src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRF_VdwLJ_GeomW4P1_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4P1_VF_sse4_1_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwLJ_GeomW4P1_VF_sse4_1_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwioffset3;
  73     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B;
  75     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  77     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  78     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  79     __m128d          dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  80     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  81     real             *charge;
  82     int              nvdwtype;
  83     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  84     int              *vdwtype;
  85     real             *vdwparam;
  86     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  87     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  88     __m128d          dummy_mask,cutoff_mask;
  89     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  90     __m128d          one     = _mm_set1_pd(1.0);
  91     __m128d          two     = _mm_set1_pd(2.0);
  92     x                = xx[0];
  93     f                = ff[0];
  94
  95     nri              = nlist->nri;
  96     iinr             = nlist->iinr;
  97     jindex           = nlist->jindex;
  98     jjnr             = nlist->jjnr;
  99     shiftidx         = nlist->shift;
 100     gid              = nlist->gid;
 101     shiftvec         = fr->shift_vec[0];
 102     fshift           = fr->fshift[0];
 103     facel            = _mm_set1_pd(fr->epsfac);
 104     charge           = mdatoms->chargeA;
 105     krf              = _mm_set1_pd(fr->ic->k_rf);
 106     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 107     crf              = _mm_set1_pd(fr->ic->c_rf);
 108     nvdwtype         = fr->ntype;
 109     vdwparam         = fr->nbfp;
 110     vdwtype          = mdatoms->typeA;
 111
 112     /* Setup water-specific parameters */
 113     inr              = nlist->iinr[0];
 114     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 115     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 116     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 117     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 118
 119     /* Avoid stupid compiler warnings */
 120     jnrA = jnrB = 0;
 121     j_coord_offsetA = 0;
 122     j_coord_offsetB = 0;
 123
 124     outeriter        = 0;
 125     inneriter        = 0;
 126
 127     /* Start outer loop over neighborlists */
 128     for(iidx=0; iidx<nri; iidx++)
 129     {
 130         /* Load shift vector for this list */
 131         i_shift_offset   = DIM*shiftidx[iidx];
 132
 133         /* Load limits for loop over neighbors */
 134         j_index_start    = jindex[iidx];
 135         j_index_end      = jindex[iidx+1];
 136
 137         /* Get outer coordinate index */
 138         inr              = iinr[iidx];
 139         i_coord_offset   = DIM*inr;
 140
 141         /* Load i particle coords and add shift vector */
 142         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 143                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 144
 145         fix0             = _mm_setzero_pd();
 146         fiy0             = _mm_setzero_pd();
 147         fiz0             = _mm_setzero_pd();
 148         fix1             = _mm_setzero_pd();
 149         fiy1             = _mm_setzero_pd();
 150         fiz1             = _mm_setzero_pd();
 151         fix2             = _mm_setzero_pd();
 152         fiy2             = _mm_setzero_pd();
 153         fiz2             = _mm_setzero_pd();
 154         fix3             = _mm_setzero_pd();
 155         fiy3             = _mm_setzero_pd();
 156         fiz3             = _mm_setzero_pd();
 157
 158         /* Reset potential sums */
 159         velecsum         = _mm_setzero_pd();
 160         vvdwsum          = _mm_setzero_pd();
 161
 162         /* Start inner kernel loop */
 163         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 164         {
 165
 166             /* Get j neighbor index, and coordinate index */
 167             jnrA             = jjnr[jidx];
 168             jnrB             = jjnr[jidx+1];
 169             j_coord_offsetA  = DIM*jnrA;
 170             j_coord_offsetB  = DIM*jnrB;
 171
 172             /* load j atom coordinates */
 173             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 174                                               &jx0,&jy0,&jz0);
 175
 176             /* Calculate displacement vector */
 177             dx00             = _mm_sub_pd(ix0,jx0);
 178             dy00             = _mm_sub_pd(iy0,jy0);
 179             dz00             = _mm_sub_pd(iz0,jz0);
 180             dx10             = _mm_sub_pd(ix1,jx0);
 181             dy10             = _mm_sub_pd(iy1,jy0);
 182             dz10             = _mm_sub_pd(iz1,jz0);
 183             dx20             = _mm_sub_pd(ix2,jx0);
 184             dy20             = _mm_sub_pd(iy2,jy0);
 185             dz20             = _mm_sub_pd(iz2,jz0);
 186             dx30             = _mm_sub_pd(ix3,jx0);
 187             dy30             = _mm_sub_pd(iy3,jy0);
 188             dz30             = _mm_sub_pd(iz3,jz0);
 189
 190             /* Calculate squared distance and things based on it */
 191             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 192             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 193             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 194             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 195
 196             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 197             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 198             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 199
 200             rinvsq00         = gmx_mm_inv_pd(rsq00);
 201             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 202             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 203             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 204
 205             /* Load parameters for j particles */
 206             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 207             vdwjidx0A        = 2*vdwtype[jnrA+0];
 208             vdwjidx0B        = 2*vdwtype[jnrB+0];
 209
 210             fjx0             = _mm_setzero_pd();
 211             fjy0             = _mm_setzero_pd();
 212             fjz0             = _mm_setzero_pd();
 213
 214             /**************************
 215              * CALCULATE INTERACTIONS *
 216              **************************/
 217
 218             /* Compute parameters for interactions between i and j atoms */
 219             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 220                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 221
 222             /* LENNARD-JONES DISPERSION/REPULSION */
 223
 224             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 225             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 226             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 227             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 228             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 229
 230             /* Update potential sum for this i atom from the interaction with this j atom. */
 231             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 232
 233             fscal            = fvdw;
 234
 235             /* Calculate temporary vectorial force */
 236             tx               = _mm_mul_pd(fscal,dx00);
 237             ty               = _mm_mul_pd(fscal,dy00);
 238             tz               = _mm_mul_pd(fscal,dz00);
 239
 240             /* Update vectorial force */
 241             fix0             = _mm_add_pd(fix0,tx);
 242             fiy0             = _mm_add_pd(fiy0,ty);
 243             fiz0             = _mm_add_pd(fiz0,tz);
 244
 245             fjx0             = _mm_add_pd(fjx0,tx);
 246             fjy0             = _mm_add_pd(fjy0,ty);
 247             fjz0             = _mm_add_pd(fjz0,tz);
 248
 249             /**************************
 250              * CALCULATE INTERACTIONS *
 251              **************************/
 252
 253             /* Compute parameters for interactions between i and j atoms */
 254             qq10             = _mm_mul_pd(iq1,jq0);
 255
 256             /* REACTION-FIELD ELECTROSTATICS */
 257             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 258             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 259
 260             /* Update potential sum for this i atom from the interaction with this j atom. */
 261             velecsum         = _mm_add_pd(velecsum,velec);
 262
 263             fscal            = felec;
 264
 265             /* Calculate temporary vectorial force */
 266             tx               = _mm_mul_pd(fscal,dx10);
 267             ty               = _mm_mul_pd(fscal,dy10);
 268             tz               = _mm_mul_pd(fscal,dz10);
 269
 270             /* Update vectorial force */
 271             fix1             = _mm_add_pd(fix1,tx);
 272             fiy1             = _mm_add_pd(fiy1,ty);
 273             fiz1             = _mm_add_pd(fiz1,tz);
 274
 275             fjx0             = _mm_add_pd(fjx0,tx);
 276             fjy0             = _mm_add_pd(fjy0,ty);
 277             fjz0             = _mm_add_pd(fjz0,tz);
 278
 279             /**************************
 280              * CALCULATE INTERACTIONS *
 281              **************************/
 282
 283             /* Compute parameters for interactions between i and j atoms */
 284             qq20             = _mm_mul_pd(iq2,jq0);
 285
 286             /* REACTION-FIELD ELECTROSTATICS */
 287             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 288             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 289
 290             /* Update potential sum for this i atom from the interaction with this j atom. */
 291             velecsum         = _mm_add_pd(velecsum,velec);
 292
 293             fscal            = felec;
 294
 295             /* Calculate temporary vectorial force */
 296             tx               = _mm_mul_pd(fscal,dx20);
 297             ty               = _mm_mul_pd(fscal,dy20);
 298             tz               = _mm_mul_pd(fscal,dz20);
 299
 300             /* Update vectorial force */
 301             fix2             = _mm_add_pd(fix2,tx);
 302             fiy2             = _mm_add_pd(fiy2,ty);
 303             fiz2             = _mm_add_pd(fiz2,tz);
 304
 305             fjx0             = _mm_add_pd(fjx0,tx);
 306             fjy0             = _mm_add_pd(fjy0,ty);
 307             fjz0             = _mm_add_pd(fjz0,tz);
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             /* Compute parameters for interactions between i and j atoms */
 314             qq30             = _mm_mul_pd(iq3,jq0);
 315
 316             /* REACTION-FIELD ELECTROSTATICS */
 317             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_add_pd(rinv30,_mm_mul_pd(krf,rsq30)),crf));
 318             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 319
 320             /* Update potential sum for this i atom from the interaction with this j atom. */
 321             velecsum         = _mm_add_pd(velecsum,velec);
 322
 323             fscal            = felec;
 324
 325             /* Calculate temporary vectorial force */
 326             tx               = _mm_mul_pd(fscal,dx30);
 327             ty               = _mm_mul_pd(fscal,dy30);
 328             tz               = _mm_mul_pd(fscal,dz30);
 329
 330             /* Update vectorial force */
 331             fix3             = _mm_add_pd(fix3,tx);
 332             fiy3             = _mm_add_pd(fiy3,ty);
 333             fiz3             = _mm_add_pd(fiz3,tz);
 334
 335             fjx0             = _mm_add_pd(fjx0,tx);
 336             fjy0             = _mm_add_pd(fjy0,ty);
 337             fjz0             = _mm_add_pd(fjz0,tz);
 338
 339             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 340
 341             /* Inner loop uses 131 flops */
 342         }
 343
 344         if(jidx<j_index_end)
 345         {
 346
 347             jnrA             = jjnr[jidx];
 348             j_coord_offsetA  = DIM*jnrA;
 349
 350             /* load j atom coordinates */
 351             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 352                                               &jx0,&jy0,&jz0);
 353
 354             /* Calculate displacement vector */
 355             dx00             = _mm_sub_pd(ix0,jx0);
 356             dy00             = _mm_sub_pd(iy0,jy0);
 357             dz00             = _mm_sub_pd(iz0,jz0);
 358             dx10             = _mm_sub_pd(ix1,jx0);
 359             dy10             = _mm_sub_pd(iy1,jy0);
 360             dz10             = _mm_sub_pd(iz1,jz0);
 361             dx20             = _mm_sub_pd(ix2,jx0);
 362             dy20             = _mm_sub_pd(iy2,jy0);
 363             dz20             = _mm_sub_pd(iz2,jz0);
 364             dx30             = _mm_sub_pd(ix3,jx0);
 365             dy30             = _mm_sub_pd(iy3,jy0);
 366             dz30             = _mm_sub_pd(iz3,jz0);
 367
 368             /* Calculate squared distance and things based on it */
 369             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 370             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 371             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 372             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 373
 374             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 375             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 376             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 377
 378             rinvsq00         = gmx_mm_inv_pd(rsq00);
 379             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 380             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 381             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 382
 383             /* Load parameters for j particles */
 384             jq0              = _mm_load_sd(charge+jnrA+0);
 385             vdwjidx0A        = 2*vdwtype[jnrA+0];
 386
 387             fjx0             = _mm_setzero_pd();
 388             fjy0             = _mm_setzero_pd();
 389             fjz0             = _mm_setzero_pd();
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             /* Compute parameters for interactions between i and j atoms */
 396             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 397
 398             /* LENNARD-JONES DISPERSION/REPULSION */
 399
 400             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 401             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 402             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 403             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 404             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 408             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 409
 410             fscal            = fvdw;
 411
 412             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 413
 414             /* Calculate temporary vectorial force */
 415             tx               = _mm_mul_pd(fscal,dx00);
 416             ty               = _mm_mul_pd(fscal,dy00);
 417             tz               = _mm_mul_pd(fscal,dz00);
 418
 419             /* Update vectorial force */
 420             fix0             = _mm_add_pd(fix0,tx);
 421             fiy0             = _mm_add_pd(fiy0,ty);
 422             fiz0             = _mm_add_pd(fiz0,tz);
 423
 424             fjx0             = _mm_add_pd(fjx0,tx);
 425             fjy0             = _mm_add_pd(fjy0,ty);
 426             fjz0             = _mm_add_pd(fjz0,tz);
 427
 428             /**************************
 429              * CALCULATE INTERACTIONS *
 430              **************************/
 431
 432             /* Compute parameters for interactions between i and j atoms */
 433             qq10             = _mm_mul_pd(iq1,jq0);
 434
 435             /* REACTION-FIELD ELECTROSTATICS */
 436             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 437             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 438
 439             /* Update potential sum for this i atom from the interaction with this j atom. */
 440             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 441             velecsum         = _mm_add_pd(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 446
 447             /* Calculate temporary vectorial force */
 448             tx               = _mm_mul_pd(fscal,dx10);
 449             ty               = _mm_mul_pd(fscal,dy10);
 450             tz               = _mm_mul_pd(fscal,dz10);
 451
 452             /* Update vectorial force */
 453             fix1             = _mm_add_pd(fix1,tx);
 454             fiy1             = _mm_add_pd(fiy1,ty);
 455             fiz1             = _mm_add_pd(fiz1,tz);
 456
 457             fjx0             = _mm_add_pd(fjx0,tx);
 458             fjy0             = _mm_add_pd(fjy0,ty);
 459             fjz0             = _mm_add_pd(fjz0,tz);
 460
 461             /**************************
 462              * CALCULATE INTERACTIONS *
 463              **************************/
 464
 465             /* Compute parameters for interactions between i and j atoms */
 466             qq20             = _mm_mul_pd(iq2,jq0);
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 470             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 474             velecsum         = _mm_add_pd(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 479
 480             /* Calculate temporary vectorial force */
 481             tx               = _mm_mul_pd(fscal,dx20);
 482             ty               = _mm_mul_pd(fscal,dy20);
 483             tz               = _mm_mul_pd(fscal,dz20);
 484
 485             /* Update vectorial force */
 486             fix2             = _mm_add_pd(fix2,tx);
 487             fiy2             = _mm_add_pd(fiy2,ty);
 488             fiz2             = _mm_add_pd(fiz2,tz);
 489
 490             fjx0             = _mm_add_pd(fjx0,tx);
 491             fjy0             = _mm_add_pd(fjy0,ty);
 492             fjz0             = _mm_add_pd(fjz0,tz);
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             /* Compute parameters for interactions between i and j atoms */
 499             qq30             = _mm_mul_pd(iq3,jq0);
 500
 501             /* REACTION-FIELD ELECTROSTATICS */
 502             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_add_pd(rinv30,_mm_mul_pd(krf,rsq30)),crf));
 503             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 504
 505             /* Update potential sum for this i atom from the interaction with this j atom. */
 506             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 507             velecsum         = _mm_add_pd(velecsum,velec);
 508
 509             fscal            = felec;
 510
 511             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 512
 513             /* Calculate temporary vectorial force */
 514             tx               = _mm_mul_pd(fscal,dx30);
 515             ty               = _mm_mul_pd(fscal,dy30);
 516             tz               = _mm_mul_pd(fscal,dz30);
 517
 518             /* Update vectorial force */
 519             fix3             = _mm_add_pd(fix3,tx);
 520             fiy3             = _mm_add_pd(fiy3,ty);
 521             fiz3             = _mm_add_pd(fiz3,tz);
 522
 523             fjx0             = _mm_add_pd(fjx0,tx);
 524             fjy0             = _mm_add_pd(fjy0,ty);
 525             fjz0             = _mm_add_pd(fjz0,tz);
 526
 527             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 528
 529             /* Inner loop uses 131 flops */
 530         }
 531
 532         /* End of innermost loop */
 533
 534         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 535                                               f+i_coord_offset,fshift+i_shift_offset);
 536
 537         ggid                        = gid[iidx];
 538         /* Update potential energies */
 539         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 540         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 541
 542         /* Increment number of inner iterations */
 543         inneriter                  += j_index_end - j_index_start;
 544
 545         /* Outer loop uses 26 flops */
 546     }
 547
 548     /* Increment number of outer iterations */
 549     outeriter        += nri;
 550
 551     /* Update outer/inner flops */
 552
 553     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*131);
 554 }
 555 /*
 556  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sse4_1_double
 557  * Electrostatics interaction: ReactionField
 558  * VdW interaction:            LennardJones
 559  * Geometry:                   Water4-Particle
 560  * Calculate force/pot:        Force
 561  */
 562 void
 563 nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sse4_1_double
 564                     (t_nblist * gmx_restrict                nlist,
 565                      rvec * gmx_restrict                    xx,
 566                      rvec * gmx_restrict                    ff,
 567                      t_forcerec * gmx_restrict              fr,
 568                      t_mdatoms * gmx_restrict               mdatoms,
 569                      nb_kernel_data_t * gmx_restrict        kernel_data,
 570                      t_nrnb * gmx_restrict                  nrnb)
 571 {
 572     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 573      * just 0 for non-waters.
 574      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 575      * jnr indices corresponding to data put in the four positions in the SIMD register.
 576      */
 577     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 578     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 579     int              jnrA,jnrB;
 580     int              j_coord_offsetA,j_coord_offsetB;
 581     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 582     real             rcutoff_scalar;
 583     real             *shiftvec,*fshift,*x,*f;
 584     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 585     int              vdwioffset0;
 586     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 587     int              vdwioffset1;
 588     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 589     int              vdwioffset2;
 590     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 591     int              vdwioffset3;
 592     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 593     int              vdwjidx0A,vdwjidx0B;
 594     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 595     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 596     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 597     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 598     __m128d          dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 599     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 600     real             *charge;
 601     int              nvdwtype;
 602     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 603     int              *vdwtype;
 604     real             *vdwparam;
 605     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 606     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 607     __m128d          dummy_mask,cutoff_mask;
 608     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 609     __m128d          one     = _mm_set1_pd(1.0);
 610     __m128d          two     = _mm_set1_pd(2.0);
 611     x                = xx[0];
 612     f                = ff[0];
 613
 614     nri              = nlist->nri;
 615     iinr             = nlist->iinr;
 616     jindex           = nlist->jindex;
 617     jjnr             = nlist->jjnr;
 618     shiftidx         = nlist->shift;
 619     gid              = nlist->gid;
 620     shiftvec         = fr->shift_vec[0];
 621     fshift           = fr->fshift[0];
 622     facel            = _mm_set1_pd(fr->epsfac);
 623     charge           = mdatoms->chargeA;
 624     krf              = _mm_set1_pd(fr->ic->k_rf);
 625     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 626     crf              = _mm_set1_pd(fr->ic->c_rf);
 627     nvdwtype         = fr->ntype;
 628     vdwparam         = fr->nbfp;
 629     vdwtype          = mdatoms->typeA;
 630
 631     /* Setup water-specific parameters */
 632     inr              = nlist->iinr[0];
 633     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 634     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 635     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 636     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 637
 638     /* Avoid stupid compiler warnings */
 639     jnrA = jnrB = 0;
 640     j_coord_offsetA = 0;
 641     j_coord_offsetB = 0;
 642
 643     outeriter        = 0;
 644     inneriter        = 0;
 645
 646     /* Start outer loop over neighborlists */
 647     for(iidx=0; iidx<nri; iidx++)
 648     {
 649         /* Load shift vector for this list */
 650         i_shift_offset   = DIM*shiftidx[iidx];
 651
 652         /* Load limits for loop over neighbors */
 653         j_index_start    = jindex[iidx];
 654         j_index_end      = jindex[iidx+1];
 655
 656         /* Get outer coordinate index */
 657         inr              = iinr[iidx];
 658         i_coord_offset   = DIM*inr;
 659
 660         /* Load i particle coords and add shift vector */
 661         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 662                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 663
 664         fix0             = _mm_setzero_pd();
 665         fiy0             = _mm_setzero_pd();
 666         fiz0             = _mm_setzero_pd();
 667         fix1             = _mm_setzero_pd();
 668         fiy1             = _mm_setzero_pd();
 669         fiz1             = _mm_setzero_pd();
 670         fix2             = _mm_setzero_pd();
 671         fiy2             = _mm_setzero_pd();
 672         fiz2             = _mm_setzero_pd();
 673         fix3             = _mm_setzero_pd();
 674         fiy3             = _mm_setzero_pd();
 675         fiz3             = _mm_setzero_pd();
 676
 677         /* Start inner kernel loop */
 678         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 679         {
 680
 681             /* Get j neighbor index, and coordinate index */
 682             jnrA             = jjnr[jidx];
 683             jnrB             = jjnr[jidx+1];
 684             j_coord_offsetA  = DIM*jnrA;
 685             j_coord_offsetB  = DIM*jnrB;
 686
 687             /* load j atom coordinates */
 688             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 689                                               &jx0,&jy0,&jz0);
 690
 691             /* Calculate displacement vector */
 692             dx00             = _mm_sub_pd(ix0,jx0);
 693             dy00             = _mm_sub_pd(iy0,jy0);
 694             dz00             = _mm_sub_pd(iz0,jz0);
 695             dx10             = _mm_sub_pd(ix1,jx0);
 696             dy10             = _mm_sub_pd(iy1,jy0);
 697             dz10             = _mm_sub_pd(iz1,jz0);
 698             dx20             = _mm_sub_pd(ix2,jx0);
 699             dy20             = _mm_sub_pd(iy2,jy0);
 700             dz20             = _mm_sub_pd(iz2,jz0);
 701             dx30             = _mm_sub_pd(ix3,jx0);
 702             dy30             = _mm_sub_pd(iy3,jy0);
 703             dz30             = _mm_sub_pd(iz3,jz0);
 704
 705             /* Calculate squared distance and things based on it */
 706             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 707             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 708             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 709             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 710
 711             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 712             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 713             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 714
 715             rinvsq00         = gmx_mm_inv_pd(rsq00);
 716             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 717             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 718             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 719
 720             /* Load parameters for j particles */
 721             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 722             vdwjidx0A        = 2*vdwtype[jnrA+0];
 723             vdwjidx0B        = 2*vdwtype[jnrB+0];
 724
 725             fjx0             = _mm_setzero_pd();
 726             fjy0             = _mm_setzero_pd();
 727             fjz0             = _mm_setzero_pd();
 728
 729             /**************************
 730              * CALCULATE INTERACTIONS *
 731              **************************/
 732
 733             /* Compute parameters for interactions between i and j atoms */
 734             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 735                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 736
 737             /* LENNARD-JONES DISPERSION/REPULSION */
 738
 739             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 740             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 741
 742             fscal            = fvdw;
 743
 744             /* Calculate temporary vectorial force */
 745             tx               = _mm_mul_pd(fscal,dx00);
 746             ty               = _mm_mul_pd(fscal,dy00);
 747             tz               = _mm_mul_pd(fscal,dz00);
 748
 749             /* Update vectorial force */
 750             fix0             = _mm_add_pd(fix0,tx);
 751             fiy0             = _mm_add_pd(fiy0,ty);
 752             fiz0             = _mm_add_pd(fiz0,tz);
 753
 754             fjx0             = _mm_add_pd(fjx0,tx);
 755             fjy0             = _mm_add_pd(fjy0,ty);
 756             fjz0             = _mm_add_pd(fjz0,tz);
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             /* Compute parameters for interactions between i and j atoms */
 763             qq10             = _mm_mul_pd(iq1,jq0);
 764
 765             /* REACTION-FIELD ELECTROSTATICS */
 766             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 767
 768             fscal            = felec;
 769
 770             /* Calculate temporary vectorial force */
 771             tx               = _mm_mul_pd(fscal,dx10);
 772             ty               = _mm_mul_pd(fscal,dy10);
 773             tz               = _mm_mul_pd(fscal,dz10);
 774
 775             /* Update vectorial force */
 776             fix1             = _mm_add_pd(fix1,tx);
 777             fiy1             = _mm_add_pd(fiy1,ty);
 778             fiz1             = _mm_add_pd(fiz1,tz);
 779
 780             fjx0             = _mm_add_pd(fjx0,tx);
 781             fjy0             = _mm_add_pd(fjy0,ty);
 782             fjz0             = _mm_add_pd(fjz0,tz);
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             /* Compute parameters for interactions between i and j atoms */
 789             qq20             = _mm_mul_pd(iq2,jq0);
 790
 791             /* REACTION-FIELD ELECTROSTATICS */
 792             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 793
 794             fscal            = felec;
 795
 796             /* Calculate temporary vectorial force */
 797             tx               = _mm_mul_pd(fscal,dx20);
 798             ty               = _mm_mul_pd(fscal,dy20);
 799             tz               = _mm_mul_pd(fscal,dz20);
 800
 801             /* Update vectorial force */
 802             fix2             = _mm_add_pd(fix2,tx);
 803             fiy2             = _mm_add_pd(fiy2,ty);
 804             fiz2             = _mm_add_pd(fiz2,tz);
 805
 806             fjx0             = _mm_add_pd(fjx0,tx);
 807             fjy0             = _mm_add_pd(fjy0,ty);
 808             fjz0             = _mm_add_pd(fjz0,tz);
 809
 810             /**************************
 811              * CALCULATE INTERACTIONS *
 812              **************************/
 813
 814             /* Compute parameters for interactions between i and j atoms */
 815             qq30             = _mm_mul_pd(iq3,jq0);
 816
 817             /* REACTION-FIELD ELECTROSTATICS */
 818             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 819
 820             fscal            = felec;
 821
 822             /* Calculate temporary vectorial force */
 823             tx               = _mm_mul_pd(fscal,dx30);
 824             ty               = _mm_mul_pd(fscal,dy30);
 825             tz               = _mm_mul_pd(fscal,dz30);
 826
 827             /* Update vectorial force */
 828             fix3             = _mm_add_pd(fix3,tx);
 829             fiy3             = _mm_add_pd(fiy3,ty);
 830             fiz3             = _mm_add_pd(fiz3,tz);
 831
 832             fjx0             = _mm_add_pd(fjx0,tx);
 833             fjy0             = _mm_add_pd(fjy0,ty);
 834             fjz0             = _mm_add_pd(fjz0,tz);
 835
 836             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 837
 838             /* Inner loop uses 111 flops */
 839         }
 840
 841         if(jidx<j_index_end)
 842         {
 843
 844             jnrA             = jjnr[jidx];
 845             j_coord_offsetA  = DIM*jnrA;
 846
 847             /* load j atom coordinates */
 848             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 849                                               &jx0,&jy0,&jz0);
 850
 851             /* Calculate displacement vector */
 852             dx00             = _mm_sub_pd(ix0,jx0);
 853             dy00             = _mm_sub_pd(iy0,jy0);
 854             dz00             = _mm_sub_pd(iz0,jz0);
 855             dx10             = _mm_sub_pd(ix1,jx0);
 856             dy10             = _mm_sub_pd(iy1,jy0);
 857             dz10             = _mm_sub_pd(iz1,jz0);
 858             dx20             = _mm_sub_pd(ix2,jx0);
 859             dy20             = _mm_sub_pd(iy2,jy0);
 860             dz20             = _mm_sub_pd(iz2,jz0);
 861             dx30             = _mm_sub_pd(ix3,jx0);
 862             dy30             = _mm_sub_pd(iy3,jy0);
 863             dz30             = _mm_sub_pd(iz3,jz0);
 864
 865             /* Calculate squared distance and things based on it */
 866             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 867             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 868             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 869             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 870
 871             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 872             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 873             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 874
 875             rinvsq00         = gmx_mm_inv_pd(rsq00);
 876             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 877             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 878             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 879
 880             /* Load parameters for j particles */
 881             jq0              = _mm_load_sd(charge+jnrA+0);
 882             vdwjidx0A        = 2*vdwtype[jnrA+0];
 883
 884             fjx0             = _mm_setzero_pd();
 885             fjy0             = _mm_setzero_pd();
 886             fjz0             = _mm_setzero_pd();
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             /* Compute parameters for interactions between i and j atoms */
 893             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 894
 895             /* LENNARD-JONES DISPERSION/REPULSION */
 896
 897             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 898             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 899
 900             fscal            = fvdw;
 901
 902             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 903
 904             /* Calculate temporary vectorial force */
 905             tx               = _mm_mul_pd(fscal,dx00);
 906             ty               = _mm_mul_pd(fscal,dy00);
 907             tz               = _mm_mul_pd(fscal,dz00);
 908
 909             /* Update vectorial force */
 910             fix0             = _mm_add_pd(fix0,tx);
 911             fiy0             = _mm_add_pd(fiy0,ty);
 912             fiz0             = _mm_add_pd(fiz0,tz);
 913
 914             fjx0             = _mm_add_pd(fjx0,tx);
 915             fjy0             = _mm_add_pd(fjy0,ty);
 916             fjz0             = _mm_add_pd(fjz0,tz);
 917
 918             /**************************
 919              * CALCULATE INTERACTIONS *
 920              **************************/
 921
 922             /* Compute parameters for interactions between i and j atoms */
 923             qq10             = _mm_mul_pd(iq1,jq0);
 924
 925             /* REACTION-FIELD ELECTROSTATICS */
 926             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 927
 928             fscal            = felec;
 929
 930             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 931
 932             /* Calculate temporary vectorial force */
 933             tx               = _mm_mul_pd(fscal,dx10);
 934             ty               = _mm_mul_pd(fscal,dy10);
 935             tz               = _mm_mul_pd(fscal,dz10);
 936
 937             /* Update vectorial force */
 938             fix1             = _mm_add_pd(fix1,tx);
 939             fiy1             = _mm_add_pd(fiy1,ty);
 940             fiz1             = _mm_add_pd(fiz1,tz);
 941
 942             fjx0             = _mm_add_pd(fjx0,tx);
 943             fjy0             = _mm_add_pd(fjy0,ty);
 944             fjz0             = _mm_add_pd(fjz0,tz);
 945
 946             /**************************
 947              * CALCULATE INTERACTIONS *
 948              **************************/
 949
 950             /* Compute parameters for interactions between i and j atoms */
 951             qq20             = _mm_mul_pd(iq2,jq0);
 952
 953             /* REACTION-FIELD ELECTROSTATICS */
 954             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 955
 956             fscal            = felec;
 957
 958             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 959
 960             /* Calculate temporary vectorial force */
 961             tx               = _mm_mul_pd(fscal,dx20);
 962             ty               = _mm_mul_pd(fscal,dy20);
 963             tz               = _mm_mul_pd(fscal,dz20);
 964
 965             /* Update vectorial force */
 966             fix2             = _mm_add_pd(fix2,tx);
 967             fiy2             = _mm_add_pd(fiy2,ty);
 968             fiz2             = _mm_add_pd(fiz2,tz);
 969
 970             fjx0             = _mm_add_pd(fjx0,tx);
 971             fjy0             = _mm_add_pd(fjy0,ty);
 972             fjz0             = _mm_add_pd(fjz0,tz);
 973
 974             /**************************
 975              * CALCULATE INTERACTIONS *
 976              **************************/
 977
 978             /* Compute parameters for interactions between i and j atoms */
 979             qq30             = _mm_mul_pd(iq3,jq0);
 980
 981             /* REACTION-FIELD ELECTROSTATICS */
 982             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 983
 984             fscal            = felec;
 985
 986             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 987
 988             /* Calculate temporary vectorial force */
 989             tx               = _mm_mul_pd(fscal,dx30);
 990             ty               = _mm_mul_pd(fscal,dy30);
 991             tz               = _mm_mul_pd(fscal,dz30);
 992
 993             /* Update vectorial force */
 994             fix3             = _mm_add_pd(fix3,tx);
 995             fiy3             = _mm_add_pd(fiy3,ty);
 996             fiz3             = _mm_add_pd(fiz3,tz);
 997
 998             fjx0             = _mm_add_pd(fjx0,tx);
 999             fjy0             = _mm_add_pd(fjy0,ty);
1000             fjz0             = _mm_add_pd(fjz0,tz);
1001
1002             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1003
1004             /* Inner loop uses 111 flops */
1005         }
1006
1007         /* End of innermost loop */
1008
1009         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1010                                               f+i_coord_offset,fshift+i_shift_offset);
1011
1012         /* Increment number of inner iterations */
1013         inneriter                  += j_index_end - j_index_start;
1014
1015         /* Outer loop uses 24 flops */
1016     }
1017
1018     /* Increment number of outer iterations */
1019     outeriter        += nri;
1020
1021     /* Update outer/inner flops */
1022
1023     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*111);
1024 }