src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRFCut_VdwLJSh_GeomW3P1_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sparc64_hpc_ace_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3P1_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: ReactionField
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water3-Particle
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecRFCut_VdwLJSh_GeomW3P1_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  91     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  92     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
  99     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 100     _fjsp_v2r8       itab_tmp;
 101     _fjsp_v2r8       dummy_mask,cutoff_mask;
 102     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 103     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 104     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 105
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 120     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 121     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 129     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 130     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 134     rcutoff_scalar   = fr->rcoulomb;
 135     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 136     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 137
 138     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 139     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 140
 141     /* Avoid stupid compiler warnings */
 142     jnrA = jnrB = 0;
 143     j_coord_offsetA = 0;
 144     j_coord_offsetB = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 165                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 166
 167         fix0             = _fjsp_setzero_v2r8();
 168         fiy0             = _fjsp_setzero_v2r8();
 169         fiz0             = _fjsp_setzero_v2r8();
 170         fix1             = _fjsp_setzero_v2r8();
 171         fiy1             = _fjsp_setzero_v2r8();
 172         fiz1             = _fjsp_setzero_v2r8();
 173         fix2             = _fjsp_setzero_v2r8();
 174         fiy2             = _fjsp_setzero_v2r8();
 175         fiz2             = _fjsp_setzero_v2r8();
 176
 177         /* Reset potential sums */
 178         velecsum         = _fjsp_setzero_v2r8();
 179         vvdwsum          = _fjsp_setzero_v2r8();
 180
 181         /* Start inner kernel loop */
 182         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 183         {
 184
 185             /* Get j neighbor index, and coordinate index */
 186             jnrA             = jjnr[jidx];
 187             jnrB             = jjnr[jidx+1];
 188             j_coord_offsetA  = DIM*jnrA;
 189             j_coord_offsetB  = DIM*jnrB;
 190
 191             /* load j atom coordinates */
 192             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 193                                               &jx0,&jy0,&jz0);
 194
 195             /* Calculate displacement vector */
 196             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 197             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 198             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 199             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 200             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 201             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 202             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 203             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 204             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 205
 206             /* Calculate squared distance and things based on it */
 207             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 208             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 209             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 210
 211             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 212             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 213             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 214
 215             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 216             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 217             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 218
 219             /* Load parameters for j particles */
 220             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 221             vdwjidx0A        = 2*vdwtype[jnrA+0];
 222             vdwjidx0B        = 2*vdwtype[jnrB+0];
 223
 224             fjx0             = _fjsp_setzero_v2r8();
 225             fjy0             = _fjsp_setzero_v2r8();
 226             fjz0             = _fjsp_setzero_v2r8();
 227
 228             /**************************
 229              * CALCULATE INTERACTIONS *
 230              **************************/
 231
 232             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 233             {
 234
 235             /* Compute parameters for interactions between i and j atoms */
 236             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 237             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 238                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 239
 240             /* REACTION-FIELD ELECTROSTATICS */
 241             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 242             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 243
 244             /* LENNARD-JONES DISPERSION/REPULSION */
 245
 246             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 247             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 248             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 249             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 250                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 251             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 252
 253             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 254
 255             /* Update potential sum for this i atom from the interaction with this j atom. */
 256             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 257             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 258             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 259             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 260
 261             fscal            = _fjsp_add_v2r8(felec,fvdw);
 262
 263             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 264
 265             /* Update vectorial force */
 266             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 267             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 268             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 269
 270             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 271             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 272             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 273
 274             }
 275
 276             /**************************
 277              * CALCULATE INTERACTIONS *
 278              **************************/
 279
 280             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 281             {
 282
 283             /* Compute parameters for interactions between i and j atoms */
 284             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 285
 286             /* REACTION-FIELD ELECTROSTATICS */
 287             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 288             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 289
 290             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 291
 292             /* Update potential sum for this i atom from the interaction with this j atom. */
 293             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 294             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 295
 296             fscal            = felec;
 297
 298             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 299
 300             /* Update vectorial force */
 301             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 302             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 303             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 304
 305             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 306             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 307             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 308
 309             }
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 316             {
 317
 318             /* Compute parameters for interactions between i and j atoms */
 319             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 320
 321             /* REACTION-FIELD ELECTROSTATICS */
 322             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 323             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 324
 325             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 326
 327             /* Update potential sum for this i atom from the interaction with this j atom. */
 328             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 329             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 330
 331             fscal            = felec;
 332
 333             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 334
 335             /* Update vectorial force */
 336             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 337             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 338             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 339
 340             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 341             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 342             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 343
 344             }
 345
 346             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 347
 348             /* Inner loop uses 138 flops */
 349         }
 350
 351         if(jidx<j_index_end)
 352         {
 353
 354             jnrA             = jjnr[jidx];
 355             j_coord_offsetA  = DIM*jnrA;
 356
 357             /* load j atom coordinates */
 358             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 359                                               &jx0,&jy0,&jz0);
 360
 361             /* Calculate displacement vector */
 362             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 363             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 364             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 365             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 366             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 367             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 368             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 369             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 370             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 371
 372             /* Calculate squared distance and things based on it */
 373             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 374             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 375             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 376
 377             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 378             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 379             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 380
 381             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 382             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 383             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 384
 385             /* Load parameters for j particles */
 386             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 387             vdwjidx0A        = 2*vdwtype[jnrA+0];
 388
 389             fjx0             = _fjsp_setzero_v2r8();
 390             fjy0             = _fjsp_setzero_v2r8();
 391             fjz0             = _fjsp_setzero_v2r8();
 392
 393             /**************************
 394              * CALCULATE INTERACTIONS *
 395              **************************/
 396
 397             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 398             {
 399
 400             /* Compute parameters for interactions between i and j atoms */
 401             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 402             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 403
 404             /* REACTION-FIELD ELECTROSTATICS */
 405             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 406             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 407
 408             /* LENNARD-JONES DISPERSION/REPULSION */
 409
 410             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 411             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 412             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 413             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 414                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 415             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 416
 417             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 421             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 422             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 423             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 424             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 425             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 426
 427             fscal            = _fjsp_add_v2r8(felec,fvdw);
 428
 429             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 430
 431             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 432
 433             /* Update vectorial force */
 434             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 435             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 436             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 437
 438             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 439             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 440             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 441
 442             }
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 449             {
 450
 451             /* Compute parameters for interactions between i and j atoms */
 452             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 453
 454             /* REACTION-FIELD ELECTROSTATICS */
 455             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 456             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 457
 458             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 462             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 463             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 468
 469             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 470
 471             /* Update vectorial force */
 472             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 473             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 474             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 475
 476             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 477             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 478             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 479
 480             }
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 487             {
 488
 489             /* Compute parameters for interactions between i and j atoms */
 490             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 491
 492             /* REACTION-FIELD ELECTROSTATICS */
 493             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 494             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 495
 496             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 497
 498             /* Update potential sum for this i atom from the interaction with this j atom. */
 499             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 500             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 501             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 502
 503             fscal            = felec;
 504
 505             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 506
 507             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 508
 509             /* Update vectorial force */
 510             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 511             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 512             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 513
 514             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 515             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 516             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 517
 518             }
 519
 520             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 521
 522             /* Inner loop uses 138 flops */
 523         }
 524
 525         /* End of innermost loop */
 526
 527         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 528                                               f+i_coord_offset,fshift+i_shift_offset);
 529
 530         ggid                        = gid[iidx];
 531         /* Update potential energies */
 532         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 533         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 534
 535         /* Increment number of inner iterations */
 536         inneriter                  += j_index_end - j_index_start;
 537
 538         /* Outer loop uses 20 flops */
 539     }
 540
 541     /* Increment number of outer iterations */
 542     outeriter        += nri;
 543
 544     /* Update outer/inner flops */
 545
 546     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*138);
 547 }
 548 /*
 549  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3P1_F_sparc64_hpc_ace_double
 550  * Electrostatics interaction: ReactionField
 551  * VdW interaction:            LennardJones
 552  * Geometry:                   Water3-Particle
 553  * Calculate force/pot:        Force
 554  */
 555 void
 556 nb_kernel_ElecRFCut_VdwLJSh_GeomW3P1_F_sparc64_hpc_ace_double
 557                     (t_nblist                    * gmx_restrict       nlist,
 558                      rvec                        * gmx_restrict          xx,
 559                      rvec                        * gmx_restrict          ff,
 560                      t_forcerec                  * gmx_restrict          fr,
 561                      t_mdatoms                   * gmx_restrict     mdatoms,
 562                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 563                      t_nrnb                      * gmx_restrict        nrnb)
 564 {
 565     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 566      * just 0 for non-waters.
 567      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 568      * jnr indices corresponding to data put in the four positions in the SIMD register.
 569      */
 570     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 571     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 572     int              jnrA,jnrB;
 573     int              j_coord_offsetA,j_coord_offsetB;
 574     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 575     real             rcutoff_scalar;
 576     real             *shiftvec,*fshift,*x,*f;
 577     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 578     int              vdwioffset0;
 579     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 580     int              vdwioffset1;
 581     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 582     int              vdwioffset2;
 583     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 584     int              vdwjidx0A,vdwjidx0B;
 585     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 586     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 587     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 588     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 589     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 590     real             *charge;
 591     int              nvdwtype;
 592     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 593     int              *vdwtype;
 594     real             *vdwparam;
 595     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 596     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 597     _fjsp_v2r8       itab_tmp;
 598     _fjsp_v2r8       dummy_mask,cutoff_mask;
 599     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 600     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 601     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 602
 603     x                = xx[0];
 604     f                = ff[0];
 605
 606     nri              = nlist->nri;
 607     iinr             = nlist->iinr;
 608     jindex           = nlist->jindex;
 609     jjnr             = nlist->jjnr;
 610     shiftidx         = nlist->shift;
 611     gid              = nlist->gid;
 612     shiftvec         = fr->shift_vec[0];
 613     fshift           = fr->fshift[0];
 614     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 615     charge           = mdatoms->chargeA;
 616     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 617     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 618     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 619     nvdwtype         = fr->ntype;
 620     vdwparam         = fr->nbfp;
 621     vdwtype          = mdatoms->typeA;
 622
 623     /* Setup water-specific parameters */
 624     inr              = nlist->iinr[0];
 625     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 626     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 627     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 628     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 629
 630     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 631     rcutoff_scalar   = fr->rcoulomb;
 632     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 633     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 634
 635     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 636     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 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_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 662                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 663
 664         fix0             = _fjsp_setzero_v2r8();
 665         fiy0             = _fjsp_setzero_v2r8();
 666         fiz0             = _fjsp_setzero_v2r8();
 667         fix1             = _fjsp_setzero_v2r8();
 668         fiy1             = _fjsp_setzero_v2r8();
 669         fiz1             = _fjsp_setzero_v2r8();
 670         fix2             = _fjsp_setzero_v2r8();
 671         fiy2             = _fjsp_setzero_v2r8();
 672         fiz2             = _fjsp_setzero_v2r8();
 673
 674         /* Start inner kernel loop */
 675         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 676         {
 677
 678             /* Get j neighbor index, and coordinate index */
 679             jnrA             = jjnr[jidx];
 680             jnrB             = jjnr[jidx+1];
 681             j_coord_offsetA  = DIM*jnrA;
 682             j_coord_offsetB  = DIM*jnrB;
 683
 684             /* load j atom coordinates */
 685             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 686                                               &jx0,&jy0,&jz0);
 687
 688             /* Calculate displacement vector */
 689             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 690             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 691             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 692             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 693             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 694             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 695             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 696             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 697             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 698
 699             /* Calculate squared distance and things based on it */
 700             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 701             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 702             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 703
 704             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 705             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 706             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 707
 708             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 709             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 710             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 711
 712             /* Load parameters for j particles */
 713             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 714             vdwjidx0A        = 2*vdwtype[jnrA+0];
 715             vdwjidx0B        = 2*vdwtype[jnrB+0];
 716
 717             fjx0             = _fjsp_setzero_v2r8();
 718             fjy0             = _fjsp_setzero_v2r8();
 719             fjz0             = _fjsp_setzero_v2r8();
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 726             {
 727
 728             /* Compute parameters for interactions between i and j atoms */
 729             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 730             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 731                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 732
 733             /* REACTION-FIELD ELECTROSTATICS */
 734             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 735
 736             /* LENNARD-JONES DISPERSION/REPULSION */
 737
 738             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 739             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 740
 741             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 742
 743             fscal            = _fjsp_add_v2r8(felec,fvdw);
 744
 745             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 746
 747             /* Update vectorial force */
 748             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 749             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 750             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 751
 752             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 753             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 754             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 755
 756             }
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 763             {
 764
 765             /* Compute parameters for interactions between i and j atoms */
 766             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 767
 768             /* REACTION-FIELD ELECTROSTATICS */
 769             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 770
 771             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 772
 773             fscal            = felec;
 774
 775             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 776
 777             /* Update vectorial force */
 778             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 779             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 780             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 781
 782             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 783             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 784             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 785
 786             }
 787
 788             /**************************
 789              * CALCULATE INTERACTIONS *
 790              **************************/
 791
 792             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 793             {
 794
 795             /* Compute parameters for interactions between i and j atoms */
 796             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 800
 801             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 802
 803             fscal            = felec;
 804
 805             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 806
 807             /* Update vectorial force */
 808             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 809             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 810             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 811
 812             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 813             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 814             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 815
 816             }
 817
 818             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 819
 820             /* Inner loop uses 109 flops */
 821         }
 822
 823         if(jidx<j_index_end)
 824         {
 825
 826             jnrA             = jjnr[jidx];
 827             j_coord_offsetA  = DIM*jnrA;
 828
 829             /* load j atom coordinates */
 830             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 831                                               &jx0,&jy0,&jz0);
 832
 833             /* Calculate displacement vector */
 834             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 835             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 836             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 837             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 838             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 839             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 840             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 841             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 842             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 843
 844             /* Calculate squared distance and things based on it */
 845             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 846             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 847             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 848
 849             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 850             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 851             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 852
 853             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 854             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 855             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 856
 857             /* Load parameters for j particles */
 858             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 859             vdwjidx0A        = 2*vdwtype[jnrA+0];
 860
 861             fjx0             = _fjsp_setzero_v2r8();
 862             fjy0             = _fjsp_setzero_v2r8();
 863             fjz0             = _fjsp_setzero_v2r8();
 864
 865             /**************************
 866              * CALCULATE INTERACTIONS *
 867              **************************/
 868
 869             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 870             {
 871
 872             /* Compute parameters for interactions between i and j atoms */
 873             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 874             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 875
 876             /* REACTION-FIELD ELECTROSTATICS */
 877             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 878
 879             /* LENNARD-JONES DISPERSION/REPULSION */
 880
 881             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 882             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 883
 884             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 885
 886             fscal            = _fjsp_add_v2r8(felec,fvdw);
 887
 888             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 889
 890             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 891
 892             /* Update vectorial force */
 893             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 894             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 895             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 896
 897             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 898             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 899             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 900
 901             }
 902
 903             /**************************
 904              * CALCULATE INTERACTIONS *
 905              **************************/
 906
 907             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 908             {
 909
 910             /* Compute parameters for interactions between i and j atoms */
 911             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 912
 913             /* REACTION-FIELD ELECTROSTATICS */
 914             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 915
 916             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 917
 918             fscal            = felec;
 919
 920             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 921
 922             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 923
 924             /* Update vectorial force */
 925             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 926             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 927             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 928
 929             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 930             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 931             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 932
 933             }
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 940             {
 941
 942             /* Compute parameters for interactions between i and j atoms */
 943             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 944
 945             /* REACTION-FIELD ELECTROSTATICS */
 946             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 947
 948             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 949
 950             fscal            = felec;
 951
 952             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 953
 954             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 955
 956             /* Update vectorial force */
 957             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 958             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 959             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 960
 961             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 962             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 963             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 964
 965             }
 966
 967             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 968
 969             /* Inner loop uses 109 flops */
 970         }
 971
 972         /* End of innermost loop */
 973
 974         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 975                                               f+i_coord_offset,fshift+i_shift_offset);
 976
 977         /* Increment number of inner iterations */
 978         inneriter                  += j_index_end - j_index_start;
 979
 980         /* Outer loop uses 18 flops */
 981     }
 982
 983     /* Increment number of outer iterations */
 984     outeriter        += nri;
 985
 986     /* Update outer/inner flops */
 987
 988     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*109);
 989 }