src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRF_VdwLJ_GeomW4P1_sse2_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 sse2_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_double.h"
  50 #include "kernelutil_x86_sse2_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4P1_VF_sse2_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water4-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwLJ_GeomW4P1_VF_sse2_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     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  94     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  95     __m128d          dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  96     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     int              nvdwtype;
  99     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 103     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 104     __m128d          dummy_mask,cutoff_mask;
 105     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 106     __m128d          one     = _mm_set1_pd(1.0);
 107     __m128d          two     = _mm_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     krf              = _mm_set1_pd(fr->ic->k_rf);
 122     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 123     crf              = _mm_set1_pd(fr->ic->c_rf);
 124     nvdwtype         = fr->ntype;
 125     vdwparam         = fr->nbfp;
 126     vdwtype          = mdatoms->typeA;
 127
 128     /* Setup water-specific parameters */
 129     inr              = nlist->iinr[0];
 130     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 131     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 132     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 133     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 134
 135     /* Avoid stupid compiler warnings */
 136     jnrA = jnrB = 0;
 137     j_coord_offsetA = 0;
 138     j_coord_offsetB = 0;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     /* Start outer loop over neighborlists */
 144     for(iidx=0; iidx<nri; iidx++)
 145     {
 146         /* Load shift vector for this list */
 147         i_shift_offset   = DIM*shiftidx[iidx];
 148
 149         /* Load limits for loop over neighbors */
 150         j_index_start    = jindex[iidx];
 151         j_index_end      = jindex[iidx+1];
 152
 153         /* Get outer coordinate index */
 154         inr              = iinr[iidx];
 155         i_coord_offset   = DIM*inr;
 156
 157         /* Load i particle coords and add shift vector */
 158         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 159                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 160
 161         fix0             = _mm_setzero_pd();
 162         fiy0             = _mm_setzero_pd();
 163         fiz0             = _mm_setzero_pd();
 164         fix1             = _mm_setzero_pd();
 165         fiy1             = _mm_setzero_pd();
 166         fiz1             = _mm_setzero_pd();
 167         fix2             = _mm_setzero_pd();
 168         fiy2             = _mm_setzero_pd();
 169         fiz2             = _mm_setzero_pd();
 170         fix3             = _mm_setzero_pd();
 171         fiy3             = _mm_setzero_pd();
 172         fiz3             = _mm_setzero_pd();
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_pd();
 176         vvdwsum          = _mm_setzero_pd();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187
 188             /* load j atom coordinates */
 189             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 190                                               &jx0,&jy0,&jz0);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm_sub_pd(ix0,jx0);
 194             dy00             = _mm_sub_pd(iy0,jy0);
 195             dz00             = _mm_sub_pd(iz0,jz0);
 196             dx10             = _mm_sub_pd(ix1,jx0);
 197             dy10             = _mm_sub_pd(iy1,jy0);
 198             dz10             = _mm_sub_pd(iz1,jz0);
 199             dx20             = _mm_sub_pd(ix2,jx0);
 200             dy20             = _mm_sub_pd(iy2,jy0);
 201             dz20             = _mm_sub_pd(iz2,jz0);
 202             dx30             = _mm_sub_pd(ix3,jx0);
 203             dy30             = _mm_sub_pd(iy3,jy0);
 204             dz30             = _mm_sub_pd(iz3,jz0);
 205
 206             /* Calculate squared distance and things based on it */
 207             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 208             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 209             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 210             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 211
 212             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 213             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 214             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 215
 216             rinvsq00         = gmx_mm_inv_pd(rsq00);
 217             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 218             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 219             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 220
 221             /* Load parameters for j particles */
 222             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 223             vdwjidx0A        = 2*vdwtype[jnrA+0];
 224             vdwjidx0B        = 2*vdwtype[jnrB+0];
 225
 226             fjx0             = _mm_setzero_pd();
 227             fjy0             = _mm_setzero_pd();
 228             fjz0             = _mm_setzero_pd();
 229
 230             /**************************
 231              * CALCULATE INTERACTIONS *
 232              **************************/
 233
 234             /* Compute parameters for interactions between i and j atoms */
 235             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 236                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 237
 238             /* LENNARD-JONES DISPERSION/REPULSION */
 239
 240             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 241             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 242             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 243             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 244             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 245
 246             /* Update potential sum for this i atom from the interaction with this j atom. */
 247             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 248
 249             fscal            = fvdw;
 250
 251             /* Calculate temporary vectorial force */
 252             tx               = _mm_mul_pd(fscal,dx00);
 253             ty               = _mm_mul_pd(fscal,dy00);
 254             tz               = _mm_mul_pd(fscal,dz00);
 255
 256             /* Update vectorial force */
 257             fix0             = _mm_add_pd(fix0,tx);
 258             fiy0             = _mm_add_pd(fiy0,ty);
 259             fiz0             = _mm_add_pd(fiz0,tz);
 260
 261             fjx0             = _mm_add_pd(fjx0,tx);
 262             fjy0             = _mm_add_pd(fjy0,ty);
 263             fjz0             = _mm_add_pd(fjz0,tz);
 264
 265             /**************************
 266              * CALCULATE INTERACTIONS *
 267              **************************/
 268
 269             /* Compute parameters for interactions between i and j atoms */
 270             qq10             = _mm_mul_pd(iq1,jq0);
 271
 272             /* REACTION-FIELD ELECTROSTATICS */
 273             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 274             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 275
 276             /* Update potential sum for this i atom from the interaction with this j atom. */
 277             velecsum         = _mm_add_pd(velecsum,velec);
 278
 279             fscal            = felec;
 280
 281             /* Calculate temporary vectorial force */
 282             tx               = _mm_mul_pd(fscal,dx10);
 283             ty               = _mm_mul_pd(fscal,dy10);
 284             tz               = _mm_mul_pd(fscal,dz10);
 285
 286             /* Update vectorial force */
 287             fix1             = _mm_add_pd(fix1,tx);
 288             fiy1             = _mm_add_pd(fiy1,ty);
 289             fiz1             = _mm_add_pd(fiz1,tz);
 290
 291             fjx0             = _mm_add_pd(fjx0,tx);
 292             fjy0             = _mm_add_pd(fjy0,ty);
 293             fjz0             = _mm_add_pd(fjz0,tz);
 294
 295             /**************************
 296              * CALCULATE INTERACTIONS *
 297              **************************/
 298
 299             /* Compute parameters for interactions between i and j atoms */
 300             qq20             = _mm_mul_pd(iq2,jq0);
 301
 302             /* REACTION-FIELD ELECTROSTATICS */
 303             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 304             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 305
 306             /* Update potential sum for this i atom from the interaction with this j atom. */
 307             velecsum         = _mm_add_pd(velecsum,velec);
 308
 309             fscal            = felec;
 310
 311             /* Calculate temporary vectorial force */
 312             tx               = _mm_mul_pd(fscal,dx20);
 313             ty               = _mm_mul_pd(fscal,dy20);
 314             tz               = _mm_mul_pd(fscal,dz20);
 315
 316             /* Update vectorial force */
 317             fix2             = _mm_add_pd(fix2,tx);
 318             fiy2             = _mm_add_pd(fiy2,ty);
 319             fiz2             = _mm_add_pd(fiz2,tz);
 320
 321             fjx0             = _mm_add_pd(fjx0,tx);
 322             fjy0             = _mm_add_pd(fjy0,ty);
 323             fjz0             = _mm_add_pd(fjz0,tz);
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             /* Compute parameters for interactions between i and j atoms */
 330             qq30             = _mm_mul_pd(iq3,jq0);
 331
 332             /* REACTION-FIELD ELECTROSTATICS */
 333             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_add_pd(rinv30,_mm_mul_pd(krf,rsq30)),crf));
 334             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 335
 336             /* Update potential sum for this i atom from the interaction with this j atom. */
 337             velecsum         = _mm_add_pd(velecsum,velec);
 338
 339             fscal            = felec;
 340
 341             /* Calculate temporary vectorial force */
 342             tx               = _mm_mul_pd(fscal,dx30);
 343             ty               = _mm_mul_pd(fscal,dy30);
 344             tz               = _mm_mul_pd(fscal,dz30);
 345
 346             /* Update vectorial force */
 347             fix3             = _mm_add_pd(fix3,tx);
 348             fiy3             = _mm_add_pd(fiy3,ty);
 349             fiz3             = _mm_add_pd(fiz3,tz);
 350
 351             fjx0             = _mm_add_pd(fjx0,tx);
 352             fjy0             = _mm_add_pd(fjy0,ty);
 353             fjz0             = _mm_add_pd(fjz0,tz);
 354
 355             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 356
 357             /* Inner loop uses 131 flops */
 358         }
 359
 360         if(jidx<j_index_end)
 361         {
 362
 363             jnrA             = jjnr[jidx];
 364             j_coord_offsetA  = DIM*jnrA;
 365
 366             /* load j atom coordinates */
 367             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 368                                               &jx0,&jy0,&jz0);
 369
 370             /* Calculate displacement vector */
 371             dx00             = _mm_sub_pd(ix0,jx0);
 372             dy00             = _mm_sub_pd(iy0,jy0);
 373             dz00             = _mm_sub_pd(iz0,jz0);
 374             dx10             = _mm_sub_pd(ix1,jx0);
 375             dy10             = _mm_sub_pd(iy1,jy0);
 376             dz10             = _mm_sub_pd(iz1,jz0);
 377             dx20             = _mm_sub_pd(ix2,jx0);
 378             dy20             = _mm_sub_pd(iy2,jy0);
 379             dz20             = _mm_sub_pd(iz2,jz0);
 380             dx30             = _mm_sub_pd(ix3,jx0);
 381             dy30             = _mm_sub_pd(iy3,jy0);
 382             dz30             = _mm_sub_pd(iz3,jz0);
 383
 384             /* Calculate squared distance and things based on it */
 385             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 386             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 387             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 388             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 389
 390             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 391             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 392             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 393
 394             rinvsq00         = gmx_mm_inv_pd(rsq00);
 395             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 396             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 397             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 398
 399             /* Load parameters for j particles */
 400             jq0              = _mm_load_sd(charge+jnrA+0);
 401             vdwjidx0A        = 2*vdwtype[jnrA+0];
 402
 403             fjx0             = _mm_setzero_pd();
 404             fjy0             = _mm_setzero_pd();
 405             fjz0             = _mm_setzero_pd();
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             /* Compute parameters for interactions between i and j atoms */
 412             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 413
 414             /* LENNARD-JONES DISPERSION/REPULSION */
 415
 416             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 417             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 418             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 419             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 420             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 421
 422             /* Update potential sum for this i atom from the interaction with this j atom. */
 423             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 424             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 425
 426             fscal            = fvdw;
 427
 428             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 429
 430             /* Calculate temporary vectorial force */
 431             tx               = _mm_mul_pd(fscal,dx00);
 432             ty               = _mm_mul_pd(fscal,dy00);
 433             tz               = _mm_mul_pd(fscal,dz00);
 434
 435             /* Update vectorial force */
 436             fix0             = _mm_add_pd(fix0,tx);
 437             fiy0             = _mm_add_pd(fiy0,ty);
 438             fiz0             = _mm_add_pd(fiz0,tz);
 439
 440             fjx0             = _mm_add_pd(fjx0,tx);
 441             fjy0             = _mm_add_pd(fjy0,ty);
 442             fjz0             = _mm_add_pd(fjz0,tz);
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             /* Compute parameters for interactions between i and j atoms */
 449             qq10             = _mm_mul_pd(iq1,jq0);
 450
 451             /* REACTION-FIELD ELECTROSTATICS */
 452             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 453             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 457             velecsum         = _mm_add_pd(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 462
 463             /* Calculate temporary vectorial force */
 464             tx               = _mm_mul_pd(fscal,dx10);
 465             ty               = _mm_mul_pd(fscal,dy10);
 466             tz               = _mm_mul_pd(fscal,dz10);
 467
 468             /* Update vectorial force */
 469             fix1             = _mm_add_pd(fix1,tx);
 470             fiy1             = _mm_add_pd(fiy1,ty);
 471             fiz1             = _mm_add_pd(fiz1,tz);
 472
 473             fjx0             = _mm_add_pd(fjx0,tx);
 474             fjy0             = _mm_add_pd(fjy0,ty);
 475             fjz0             = _mm_add_pd(fjz0,tz);
 476
 477             /**************************
 478              * CALCULATE INTERACTIONS *
 479              **************************/
 480
 481             /* Compute parameters for interactions between i and j atoms */
 482             qq20             = _mm_mul_pd(iq2,jq0);
 483
 484             /* REACTION-FIELD ELECTROSTATICS */
 485             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 486             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 487
 488             /* Update potential sum for this i atom from the interaction with this j atom. */
 489             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 490             velecsum         = _mm_add_pd(velecsum,velec);
 491
 492             fscal            = felec;
 493
 494             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 495
 496             /* Calculate temporary vectorial force */
 497             tx               = _mm_mul_pd(fscal,dx20);
 498             ty               = _mm_mul_pd(fscal,dy20);
 499             tz               = _mm_mul_pd(fscal,dz20);
 500
 501             /* Update vectorial force */
 502             fix2             = _mm_add_pd(fix2,tx);
 503             fiy2             = _mm_add_pd(fiy2,ty);
 504             fiz2             = _mm_add_pd(fiz2,tz);
 505
 506             fjx0             = _mm_add_pd(fjx0,tx);
 507             fjy0             = _mm_add_pd(fjy0,ty);
 508             fjz0             = _mm_add_pd(fjz0,tz);
 509
 510             /**************************
 511              * CALCULATE INTERACTIONS *
 512              **************************/
 513
 514             /* Compute parameters for interactions between i and j atoms */
 515             qq30             = _mm_mul_pd(iq3,jq0);
 516
 517             /* REACTION-FIELD ELECTROSTATICS */
 518             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_add_pd(rinv30,_mm_mul_pd(krf,rsq30)),crf));
 519             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 520
 521             /* Update potential sum for this i atom from the interaction with this j atom. */
 522             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 523             velecsum         = _mm_add_pd(velecsum,velec);
 524
 525             fscal            = felec;
 526
 527             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 528
 529             /* Calculate temporary vectorial force */
 530             tx               = _mm_mul_pd(fscal,dx30);
 531             ty               = _mm_mul_pd(fscal,dy30);
 532             tz               = _mm_mul_pd(fscal,dz30);
 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             fjx0             = _mm_add_pd(fjx0,tx);
 540             fjy0             = _mm_add_pd(fjy0,ty);
 541             fjz0             = _mm_add_pd(fjz0,tz);
 542
 543             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 544
 545             /* Inner loop uses 131 flops */
 546         }
 547
 548         /* End of innermost loop */
 549
 550         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 551                                               f+i_coord_offset,fshift+i_shift_offset);
 552
 553         ggid                        = gid[iidx];
 554         /* Update potential energies */
 555         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 556         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 557
 558         /* Increment number of inner iterations */
 559         inneriter                  += j_index_end - j_index_start;
 560
 561         /* Outer loop uses 26 flops */
 562     }
 563
 564     /* Increment number of outer iterations */
 565     outeriter        += nri;
 566
 567     /* Update outer/inner flops */
 568
 569     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*131);
 570 }
 571 /*
 572  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sse2_double
 573  * Electrostatics interaction: ReactionField
 574  * VdW interaction:            LennardJones
 575  * Geometry:                   Water4-Particle
 576  * Calculate force/pot:        Force
 577  */
 578 void
 579 nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sse2_double
 580                     (t_nblist                    * gmx_restrict       nlist,
 581                      rvec                        * gmx_restrict          xx,
 582                      rvec                        * gmx_restrict          ff,
 583                      t_forcerec                  * gmx_restrict          fr,
 584                      t_mdatoms                   * gmx_restrict     mdatoms,
 585                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 586                      t_nrnb                      * gmx_restrict        nrnb)
 587 {
 588     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 589      * just 0 for non-waters.
 590      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 591      * jnr indices corresponding to data put in the four positions in the SIMD register.
 592      */
 593     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 594     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 595     int              jnrA,jnrB;
 596     int              j_coord_offsetA,j_coord_offsetB;
 597     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 598     real             rcutoff_scalar;
 599     real             *shiftvec,*fshift,*x,*f;
 600     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 601     int              vdwioffset0;
 602     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 603     int              vdwioffset1;
 604     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 605     int              vdwioffset2;
 606     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 607     int              vdwioffset3;
 608     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 609     int              vdwjidx0A,vdwjidx0B;
 610     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 611     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 612     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 613     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 614     __m128d          dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 615     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 616     real             *charge;
 617     int              nvdwtype;
 618     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 619     int              *vdwtype;
 620     real             *vdwparam;
 621     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 622     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 623     __m128d          dummy_mask,cutoff_mask;
 624     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 625     __m128d          one     = _mm_set1_pd(1.0);
 626     __m128d          two     = _mm_set1_pd(2.0);
 627     x                = xx[0];
 628     f                = ff[0];
 629
 630     nri              = nlist->nri;
 631     iinr             = nlist->iinr;
 632     jindex           = nlist->jindex;
 633     jjnr             = nlist->jjnr;
 634     shiftidx         = nlist->shift;
 635     gid              = nlist->gid;
 636     shiftvec         = fr->shift_vec[0];
 637     fshift           = fr->fshift[0];
 638     facel            = _mm_set1_pd(fr->epsfac);
 639     charge           = mdatoms->chargeA;
 640     krf              = _mm_set1_pd(fr->ic->k_rf);
 641     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 642     crf              = _mm_set1_pd(fr->ic->c_rf);
 643     nvdwtype         = fr->ntype;
 644     vdwparam         = fr->nbfp;
 645     vdwtype          = mdatoms->typeA;
 646
 647     /* Setup water-specific parameters */
 648     inr              = nlist->iinr[0];
 649     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 650     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 651     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 652     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 653
 654     /* Avoid stupid compiler warnings */
 655     jnrA = jnrB = 0;
 656     j_coord_offsetA = 0;
 657     j_coord_offsetB = 0;
 658
 659     outeriter        = 0;
 660     inneriter        = 0;
 661
 662     /* Start outer loop over neighborlists */
 663     for(iidx=0; iidx<nri; iidx++)
 664     {
 665         /* Load shift vector for this list */
 666         i_shift_offset   = DIM*shiftidx[iidx];
 667
 668         /* Load limits for loop over neighbors */
 669         j_index_start    = jindex[iidx];
 670         j_index_end      = jindex[iidx+1];
 671
 672         /* Get outer coordinate index */
 673         inr              = iinr[iidx];
 674         i_coord_offset   = DIM*inr;
 675
 676         /* Load i particle coords and add shift vector */
 677         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 678                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 679
 680         fix0             = _mm_setzero_pd();
 681         fiy0             = _mm_setzero_pd();
 682         fiz0             = _mm_setzero_pd();
 683         fix1             = _mm_setzero_pd();
 684         fiy1             = _mm_setzero_pd();
 685         fiz1             = _mm_setzero_pd();
 686         fix2             = _mm_setzero_pd();
 687         fiy2             = _mm_setzero_pd();
 688         fiz2             = _mm_setzero_pd();
 689         fix3             = _mm_setzero_pd();
 690         fiy3             = _mm_setzero_pd();
 691         fiz3             = _mm_setzero_pd();
 692
 693         /* Start inner kernel loop */
 694         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 695         {
 696
 697             /* Get j neighbor index, and coordinate index */
 698             jnrA             = jjnr[jidx];
 699             jnrB             = jjnr[jidx+1];
 700             j_coord_offsetA  = DIM*jnrA;
 701             j_coord_offsetB  = DIM*jnrB;
 702
 703             /* load j atom coordinates */
 704             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 705                                               &jx0,&jy0,&jz0);
 706
 707             /* Calculate displacement vector */
 708             dx00             = _mm_sub_pd(ix0,jx0);
 709             dy00             = _mm_sub_pd(iy0,jy0);
 710             dz00             = _mm_sub_pd(iz0,jz0);
 711             dx10             = _mm_sub_pd(ix1,jx0);
 712             dy10             = _mm_sub_pd(iy1,jy0);
 713             dz10             = _mm_sub_pd(iz1,jz0);
 714             dx20             = _mm_sub_pd(ix2,jx0);
 715             dy20             = _mm_sub_pd(iy2,jy0);
 716             dz20             = _mm_sub_pd(iz2,jz0);
 717             dx30             = _mm_sub_pd(ix3,jx0);
 718             dy30             = _mm_sub_pd(iy3,jy0);
 719             dz30             = _mm_sub_pd(iz3,jz0);
 720
 721             /* Calculate squared distance and things based on it */
 722             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 723             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 724             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 725             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 726
 727             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 728             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 729             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 730
 731             rinvsq00         = gmx_mm_inv_pd(rsq00);
 732             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 733             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 734             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 735
 736             /* Load parameters for j particles */
 737             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 738             vdwjidx0A        = 2*vdwtype[jnrA+0];
 739             vdwjidx0B        = 2*vdwtype[jnrB+0];
 740
 741             fjx0             = _mm_setzero_pd();
 742             fjy0             = _mm_setzero_pd();
 743             fjz0             = _mm_setzero_pd();
 744
 745             /**************************
 746              * CALCULATE INTERACTIONS *
 747              **************************/
 748
 749             /* Compute parameters for interactions between i and j atoms */
 750             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 751                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 752
 753             /* LENNARD-JONES DISPERSION/REPULSION */
 754
 755             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 756             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 757
 758             fscal            = fvdw;
 759
 760             /* Calculate temporary vectorial force */
 761             tx               = _mm_mul_pd(fscal,dx00);
 762             ty               = _mm_mul_pd(fscal,dy00);
 763             tz               = _mm_mul_pd(fscal,dz00);
 764
 765             /* Update vectorial force */
 766             fix0             = _mm_add_pd(fix0,tx);
 767             fiy0             = _mm_add_pd(fiy0,ty);
 768             fiz0             = _mm_add_pd(fiz0,tz);
 769
 770             fjx0             = _mm_add_pd(fjx0,tx);
 771             fjy0             = _mm_add_pd(fjy0,ty);
 772             fjz0             = _mm_add_pd(fjz0,tz);
 773
 774             /**************************
 775              * CALCULATE INTERACTIONS *
 776              **************************/
 777
 778             /* Compute parameters for interactions between i and j atoms */
 779             qq10             = _mm_mul_pd(iq1,jq0);
 780
 781             /* REACTION-FIELD ELECTROSTATICS */
 782             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 783
 784             fscal            = felec;
 785
 786             /* Calculate temporary vectorial force */
 787             tx               = _mm_mul_pd(fscal,dx10);
 788             ty               = _mm_mul_pd(fscal,dy10);
 789             tz               = _mm_mul_pd(fscal,dz10);
 790
 791             /* Update vectorial force */
 792             fix1             = _mm_add_pd(fix1,tx);
 793             fiy1             = _mm_add_pd(fiy1,ty);
 794             fiz1             = _mm_add_pd(fiz1,tz);
 795
 796             fjx0             = _mm_add_pd(fjx0,tx);
 797             fjy0             = _mm_add_pd(fjy0,ty);
 798             fjz0             = _mm_add_pd(fjz0,tz);
 799
 800             /**************************
 801              * CALCULATE INTERACTIONS *
 802              **************************/
 803
 804             /* Compute parameters for interactions between i and j atoms */
 805             qq20             = _mm_mul_pd(iq2,jq0);
 806
 807             /* REACTION-FIELD ELECTROSTATICS */
 808             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 809
 810             fscal            = felec;
 811
 812             /* Calculate temporary vectorial force */
 813             tx               = _mm_mul_pd(fscal,dx20);
 814             ty               = _mm_mul_pd(fscal,dy20);
 815             tz               = _mm_mul_pd(fscal,dz20);
 816
 817             /* Update vectorial force */
 818             fix2             = _mm_add_pd(fix2,tx);
 819             fiy2             = _mm_add_pd(fiy2,ty);
 820             fiz2             = _mm_add_pd(fiz2,tz);
 821
 822             fjx0             = _mm_add_pd(fjx0,tx);
 823             fjy0             = _mm_add_pd(fjy0,ty);
 824             fjz0             = _mm_add_pd(fjz0,tz);
 825
 826             /**************************
 827              * CALCULATE INTERACTIONS *
 828              **************************/
 829
 830             /* Compute parameters for interactions between i and j atoms */
 831             qq30             = _mm_mul_pd(iq3,jq0);
 832
 833             /* REACTION-FIELD ELECTROSTATICS */
 834             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 835
 836             fscal            = felec;
 837
 838             /* Calculate temporary vectorial force */
 839             tx               = _mm_mul_pd(fscal,dx30);
 840             ty               = _mm_mul_pd(fscal,dy30);
 841             tz               = _mm_mul_pd(fscal,dz30);
 842
 843             /* Update vectorial force */
 844             fix3             = _mm_add_pd(fix3,tx);
 845             fiy3             = _mm_add_pd(fiy3,ty);
 846             fiz3             = _mm_add_pd(fiz3,tz);
 847
 848             fjx0             = _mm_add_pd(fjx0,tx);
 849             fjy0             = _mm_add_pd(fjy0,ty);
 850             fjz0             = _mm_add_pd(fjz0,tz);
 851
 852             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 853
 854             /* Inner loop uses 111 flops */
 855         }
 856
 857         if(jidx<j_index_end)
 858         {
 859
 860             jnrA             = jjnr[jidx];
 861             j_coord_offsetA  = DIM*jnrA;
 862
 863             /* load j atom coordinates */
 864             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 865                                               &jx0,&jy0,&jz0);
 866
 867             /* Calculate displacement vector */
 868             dx00             = _mm_sub_pd(ix0,jx0);
 869             dy00             = _mm_sub_pd(iy0,jy0);
 870             dz00             = _mm_sub_pd(iz0,jz0);
 871             dx10             = _mm_sub_pd(ix1,jx0);
 872             dy10             = _mm_sub_pd(iy1,jy0);
 873             dz10             = _mm_sub_pd(iz1,jz0);
 874             dx20             = _mm_sub_pd(ix2,jx0);
 875             dy20             = _mm_sub_pd(iy2,jy0);
 876             dz20             = _mm_sub_pd(iz2,jz0);
 877             dx30             = _mm_sub_pd(ix3,jx0);
 878             dy30             = _mm_sub_pd(iy3,jy0);
 879             dz30             = _mm_sub_pd(iz3,jz0);
 880
 881             /* Calculate squared distance and things based on it */
 882             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 883             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 884             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 885             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 886
 887             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 888             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 889             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 890
 891             rinvsq00         = gmx_mm_inv_pd(rsq00);
 892             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 893             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 894             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 895
 896             /* Load parameters for j particles */
 897             jq0              = _mm_load_sd(charge+jnrA+0);
 898             vdwjidx0A        = 2*vdwtype[jnrA+0];
 899
 900             fjx0             = _mm_setzero_pd();
 901             fjy0             = _mm_setzero_pd();
 902             fjz0             = _mm_setzero_pd();
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             /* Compute parameters for interactions between i and j atoms */
 909             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 910
 911             /* LENNARD-JONES DISPERSION/REPULSION */
 912
 913             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 914             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 915
 916             fscal            = fvdw;
 917
 918             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 919
 920             /* Calculate temporary vectorial force */
 921             tx               = _mm_mul_pd(fscal,dx00);
 922             ty               = _mm_mul_pd(fscal,dy00);
 923             tz               = _mm_mul_pd(fscal,dz00);
 924
 925             /* Update vectorial force */
 926             fix0             = _mm_add_pd(fix0,tx);
 927             fiy0             = _mm_add_pd(fiy0,ty);
 928             fiz0             = _mm_add_pd(fiz0,tz);
 929
 930             fjx0             = _mm_add_pd(fjx0,tx);
 931             fjy0             = _mm_add_pd(fjy0,ty);
 932             fjz0             = _mm_add_pd(fjz0,tz);
 933
 934             /**************************
 935              * CALCULATE INTERACTIONS *
 936              **************************/
 937
 938             /* Compute parameters for interactions between i and j atoms */
 939             qq10             = _mm_mul_pd(iq1,jq0);
 940
 941             /* REACTION-FIELD ELECTROSTATICS */
 942             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 947
 948             /* Calculate temporary vectorial force */
 949             tx               = _mm_mul_pd(fscal,dx10);
 950             ty               = _mm_mul_pd(fscal,dy10);
 951             tz               = _mm_mul_pd(fscal,dz10);
 952
 953             /* Update vectorial force */
 954             fix1             = _mm_add_pd(fix1,tx);
 955             fiy1             = _mm_add_pd(fiy1,ty);
 956             fiz1             = _mm_add_pd(fiz1,tz);
 957
 958             fjx0             = _mm_add_pd(fjx0,tx);
 959             fjy0             = _mm_add_pd(fjy0,ty);
 960             fjz0             = _mm_add_pd(fjz0,tz);
 961
 962             /**************************
 963              * CALCULATE INTERACTIONS *
 964              **************************/
 965
 966             /* Compute parameters for interactions between i and j atoms */
 967             qq20             = _mm_mul_pd(iq2,jq0);
 968
 969             /* REACTION-FIELD ELECTROSTATICS */
 970             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 971
 972             fscal            = felec;
 973
 974             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 975
 976             /* Calculate temporary vectorial force */
 977             tx               = _mm_mul_pd(fscal,dx20);
 978             ty               = _mm_mul_pd(fscal,dy20);
 979             tz               = _mm_mul_pd(fscal,dz20);
 980
 981             /* Update vectorial force */
 982             fix2             = _mm_add_pd(fix2,tx);
 983             fiy2             = _mm_add_pd(fiy2,ty);
 984             fiz2             = _mm_add_pd(fiz2,tz);
 985
 986             fjx0             = _mm_add_pd(fjx0,tx);
 987             fjy0             = _mm_add_pd(fjy0,ty);
 988             fjz0             = _mm_add_pd(fjz0,tz);
 989
 990             /**************************
 991              * CALCULATE INTERACTIONS *
 992              **************************/
 993
 994             /* Compute parameters for interactions between i and j atoms */
 995             qq30             = _mm_mul_pd(iq3,jq0);
 996
 997             /* REACTION-FIELD ELECTROSTATICS */
 998             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 999
1000             fscal            = felec;
1001
1002             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1003
1004             /* Calculate temporary vectorial force */
1005             tx               = _mm_mul_pd(fscal,dx30);
1006             ty               = _mm_mul_pd(fscal,dy30);
1007             tz               = _mm_mul_pd(fscal,dz30);
1008
1009             /* Update vectorial force */
1010             fix3             = _mm_add_pd(fix3,tx);
1011             fiy3             = _mm_add_pd(fiy3,ty);
1012             fiz3             = _mm_add_pd(fiz3,tz);
1013
1014             fjx0             = _mm_add_pd(fjx0,tx);
1015             fjy0             = _mm_add_pd(fjy0,ty);
1016             fjz0             = _mm_add_pd(fjz0,tz);
1017
1018             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1019
1020             /* Inner loop uses 111 flops */
1021         }
1022
1023         /* End of innermost loop */
1024
1025         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1026                                               f+i_coord_offset,fshift+i_shift_offset);
1027
1028         /* Increment number of inner iterations */
1029         inneriter                  += j_index_end - j_index_start;
1030
1031         /* Outer loop uses 24 flops */
1032     }
1033
1034     /* Increment number of outer iterations */
1035     outeriter        += nri;
1036
1037     /* Update outer/inner flops */
1038
1039     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*111);
1040 }