             felec            = _mm_mul_pd(velec,rinvsq32);
1597
1598             fscal            = felec;
1599
1600             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1601
1602             /* Update vectorial force */
1603             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1604             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1605             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1606
1607             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1608             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1609             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1610
1611             /**************************
1612              * CALCULATE INTERACTIONS *
1613              **************************/
1614
1615             /* COULOMB ELECTROSTATICS */
1616             velec            = _mm_mul_pd(qq33,rinv33);
1617             felec            = _mm_mul_pd(velec,rinvsq33);
1618
1619             fscal            = felec;
1620
1621             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1622
1623             /* Update vectorial force */
1624             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1625             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1626             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1627
1628             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1629             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1630             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1631
1632             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1633
1634             /* Inner loop uses 303 flops */
1635         }
1636
1637         /* End of innermost loop */
1638
1639         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1640                                               f+i_coord_offset,fshift+i_shift_offset);
1641
1642         /* Increment number of inner iterations */
1643         inneriter                  += j_index_end - j_index_start;
1644
1645         /* Outer loop uses 24 flops */
1646     }
1647
1648     /* Increment number of outer iterations */
1649     outeriter        += nri;
1650
1651     /* Update outer/inner flops */
1652
1653     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1654 }