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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwioffset3;
  86     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  91     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  92     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  93     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 100     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 101     _fjsp_v2r8       itab_tmp;
 102     _fjsp_v2r8       dummy_mask,cutoff_mask;
 103     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 104     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 105     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 106
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 121     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 122     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 130     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 131     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 132     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 133
 134     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 135     rcutoff_scalar   = fr->rcoulomb;
 136     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 137     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 138
 139     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 140     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146
 147     outeriter        = 0;
 148     inneriter        = 0;
 149
 150     /* Start outer loop over neighborlists */
 151     for(iidx=0; iidx<nri; iidx++)
 152     {
 153         /* Load shift vector for this list */
 154         i_shift_offset   = DIM*shiftidx[iidx];
 155
 156         /* Load limits for loop over neighbors */
 157         j_index_start    = jindex[iidx];
 158         j_index_end      = jindex[iidx+1];
 159
 160         /* Get outer coordinate index */
 161         inr              = iinr[iidx];
 162         i_coord_offset   = DIM*inr;
 163
 164         /* Load i particle coords and add shift vector */
 165         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 166                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 167
 168         fix0             = _fjsp_setzero_v2r8();
 169         fiy0             = _fjsp_setzero_v2r8();
 170         fiz0             = _fjsp_setzero_v2r8();
 171         fix1             = _fjsp_setzero_v2r8();
 172         fiy1             = _fjsp_setzero_v2r8();
 173         fiz1             = _fjsp_setzero_v2r8();
 174         fix2             = _fjsp_setzero_v2r8();
 175         fiy2             = _fjsp_setzero_v2r8();
 176         fiz2             = _fjsp_setzero_v2r8();
 177         fix3             = _fjsp_setzero_v2r8();
 178         fiy3             = _fjsp_setzero_v2r8();
 179         fiz3             = _fjsp_setzero_v2r8();
 180
 181         /* Reset potential sums */
 182         velecsum         = _fjsp_setzero_v2r8();
 183         vvdwsum          = _fjsp_setzero_v2r8();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194
 195             /* load j atom coordinates */
 196             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 197                                               &jx0,&jy0,&jz0);
 198
 199             /* Calculate displacement vector */
 200             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 201             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 202             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 203             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 204             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 205             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 206             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 207             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 208             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 209             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 210             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 211             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 212
 213             /* Calculate squared distance and things based on it */
 214             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 215             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 216             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 217             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 218
 219             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 220             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 221             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 222
 223             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 224             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 225             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 226             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 227
 228             /* Load parameters for j particles */
 229             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 230             vdwjidx0A        = 2*vdwtype[jnrA+0];
 231             vdwjidx0B        = 2*vdwtype[jnrB+0];
 232
 233             fjx0             = _fjsp_setzero_v2r8();
 234             fjy0             = _fjsp_setzero_v2r8();
 235             fjz0             = _fjsp_setzero_v2r8();
 236
 237             /**************************
 238              * CALCULATE INTERACTIONS *
 239              **************************/
 240
 241             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 242             {
 243
 244             /* Compute parameters for interactions between i and j atoms */
 245             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 246                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 247
 248             /* LENNARD-JONES DISPERSION/REPULSION */
 249
 250             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 251             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 252             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 253             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 254                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 255             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 256
 257             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 258
 259             /* Update potential sum for this i atom from the interaction with this j atom. */
 260             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 261             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 262
 263             fscal            = fvdw;
 264
 265             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 266
 267             /* Update vectorial force */
 268             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 269             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 270             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 271
 272             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 273             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 274             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 275
 276             }
 277
 278             /**************************
 279              * CALCULATE INTERACTIONS *
 280              **************************/
 281
 282             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 283             {
 284
 285             /* Compute parameters for interactions between i and j atoms */
 286             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 287
 288             /* REACTION-FIELD ELECTROSTATICS */
 289             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 290             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 291
 292             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 293
 294             /* Update potential sum for this i atom from the interaction with this j atom. */
 295             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 296             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 297
 298             fscal            = felec;
 299
 300             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 301
 302             /* Update vectorial force */
 303             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 304             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 305             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 306
 307             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 308             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 309             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 310
 311             }
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 318             {
 319
 320             /* Compute parameters for interactions between i and j atoms */
 321             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 322
 323             /* REACTION-FIELD ELECTROSTATICS */
 324             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 325             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 326
 327             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 328
 329             /* Update potential sum for this i atom from the interaction with this j atom. */
 330             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 331             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 332
 333             fscal            = felec;
 334
 335             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 336
 337             /* Update vectorial force */
 338             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 339             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 340             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 341
 342             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 343             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 344             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 345
 346             }
 347
 348             /**************************
 349              * CALCULATE INTERACTIONS *
 350              **************************/
 351
 352             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 353             {
 354
 355             /* Compute parameters for interactions between i and j atoms */
 356             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 357
 358             /* REACTION-FIELD ELECTROSTATICS */
 359             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 360             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 361
 362             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 363
 364             /* Update potential sum for this i atom from the interaction with this j atom. */
 365             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 366             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 371
 372             /* Update vectorial force */
 373             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 374             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 375             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 376
 377             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 378             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 379             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 380
 381             }
 382
 383             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 384
 385             /* Inner loop uses 164 flops */
 386         }
 387
 388         if(jidx<j_index_end)
 389         {
 390
 391             jnrA             = jjnr[jidx];
 392             j_coord_offsetA  = DIM*jnrA;
 393
 394             /* load j atom coordinates */
 395             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 396                                               &jx0,&jy0,&jz0);
 397
 398             /* Calculate displacement vector */
 399             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 400             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 401             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 402             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 403             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 404             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 405             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 406             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 407             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 408             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 409             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 410             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 411
 412             /* Calculate squared distance and things based on it */
 413             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 414             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 415             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 416             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 417
 418             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 419             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 420             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 421
 422             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 423             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 424             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 425             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 426
 427             /* Load parameters for j particles */
 428             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 429             vdwjidx0A        = 2*vdwtype[jnrA+0];
 430
 431             fjx0             = _fjsp_setzero_v2r8();
 432             fjy0             = _fjsp_setzero_v2r8();
 433             fjz0             = _fjsp_setzero_v2r8();
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 440             {
 441
 442             /* Compute parameters for interactions between i and j atoms */
 443             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 444
 445             /* LENNARD-JONES DISPERSION/REPULSION */
 446
 447             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 448             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 449             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 450             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 451                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 452             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 453
 454             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 458             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 459             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 460
 461             fscal            = fvdw;
 462
 463             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 464
 465             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 466
 467             /* Update vectorial force */
 468             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 469             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 470             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 471
 472             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 473             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 474             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 475
 476             }
 477
 478             /**************************
 479              * CALCULATE INTERACTIONS *
 480              **************************/
 481
 482             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 483             {
 484
 485             /* Compute parameters for interactions between i and j atoms */
 486             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 487
 488             /* REACTION-FIELD ELECTROSTATICS */
 489             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 490             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 491
 492             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 493
 494             /* Update potential sum for this i atom from the interaction with this j atom. */
 495             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 496             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 497             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 498
 499             fscal            = felec;
 500
 501             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 502
 503             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 504
 505             /* Update vectorial force */
 506             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 507             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 508             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 509
 510             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 511             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 512             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 513
 514             }
 515
 516             /**************************
 517              * CALCULATE INTERACTIONS *
 518              **************************/
 519
 520             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 521             {
 522
 523             /* Compute parameters for interactions between i and j atoms */
 524             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 525
 526             /* REACTION-FIELD ELECTROSTATICS */
 527             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 528             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 529
 530             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 531
 532             /* Update potential sum for this i atom from the interaction with this j atom. */
 533             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 534             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 535             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 540
 541             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 542
 543             /* Update vectorial force */
 544             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 545             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 546             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 547
 548             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 549             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 550             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 551
 552             }
 553
 554             /**************************
 555              * CALCULATE INTERACTIONS *
 556              **************************/
 557
 558             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 559             {
 560
 561             /* Compute parameters for interactions between i and j atoms */
 562             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq30,rinv30),crf));
 566             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 567
 568             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 572             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 573             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 574
 575             fscal            = felec;
 576
 577             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 578
 579             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 580
 581             /* Update vectorial force */
 582             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 583             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 584             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 585
 586             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 587             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 588             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 589
 590             }
 591
 592             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 593
 594             /* Inner loop uses 164 flops */
 595         }
 596
 597         /* End of innermost loop */
 598
 599         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 600                                               f+i_coord_offset,fshift+i_shift_offset);
 601
 602         ggid                        = gid[iidx];
 603         /* Update potential energies */
 604         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 605         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 606
 607         /* Increment number of inner iterations */
 608         inneriter                  += j_index_end - j_index_start;
 609
 610         /* Outer loop uses 26 flops */
 611     }
 612
 613     /* Increment number of outer iterations */
 614     outeriter        += nri;
 615
 616     /* Update outer/inner flops */
 617
 618     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*164);
 619 }
 620 /*
 621  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sparc64_hpc_ace_double
 622  * Electrostatics interaction: ReactionField
 623  * VdW interaction:            LennardJones
 624  * Geometry:                   Water4-Particle
 625  * Calculate force/pot:        Force
 626  */
 627 void
 628 nb_kernel_ElecRFCut_VdwLJSh_GeomW4P1_F_sparc64_hpc_ace_double
 629                     (t_nblist                    * gmx_restrict       nlist,
 630                      rvec                        * gmx_restrict          xx,
 631                      rvec                        * gmx_restrict          ff,
 632                      t_forcerec                  * gmx_restrict          fr,
 633                      t_mdatoms                   * gmx_restrict     mdatoms,
 634                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 635                      t_nrnb                      * gmx_restrict        nrnb)
 636 {
 637     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 638      * just 0 for non-waters.
 639      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 640      * jnr indices corresponding to data put in the four positions in the SIMD register.
 641      */
 642     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 643     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 644     int              jnrA,jnrB;
 645     int              j_coord_offsetA,j_coord_offsetB;
 646     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 647     real             rcutoff_scalar;
 648     real             *shiftvec,*fshift,*x,*f;
 649     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 650     int              vdwioffset0;
 651     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 652     int              vdwioffset1;
 653     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 654     int              vdwioffset2;
 655     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 656     int              vdwioffset3;
 657     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 658     int              vdwjidx0A,vdwjidx0B;
 659     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 660     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 661     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 662     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 663     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 664     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 665     real             *charge;
 666     int              nvdwtype;
 667     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 668     int              *vdwtype;
 669     real             *vdwparam;
 670     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 671     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 672     _fjsp_v2r8       itab_tmp;
 673     _fjsp_v2r8       dummy_mask,cutoff_mask;
 674     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 675     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 676     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 677
 678     x                = xx[0];
 679     f                = ff[0];
 680
 681     nri              = nlist->nri;
 682     iinr             = nlist->iinr;
 683     jindex           = nlist->jindex;
 684     jjnr             = nlist->jjnr;
 685     shiftidx         = nlist->shift;
 686     gid              = nlist->gid;
 687     shiftvec         = fr->shift_vec[0];
 688     fshift           = fr->fshift[0];
 689     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 690     charge           = mdatoms->chargeA;
 691     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 692     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 693     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 694     nvdwtype         = fr->ntype;
 695     vdwparam         = fr->nbfp;
 696     vdwtype          = mdatoms->typeA;
 697
 698     /* Setup water-specific parameters */
 699     inr              = nlist->iinr[0];
 700     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 701     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 702     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 703     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 704
 705     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 706     rcutoff_scalar   = fr->rcoulomb;
 707     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 708     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 709
 710     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 711     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 712
 713     /* Avoid stupid compiler warnings */
 714     jnrA = jnrB = 0;
 715     j_coord_offsetA = 0;
 716     j_coord_offsetB = 0;
 717
 718     outeriter        = 0;
 719     inneriter        = 0;
 720
 721     /* Start outer loop over neighborlists */
 722     for(iidx=0; iidx<nri; iidx++)
 723     {
 724         /* Load shift vector for this list */
 725         i_shift_offset   = DIM*shiftidx[iidx];
 726
 727         /* Load limits for loop over neighbors */
 728         j_index_start    = jindex[iidx];
 729         j_index_end      = jindex[iidx+1];
 730
 731         /* Get outer coordinate index */
 732         inr              = iinr[iidx];
 733         i_coord_offset   = DIM*inr;
 734
 735         /* Load i particle coords and add shift vector */
 736         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 737                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 738
 739         fix0             = _fjsp_setzero_v2r8();
 740         fiy0             = _fjsp_setzero_v2r8();
 741         fiz0             = _fjsp_setzero_v2r8();
 742         fix1             = _fjsp_setzero_v2r8();
 743         fiy1             = _fjsp_setzero_v2r8();
 744         fiz1             = _fjsp_setzero_v2r8();
 745         fix2             = _fjsp_setzero_v2r8();
 746         fiy2             = _fjsp_setzero_v2r8();
 747         fiz2             = _fjsp_setzero_v2r8();
 748         fix3             = _fjsp_setzero_v2r8();
 749         fiy3             = _fjsp_setzero_v2r8();
 750         fiz3             = _fjsp_setzero_v2r8();
 751
 752         /* Start inner kernel loop */
 753         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 754         {
 755
 756             /* Get j neighbor index, and coordinate index */
 757             jnrA             = jjnr[jidx];
 758             jnrB             = jjnr[jidx+1];
 759             j_coord_offsetA  = DIM*jnrA;
 760             j_coord_offsetB  = DIM*jnrB;
 761
 762             /* load j atom coordinates */
 763             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 764                                               &jx0,&jy0,&jz0);
 765
 766             /* Calculate displacement vector */
 767             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 768             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 769             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 770             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 771             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 772             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 773             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 774             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 775             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 776             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 777             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 778             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 779
 780             /* Calculate squared distance and things based on it */
 781             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 782             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 783             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 784             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 785
 786             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 787             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 788             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 789
 790             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 791             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 792             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 793             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 794
 795             /* Load parameters for j particles */
 796             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 797             vdwjidx0A        = 2*vdwtype[jnrA+0];
 798             vdwjidx0B        = 2*vdwtype[jnrB+0];
 799
 800             fjx0             = _fjsp_setzero_v2r8();
 801             fjy0             = _fjsp_setzero_v2r8();
 802             fjz0             = _fjsp_setzero_v2r8();
 803
 804             /**************************
 805              * CALCULATE INTERACTIONS *
 806              **************************/
 807
 808             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 809             {
 810
 811             /* Compute parameters for interactions between i and j atoms */
 812             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 813                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 814
 815             /* LENNARD-JONES DISPERSION/REPULSION */
 816
 817             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 818             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 819
 820             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 821
 822             fscal            = fvdw;
 823
 824             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 825
 826             /* Update vectorial force */
 827             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 828             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 829             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 830
 831             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 832             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 833             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 834
 835             }
 836
 837             /**************************
 838              * CALCULATE INTERACTIONS *
 839              **************************/
 840
 841             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 842             {
 843
 844             /* Compute parameters for interactions between i and j atoms */
 845             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 846
 847             /* REACTION-FIELD ELECTROSTATICS */
 848             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 849
 850             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 851
 852             fscal            = felec;
 853
 854             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 855
 856             /* Update vectorial force */
 857             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 858             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 859             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 860
 861             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 862             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 863             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 864
 865             }
 866
 867             /**************************
 868              * CALCULATE INTERACTIONS *
 869              **************************/
 870
 871             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 872             {
 873
 874             /* Compute parameters for interactions between i and j atoms */
 875             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 876
 877             /* REACTION-FIELD ELECTROSTATICS */
 878             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 879
 880             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 881
 882             fscal            = felec;
 883
 884             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 885
 886             /* Update vectorial force */
 887             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 888             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 889             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 890
 891             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 892             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 893             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 894
 895             }
 896
 897             /**************************
 898              * CALCULATE INTERACTIONS *
 899              **************************/
 900
 901             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
 902             {
 903
 904             /* Compute parameters for interactions between i and j atoms */
 905             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 906
 907             /* REACTION-FIELD ELECTROSTATICS */
 908             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
 909
 910             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
 911
 912             fscal            = felec;
 913
 914             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 915
 916             /* Update vectorial force */
 917             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 918             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 919             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 920
 921             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 922             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 923             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 924
 925             }
 926
 927             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 928
 929             /* Inner loop uses 135 flops */
 930         }
 931
 932         if(jidx<j_index_end)
 933         {
 934
 935             jnrA             = jjnr[jidx];
 936             j_coord_offsetA  = DIM*jnrA;
 937
 938             /* load j atom coordinates */
 939             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 940                                               &jx0,&jy0,&jz0);
 941
 942             /* Calculate displacement vector */
 943             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 944             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 945             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 946             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 947             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 948             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 949             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 950             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 951             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 952             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 953             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 954             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 955
 956             /* Calculate squared distance and things based on it */
 957             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 958             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 959             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 960             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 961
 962             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 963             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 964             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 965
 966             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 967             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 968             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 969             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 970
 971             /* Load parameters for j particles */
 972             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 973             vdwjidx0A        = 2*vdwtype[jnrA+0];
 974
 975             fjx0             = _fjsp_setzero_v2r8();
 976             fjy0             = _fjsp_setzero_v2r8();
 977             fjz0             = _fjsp_setzero_v2r8();
 978
 979             /**************************
 980              * CALCULATE INTERACTIONS *
 981              **************************/
 982
 983             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 984             {
 985
 986             /* Compute parameters for interactions between i and j atoms */
 987             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 988
 989             /* LENNARD-JONES DISPERSION/REPULSION */
 990
 991             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 992             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 993
 994             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 995
 996             fscal            = fvdw;
 997
 998             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 999
1000             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1001
1002             /* Update vectorial force */
1003             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1004             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1005             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1006
1007             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1008             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1009             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1010
1011             }
1012
1013             /**************************
1014              * CALCULATE INTERACTIONS *
1015              **************************/
1016
1017             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1018             {
1019
1020             /* Compute parameters for interactions between i and j atoms */
1021             qq10             = _fjsp_mul_v2r8(iq1,jq0);
1022
1023             /* REACTION-FIELD ELECTROSTATICS */
1024             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1025
1026             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1027
1028             fscal            = felec;
1029
1030             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1031
1032             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1033
1034             /* Update vectorial force */
1035             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1036             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1037             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1038
1039             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1040             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1041             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1042
1043             }
1044
1045             /**************************
1046              * CALCULATE INTERACTIONS *
1047              **************************/
1048
1049             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1050             {
1051
1052             /* Compute parameters for interactions between i and j atoms */
1053             qq20             = _fjsp_mul_v2r8(iq2,jq0);
1054
1055             /* REACTION-FIELD ELECTROSTATICS */
1056             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1057
1058             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1059
1060             fscal            = felec;
1061
1062             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1063
1064             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1065
1066             /* Update vectorial force */
1067             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1068             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1069             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1070
1071             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1072             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1073             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1074
1075             }
1076
1077             /**************************
1078              * CALCULATE INTERACTIONS *
1079              **************************/
1080
1081             if (gmx_fjsp_any_lt_v2r8(rsq30,rcutoff2))
1082             {
1083
1084             /* Compute parameters for interactions between i and j atoms */
1085             qq30             = _fjsp_mul_v2r8(iq3,jq0);
1086
1087             /* REACTION-FIELD ELECTROSTATICS */
1088             felec            = _fjsp_mul_v2r8(qq30,_fjsp_msub_v2r8(rinv30,rinvsq30,krf2));
1089
1090             cutoff_mask      = _fjsp_cmplt_v2r8(rsq30,rcutoff2);
1091
1092             fscal            = felec;
1093
1094             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1095
1096             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1097
1098             /* Update vectorial force */
1099             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
1100             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
1101             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
1102
1103             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
1104             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
1105             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
1106
1107             }
1108
1109             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1110
1111             /* Inner loop uses 135 flops */
1112         }
1113
1114         /* End of innermost loop */
1115
1116         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1117                                               f+i_coord_offset,fshift+i_shift_offset);
1118
1119         /* Increment number of inner iterations */
1120         inneriter                  += j_index_end - j_index_start;
1121
1122         /* Outer loop uses 24 flops */
1123     }
1124
1125     /* Increment number of outer iterations */
1126     outeriter        += nri;
1127
1128     /* Update outer/inner flops */
1129
1130     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*135);
1131 }