src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRF_VdwLJ_GeomW4P1_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS 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 "gromacs/legacyheaders/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4P1_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: ReactionField
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water4-Particle
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecRF_VdwLJ_GeomW4P1_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              vdwioffset3;
  88     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  89     int              vdwjidx0A,vdwjidx0B;
  90     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  93     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  94     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  95     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 102     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 103     _fjsp_v2r8       itab_tmp;
 104     _fjsp_v2r8       dummy_mask,cutoff_mask;
 105     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 106     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 107     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 108
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 123     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 124     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 132     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 133     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 134     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 135
 136     /* Avoid stupid compiler warnings */
 137     jnrA = jnrB = 0;
 138     j_coord_offsetA = 0;
 139     j_coord_offsetB = 0;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 160                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 161
 162         fix0             = _fjsp_setzero_v2r8();
 163         fiy0             = _fjsp_setzero_v2r8();
 164         fiz0             = _fjsp_setzero_v2r8();
 165         fix1             = _fjsp_setzero_v2r8();
 166         fiy1             = _fjsp_setzero_v2r8();
 167         fiz1             = _fjsp_setzero_v2r8();
 168         fix2             = _fjsp_setzero_v2r8();
 169         fiy2             = _fjsp_setzero_v2r8();
 170         fiz2             = _fjsp_setzero_v2r8();
 171         fix3             = _fjsp_setzero_v2r8();
 172         fiy3             = _fjsp_setzero_v2r8();
 173         fiz3             = _fjsp_setzero_v2r8();
 174
 175         /* Reset potential sums */
 176         velecsum         = _fjsp_setzero_v2r8();
 177         vvdwsum          = _fjsp_setzero_v2r8();
 178
 179         /* Start inner kernel loop */
 180         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 181         {
 182
 183             /* Get j neighbor index, and coordinate index */
 184             jnrA             = jjnr[jidx];
 185             jnrB             = jjnr[jidx+1];
 186             j_coord_offsetA  = DIM*jnrA;
 187             j_coord_offsetB  = DIM*jnrB;
 188
 189             /* load j atom coordinates */
 190             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                               &jx0,&jy0,&jz0);
 192
 193             /* Calculate displacement vector */
 194             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 195             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 196             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 197             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 198             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 199             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 200             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 201             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 202             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 203             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 204             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 205             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 206
 207             /* Calculate squared distance and things based on it */
 208             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 209             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 210             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 211             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 212
 213             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 214             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 215             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 216
 217             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 218             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 219             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 220             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 221
 222             /* Load parameters for j particles */
 223             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 224             vdwjidx0A        = 2*vdwtype[jnrA+0];
 225             vdwjidx0B        = 2*vdwtype[jnrB+0];
 226
 227             fjx0             = _fjsp_setzero_v2r8();
 228             fjy0             = _fjsp_setzero_v2r8();
 229             fjz0             = _fjsp_setzero_v2r8();
 230
 231             /**************************
 232              * CALCULATE INTERACTIONS *
 233              **************************/
 234
 235             /* Compute parameters for interactions between i and j atoms */
 236             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 237                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 238
 239             /* LENNARD-JONES DISPERSION/REPULSION */
 240
 241             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 242             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 243             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 244             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 245             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 246
 247             /* Update potential sum for this i atom from the interaction with this j atom. */
 248             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 249
 250             fscal            = fvdw;
 251
 252             /* Update vectorial force */
 253             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 254             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 255             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 256
 257             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 258             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 259             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 260
 261             /**************************
 262              * CALCULATE INTERACTIONS *
 263              **************************/
 264
 265             /* Compute parameters for interactions between i and j atoms */
 266             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 267
 268             /* REACTION-FIELD ELECTROSTATICS */
 269             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 270             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 271
 272             /* Update potential sum for this i atom from the interaction with this j atom. */
 273             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 274
 275             fscal            = felec;
 276
 277             /* Update vectorial force */
 278             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 279             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 280             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 281
 282             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 283             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 284             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 285
 286             /**************************
 287              * CALCULATE INTERACTIONS *
 288              **************************/
 289
 290             /* Compute parameters for interactions between i and j atoms */
 291             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 292
 293             /* REACTION-FIELD ELECTROSTATICS */
 294             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 295             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 299
 300             fscal            = felec;
 301
 302             /* Update vectorial force */
 303             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 304             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 305             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 306
 307             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 308             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 309             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             /* Compute parameters for interactions between i and j atoms */
 316             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 317
 318             /* REACTION-FIELD ELECTROSTATICS */
 319             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 320             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 321
 322             /* Update potential sum for this i atom from the interaction with this j atom. */
 323             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 324
 325             fscal            = felec;
 326
 327             /* Update vectorial force */
 328             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 329             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 330             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 331
 332             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 333             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 334             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 335
 336             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 337
 338             /* Inner loop uses 143 flops */
 339         }
 340
 341         if(jidx<j_index_end)
 342         {
 343
 344             jnrA             = jjnr[jidx];
 345             j_coord_offsetA  = DIM*jnrA;
 346
 347             /* load j atom coordinates */
 348             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 349                                               &jx0,&jy0,&jz0);
 350
 351             /* Calculate displacement vector */
 352             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 353             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 354             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 355             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 356             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 357             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 358             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 359             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 360             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 361             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 362             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 363             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 364
 365             /* Calculate squared distance and things based on it */
 366             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 367             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 368             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 369             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 370
 371             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 372             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 373             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 374
 375             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 376             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 377             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 378             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 379
 380             /* Load parameters for j particles */
 381             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 382             vdwjidx0A        = 2*vdwtype[jnrA+0];
 383
 384             fjx0             = _fjsp_setzero_v2r8();
 385             fjy0             = _fjsp_setzero_v2r8();
 386             fjz0             = _fjsp_setzero_v2r8();
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             /* Compute parameters for interactions between i and j atoms */
 393             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 394                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 395
 396             /* LENNARD-JONES DISPERSION/REPULSION */
 397
 398             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 399             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 400             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 401             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 402             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 406             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 407
 408             fscal            = fvdw;
 409
 410             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 411
 412             /* Update vectorial force */
 413             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 414             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 415             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 416
 417             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 418             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 419             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             /* Compute parameters for interactions between i and j atoms */
 426             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 427
 428             /* REACTION-FIELD ELECTROSTATICS */
 429             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 430             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 434             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 439
 440             /* Update vectorial force */
 441             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 442             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 443             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 444
 445             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 446             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 447             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             /* Compute parameters for interactions between i and j atoms */
 454             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 455
 456             /* REACTION-FIELD ELECTROSTATICS */
 457             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 458             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 462             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 467
 468             /* Update vectorial force */
 469             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 470             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 471             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 472
 473             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 474             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 475             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 476
 477             /**************************
 478              * CALCULATE INTERACTIONS *
 479              **************************/
 480
 481             /* Compute parameters for interactions between i and j atoms */
 482             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 483
 484             /* REACTION-FIELD ELECTROSTATICS */
 485             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 486             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 487
 488             /* Update potential sum for this i atom from the interaction with this j atom. */
 489             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 490             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 491
 492             fscal            = felec;
 493
 494             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 495
 496             /* Update vectorial force */
 497             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 498             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 499             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 500
 501             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 502             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 503             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 504
 505             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 506
 507             /* Inner loop uses 143 flops */
 508         }
 509
 510         /* End of innermost loop */
 511
 512         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 513                                               f+i_coord_offset,fshift+i_shift_offset);
 514
 515         ggid                        = gid[iidx];
 516         /* Update potential energies */
 517         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 518         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 519
 520         /* Increment number of inner iterations */
 521         inneriter                  += j_index_end - j_index_start;
 522
 523         /* Outer loop uses 26 flops */
 524     }
 525
 526     /* Increment number of outer iterations */
 527     outeriter        += nri;
 528
 529     /* Update outer/inner flops */
 530
 531     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*143);
 532 }
 533 /*
 534  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 535  * Electrostatics interaction: ReactionField
 536  * VdW interaction:            LennardJones
 537  * Geometry:                   Water4-Particle
 538  * Calculate force/pot:        Force
 539  */
 540 void
 541 nb_kernel_ElecRF_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 542                     (t_nblist                    * gmx_restrict       nlist,
 543                      rvec                        * gmx_restrict          xx,
 544                      rvec                        * gmx_restrict          ff,
 545                      t_forcerec                  * gmx_restrict          fr,
 546                      t_mdatoms                   * gmx_restrict     mdatoms,
 547                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 548                      t_nrnb                      * gmx_restrict        nrnb)
 549 {
 550     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 551      * just 0 for non-waters.
 552      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 553      * jnr indices corresponding to data put in the four positions in the SIMD register.
 554      */
 555     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 556     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 557     int              jnrA,jnrB;
 558     int              j_coord_offsetA,j_coord_offsetB;
 559     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 560     real             rcutoff_scalar;
 561     real             *shiftvec,*fshift,*x,*f;
 562     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 563     int              vdwioffset0;
 564     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 565     int              vdwioffset1;
 566     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 567     int              vdwioffset2;
 568     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 569     int              vdwioffset3;
 570     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 571     int              vdwjidx0A,vdwjidx0B;
 572     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 573     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 574     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 575     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 576     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 577     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 578     real             *charge;
 579     int              nvdwtype;
 580     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 581     int              *vdwtype;
 582     real             *vdwparam;
 583     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 584     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 585     _fjsp_v2r8       itab_tmp;
 586     _fjsp_v2r8       dummy_mask,cutoff_mask;
 587     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 588     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 589     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 590
 591     x                = xx[0];
 592     f                = ff[0];
 593
 594     nri              = nlist->nri;
 595     iinr             = nlist->iinr;
 596     jindex           = nlist->jindex;
 597     jjnr             = nlist->jjnr;
 598     shiftidx         = nlist->shift;
 599     gid              = nlist->gid;
 600     shiftvec         = fr->shift_vec[0];
 601     fshift           = fr->fshift[0];
 602     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 603     charge           = mdatoms->chargeA;
 604     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 605     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 606     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 607     nvdwtype         = fr->ntype;
 608     vdwparam         = fr->nbfp;
 609     vdwtype          = mdatoms->typeA;
 610
 611     /* Setup water-specific parameters */
 612     inr              = nlist->iinr[0];
 613     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 614     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 615     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 616     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 617
 618     /* Avoid stupid compiler warnings */
 619     jnrA = jnrB = 0;
 620     j_coord_offsetA = 0;
 621     j_coord_offsetB = 0;
 622
 623     outeriter        = 0;
 624     inneriter        = 0;
 625
 626     /* Start outer loop over neighborlists */
 627     for(iidx=0; iidx<nri; iidx++)
 628     {
 629         /* Load shift vector for this list */
 630         i_shift_offset   = DIM*shiftidx[iidx];
 631
 632         /* Load limits for loop over neighbors */
 633         j_index_start    = jindex[iidx];
 634         j_index_end      = jindex[iidx+1];
 635
 636         /* Get outer coordinate index */
 637         inr              = iinr[iidx];
 638         i_coord_offset   = DIM*inr;
 639
 640         /* Load i particle coords and add shift vector */
 641         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 642                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 643
 644         fix0             = _fjsp_setzero_v2r8();
 645         fiy0             = _fjsp_setzero_v2r8();
 646         fiz0             = _fjsp_setzero_v2r8();
 647         fix1             = _fjsp_setzero_v2r8();
 648         fiy1             = _fjsp_setzero_v2r8();
 649         fiz1             = _fjsp_setzero_v2r8();
 650         fix2             = _fjsp_setzero_v2r8();
 651         fiy2             = _fjsp_setzero_v2r8();
 652         fiz2             = _fjsp_setzero_v2r8();
 653         fix3             = _fjsp_setzero_v2r8();
 654         fiy3             = _fjsp_setzero_v2r8();
 655         fiz3             = _fjsp_setzero_v2r8();
 656
 657         /* Start inner kernel loop */
 658         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 659         {
 660
 661             /* Get j neighbor index, and coordinate index */
 662             jnrA             = jjnr[jidx];
 663             jnrB             = jjnr[jidx+1];
 664             j_coord_offsetA  = DIM*jnrA;
 665             j_coord_offsetB  = DIM*jnrB;
 666
 667             /* load j atom coordinates */
 668             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 669                                               &jx0,&jy0,&jz0);
 670
 671             /* Calculate displacement vector */
 672             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 673             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 674             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 675             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 676             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 677             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 678             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 679             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 680             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 681             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 682             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 683             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 684
 685             /* Calculate squared distance and things based on it */
 686             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 687             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 688             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 689             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 690
 691             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 692             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 693             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 694
 695             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 696             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 697             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 698             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 699
 700             /* Load parameters for j particles */
 701             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 702             vdwjidx0A        = 2*vdwtype[jnrA+0];
 703             vdwjidx0B        = 2*vdwtype[jnrB+0];
 704
 705             fjx0             = _fjsp_setzero_v2r8();
 706             fjy0             = _fjsp_setzero_v2r8();
 707             fjz0             = _fjsp_setzero_v2r8();
 708
 709             /**************************
 710              * CALCULATE INTERACTIONS *
 711              **************************/
 712
 713             /* Compute parameters for interactions between i and j atoms */
 714             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 715                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 716
 717             /* LENNARD-JONES DISPERSION/REPULSION */
 718
 719             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 720             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 721
 722             fscal            = fvdw;
 723
 724             /* Update vectorial force */
 725             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 726             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 727             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 728
 729             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 730             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 731             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 732
 733             /**************************
 734              * CALCULATE INTERACTIONS *
 735              **************************/
 736
 737             /* Compute parameters for interactions between i and j atoms */
 738             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 739
 740             /* REACTION-FIELD ELECTROSTATICS */
 741             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 742
 743             fscal            = felec;
 744
 745             /* Update vectorial force */
 746             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 747             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 748             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 749
 750             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 751             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 752             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             /* Compute parameters for interactions between i and j atoms */
 759             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 760
 761             /* REACTION-FIELD ELECTROSTATICS */
 762             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 763
 764             fscal            = felec;
 765
 766             /* Update vectorial force */
 767             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 768             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 769             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 770
 771             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 772             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 773             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 774
 775             /**************************
 776              * CALCULATE INTERACTIONS *
 777              **************************/
 778
 779             /* Compute parameters for interactions between i and j atoms */
 780             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 781
 782             /* REACTION-FIELD ELECTROSTATICS */
 783             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 784
 785             fscal            = felec;
 786
 787             /* Update vectorial force */
 788             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 789             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 790             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 791
 792             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 793             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 794             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 795
 796             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 797
 798             /* Inner loop uses 123 flops */
 799         }
 800
 801         if(jidx<j_index_end)
 802         {
 803
 804             jnrA             = jjnr[jidx];
 805             j_coord_offsetA  = DIM*jnrA;
 806
 807             /* load j atom coordinates */
 808             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 809                                               &jx0,&jy0,&jz0);
 810
 811             /* Calculate displacement vector */
 812             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 813             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 814             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 815             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 816             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 817             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 818             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 819             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 820             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 821             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 822             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 823             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 824
 825             /* Calculate squared distance and things based on it */
 826             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 827             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 828             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 829             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 830
 831             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 832             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 833             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 834
 835             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 836             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 837             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 838             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 839
 840             /* Load parameters for j particles */
 841             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 842             vdwjidx0A        = 2*vdwtype[jnrA+0];
 843
 844             fjx0             = _fjsp_setzero_v2r8();
 845             fjy0             = _fjsp_setzero_v2r8();
 846             fjz0             = _fjsp_setzero_v2r8();
 847
 848             /**************************
 849              * CALCULATE INTERACTIONS *
 850              **************************/
 851
 852             /* Compute parameters for interactions between i and j atoms */
 853             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 854                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 855
 856             /* LENNARD-JONES DISPERSION/REPULSION */
 857
 858             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 859             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 860
 861             fscal            = fvdw;
 862
 863             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 864
 865             /* Update vectorial force */
 866             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 867             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 868             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 869
 870             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 871             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 872             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 873
 874             /**************************
 875              * CALCULATE INTERACTIONS *
 876              **************************/
 877
 878             /* Compute parameters for interactions between i and j atoms */
 879             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 880
 881             /* REACTION-FIELD ELECTROSTATICS */
 882             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 883
 884             fscal            = felec;
 885
 886             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 887
 888             /* Update vectorial force */
 889             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 890             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 891             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 892
 893             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 894             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 895             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 896
 897             /**************************
 898              * CALCULATE INTERACTIONS *
 899              **************************/
 900
 901             /* Compute parameters for interactions between i and j atoms */
 902             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 903
 904             /* REACTION-FIELD ELECTROSTATICS */
 905             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 906
 907             fscal            = felec;
 908
 909             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 910
 911             /* Update vectorial force */
 912             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 913             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 914             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 915
 916             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 917             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 918             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 919
 920             /**************************
 921              * CALCULATE INTERACTIONS *
 922              **************************/
 923
 924             /* Compute parameters for interactions between i and j atoms */
 925             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 926
 927             /* REACTION-FIELD ELECTROSTATICS */
 928             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 929
 930             fscal            = felec;
 931
 932             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 933
 934             /* Update vectorial force */
 935             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 936             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 937             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 938
 939             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 940             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 941             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 942
 943             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 944
 945             /* Inner loop uses 123 flops */
 946         }
 947
 948         /* End of innermost loop */
 949
 950         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 951                                               f+i_coord_offset,fshift+i_shift_offset);
 952
 953         /* Increment number of inner iterations */
 954         inneriter                  += j_index_end - j_index_start;
 955
 956         /* Outer loop uses 24 flops */
 957     }
 958
 959     /* Increment number of outer iterations */
 960     outeriter        += nri;
 961
 962     /* Update outer/inner flops */
 963
 964     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*123);
 965 }