src/gromacs/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) 2012,2013,2014, 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 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/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_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     /* 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_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 446                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 447
 448             /* LENNARD-JONES DISPERSION/REPULSION */
 449
 450             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 451             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 452             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 453             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 454                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 455             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 456
 457             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 458
 459             /* Update potential sum for this i atom from the interaction with this j atom. */
 460             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 461             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 462             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 463
 464             fscal            = fvdw;
 465
 466             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 467
 468             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 469
 470             /* Update vectorial force */
 471             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 472             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 473             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 474
 475             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 476             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 477             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 478
 479             }
 480
 481             /**************************
 482              * CALCULATE INTERACTIONS *
 483              **************************/
 484
 485             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 486             {
 487
 488             /* Compute parameters for interactions between i and j atoms */
 489             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 490
 491             /* REACTION-FIELD ELECTROSTATICS */
 492             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 493             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 494
 495             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 499             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 500             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 501
 502             fscal            = felec;
 503
 504             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 505
 506             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 507
 508             /* Update vectorial force */
 509             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 510             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 511             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 512
 513             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 514             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 515             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 516
 517             }
 518
 519             /**************************
 520              * CALCULATE INTERACTIONS *
 521              **************************/
 522
 523             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 524             {
 525
 526             /* Compute parameters for interactions between i and j atoms */
 527             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 528
 529             /* REACTION-FIELD ELECTROSTATICS */
 530             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 531             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 532
 533             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 534
 535             /* Update potential sum for this i atom from the interaction with this j atom. */
 536             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 537             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 538             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 539
 540             fscal            = felec;
 541
 542             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 543
 544             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 545
 546             /* Update vectorial force */
 547             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 548             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 549             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 550
 551             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 552             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 553             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 554
 555             }
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 562             {
 563
 564             /* Compute parameters for interactions between i and j atoms */
 565             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 566
 567             /* REACTION-FIELD ELECTROSTATICS */
 568             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 569             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 570
 571             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 572
 573             /* Update potential sum for this i atom from the interaction with this j atom. */
 574             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 575             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 576             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 577
 578             fscal            = felec;
 579
 580             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 581
 582             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 583
 584             /* Update vectorial force */
 585             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 586             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 587             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 588
 589             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 590             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 591             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 592
 593             }
 594
 595             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 596
 597             /* Inner loop uses 164 flops */
 598         }
 599
 600         /* End of innermost loop */
 601
 602         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 603                                               f+i_coord_offset,fshift+i_shift_offset);
 604
 605         ggid                        = gid[iidx];
 606         /* Update potential energies */
 607         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 608         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 609
 610         /* Increment number of inner iterations */
 611         inneriter                  += j_index_end - j_index_start;
 612
 613         /* Outer loop uses 26 flops */
 614     }
 615
 616     /* Increment number of outer iterations */
 617     outeriter        += nri;
 618
 619     /* Update outer/inner flops */
 620
 621     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*164);
 622 }
 623 /*
 624  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sparc64_hpc_ace_double
 625  * Electrostatics interaction: ReactionField
 626  * VdW interaction:            LennardJones
 627  * Geometry:                   Water4-Particle
 628  * Calculate force/pot:        Force
 629  */
 630 void
 631 nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sparc64_hpc_ace_double
 632                     (t_nblist                    * gmx_restrict       nlist,
 633                      rvec                        * gmx_restrict          xx,
 634                      rvec                        * gmx_restrict          ff,
 635                      t_forcerec                  * gmx_restrict          fr,
 636                      t_mdatoms                   * gmx_restrict     mdatoms,
 637                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 638                      t_nrnb                      * gmx_restrict        nrnb)
 639 {
 640     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 641      * just 0 for non-waters.
 642      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 643      * jnr indices corresponding to data put in the four positions in the SIMD register.
 644      */
 645     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 646     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 647     int              jnrA,jnrB;
 648     int              j_coord_offsetA,j_coord_offsetB;
 649     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 650     real             rcutoff_scalar;
 651     real             *shiftvec,*fshift,*x,*f;
 652     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 653     int              vdwioffset0;
 654     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 655     int              vdwioffset1;
 656     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 657     int              vdwioffset2;
 658     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 659     int              vdwioffset3;
 660     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 661     int              vdwjidx0A,vdwjidx0B;
 662     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 663     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 664     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 665     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 666     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 667     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 668     real             *charge;
 669     int              nvdwtype;
 670     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 671     int              *vdwtype;
 672     real             *vdwparam;
 673     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 674     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 675     _fjsp_v2r8       itab_tmp;
 676     _fjsp_v2r8       dummy_mask,cutoff_mask;
 677     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 678     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 679     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 680
 681     x                = xx[0];
 682     f                = ff[0];
 683
 684     nri              = nlist->nri;
 685     iinr             = nlist->iinr;
 686     jindex           = nlist->jindex;
 687     jjnr             = nlist->jjnr;
 688     shiftidx         = nlist->shift;
 689     gid              = nlist->gid;
 690     shiftvec         = fr->shift_vec[0];
 691     fshift           = fr->fshift[0];
 692     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 693     charge           = mdatoms->chargeA;
 694     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 695     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 696     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 697     nvdwtype         = fr->ntype;
 698     vdwparam         = fr->nbfp;
 699     vdwtype          = mdatoms->typeA;
 700
 701     /* Setup water-specific parameters */
 702     inr              = nlist->iinr[0];
 703     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 704     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 705     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 706     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 707
 708     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 709     rcutoff_scalar   = fr->rcoulomb;
 710     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 711     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 712
 713     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 714     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 715
 716     /* Avoid stupid compiler warnings */
 717     jnrA = jnrB = 0;
 718     j_coord_offsetA = 0;
 719     j_coord_offsetB = 0;
 720
 721     outeriter        = 0;
 722     inneriter        = 0;
 723
 724     /* Start outer loop over neighborlists */
 725     for(iidx=0; iidx<nri; iidx++)
 726     {
 727         /* Load shift vector for this list */
 728         i_shift_offset   = DIM*shiftidx[iidx];
 729
 730         /* Load limits for loop over neighbors */
 731         j_index_start    = jindex[iidx];
 732         j_index_end      = jindex[iidx+1];
 733
 734         /* Get outer coordinate index */
 735         inr              = iinr[iidx];
 736         i_coord_offset   = DIM*inr;
 737
 738         /* Load i particle coords and add shift vector */
 739         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 740                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 741
 742         fix0             = _fjsp_setzero_v2r8();
 743         fiy0             = _fjsp_setzero_v2r8();
 744         fiz0             = _fjsp_setzero_v2r8();
 745         fix1             = _fjsp_setzero_v2r8();
 746         fiy1             = _fjsp_setzero_v2r8();
 747         fiz1             = _fjsp_setzero_v2r8();
 748         fix2             = _fjsp_setzero_v2r8();
 749         fiy2             = _fjsp_setzero_v2r8();
 750         fiz2             = _fjsp_setzero_v2r8();
 751         fix3             = _fjsp_setzero_v2r8();
 752         fiy3             = _fjsp_setzero_v2r8();
 753         fiz3             = _fjsp_setzero_v2r8();
 754
 755         /* Start inner kernel loop */
 756         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 757         {
 758
 759             /* Get j neighbor index, and coordinate index */
 760             jnrA             = jjnr[jidx];
 761             jnrB             = jjnr[jidx+1];
 762             j_coord_offsetA  = DIM*jnrA;
 763             j_coord_offsetB  = DIM*jnrB;
 764
 765             /* load j atom coordinates */
 766             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 767                                               &jx0,&jy0,&jz0);
 768
 769             /* Calculate displacement vector */
 770             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 771             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 772             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 773             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 774             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 775             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 776             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 777             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 778             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 779             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 780             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 781             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 782
 783             /* Calculate squared distance and things based on it */
 784             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 785             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 786             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 787             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 788
 789             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 790             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 791             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 792
 793             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 794             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 795             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 796             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 797
 798             /* Load parameters for j particles */
 799             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 800             vdwjidx0A        = 2*vdwtype[jnrA+0];
 801             vdwjidx0B        = 2*vdwtype[jnrB+0];
 802
 803             fjx0             = _fjsp_setzero_v2r8();
 804             fjy0             = _fjsp_setzero_v2r8();
 805             fjz0             = _fjsp_setzero_v2r8();
 806
 807             /**************************
 808              * CALCULATE INTERACTIONS *
 809              **************************/
 810
 811             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 812             {
 813
 814             /* Compute parameters for interactions between i and j atoms */
 815             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 816                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 817
 818             /* LENNARD-JONES DISPERSION/REPULSION */
 819
 820             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 821             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 822
 823             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 824
 825             fscal            = fvdw;
 826
 827             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 828
 829             /* Update vectorial force */
 830             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 831             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 832             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 833
 834             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 835             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 836             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 837
 838             }
 839
 840             /**************************
 841              * CALCULATE INTERACTIONS *
 842              **************************/
 843
 844             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 845             {
 846
 847             /* Compute parameters for interactions between i and j atoms */
 848             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 849
 850             /* REACTION-FIELD ELECTROSTATICS */
 851             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 852
 853             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 854
 855             fscal            = felec;
 856
 857             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 858
 859             /* Update vectorial force */
 860             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 861             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 862             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 863
 864             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 865             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 866             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 867
 868             }
 869
 870             /**************************
 871              * CALCULATE INTERACTIONS *
 872              **************************/
 873
 874             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 875             {
 876
 877             /* Compute parameters for interactions between i and j atoms */
 878             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 879
 880             /* REACTION-FIELD ELECTROSTATICS */
 881             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 882
 883             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 884
 885             fscal            = felec;
 886
 887             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 888
 889             /* Update vectorial force */
 890             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 891             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 892             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 893
 894             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 895             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 896             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 897
 898             }
 899
 900             /**************************
 901              * CALCULATE INTERACTIONS *
 902              **************************/
 903
 904             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 905             {
 906
 907             /* Compute parameters for interactions between i and j atoms */
 908             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 909
 910             /* REACTION-FIELD ELECTROSTATICS */
 911             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 912
 913             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 914
 915             fscal            = felec;
 916
 917             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 918
 919             /* Update vectorial force */
 920             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 921             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 922             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 923
 924             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 925             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 926             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 927
 928             }
 929
 930             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 931
 932             /* Inner loop uses 135 flops */
 933         }
 934
 935         if(jidx<j_index_end)
 936         {
 937
 938             jnrA             = jjnr[jidx];
 939             j_coord_offsetA  = DIM*jnrA;
 940
 941             /* load j atom coordinates */
 942             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 943                                               &jx0,&jy0,&jz0);
 944
 945             /* Calculate displacement vector */
 946             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 947             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 948             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 949             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 950             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 951             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 952             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 953             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 954             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 955             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 956             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 957             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 958
 959             /* Calculate squared distance and things based on it */
 960             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 961             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 962             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 963             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 964
 965             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 966             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 967             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 968
 969             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 970             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 971             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 972             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 973
 974             /* Load parameters for j particles */
 975             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 976             vdwjidx0A        = 2*vdwtype[jnrA+0];
 977
 978             fjx0             = _fjsp_setzero_v2r8();
 979             fjy0             = _fjsp_setzero_v2r8();
 980             fjz0             = _fjsp_setzero_v2r8();
 981
 982             /**************************
 983              * CALCULATE INTERACTIONS *
 984              **************************/
 985
 986             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 987             {
 988
 989             /* Compute parameters for interactions between i and j atoms */
 990             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 991                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 992
 993             /* LENNARD-JONES DISPERSION/REPULSION */
 994
 995             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 996             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 997
 998             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 999
1000             fscal            = fvdw;
1001
1002             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1003
1004             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1005
1006             /* Update vectorial force */
1007             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1008             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1009             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1010
1011             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1012             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1013             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1014
1015             }
1016
1017             /**************************
1018              * CALCULATE INTERACTIONS *
1019              **************************/
1020
1021             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1022             {
1023
1024             /* Compute parameters for interactions between i and j atoms */
1025             qq10             = _fjsp_mul_v2r8(iq1,jq0);
1026
1027             /* REACTION-FIELD ELECTROSTATICS */
1028             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1029
1030             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1031
1032             fscal            = felec;
1033
1034             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1035
1036             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1037
1038             /* Update vectorial force */
1039             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1040             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1041             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1042
1043             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1044             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1045             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1046
1047             }
1048
1049             /**************************
1050              * CALCULATE INTERACTIONS *
1051              **************************/
1052
1053             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1054             {
1055
1056             /* Compute parameters for interactions between i and j atoms */
1057             qq20             = _fjsp_mul_v2r8(iq2,jq0);
1058
1059             /* REACTION-FIELD ELECTROSTATICS */
1060             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1061
1062             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1063
1064             fscal            = felec;
1065
1066             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1067
1068             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1069
1070             /* Update vectorial force */
1071             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1072             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1073             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1074
1075             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1076             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1077             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1078
1079             }
1080
1081             /**************************
1082              * CALCULATE INTERACTIONS *
1083              **************************/
1084
1085             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
1086             {
1087
1088             /* Compute parameters for interactions between i and j atoms */
1089             qq30             = _fjsp_mul_v2r8(iq3,jq0);
1090
1091             /* REACTION-FIELD ELECTROSTATICS */
1092             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
1093
1094             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
1095
1096             fscal            = felec;
1097
1098             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1099
1100             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1101
1102             /* Update vectorial force */
1103             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
1104             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
1105             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
1106
1107             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
1108             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
1109             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
1110
1111             }
1112
1113             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1114
1115             /* Inner loop uses 135 flops */
1116         }
1117
1118         /* End of innermost loop */
1119
1120         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1121                                               f+i_coord_offset,fshift+i_shift_offset);
1122
1123         /* Increment number of inner iterations */
1124         inneriter                  += j_index_end - j_index_start;
1125
1126         /* Outer loop uses 24 flops */
1127     }
1128
1129     /* Increment number of outer iterations */
1130     outeriter        += nri;
1131
1132     /* Update outer/inner flops */
1133
1134     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*135);
1135 }