src/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRFCut_VdwLJSh_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_ElecRFCut_VdwLJSh_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_ElecRFCut_VdwLJSh_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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 120     rcutoff_scalar   = fr->rcoulomb;
 121     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 122     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 123
 124     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 125     rvdw             = _mm_set1_pd(fr->rvdw);
 126
 127     /* Avoid stupid compiler warnings */
 128     jnrA = jnrB = 0;
 129     j_coord_offsetA = 0;
 130     j_coord_offsetB = 0;
 131
 132     outeriter        = 0;
 133     inneriter        = 0;
 134
 135     /* Start outer loop over neighborlists */
 136     for(iidx=0; iidx<nri; iidx++)
 137     {
 138         /* Load shift vector for this list */
 139         i_shift_offset   = DIM*shiftidx[iidx];
 140
 141         /* Load limits for loop over neighbors */
 142         j_index_start    = jindex[iidx];
 143         j_index_end      = jindex[iidx+1];
 144
 145         /* Get outer coordinate index */
 146         inr              = iinr[iidx];
 147         i_coord_offset   = DIM*inr;
 148
 149         /* Load i particle coords and add shift vector */
 150         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 151                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 152
 153         fix0             = _mm_setzero_pd();
 154         fiy0             = _mm_setzero_pd();
 155         fiz0             = _mm_setzero_pd();
 156         fix1             = _mm_setzero_pd();
 157         fiy1             = _mm_setzero_pd();
 158         fiz1             = _mm_setzero_pd();
 159         fix2             = _mm_setzero_pd();
 160         fiy2             = _mm_setzero_pd();
 161         fiz2             = _mm_setzero_pd();
 162         fix3             = _mm_setzero_pd();
 163         fiy3             = _mm_setzero_pd();
 164         fiz3             = _mm_setzero_pd();
 165
 166         /* Reset potential sums */
 167         velecsum         = _mm_setzero_pd();
 168         vvdwsum          = _mm_setzero_pd();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179
 180             /* load j atom coordinates */
 181             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 182                                               &jx0,&jy0,&jz0);
 183
 184             /* Calculate displacement vector */
 185             dx00             = _mm_sub_pd(ix0,jx0);
 186             dy00             = _mm_sub_pd(iy0,jy0);
 187             dz00             = _mm_sub_pd(iz0,jz0);
 188             dx10             = _mm_sub_pd(ix1,jx0);
 189             dy10             = _mm_sub_pd(iy1,jy0);
 190             dz10             = _mm_sub_pd(iz1,jz0);
 191             dx20             = _mm_sub_pd(ix2,jx0);
 192             dy20             = _mm_sub_pd(iy2,jy0);
 193             dz20             = _mm_sub_pd(iz2,jz0);
 194             dx30             = _mm_sub_pd(ix3,jx0);
 195             dy30             = _mm_sub_pd(iy3,jy0);
 196             dz30             = _mm_sub_pd(iz3,jz0);
 197
 198             /* Calculate squared distance and things based on it */
 199             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 200             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 201             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 202             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 203
 204             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 205             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 206             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 207
 208             rinvsq00         = gmx_mm_inv_pd(rsq00);
 209             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 210             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 211             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 212
 213             /* Load parameters for j particles */
 214             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 215             vdwjidx0A        = 2*vdwtype[jnrA+0];
 216             vdwjidx0B        = 2*vdwtype[jnrB+0];
 217
 218             fjx0             = _mm_setzero_pd();
 219             fjy0             = _mm_setzero_pd();
 220             fjz0             = _mm_setzero_pd();
 221
 222             /**************************
 223              * CALCULATE INTERACTIONS *
 224              **************************/
 225
 226             if (gmx_mm_any_lt(rsq00,rcutoff2))
 227             {
 228
 229             /* Compute parameters for interactions between i and j atoms */
 230             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 231                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 232
 233             /* LENNARD-JONES DISPERSION/REPULSION */
 234
 235             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 236             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 237             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 238             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 239                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 240             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 241
 242             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 243
 244             /* Update potential sum for this i atom from the interaction with this j atom. */
 245             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 246             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 247
 248             fscal            = fvdw;
 249
 250             fscal            = _mm_and_pd(fscal,cutoff_mask);
 251
 252             /* Calculate temporary vectorial force */
 253             tx               = _mm_mul_pd(fscal,dx00);
 254             ty               = _mm_mul_pd(fscal,dy00);
 255             tz               = _mm_mul_pd(fscal,dz00);
 256
 257             /* Update vectorial force */
 258             fix0             = _mm_add_pd(fix0,tx);
 259             fiy0             = _mm_add_pd(fiy0,ty);
 260             fiz0             = _mm_add_pd(fiz0,tz);
 261
 262             fjx0             = _mm_add_pd(fjx0,tx);
 263             fjy0             = _mm_add_pd(fjy0,ty);
 264             fjz0             = _mm_add_pd(fjz0,tz);
 265
 266             }
 267
 268             /**************************
 269              * CALCULATE INTERACTIONS *
 270              **************************/
 271
 272             if (gmx_mm_any_lt(rsq10,rcutoff2))
 273             {
 274
 275             /* Compute parameters for interactions between i and j atoms */
 276             qq10             = _mm_mul_pd(iq1,jq0);
 277
 278             /* REACTION-FIELD ELECTROSTATICS */
 279             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 280             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 281
 282             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 283
 284             /* Update potential sum for this i atom from the interaction with this j atom. */
 285             velec            = _mm_and_pd(velec,cutoff_mask);
 286             velecsum         = _mm_add_pd(velecsum,velec);
 287
 288             fscal            = felec;
 289
 290             fscal            = _mm_and_pd(fscal,cutoff_mask);
 291
 292             /* Calculate temporary vectorial force */
 293             tx               = _mm_mul_pd(fscal,dx10);
 294             ty               = _mm_mul_pd(fscal,dy10);
 295             tz               = _mm_mul_pd(fscal,dz10);
 296
 297             /* Update vectorial force */
 298             fix1             = _mm_add_pd(fix1,tx);
 299             fiy1             = _mm_add_pd(fiy1,ty);
 300             fiz1             = _mm_add_pd(fiz1,tz);
 301
 302             fjx0             = _mm_add_pd(fjx0,tx);
 303             fjy0             = _mm_add_pd(fjy0,ty);
 304             fjz0             = _mm_add_pd(fjz0,tz);
 305
 306             }
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             if (gmx_mm_any_lt(rsq20,rcutoff2))
 313             {
 314
 315             /* Compute parameters for interactions between i and j atoms */
 316             qq20             = _mm_mul_pd(iq2,jq0);
 317
 318             /* REACTION-FIELD ELECTROSTATICS */
 319             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 320             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 321
 322             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 323
 324             /* Update potential sum for this i atom from the interaction with this j atom. */
 325             velec            = _mm_and_pd(velec,cutoff_mask);
 326             velecsum         = _mm_add_pd(velecsum,velec);
 327
 328             fscal            = felec;
 329
 330             fscal            = _mm_and_pd(fscal,cutoff_mask);
 331
 332             /* Calculate temporary vectorial force */
 333             tx               = _mm_mul_pd(fscal,dx20);
 334             ty               = _mm_mul_pd(fscal,dy20);
 335             tz               = _mm_mul_pd(fscal,dz20);
 336
 337             /* Update vectorial force */
 338             fix2             = _mm_add_pd(fix2,tx);
 339             fiy2             = _mm_add_pd(fiy2,ty);
 340             fiz2             = _mm_add_pd(fiz2,tz);
 341
 342             fjx0             = _mm_add_pd(fjx0,tx);
 343             fjy0             = _mm_add_pd(fjy0,ty);
 344             fjz0             = _mm_add_pd(fjz0,tz);
 345
 346             }
 347
 348             /**************************
 349              * CALCULATE INTERACTIONS *
 350              **************************/
 351
 352             if (gmx_mm_any_lt(rsq30,rcutoff2))
 353             {
 354
 355             /* Compute parameters for interactions between i and j atoms */
 356             qq30             = _mm_mul_pd(iq3,jq0);
 357
 358             /* REACTION-FIELD ELECTROSTATICS */
 359             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_add_pd(rinv30,_mm_mul_pd(krf,rsq30)),crf));
 360             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 361
 362             cutoff_mask      = _mm_cmplt_pd(rsq30,rcutoff2);
 363
 364             /* Update potential sum for this i atom from the interaction with this j atom. */
 365             velec            = _mm_and_pd(velec,cutoff_mask);
 366             velecsum         = _mm_add_pd(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370             fscal            = _mm_and_pd(fscal,cutoff_mask);
 371
 372             /* Calculate temporary vectorial force */
 373             tx               = _mm_mul_pd(fscal,dx30);
 374             ty               = _mm_mul_pd(fscal,dy30);
 375             tz               = _mm_mul_pd(fscal,dz30);
 376
 377             /* Update vectorial force */
 378             fix3             = _mm_add_pd(fix3,tx);
 379             fiy3             = _mm_add_pd(fiy3,ty);
 380             fiz3             = _mm_add_pd(fiz3,tz);
 381
 382             fjx0             = _mm_add_pd(fjx0,tx);
 383             fjy0             = _mm_add_pd(fjy0,ty);
 384             fjz0             = _mm_add_pd(fjz0,tz);
 385
 386             }
 387
 388             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 389
 390             /* Inner loop uses 152 flops */
 391         }
 392
 393         if(jidx<j_index_end)
 394         {
 395
 396             jnrA             = jjnr[jidx];
 397             j_coord_offsetA  = DIM*jnrA;
 398
 399             /* load j atom coordinates */
 400             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 401                                               &jx0,&jy0,&jz0);
 402
 403             /* Calculate displacement vector */
 404             dx00             = _mm_sub_pd(ix0,jx0);
 405             dy00             = _mm_sub_pd(iy0,jy0);
 406             dz00             = _mm_sub_pd(iz0,jz0);
 407             dx10             = _mm_sub_pd(ix1,jx0);
 408             dy10             = _mm_sub_pd(iy1,jy0);
 409             dz10             = _mm_sub_pd(iz1,jz0);
 410             dx20             = _mm_sub_pd(ix2,jx0);
 411             dy20             = _mm_sub_pd(iy2,jy0);
 412             dz20             = _mm_sub_pd(iz2,jz0);
 413             dx30             = _mm_sub_pd(ix3,jx0);
 414             dy30             = _mm_sub_pd(iy3,jy0);
 415             dz30             = _mm_sub_pd(iz3,jz0);
 416
 417             /* Calculate squared distance and things based on it */
 418             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 419             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 420             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 421             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 422
 423             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 424             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 425             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 426
 427             rinvsq00         = gmx_mm_inv_pd(rsq00);
 428             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 429             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 430             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 431
 432             /* Load parameters for j particles */
 433             jq0              = _mm_load_sd(charge+jnrA+0);
 434             vdwjidx0A        = 2*vdwtype[jnrA+0];
 435
 436             fjx0             = _mm_setzero_pd();
 437             fjy0             = _mm_setzero_pd();
 438             fjz0             = _mm_setzero_pd();
 439
 440             /**************************
 441              * CALCULATE INTERACTIONS *
 442              **************************/
 443
 444             if (gmx_mm_any_lt(rsq00,rcutoff2))
 445             {
 446
 447             /* Compute parameters for interactions between i and j atoms */
 448             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 449
 450             /* LENNARD-JONES DISPERSION/REPULSION */
 451
 452             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 453             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 454             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 455             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 456                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 457             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 458
 459             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 463             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 464             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 465
 466             fscal            = fvdw;
 467
 468             fscal            = _mm_and_pd(fscal,cutoff_mask);
 469
 470             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 471
 472             /* Calculate temporary vectorial force */
 473             tx               = _mm_mul_pd(fscal,dx00);
 474             ty               = _mm_mul_pd(fscal,dy00);
 475             tz               = _mm_mul_pd(fscal,dz00);
 476
 477             /* Update vectorial force */
 478             fix0             = _mm_add_pd(fix0,tx);
 479             fiy0             = _mm_add_pd(fiy0,ty);
 480             fiz0             = _mm_add_pd(fiz0,tz);
 481
 482             fjx0             = _mm_add_pd(fjx0,tx);
 483             fjy0             = _mm_add_pd(fjy0,ty);
 484             fjz0             = _mm_add_pd(fjz0,tz);
 485
 486             }
 487
 488             /**************************
 489              * CALCULATE INTERACTIONS *
 490              **************************/
 491
 492             if (gmx_mm_any_lt(rsq10,rcutoff2))
 493             {
 494
 495             /* Compute parameters for interactions between i and j atoms */
 496             qq10             = _mm_mul_pd(iq1,jq0);
 497
 498             /* REACTION-FIELD ELECTROSTATICS */
 499             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 500             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 501
 502             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 503
 504             /* Update potential sum for this i atom from the interaction with this j atom. */
 505             velec            = _mm_and_pd(velec,cutoff_mask);
 506             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 507             velecsum         = _mm_add_pd(velecsum,velec);
 508
 509             fscal            = felec;
 510
 511             fscal            = _mm_and_pd(fscal,cutoff_mask);
 512
 513             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 514
 515             /* Calculate temporary vectorial force */
 516             tx               = _mm_mul_pd(fscal,dx10);
 517             ty               = _mm_mul_pd(fscal,dy10);
 518             tz               = _mm_mul_pd(fscal,dz10);
 519
 520             /* Update vectorial force */
 521             fix1             = _mm_add_pd(fix1,tx);
 522             fiy1             = _mm_add_pd(fiy1,ty);
 523             fiz1             = _mm_add_pd(fiz1,tz);
 524
 525             fjx0             = _mm_add_pd(fjx0,tx);
 526             fjy0             = _mm_add_pd(fjy0,ty);
 527             fjz0             = _mm_add_pd(fjz0,tz);
 528
 529             }
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             if (gmx_mm_any_lt(rsq20,rcutoff2))
 536             {
 537
 538             /* Compute parameters for interactions between i and j atoms */
 539             qq20             = _mm_mul_pd(iq2,jq0);
 540
 541             /* REACTION-FIELD ELECTROSTATICS */
 542             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 543             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 544
 545             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 546
 547             /* Update potential sum for this i atom from the interaction with this j atom. */
 548             velec            = _mm_and_pd(velec,cutoff_mask);
 549             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 550             velecsum         = _mm_add_pd(velecsum,velec);
 551
 552             fscal            = felec;
 553
 554             fscal            = _mm_and_pd(fscal,cutoff_mask);
 555
 556             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 557
 558             /* Calculate temporary vectorial force */
 559             tx               = _mm_mul_pd(fscal,dx20);
 560             ty               = _mm_mul_pd(fscal,dy20);
 561             tz               = _mm_mul_pd(fscal,dz20);
 562
 563             /* Update vectorial force */
 564             fix2             = _mm_add_pd(fix2,tx);
 565             fiy2             = _mm_add_pd(fiy2,ty);
 566             fiz2             = _mm_add_pd(fiz2,tz);
 567
 568             fjx0             = _mm_add_pd(fjx0,tx);
 569             fjy0             = _mm_add_pd(fjy0,ty);
 570             fjz0             = _mm_add_pd(fjz0,tz);
 571
 572             }
 573
 574             /**************************
 575              * CALCULATE INTERACTIONS *
 576              **************************/
 577
 578             if (gmx_mm_any_lt(rsq30,rcutoff2))
 579             {
 580
 581             /* Compute parameters for interactions between i and j atoms */
 582             qq30             = _mm_mul_pd(iq3,jq0);
 583
 584             /* REACTION-FIELD ELECTROSTATICS */
 585             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_add_pd(rinv30,_mm_mul_pd(krf,rsq30)),crf));
 586             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 587
 588             cutoff_mask      = _mm_cmplt_pd(rsq30,rcutoff2);
 589
 590             /* Update potential sum for this i atom from the interaction with this j atom. */
 591             velec            = _mm_and_pd(velec,cutoff_mask);
 592             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 593             velecsum         = _mm_add_pd(velecsum,velec);
 594
 595             fscal            = felec;
 596
 597             fscal            = _mm_and_pd(fscal,cutoff_mask);
 598
 599             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 600
 601             /* Calculate temporary vectorial force */
 602             tx               = _mm_mul_pd(fscal,dx30);
 603             ty               = _mm_mul_pd(fscal,dy30);
 604             tz               = _mm_mul_pd(fscal,dz30);
 605
 606             /* Update vectorial force */
 607             fix3             = _mm_add_pd(fix3,tx);
 608             fiy3             = _mm_add_pd(fiy3,ty);
 609             fiz3             = _mm_add_pd(fiz3,tz);
 610
 611             fjx0             = _mm_add_pd(fjx0,tx);
 612             fjy0             = _mm_add_pd(fjy0,ty);
 613             fjz0             = _mm_add_pd(fjz0,tz);
 614
 615             }
 616
 617             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 618
 619             /* Inner loop uses 152 flops */
 620         }
 621
 622         /* End of innermost loop */
 623
 624         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 625                                               f+i_coord_offset,fshift+i_shift_offset);
 626
 627         ggid                        = gid[iidx];
 628         /* Update potential energies */
 629         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 630         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 631
 632         /* Increment number of inner iterations */
 633         inneriter                  += j_index_end - j_index_start;
 634
 635         /* Outer loop uses 26 flops */
 636     }
 637
 638     /* Increment number of outer iterations */
 639     outeriter        += nri;
 640
 641     /* Update outer/inner flops */
 642
 643     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*152);
 644 }
 645 /*
 646  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sse4_1_double
 647  * Electrostatics interaction: ReactionField
 648  * VdW interaction:            LennardJones
 649  * Geometry:                   Water4-Particle
 650  * Calculate force/pot:        Force
 651  */
 652 void
 653 nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sse4_1_double
 654                     (t_nblist * gmx_restrict                nlist,
 655                      rvec * gmx_restrict                    xx,
 656                      rvec * gmx_restrict                    ff,
 657                      t_forcerec * gmx_restrict              fr,
 658                      t_mdatoms * gmx_restrict               mdatoms,
 659                      nb_kernel_data_t * gmx_restrict        kernel_data,
 660                      t_nrnb * gmx_restrict                  nrnb)
 661 {
 662     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 663      * just 0 for non-waters.
 664      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 665      * jnr indices corresponding to data put in the four positions in the SIMD register.
 666      */
 667     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 668     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 669     int              jnrA,jnrB;
 670     int              j_coord_offsetA,j_coord_offsetB;
 671     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 672     real             rcutoff_scalar;
 673     real             *shiftvec,*fshift,*x,*f;
 674     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 675     int              vdwioffset0;
 676     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 677     int              vdwioffset1;
 678     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 679     int              vdwioffset2;
 680     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 681     int              vdwioffset3;
 682     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 683     int              vdwjidx0A,vdwjidx0B;
 684     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 685     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 686     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 687     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 688     __m128d          dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 689     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 690     real             *charge;
 691     int              nvdwtype;
 692     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 693     int              *vdwtype;
 694     real             *vdwparam;
 695     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 696     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 697     __m128d          dummy_mask,cutoff_mask;
 698     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 699     __m128d          one     = _mm_set1_pd(1.0);
 700     __m128d          two     = _mm_set1_pd(2.0);
 701     x                = xx[0];
 702     f                = ff[0];
 703
 704     nri              = nlist->nri;
 705     iinr             = nlist->iinr;
 706     jindex           = nlist->jindex;
 707     jjnr             = nlist->jjnr;
 708     shiftidx         = nlist->shift;
 709     gid              = nlist->gid;
 710     shiftvec         = fr->shift_vec[0];
 711     fshift           = fr->fshift[0];
 712     facel            = _mm_set1_pd(fr->epsfac);
 713     charge           = mdatoms->chargeA;
 714     krf              = _mm_set1_pd(fr->ic->k_rf);
 715     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 716     crf              = _mm_set1_pd(fr->ic->c_rf);
 717     nvdwtype         = fr->ntype;
 718     vdwparam         = fr->nbfp;
 719     vdwtype          = mdatoms->typeA;
 720
 721     /* Setup water-specific parameters */
 722     inr              = nlist->iinr[0];
 723     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 724     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 725     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 726     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 727
 728     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 729     rcutoff_scalar   = fr->rcoulomb;
 730     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 731     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 732
 733     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 734     rvdw             = _mm_set1_pd(fr->rvdw);
 735
 736     /* Avoid stupid compiler warnings */
 737     jnrA = jnrB = 0;
 738     j_coord_offsetA = 0;
 739     j_coord_offsetB = 0;
 740
 741     outeriter        = 0;
 742     inneriter        = 0;
 743
 744     /* Start outer loop over neighborlists */
 745     for(iidx=0; iidx<nri; iidx++)
 746     {
 747         /* Load shift vector for this list */
 748         i_shift_offset   = DIM*shiftidx[iidx];
 749
 750         /* Load limits for loop over neighbors */
 751         j_index_start    = jindex[iidx];
 752         j_index_end      = jindex[iidx+1];
 753
 754         /* Get outer coordinate index */
 755         inr              = iinr[iidx];
 756         i_coord_offset   = DIM*inr;
 757
 758         /* Load i particle coords and add shift vector */
 759         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 760                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 761
 762         fix0             = _mm_setzero_pd();
 763         fiy0             = _mm_setzero_pd();
 764         fiz0             = _mm_setzero_pd();
 765         fix1             = _mm_setzero_pd();
 766         fiy1             = _mm_setzero_pd();
 767         fiz1             = _mm_setzero_pd();
 768         fix2             = _mm_setzero_pd();
 769         fiy2             = _mm_setzero_pd();
 770         fiz2             = _mm_setzero_pd();
 771         fix3             = _mm_setzero_pd();
 772         fiy3             = _mm_setzero_pd();
 773         fiz3             = _mm_setzero_pd();
 774
 775         /* Start inner kernel loop */
 776         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 777         {
 778
 779             /* Get j neighbor index, and coordinate index */
 780             jnrA             = jjnr[jidx];
 781             jnrB             = jjnr[jidx+1];
 782             j_coord_offsetA  = DIM*jnrA;
 783             j_coord_offsetB  = DIM*jnrB;
 784
 785             /* load j atom coordinates */
 786             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 787                                               &jx0,&jy0,&jz0);
 788
 789             /* Calculate displacement vector */
 790             dx00             = _mm_sub_pd(ix0,jx0);
 791             dy00             = _mm_sub_pd(iy0,jy0);
 792             dz00             = _mm_sub_pd(iz0,jz0);
 793             dx10             = _mm_sub_pd(ix1,jx0);
 794             dy10             = _mm_sub_pd(iy1,jy0);
 795             dz10             = _mm_sub_pd(iz1,jz0);
 796             dx20             = _mm_sub_pd(ix2,jx0);
 797             dy20             = _mm_sub_pd(iy2,jy0);
 798             dz20             = _mm_sub_pd(iz2,jz0);
 799             dx30             = _mm_sub_pd(ix3,jx0);
 800             dy30             = _mm_sub_pd(iy3,jy0);
 801             dz30             = _mm_sub_pd(iz3,jz0);
 802
 803             /* Calculate squared distance and things based on it */
 804             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 805             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 806             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 807             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 808
 809             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 810             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 811             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 812
 813             rinvsq00         = gmx_mm_inv_pd(rsq00);
 814             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 815             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 816             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 817
 818             /* Load parameters for j particles */
 819             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 820             vdwjidx0A        = 2*vdwtype[jnrA+0];
 821             vdwjidx0B        = 2*vdwtype[jnrB+0];
 822
 823             fjx0             = _mm_setzero_pd();
 824             fjy0             = _mm_setzero_pd();
 825             fjz0             = _mm_setzero_pd();
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             if (gmx_mm_any_lt(rsq00,rcutoff2))
 832             {
 833
 834             /* Compute parameters for interactions between i and j atoms */
 835             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 836                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 837
 838             /* LENNARD-JONES DISPERSION/REPULSION */
 839
 840             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 841             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 842
 843             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 844
 845             fscal            = fvdw;
 846
 847             fscal            = _mm_and_pd(fscal,cutoff_mask);
 848
 849             /* Calculate temporary vectorial force */
 850             tx               = _mm_mul_pd(fscal,dx00);
 851             ty               = _mm_mul_pd(fscal,dy00);
 852             tz               = _mm_mul_pd(fscal,dz00);
 853
 854             /* Update vectorial force */
 855             fix0             = _mm_add_pd(fix0,tx);
 856             fiy0             = _mm_add_pd(fiy0,ty);
 857             fiz0             = _mm_add_pd(fiz0,tz);
 858
 859             fjx0             = _mm_add_pd(fjx0,tx);
 860             fjy0             = _mm_add_pd(fjy0,ty);
 861             fjz0             = _mm_add_pd(fjz0,tz);
 862
 863             }
 864
 865             /**************************
 866              * CALCULATE INTERACTIONS *
 867              **************************/
 868
 869             if (gmx_mm_any_lt(rsq10,rcutoff2))
 870             {
 871
 872             /* Compute parameters for interactions between i and j atoms */
 873             qq10             = _mm_mul_pd(iq1,jq0);
 874
 875             /* REACTION-FIELD ELECTROSTATICS */
 876             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 877
 878             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 879
 880             fscal            = felec;
 881
 882             fscal            = _mm_and_pd(fscal,cutoff_mask);
 883
 884             /* Calculate temporary vectorial force */
 885             tx               = _mm_mul_pd(fscal,dx10);
 886             ty               = _mm_mul_pd(fscal,dy10);
 887             tz               = _mm_mul_pd(fscal,dz10);
 888
 889             /* Update vectorial force */
 890             fix1             = _mm_add_pd(fix1,tx);
 891             fiy1             = _mm_add_pd(fiy1,ty);
 892             fiz1             = _mm_add_pd(fiz1,tz);
 893
 894             fjx0             = _mm_add_pd(fjx0,tx);
 895             fjy0             = _mm_add_pd(fjy0,ty);
 896             fjz0             = _mm_add_pd(fjz0,tz);
 897
 898             }
 899
 900             /**************************
 901              * CALCULATE INTERACTIONS *
 902              **************************/
 903
 904             if (gmx_mm_any_lt(rsq20,rcutoff2))
 905             {
 906
 907             /* Compute parameters for interactions between i and j atoms */
 908             qq20             = _mm_mul_pd(iq2,jq0);
 909
 910             /* REACTION-FIELD ELECTROSTATICS */
 911             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 912
 913             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 914
 915             fscal            = felec;
 916
 917             fscal            = _mm_and_pd(fscal,cutoff_mask);
 918
 919             /* Calculate temporary vectorial force */
 920             tx               = _mm_mul_pd(fscal,dx20);
 921             ty               = _mm_mul_pd(fscal,dy20);
 922             tz               = _mm_mul_pd(fscal,dz20);
 923
 924             /* Update vectorial force */
 925             fix2             = _mm_add_pd(fix2,tx);
 926             fiy2             = _mm_add_pd(fiy2,ty);
 927             fiz2             = _mm_add_pd(fiz2,tz);
 928
 929             fjx0             = _mm_add_pd(fjx0,tx);
 930             fjy0             = _mm_add_pd(fjy0,ty);
 931             fjz0             = _mm_add_pd(fjz0,tz);
 932
 933             }
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             if (gmx_mm_any_lt(rsq30,rcutoff2))
 940             {
 941
 942             /* Compute parameters for interactions between i and j atoms */
 943             qq30             = _mm_mul_pd(iq3,jq0);
 944
 945             /* REACTION-FIELD ELECTROSTATICS */
 946             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
 947
 948             cutoff_mask      = _mm_cmplt_pd(rsq30,rcutoff2);
 949
 950             fscal            = felec;
 951
 952             fscal            = _mm_and_pd(fscal,cutoff_mask);
 953
 954             /* Calculate temporary vectorial force */
 955             tx               = _mm_mul_pd(fscal,dx30);
 956             ty               = _mm_mul_pd(fscal,dy30);
 957             tz               = _mm_mul_pd(fscal,dz30);
 958
 959             /* Update vectorial force */
 960             fix3             = _mm_add_pd(fix3,tx);
 961             fiy3             = _mm_add_pd(fiy3,ty);
 962             fiz3             = _mm_add_pd(fiz3,tz);
 963
 964             fjx0             = _mm_add_pd(fjx0,tx);
 965             fjy0             = _mm_add_pd(fjy0,ty);
 966             fjz0             = _mm_add_pd(fjz0,tz);
 967
 968             }
 969
 970             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 971
 972             /* Inner loop uses 123 flops */
 973         }
 974
 975         if(jidx<j_index_end)
 976         {
 977
 978             jnrA             = jjnr[jidx];
 979             j_coord_offsetA  = DIM*jnrA;
 980
 981             /* load j atom coordinates */
 982             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 983                                               &jx0,&jy0,&jz0);
 984
 985             /* Calculate displacement vector */
 986             dx00             = _mm_sub_pd(ix0,jx0);
 987             dy00             = _mm_sub_pd(iy0,jy0);
 988             dz00             = _mm_sub_pd(iz0,jz0);
 989             dx10             = _mm_sub_pd(ix1,jx0);
 990             dy10             = _mm_sub_pd(iy1,jy0);
 991             dz10             = _mm_sub_pd(iz1,jz0);
 992             dx20             = _mm_sub_pd(ix2,jx0);
 993             dy20             = _mm_sub_pd(iy2,jy0);
 994             dz20             = _mm_sub_pd(iz2,jz0);
 995             dx30             = _mm_sub_pd(ix3,jx0);
 996             dy30             = _mm_sub_pd(iy3,jy0);
 997             dz30             = _mm_sub_pd(iz3,jz0);
 998
 999             /* Calculate squared distance and things based on it */
1000             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1001             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1002             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1003             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
1004
1005             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1006             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1007             rinv30           = gmx_mm_invsqrt_pd(rsq30);
1008
1009             rinvsq00         = gmx_mm_inv_pd(rsq00);
1010             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1011             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1012             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
1013
1014             /* Load parameters for j particles */
1015             jq0              = _mm_load_sd(charge+jnrA+0);
1016             vdwjidx0A        = 2*vdwtype[jnrA+0];
1017
1018             fjx0             = _mm_setzero_pd();
1019             fjy0             = _mm_setzero_pd();
1020             fjz0             = _mm_setzero_pd();
1021
1022             /**************************
1023              * CALCULATE INTERACTIONS *
1024              **************************/
1025
1026             if (gmx_mm_any_lt(rsq00,rcutoff2))
1027             {
1028
1029             /* Compute parameters for interactions between i and j atoms */
1030             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
1031
1032             /* LENNARD-JONES DISPERSION/REPULSION */
1033
1034             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1035             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1036
1037             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1038
1039             fscal            = fvdw;
1040
1041             fscal            = _mm_and_pd(fscal,cutoff_mask);
1042
1043             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1044
1045             /* Calculate temporary vectorial force */
1046             tx               = _mm_mul_pd(fscal,dx00);
1047             ty               = _mm_mul_pd(fscal,dy00);
1048             tz               = _mm_mul_pd(fscal,dz00);
1049
1050             /* Update vectorial force */
1051             fix0             = _mm_add_pd(fix0,tx);
1052             fiy0             = _mm_add_pd(fiy0,ty);
1053             fiz0             = _mm_add_pd(fiz0,tz);
1054
1055             fjx0             = _mm_add_pd(fjx0,tx);
1056             fjy0             = _mm_add_pd(fjy0,ty);
1057             fjz0             = _mm_add_pd(fjz0,tz);
1058
1059             }
1060
1061             /**************************
1062              * CALCULATE INTERACTIONS *
1063              **************************/
1064
1065             if (gmx_mm_any_lt(rsq10,rcutoff2))
1066             {
1067
1068             /* Compute parameters for interactions between i and j atoms */
1069             qq10             = _mm_mul_pd(iq1,jq0);
1070
1071             /* REACTION-FIELD ELECTROSTATICS */
1072             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1073
1074             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1075
1076             fscal            = felec;
1077
1078             fscal            = _mm_and_pd(fscal,cutoff_mask);
1079
1080             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1081
1082             /* Calculate temporary vectorial force */
1083             tx               = _mm_mul_pd(fscal,dx10);
1084             ty               = _mm_mul_pd(fscal,dy10);
1085             tz               = _mm_mul_pd(fscal,dz10);
1086
1087             /* Update vectorial force */
1088             fix1             = _mm_add_pd(fix1,tx);
1089             fiy1             = _mm_add_pd(fiy1,ty);
1090             fiz1             = _mm_add_pd(fiz1,tz);
1091
1092             fjx0             = _mm_add_pd(fjx0,tx);
1093             fjy0             = _mm_add_pd(fjy0,ty);
1094             fjz0             = _mm_add_pd(fjz0,tz);
1095
1096             }
1097
1098             /**************************
1099              * CALCULATE INTERACTIONS *
1100              **************************/
1101
1102             if (gmx_mm_any_lt(rsq20,rcutoff2))
1103             {
1104
1105             /* Compute parameters for interactions between i and j atoms */
1106             qq20             = _mm_mul_pd(iq2,jq0);
1107
1108             /* REACTION-FIELD ELECTROSTATICS */
1109             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1110
1111             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1112
1113             fscal            = felec;
1114
1115             fscal            = _mm_and_pd(fscal,cutoff_mask);
1116
1117             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1118
1119             /* Calculate temporary vectorial force */
1120             tx               = _mm_mul_pd(fscal,dx20);
1121             ty               = _mm_mul_pd(fscal,dy20);
1122             tz               = _mm_mul_pd(fscal,dz20);
1123
1124             /* Update vectorial force */
1125             fix2             = _mm_add_pd(fix2,tx);
1126             fiy2             = _mm_add_pd(fiy2,ty);
1127             fiz2             = _mm_add_pd(fiz2,tz);
1128
1129             fjx0             = _mm_add_pd(fjx0,tx);
1130             fjy0             = _mm_add_pd(fjy0,ty);
1131             fjz0             = _mm_add_pd(fjz0,tz);
1132
1133             }
1134
1135             /**************************
1136              * CALCULATE INTERACTIONS *
1137              **************************/
1138
1139             if (gmx_mm_any_lt(rsq30,rcutoff2))
1140             {
1141
1142             /* Compute parameters for interactions between i and j atoms */
1143             qq30             = _mm_mul_pd(iq3,jq0);
1144
1145             /* REACTION-FIELD ELECTROSTATICS */
1146             felec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_mul_pd(rinv30,rinvsq30),krf2));
1147
1148             cutoff_mask      = _mm_cmplt_pd(rsq30,rcutoff2);
1149
1150             fscal            = felec;
1151
1152             fscal            = _mm_and_pd(fscal,cutoff_mask);
1153
1154             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1155
1156             /* Calculate temporary vectorial force */
1157             tx               = _mm_mul_pd(fscal,dx30);
1158             ty               = _mm_mul_pd(fscal,dy30);
1159             tz               = _mm_mul_pd(fscal,dz30);
1160
1161             /* Update vectorial force */
1162             fix3             = _mm_add_pd(fix3,tx);
1163             fiy3             = _mm_add_pd(fiy3,ty);
1164             fiz3             = _mm_add_pd(fiz3,tz);
1165
1166             fjx0             = _mm_add_pd(fjx0,tx);
1167             fjy0             = _mm_add_pd(fjy0,ty);
1168             fjz0             = _mm_add_pd(fjz0,tz);
1169
1170             }
1171
1172             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1173
1174             /* Inner loop uses 123 flops */
1175         }
1176
1177         /* End of innermost loop */
1178
1179         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1180                                               f+i_coord_offset,fshift+i_shift_offset);
1181
1182         /* Increment number of inner iterations */
1183         inneriter                  += j_index_end - j_index_start;
1184
1185         /* Outer loop uses 24 flops */
1186     }
1187
1188     /* Increment number of outer iterations */
1189     outeriter        += nri;
1190
1191     /* Update outer/inner flops */
1192
1193     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*123);
1194 }