src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEw_VdwLJ_GeomW3P1_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_ElecEw_VdwLJ_GeomW3P1_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: Ewald
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water3-Particle
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecEw_VdwLJ_GeomW3P1_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  91     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  92     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
  99     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 100     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 101     real             *ewtab;
 102     _fjsp_v2r8       itab_tmp;
 103     _fjsp_v2r8       dummy_mask,cutoff_mask;
 104     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 105     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 106     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 107
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 126     ewtab            = fr->ic->tabq_coul_FDV0;
 127     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 128     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 129
 130     /* Setup water-specific parameters */
 131     inr              = nlist->iinr[0];
 132     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 133     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 134     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 135     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 136
 137     /* Avoid stupid compiler warnings */
 138     jnrA = jnrB = 0;
 139     j_coord_offsetA = 0;
 140     j_coord_offsetB = 0;
 141
 142     outeriter        = 0;
 143     inneriter        = 0;
 144
 145     /* Start outer loop over neighborlists */
 146     for(iidx=0; iidx<nri; iidx++)
 147     {
 148         /* Load shift vector for this list */
 149         i_shift_offset   = DIM*shiftidx[iidx];
 150
 151         /* Load limits for loop over neighbors */
 152         j_index_start    = jindex[iidx];
 153         j_index_end      = jindex[iidx+1];
 154
 155         /* Get outer coordinate index */
 156         inr              = iinr[iidx];
 157         i_coord_offset   = DIM*inr;
 158
 159         /* Load i particle coords and add shift vector */
 160         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 161                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 162
 163         fix0             = _fjsp_setzero_v2r8();
 164         fiy0             = _fjsp_setzero_v2r8();
 165         fiz0             = _fjsp_setzero_v2r8();
 166         fix1             = _fjsp_setzero_v2r8();
 167         fiy1             = _fjsp_setzero_v2r8();
 168         fiz1             = _fjsp_setzero_v2r8();
 169         fix2             = _fjsp_setzero_v2r8();
 170         fiy2             = _fjsp_setzero_v2r8();
 171         fiz2             = _fjsp_setzero_v2r8();
 172
 173         /* Reset potential sums */
 174         velecsum         = _fjsp_setzero_v2r8();
 175         vvdwsum          = _fjsp_setzero_v2r8();
 176
 177         /* Start inner kernel loop */
 178         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 179         {
 180
 181             /* Get j neighbor index, and coordinate index */
 182             jnrA             = jjnr[jidx];
 183             jnrB             = jjnr[jidx+1];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186
 187             /* load j atom coordinates */
 188             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                               &jx0,&jy0,&jz0);
 190
 191             /* Calculate displacement vector */
 192             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 193             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 194             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 195             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 196             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 197             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 198             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 199             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 200             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 204             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 205             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 206
 207             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 208             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 209             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 210
 211             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 212             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 213             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 214
 215             /* Load parameters for j particles */
 216             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 217             vdwjidx0A        = 2*vdwtype[jnrA+0];
 218             vdwjidx0B        = 2*vdwtype[jnrB+0];
 219
 220             fjx0             = _fjsp_setzero_v2r8();
 221             fjy0             = _fjsp_setzero_v2r8();
 222             fjz0             = _fjsp_setzero_v2r8();
 223
 224             /**************************
 225              * CALCULATE INTERACTIONS *
 226              **************************/
 227
 228             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 229
 230             /* Compute parameters for interactions between i and j atoms */
 231             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 232             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 233                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 234
 235             /* EWALD ELECTROSTATICS */
 236
 237             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 238             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 239             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 240             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 241             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 242
 243             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 244             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 245             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 246             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 247             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 248             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 249             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 250             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 251             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 252             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 253
 254             /* LENNARD-JONES DISPERSION/REPULSION */
 255
 256             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 257             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 258             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 259             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 260             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 261
 262             /* Update potential sum for this i atom from the interaction with this j atom. */
 263             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 264             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 265
 266             fscal            = _fjsp_add_v2r8(felec,fvdw);
 267
 268             /* Update vectorial force */
 269             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 270             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 271             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 272
 273             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 274             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 275             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 276
 277             /**************************
 278              * CALCULATE INTERACTIONS *
 279              **************************/
 280
 281             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 282
 283             /* Compute parameters for interactions between i and j atoms */
 284             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 285
 286             /* EWALD ELECTROSTATICS */
 287
 288             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 289             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 290             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 291             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 292             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 293
 294             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 295             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 296             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 297             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 298             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 299             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 300             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 301             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 302             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 303             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 307
 308             fscal            = felec;
 309
 310             /* Update vectorial force */
 311             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 312             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 313             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 314
 315             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 316             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 317             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 324
 325             /* Compute parameters for interactions between i and j atoms */
 326             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 327
 328             /* EWALD ELECTROSTATICS */
 329
 330             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 331             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 332             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 333             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 334             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 335
 336             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 337             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 338             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 339             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 340             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 341             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 342             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 343             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 344             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 345             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 349
 350             fscal            = felec;
 351
 352             /* Update vectorial force */
 353             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 354             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 355             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 356
 357             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 358             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 359             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 360
 361             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 362
 363             /* Inner loop uses 147 flops */
 364         }
 365
 366         if(jidx<j_index_end)
 367         {
 368
 369             jnrA             = jjnr[jidx];
 370             j_coord_offsetA  = DIM*jnrA;
 371
 372             /* load j atom coordinates */
 373             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 374                                               &jx0,&jy0,&jz0);
 375
 376             /* Calculate displacement vector */
 377             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 378             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 379             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 380             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 381             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 382             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 383             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 384             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 385             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 386
 387             /* Calculate squared distance and things based on it */
 388             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 389             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 390             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 391
 392             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 393             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 394             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 395
 396             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 397             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 398             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 399
 400             /* Load parameters for j particles */
 401             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 402             vdwjidx0A        = 2*vdwtype[jnrA+0];
 403
 404             fjx0             = _fjsp_setzero_v2r8();
 405             fjy0             = _fjsp_setzero_v2r8();
 406             fjz0             = _fjsp_setzero_v2r8();
 407
 408             /**************************
 409              * CALCULATE INTERACTIONS *
 410              **************************/
 411
 412             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 413
 414             /* Compute parameters for interactions between i and j atoms */
 415             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 416             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 417
 418             /* EWALD ELECTROSTATICS */
 419
 420             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 421             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 422             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 423             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 424             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 425
 426             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 427             ewtabD           = _fjsp_setzero_v2r8();
 428             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 429             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 430             ewtabFn          = _fjsp_setzero_v2r8();
 431             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 432             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 433             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 434             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 435             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 436
 437             /* LENNARD-JONES DISPERSION/REPULSION */
 438
 439             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 440             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 441             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 442             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 443             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 447             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 448             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 449             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 450
 451             fscal            = _fjsp_add_v2r8(felec,fvdw);
 452
 453             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 454
 455             /* Update vectorial force */
 456             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 457             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 458             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 459
 460             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 461             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 462             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 469
 470             /* Compute parameters for interactions between i and j atoms */
 471             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 472
 473             /* EWALD ELECTROSTATICS */
 474
 475             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 476             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 477             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 478             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 479             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 480
 481             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 482             ewtabD           = _fjsp_setzero_v2r8();
 483             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 484             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 485             ewtabFn          = _fjsp_setzero_v2r8();
 486             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 487             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 488             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 489             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 490             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 491
 492             /* Update potential sum for this i atom from the interaction with this j atom. */
 493             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 494             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 495
 496             fscal            = felec;
 497
 498             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 499
 500             /* Update vectorial force */
 501             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 502             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 503             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 504
 505             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 506             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 507             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 508
 509             /**************************
 510              * CALCULATE INTERACTIONS *
 511              **************************/
 512
 513             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 514
 515             /* Compute parameters for interactions between i and j atoms */
 516             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 517
 518             /* EWALD ELECTROSTATICS */
 519
 520             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 521             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 522             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 523             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 524             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 525
 526             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 527             ewtabD           = _fjsp_setzero_v2r8();
 528             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 529             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 530             ewtabFn          = _fjsp_setzero_v2r8();
 531             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 532             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 533             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 534             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 535             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 536
 537             /* Update potential sum for this i atom from the interaction with this j atom. */
 538             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 539             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 540
 541             fscal            = felec;
 542
 543             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 544
 545             /* Update vectorial force */
 546             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 547             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 548             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 549
 550             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 551             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 552             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 553
 554             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 555
 556             /* Inner loop uses 147 flops */
 557         }
 558
 559         /* End of innermost loop */
 560
 561         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 562                                               f+i_coord_offset,fshift+i_shift_offset);
 563
 564         ggid                        = gid[iidx];
 565         /* Update potential energies */
 566         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 567         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 568
 569         /* Increment number of inner iterations */
 570         inneriter                  += j_index_end - j_index_start;
 571
 572         /* Outer loop uses 20 flops */
 573     }
 574
 575     /* Increment number of outer iterations */
 576     outeriter        += nri;
 577
 578     /* Update outer/inner flops */
 579
 580     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*147);
 581 }
 582 /*
 583  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 584  * Electrostatics interaction: Ewald
 585  * VdW interaction:            LennardJones
 586  * Geometry:                   Water3-Particle
 587  * Calculate force/pot:        Force
 588  */
 589 void
 590 nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 591                     (t_nblist                    * gmx_restrict       nlist,
 592                      rvec                        * gmx_restrict          xx,
 593                      rvec                        * gmx_restrict          ff,
 594                      t_forcerec                  * gmx_restrict          fr,
 595                      t_mdatoms                   * gmx_restrict     mdatoms,
 596                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 597                      t_nrnb                      * gmx_restrict        nrnb)
 598 {
 599     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 600      * just 0 for non-waters.
 601      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 602      * jnr indices corresponding to data put in the four positions in the SIMD register.
 603      */
 604     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 605     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 606     int              jnrA,jnrB;
 607     int              j_coord_offsetA,j_coord_offsetB;
 608     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 609     real             rcutoff_scalar;
 610     real             *shiftvec,*fshift,*x,*f;
 611     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 612     int              vdwioffset0;
 613     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 614     int              vdwioffset1;
 615     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 616     int              vdwioffset2;
 617     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 618     int              vdwjidx0A,vdwjidx0B;
 619     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 620     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 621     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 622     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 623     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 624     real             *charge;
 625     int              nvdwtype;
 626     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 627     int              *vdwtype;
 628     real             *vdwparam;
 629     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 630     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 631     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 632     real             *ewtab;
 633     _fjsp_v2r8       itab_tmp;
 634     _fjsp_v2r8       dummy_mask,cutoff_mask;
 635     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 636     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 637     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 638
 639     x                = xx[0];
 640     f                = ff[0];
 641
 642     nri              = nlist->nri;
 643     iinr             = nlist->iinr;
 644     jindex           = nlist->jindex;
 645     jjnr             = nlist->jjnr;
 646     shiftidx         = nlist->shift;
 647     gid              = nlist->gid;
 648     shiftvec         = fr->shift_vec[0];
 649     fshift           = fr->fshift[0];
 650     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 651     charge           = mdatoms->chargeA;
 652     nvdwtype         = fr->ntype;
 653     vdwparam         = fr->nbfp;
 654     vdwtype          = mdatoms->typeA;
 655
 656     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 657     ewtab            = fr->ic->tabq_coul_F;
 658     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 659     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 660
 661     /* Setup water-specific parameters */
 662     inr              = nlist->iinr[0];
 663     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 664     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 665     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 666     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 667
 668     /* Avoid stupid compiler warnings */
 669     jnrA = jnrB = 0;
 670     j_coord_offsetA = 0;
 671     j_coord_offsetB = 0;
 672
 673     outeriter        = 0;
 674     inneriter        = 0;
 675
 676     /* Start outer loop over neighborlists */
 677     for(iidx=0; iidx<nri; iidx++)
 678     {
 679         /* Load shift vector for this list */
 680         i_shift_offset   = DIM*shiftidx[iidx];
 681
 682         /* Load limits for loop over neighbors */
 683         j_index_start    = jindex[iidx];
 684         j_index_end      = jindex[iidx+1];
 685
 686         /* Get outer coordinate index */
 687         inr              = iinr[iidx];
 688         i_coord_offset   = DIM*inr;
 689
 690         /* Load i particle coords and add shift vector */
 691         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 692                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 693
 694         fix0             = _fjsp_setzero_v2r8();
 695         fiy0             = _fjsp_setzero_v2r8();
 696         fiz0             = _fjsp_setzero_v2r8();
 697         fix1             = _fjsp_setzero_v2r8();
 698         fiy1             = _fjsp_setzero_v2r8();
 699         fiz1             = _fjsp_setzero_v2r8();
 700         fix2             = _fjsp_setzero_v2r8();
 701         fiy2             = _fjsp_setzero_v2r8();
 702         fiz2             = _fjsp_setzero_v2r8();
 703
 704         /* Start inner kernel loop */
 705         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 706         {
 707
 708             /* Get j neighbor index, and coordinate index */
 709             jnrA             = jjnr[jidx];
 710             jnrB             = jjnr[jidx+1];
 711             j_coord_offsetA  = DIM*jnrA;
 712             j_coord_offsetB  = DIM*jnrB;
 713
 714             /* load j atom coordinates */
 715             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 716                                               &jx0,&jy0,&jz0);
 717
 718             /* Calculate displacement vector */
 719             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 720             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 721             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 722             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 723             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 724             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 725             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 726             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 727             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 728
 729             /* Calculate squared distance and things based on it */
 730             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 731             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 732             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 733
 734             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 735             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 736             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 737
 738             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 739             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 740             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 741
 742             /* Load parameters for j particles */
 743             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 744             vdwjidx0A        = 2*vdwtype[jnrA+0];
 745             vdwjidx0B        = 2*vdwtype[jnrB+0];
 746
 747             fjx0             = _fjsp_setzero_v2r8();
 748             fjy0             = _fjsp_setzero_v2r8();
 749             fjz0             = _fjsp_setzero_v2r8();
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 756
 757             /* Compute parameters for interactions between i and j atoms */
 758             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 759             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 760                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 761
 762             /* EWALD ELECTROSTATICS */
 763
 764             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 765             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 766             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 767             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 768             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 769
 770             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 771                                          &ewtabF,&ewtabFn);
 772             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 773             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 774
 775             /* LENNARD-JONES DISPERSION/REPULSION */
 776
 777             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 778             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 779
 780             fscal            = _fjsp_add_v2r8(felec,fvdw);
 781
 782             /* Update vectorial force */
 783             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 784             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 785             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 786
 787             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 788             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 789             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 790
 791             /**************************
 792              * CALCULATE INTERACTIONS *
 793              **************************/
 794
 795             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 796
 797             /* Compute parameters for interactions between i and j atoms */
 798             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 799
 800             /* EWALD ELECTROSTATICS */
 801
 802             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 803             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 804             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 805             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 806             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 807
 808             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 809                                          &ewtabF,&ewtabFn);
 810             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 811             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 812
 813             fscal            = felec;
 814
 815             /* Update vectorial force */
 816             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 817             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 818             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 819
 820             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 821             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 822             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 829
 830             /* Compute parameters for interactions between i and j atoms */
 831             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 832
 833             /* EWALD ELECTROSTATICS */
 834
 835             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 836             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 837             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 838             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 839             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 840
 841             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 842                                          &ewtabF,&ewtabFn);
 843             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 844             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 845
 846             fscal            = felec;
 847
 848             /* Update vectorial force */
 849             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 850             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 851             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 852
 853             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 854             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 855             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 856
 857             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 858
 859             /* Inner loop uses 127 flops */
 860         }
 861
 862         if(jidx<j_index_end)
 863         {
 864
 865             jnrA             = jjnr[jidx];
 866             j_coord_offsetA  = DIM*jnrA;
 867
 868             /* load j atom coordinates */
 869             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 870                                               &jx0,&jy0,&jz0);
 871
 872             /* Calculate displacement vector */
 873             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 874             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 875             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 876             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 877             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 878             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 879             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 880             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 881             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 882
 883             /* Calculate squared distance and things based on it */
 884             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 885             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 886             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 887
 888             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 889             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 890             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 891
 892             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 893             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 894             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 895
 896             /* Load parameters for j particles */
 897             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 898             vdwjidx0A        = 2*vdwtype[jnrA+0];
 899
 900             fjx0             = _fjsp_setzero_v2r8();
 901             fjy0             = _fjsp_setzero_v2r8();
 902             fjz0             = _fjsp_setzero_v2r8();
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 909
 910             /* Compute parameters for interactions between i and j atoms */
 911             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 912             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 913
 914             /* EWALD ELECTROSTATICS */
 915
 916             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 917             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 918             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 919             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 920             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 921
 922             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 923             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 924             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 925
 926             /* LENNARD-JONES DISPERSION/REPULSION */
 927
 928             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 929             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 930
 931             fscal            = _fjsp_add_v2r8(felec,fvdw);
 932
 933             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 934
 935             /* Update vectorial force */
 936             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 937             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 938             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 939
 940             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 941             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 942             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 943
 944             /**************************
 945              * CALCULATE INTERACTIONS *
 946              **************************/
 947
 948             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 949
 950             /* Compute parameters for interactions between i and j atoms */
 951             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 952
 953             /* EWALD ELECTROSTATICS */
 954
 955             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 956             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 957             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 958             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 959             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 960
 961             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 962             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 963             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 964
 965             fscal            = felec;
 966
 967             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 968
 969             /* Update vectorial force */
 970             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 971             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 972             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 973
 974             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 975             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 976             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 977
 978             /**************************
 979              * CALCULATE INTERACTIONS *
 980              **************************/
 981
 982             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 983
 984             /* Compute parameters for interactions between i and j atoms */
 985             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 986
 987             /* EWALD ELECTROSTATICS */
 988
 989             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 990             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 991             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 992             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 993             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 994
 995             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 996             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 997             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 998
 999             fscal            = felec;
1000
1001             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1002
1003             /* Update vectorial force */
1004             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1005             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1006             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1007
1008             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1009             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1010             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1011
1012             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1013
1014             /* Inner loop uses 127 flops */
1015         }
1016
1017         /* End of innermost loop */
1018
1019         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1020                                               f+i_coord_offset,fshift+i_shift_offset);
1021
1022         /* Increment number of inner iterations */
1023         inneriter                  += j_index_end - j_index_start;
1024
1025         /* Outer loop uses 18 flops */
1026     }
1027
1028     /* Increment number of outer iterations */
1029     outeriter        += nri;
1030
1031     /* Update outer/inner flops */
1032
1033     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*127);
1034 }