src/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012, by the GROMACS development team, led by
   5  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   6  * others, as listed in the AUTHORS file in the top-level source
   7  * 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_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_ElecRFCut_VdwLJSh_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_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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 137     rcutoff_scalar   = fr->rcoulomb;
 138     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 139     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 140
 141     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 142     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 143
 144     /* Avoid stupid compiler warnings */
 145     jnrA = jnrB = 0;
 146     j_coord_offsetA = 0;
 147     j_coord_offsetB = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157
 158         /* Load limits for loop over neighbors */
 159         j_index_start    = jindex[iidx];
 160         j_index_end      = jindex[iidx+1];
 161
 162         /* Get outer coordinate index */
 163         inr              = iinr[iidx];
 164         i_coord_offset   = DIM*inr;
 165
 166         /* Load i particle coords and add shift vector */
 167         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 168                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 169
 170         fix0             = _fjsp_setzero_v2r8();
 171         fiy0             = _fjsp_setzero_v2r8();
 172         fiz0             = _fjsp_setzero_v2r8();
 173         fix1             = _fjsp_setzero_v2r8();
 174         fiy1             = _fjsp_setzero_v2r8();
 175         fiz1             = _fjsp_setzero_v2r8();
 176         fix2             = _fjsp_setzero_v2r8();
 177         fiy2             = _fjsp_setzero_v2r8();
 178         fiz2             = _fjsp_setzero_v2r8();
 179         fix3             = _fjsp_setzero_v2r8();
 180         fiy3             = _fjsp_setzero_v2r8();
 181         fiz3             = _fjsp_setzero_v2r8();
 182
 183         /* Reset potential sums */
 184         velecsum         = _fjsp_setzero_v2r8();
 185         vvdwsum          = _fjsp_setzero_v2r8();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196
 197             /* load j atom coordinates */
 198             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                               &jx0,&jy0,&jz0);
 200
 201             /* Calculate displacement vector */
 202             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 203             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 204             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 205             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 206             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 207             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 208             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 209             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 210             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 211             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 212             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 213             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 214
 215             /* Calculate squared distance and things based on it */
 216             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 217             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 218             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 219             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 220
 221             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 222             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 223             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 224
 225             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 226             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 227             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 228             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 229
 230             /* Load parameters for j particles */
 231             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 232             vdwjidx0A        = 2*vdwtype[jnrA+0];
 233             vdwjidx0B        = 2*vdwtype[jnrB+0];
 234
 235             fjx0             = _fjsp_setzero_v2r8();
 236             fjy0             = _fjsp_setzero_v2r8();
 237             fjz0             = _fjsp_setzero_v2r8();
 238
 239             /**************************
 240              * CALCULATE INTERACTIONS *
 241              **************************/
 242
 243             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 244             {
 245
 246             /* Compute parameters for interactions between i and j atoms */
 247             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 248                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 249
 250             /* LENNARD-JONES DISPERSION/REPULSION */
 251
 252             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 253             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 254             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 255             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 256                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 257             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 258
 259             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 260
 261             /* Update potential sum for this i atom from the interaction with this j atom. */
 262             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 263             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 264
 265             fscal            = fvdw;
 266
 267             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 268
 269             /* Update vectorial force */
 270             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 271             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 272             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 273
 274             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 275             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 276             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 277
 278             }
 279
 280             /**************************
 281              * CALCULATE INTERACTIONS *
 282              **************************/
 283
 284             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 285             {
 286
 287             /* Compute parameters for interactions between i and j atoms */
 288             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 289
 290             /* REACTION-FIELD ELECTROSTATICS */
 291             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 292             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 293
 294             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 295
 296             /* Update potential sum for this i atom from the interaction with this j atom. */
 297             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 298             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 299
 300             fscal            = felec;
 301
 302             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 303
 304             /* Update vectorial force */
 305             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 306             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 307             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 308
 309             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 310             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 311             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 312
 313             }
 314
 315             /**************************
 316              * CALCULATE INTERACTIONS *
 317              **************************/
 318
 319             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 320             {
 321
 322             /* Compute parameters for interactions between i and j atoms */
 323             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 324
 325             /* REACTION-FIELD ELECTROSTATICS */
 326             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 327             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 328
 329             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 330
 331             /* Update potential sum for this i atom from the interaction with this j atom. */
 332             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 333             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 334
 335             fscal            = felec;
 336
 337             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 338
 339             /* Update vectorial force */
 340             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 341             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 342             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 343
 344             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 345             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 346             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 347
 348             }
 349
 350             /**************************
 351              * CALCULATE INTERACTIONS *
 352              **************************/
 353
 354             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 355             {
 356
 357             /* Compute parameters for interactions between i and j atoms */
 358             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 362             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 363
 364             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 365
 366             /* Update potential sum for this i atom from the interaction with this j atom. */
 367             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 368             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 369
 370             fscal            = felec;
 371
 372             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 373
 374             /* Update vectorial force */
 375             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 376             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 377             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 378
 379             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 380             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 381             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 382
 383             }
 384
 385             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 386
 387             /* Inner loop uses 164 flops */
 388         }
 389
 390         if(jidx<j_index_end)
 391         {
 392
 393             jnrA             = jjnr[jidx];
 394             j_coord_offsetA  = DIM*jnrA;
 395
 396             /* load j atom coordinates */
 397             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 398                                               &jx0,&jy0,&jz0);
 399
 400             /* Calculate displacement vector */
 401             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 402             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 403             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 404             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 405             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 406             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 407             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 408             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 409             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 410             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 411             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 412             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 413
 414             /* Calculate squared distance and things based on it */
 415             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 416             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 417             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 418             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 419
 420             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 421             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 422             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 423
 424             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 425             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 426             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 427             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 428
 429             /* Load parameters for j particles */
 430             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 431             vdwjidx0A        = 2*vdwtype[jnrA+0];
 432
 433             fjx0             = _fjsp_setzero_v2r8();
 434             fjy0             = _fjsp_setzero_v2r8();
 435             fjz0             = _fjsp_setzero_v2r8();
 436
 437             /**************************
 438              * CALCULATE INTERACTIONS *
 439              **************************/
 440
 441             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 442             {
 443
 444             /* Compute parameters for interactions between i and j atoms */
 445             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 446
 447             /* LENNARD-JONES DISPERSION/REPULSION */
 448
 449             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 450             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 451             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 452             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 453                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 454             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 455
 456             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 457
 458             /* Update potential sum for this i atom from the interaction with this j atom. */
 459             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 460             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 461             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 462
 463             fscal            = fvdw;
 464
 465             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 466
 467             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 468
 469             /* Update vectorial force */
 470             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 471             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 472             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 473
 474             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 475             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 476             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 477
 478             }
 479
 480             /**************************
 481              * CALCULATE INTERACTIONS *
 482              **************************/
 483
 484             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 485             {
 486
 487             /* Compute parameters for interactions between i and j atoms */
 488             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 489
 490             /* REACTION-FIELD ELECTROSTATICS */
 491             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 492             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 493
 494             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 498             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 499             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 500
 501             fscal            = felec;
 502
 503             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 504
 505             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 506
 507             /* Update vectorial force */
 508             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 509             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 510             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 511
 512             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 513             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 514             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 515
 516             }
 517
 518             /**************************
 519              * CALCULATE INTERACTIONS *
 520              **************************/
 521
 522             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 523             {
 524
 525             /* Compute parameters for interactions between i and j atoms */
 526             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 527
 528             /* REACTION-FIELD ELECTROSTATICS */
 529             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 530             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 531
 532             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 533
 534             /* Update potential sum for this i atom from the interaction with this j atom. */
 535             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 536             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 537             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 538
 539             fscal            = felec;
 540
 541             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 542
 543             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 544
 545             /* Update vectorial force */
 546             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 547             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 548             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 549
 550             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 551             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 552             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 553
 554             }
 555
 556             /**************************
 557              * CALCULATE INTERACTIONS *
 558              **************************/
 559
 560             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 561             {
 562
 563             /* Compute parameters for interactions between i and j atoms */
 564             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 565
 566             /* REACTION-FIELD ELECTROSTATICS */
 567             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 568             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 569
 570             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 571
 572             /* Update potential sum for this i atom from the interaction with this j atom. */
 573             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 574             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 575             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 576
 577             fscal            = felec;
 578
 579             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 580
 581             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 582
 583             /* Update vectorial force */
 584             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 585             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 586             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 587
 588             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 589             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 590             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 591
 592             }
 593
 594             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 595
 596             /* Inner loop uses 164 flops */
 597         }
 598
 599         /* End of innermost loop */
 600
 601         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 602                                               f+i_coord_offset,fshift+i_shift_offset);
 603
 604         ggid                        = gid[iidx];
 605         /* Update potential energies */
 606         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 607         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 608
 609         /* Increment number of inner iterations */
 610         inneriter                  += j_index_end - j_index_start;
 611
 612         /* Outer loop uses 26 flops */
 613     }
 614
 615     /* Increment number of outer iterations */
 616     outeriter        += nri;
 617
 618     /* Update outer/inner flops */
 619
 620     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*164);
 621 }
 622 /*
 623  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sparc64_hpc_ace_double
 624  * Electrostatics interaction: ReactionField
 625  * VdW interaction:            LennardJones
 626  * Geometry:                   Water4-Particle
 627  * Calculate force/pot:        Force
 628  */
 629 void
 630 nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sparc64_hpc_ace_double
 631                     (t_nblist * gmx_restrict                nlist,
 632                      rvec * gmx_restrict                    xx,
 633                      rvec * gmx_restrict                    ff,
 634                      t_forcerec * gmx_restrict              fr,
 635                      t_mdatoms * gmx_restrict               mdatoms,
 636                      nb_kernel_data_t * gmx_restrict        kernel_data,
 637                      t_nrnb * gmx_restrict                  nrnb)
 638 {
 639     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 640      * just 0 for non-waters.
 641      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 642      * jnr indices corresponding to data put in the four positions in the SIMD register.
 643      */
 644     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 645     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 646     int              jnrA,jnrB;
 647     int              j_coord_offsetA,j_coord_offsetB;
 648     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 649     real             rcutoff_scalar;
 650     real             *shiftvec,*fshift,*x,*f;
 651     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 652     int              vdwioffset0;
 653     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 654     int              vdwioffset1;
 655     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 656     int              vdwioffset2;
 657     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 658     int              vdwioffset3;
 659     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 660     int              vdwjidx0A,vdwjidx0B;
 661     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 662     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 663     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 664     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 665     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 666     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 667     real             *charge;
 668     int              nvdwtype;
 669     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 670     int              *vdwtype;
 671     real             *vdwparam;
 672     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 673     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 674     _fjsp_v2r8       itab_tmp;
 675     _fjsp_v2r8       dummy_mask,cutoff_mask;
 676     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 677     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 678     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 679
 680     x                = xx[0];
 681     f                = ff[0];
 682
 683     nri              = nlist->nri;
 684     iinr             = nlist->iinr;
 685     jindex           = nlist->jindex;
 686     jjnr             = nlist->jjnr;
 687     shiftidx         = nlist->shift;
 688     gid              = nlist->gid;
 689     shiftvec         = fr->shift_vec[0];
 690     fshift           = fr->fshift[0];
 691     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 692     charge           = mdatoms->chargeA;
 693     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 694     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 695     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 696     nvdwtype         = fr->ntype;
 697     vdwparam         = fr->nbfp;
 698     vdwtype          = mdatoms->typeA;
 699
 700     /* Setup water-specific parameters */
 701     inr              = nlist->iinr[0];
 702     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 703     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 704     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 705     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 706
 707     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 708     rcutoff_scalar   = fr->rcoulomb;
 709     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 710     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 711
 712     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 713     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 714
 715     /* Avoid stupid compiler warnings */
 716     jnrA = jnrB = 0;
 717     j_coord_offsetA = 0;
 718     j_coord_offsetB = 0;
 719
 720     outeriter        = 0;
 721     inneriter        = 0;
 722
 723     /* Start outer loop over neighborlists */
 724     for(iidx=0; iidx<nri; iidx++)
 725     {
 726         /* Load shift vector for this list */
 727         i_shift_offset   = DIM*shiftidx[iidx];
 728
 729         /* Load limits for loop over neighbors */
 730         j_index_start    = jindex[iidx];
 731         j_index_end      = jindex[iidx+1];
 732
 733         /* Get outer coordinate index */
 734         inr              = iinr[iidx];
 735         i_coord_offset   = DIM*inr;
 736
 737         /* Load i particle coords and add shift vector */
 738         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 739                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 740
 741         fix0             = _fjsp_setzero_v2r8();
 742         fiy0             = _fjsp_setzero_v2r8();
 743         fiz0             = _fjsp_setzero_v2r8();
 744         fix1             = _fjsp_setzero_v2r8();
 745         fiy1             = _fjsp_setzero_v2r8();
 746         fiz1             = _fjsp_setzero_v2r8();
 747         fix2             = _fjsp_setzero_v2r8();
 748         fiy2             = _fjsp_setzero_v2r8();
 749         fiz2             = _fjsp_setzero_v2r8();
 750         fix3             = _fjsp_setzero_v2r8();
 751         fiy3             = _fjsp_setzero_v2r8();
 752         fiz3             = _fjsp_setzero_v2r8();
 753
 754         /* Start inner kernel loop */
 755         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 756         {
 757
 758             /* Get j neighbor index, and coordinate index */
 759             jnrA             = jjnr[jidx];
 760             jnrB             = jjnr[jidx+1];
 761             j_coord_offsetA  = DIM*jnrA;
 762             j_coord_offsetB  = DIM*jnrB;
 763
 764             /* load j atom coordinates */
 765             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 766                                               &jx0,&jy0,&jz0);
 767
 768             /* Calculate displacement vector */
 769             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 770             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 771             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 772             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 773             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 774             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 775             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 776             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 777             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 778             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 779             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 780             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 781
 782             /* Calculate squared distance and things based on it */
 783             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 784             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 785             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 786             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 787
 788             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 789             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 790             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 791
 792             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 793             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 794             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 795             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 796
 797             /* Load parameters for j particles */
 798             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 799             vdwjidx0A        = 2*vdwtype[jnrA+0];
 800             vdwjidx0B        = 2*vdwtype[jnrB+0];
 801
 802             fjx0             = _fjsp_setzero_v2r8();
 803             fjy0             = _fjsp_setzero_v2r8();
 804             fjz0             = _fjsp_setzero_v2r8();
 805
 806             /**************************
 807              * CALCULATE INTERACTIONS *
 808              **************************/
 809
 810             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 811             {
 812
 813             /* Compute parameters for interactions between i and j atoms */
 814             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 815                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 816
 817             /* LENNARD-JONES DISPERSION/REPULSION */
 818
 819             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 820             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 821
 822             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 823
 824             fscal            = fvdw;
 825
 826             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 827
 828             /* Update vectorial force */
 829             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 830             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 831             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 832
 833             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 834             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 835             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 836
 837             }
 838
 839             /**************************
 840              * CALCULATE INTERACTIONS *
 841              **************************/
 842
 843             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 844             {
 845
 846             /* Compute parameters for interactions between i and j atoms */
 847             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 848
 849             /* REACTION-FIELD ELECTROSTATICS */
 850             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 851
 852             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 853
 854             fscal            = felec;
 855
 856             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 857
 858             /* Update vectorial force */
 859             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 860             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 861             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 862
 863             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 864             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 865             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 866
 867             }
 868
 869             /**************************
 870              * CALCULATE INTERACTIONS *
 871              **************************/
 872
 873             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 874             {
 875
 876             /* Compute parameters for interactions between i and j atoms */
 877             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 878
 879             /* REACTION-FIELD ELECTROSTATICS */
 880             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 881
 882             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 883
 884             fscal            = felec;
 885
 886             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 887
 888             /* Update vectorial force */
 889             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 890             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 891             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 892
 893             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 894             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 895             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 896
 897             }
 898
 899             /**************************
 900              * CALCULATE INTERACTIONS *
 901              **************************/
 902
 903             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 904             {
 905
 906             /* Compute parameters for interactions between i and j atoms */
 907             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 908
 909             /* REACTION-FIELD ELECTROSTATICS */
 910             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 911
 912             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 913
 914             fscal            = felec;
 915
 916             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 917
 918             /* Update vectorial force */
 919             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 920             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 921             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 922
 923             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 924             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 925             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 926
 927             }
 928
 929             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 930
 931             /* Inner loop uses 135 flops */
 932         }
 933
 934         if(jidx<j_index_end)
 935         {
 936
 937             jnrA             = jjnr[jidx];
 938             j_coord_offsetA  = DIM*jnrA;
 939
 940             /* load j atom coordinates */
 941             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 942                                               &jx0,&jy0,&jz0);
 943
 944             /* Calculate displacement vector */
 945             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 946             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 947             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 948             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 949             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 950             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 951             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 952             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 953             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 954             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 955             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 956             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 957
 958             /* Calculate squared distance and things based on it */
 959             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 960             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 961             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 962             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 963
 964             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 965             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 966             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 967
 968             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 969             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 970             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 971             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 972
 973             /* Load parameters for j particles */
 974             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 975             vdwjidx0A        = 2*vdwtype[jnrA+0];
 976
 977             fjx0             = _fjsp_setzero_v2r8();
 978             fjy0             = _fjsp_setzero_v2r8();
 979             fjz0             = _fjsp_setzero_v2r8();
 980
 981             /**************************
 982              * CALCULATE INTERACTIONS *
 983              **************************/
 984
 985             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 986             {
 987
 988             /* Compute parameters for interactions between i and j atoms */
 989             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 990
 991             /* LENNARD-JONES DISPERSION/REPULSION */
 992
 993             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 994             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 995
 996             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 997
 998             fscal            = fvdw;
 999
1000             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1001
1002             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1003
1004             /* Update vectorial force */
1005             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1006             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1007             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1008
1009             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1010             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1011             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1012
1013             }
1014
1015             /**************************
1016              * CALCULATE INTERACTIONS *
1017              **************************/
1018
1019             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1020             {
1021
1022             /* Compute parameters for interactions between i and j atoms */
1023             qq10             = _fjsp_mul_v2r8(iq1,jq0);
1024
1025             /* REACTION-FIELD ELECTROSTATICS */
1026             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1027
1028             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1029
1030             fscal            = felec;
1031
1032             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1033
1034             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1035
1036             /* Update vectorial force */
1037             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1038             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1039             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1040
1041             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1042             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1043             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1044
1045             }
1046
1047             /**************************
1048              * CALCULATE INTERACTIONS *
1049              **************************/
1050
1051             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1052             {
1053
1054             /* Compute parameters for interactions between i and j atoms */
1055             qq20             = _fjsp_mul_v2r8(iq2,jq0);
1056
1057             /* REACTION-FIELD ELECTROSTATICS */
1058             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1059
1060             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1061
1062             fscal            = felec;
1063
1064             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1065
1066             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1067
1068             /* Update vectorial force */
1069             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1070             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1071             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1072
1073             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1074             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1075             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1076
1077             }
1078
1079             /**************************
1080              * CALCULATE INTERACTIONS *
1081              **************************/
1082
1083             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
1084             {
1085
1086             /* Compute parameters for interactions between i and j atoms */
1087             qq30             = _fjsp_mul_v2r8(iq3,jq0);
1088
1089             /* REACTION-FIELD ELECTROSTATICS */
1090             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
1091
1092             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
1093
1094             fscal            = felec;
1095
1096             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1097
1098             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1099
1100             /* Update vectorial force */
1101             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
1102             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
1103             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
1104
1105             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
1106             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
1107             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
1108
1109             }
1110
1111             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1112
1113             /* Inner loop uses 135 flops */
1114         }
1115
1116         /* End of innermost loop */
1117
1118         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1119                                               f+i_coord_offset,fshift+i_shift_offset);
1120
1121         /* Increment number of inner iterations */
1122         inneriter                  += j_index_end - j_index_start;
1123
1124         /* Outer loop uses 24 flops */
1125     }
1126
1127     /* Increment number of outer iterations */
1128     outeriter        += nri;
1129
1130     /* Update outer/inner flops */
1131
1132     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*135);
1133 }