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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "gromacs/legacyheaders/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_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_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 417                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 418
 419             /* EWALD ELECTROSTATICS */
 420
 421             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 422             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 423             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 424             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 425             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 426
 427             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 428             ewtabD           = _fjsp_setzero_v2r8();
 429             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 430             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 431             ewtabFn          = _fjsp_setzero_v2r8();
 432             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 433             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 434             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 435             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 436             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 437
 438             /* LENNARD-JONES DISPERSION/REPULSION */
 439
 440             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 441             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 442             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 443             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 444             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 448             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 449             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 450             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 451
 452             fscal            = _fjsp_add_v2r8(felec,fvdw);
 453
 454             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 455
 456             /* Update vectorial force */
 457             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 458             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 459             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 460
 461             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 462             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 463             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 470
 471             /* Compute parameters for interactions between i and j atoms */
 472             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 473
 474             /* EWALD ELECTROSTATICS */
 475
 476             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 477             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 478             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 479             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 480             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 481
 482             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 483             ewtabD           = _fjsp_setzero_v2r8();
 484             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 485             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 486             ewtabFn          = _fjsp_setzero_v2r8();
 487             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 488             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 489             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 490             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 491             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 492
 493             /* Update potential sum for this i atom from the interaction with this j atom. */
 494             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 495             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 496
 497             fscal            = felec;
 498
 499             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 500
 501             /* Update vectorial force */
 502             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 503             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 504             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 505
 506             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 507             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 508             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 509
 510             /**************************
 511              * CALCULATE INTERACTIONS *
 512              **************************/
 513
 514             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 515
 516             /* Compute parameters for interactions between i and j atoms */
 517             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 518
 519             /* EWALD ELECTROSTATICS */
 520
 521             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 522             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 523             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 524             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 525             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 526
 527             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 528             ewtabD           = _fjsp_setzero_v2r8();
 529             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 530             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 531             ewtabFn          = _fjsp_setzero_v2r8();
 532             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 533             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 534             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 535             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 536             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 540             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 541
 542             fscal            = felec;
 543
 544             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 545
 546             /* Update vectorial force */
 547             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 548             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 549             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 550
 551             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 552             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 553             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 554
 555             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 556
 557             /* Inner loop uses 147 flops */
 558         }
 559
 560         /* End of innermost loop */
 561
 562         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 563                                               f+i_coord_offset,fshift+i_shift_offset);
 564
 565         ggid                        = gid[iidx];
 566         /* Update potential energies */
 567         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 568         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 569
 570         /* Increment number of inner iterations */
 571         inneriter                  += j_index_end - j_index_start;
 572
 573         /* Outer loop uses 20 flops */
 574     }
 575
 576     /* Increment number of outer iterations */
 577     outeriter        += nri;
 578
 579     /* Update outer/inner flops */
 580
 581     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*147);
 582 }
 583 /*
 584  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 585  * Electrostatics interaction: Ewald
 586  * VdW interaction:            LennardJones
 587  * Geometry:                   Water3-Particle
 588  * Calculate force/pot:        Force
 589  */
 590 void
 591 nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 592                     (t_nblist                    * gmx_restrict       nlist,
 593                      rvec                        * gmx_restrict          xx,
 594                      rvec                        * gmx_restrict          ff,
 595                      t_forcerec                  * gmx_restrict          fr,
 596                      t_mdatoms                   * gmx_restrict     mdatoms,
 597                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 598                      t_nrnb                      * gmx_restrict        nrnb)
 599 {
 600     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 601      * just 0 for non-waters.
 602      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 603      * jnr indices corresponding to data put in the four positions in the SIMD register.
 604      */
 605     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 606     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 607     int              jnrA,jnrB;
 608     int              j_coord_offsetA,j_coord_offsetB;
 609     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 610     real             rcutoff_scalar;
 611     real             *shiftvec,*fshift,*x,*f;
 612     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 613     int              vdwioffset0;
 614     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 615     int              vdwioffset1;
 616     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 617     int              vdwioffset2;
 618     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 619     int              vdwjidx0A,vdwjidx0B;
 620     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 621     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 622     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 623     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 624     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 625     real             *charge;
 626     int              nvdwtype;
 627     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 628     int              *vdwtype;
 629     real             *vdwparam;
 630     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 631     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 632     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 633     real             *ewtab;
 634     _fjsp_v2r8       itab_tmp;
 635     _fjsp_v2r8       dummy_mask,cutoff_mask;
 636     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 637     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 638     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 639
 640     x                = xx[0];
 641     f                = ff[0];
 642
 643     nri              = nlist->nri;
 644     iinr             = nlist->iinr;
 645     jindex           = nlist->jindex;
 646     jjnr             = nlist->jjnr;
 647     shiftidx         = nlist->shift;
 648     gid              = nlist->gid;
 649     shiftvec         = fr->shift_vec[0];
 650     fshift           = fr->fshift[0];
 651     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 652     charge           = mdatoms->chargeA;
 653     nvdwtype         = fr->ntype;
 654     vdwparam         = fr->nbfp;
 655     vdwtype          = mdatoms->typeA;
 656
 657     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 658     ewtab            = fr->ic->tabq_coul_F;
 659     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 660     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 661
 662     /* Setup water-specific parameters */
 663     inr              = nlist->iinr[0];
 664     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 665     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 666     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 667     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 668
 669     /* Avoid stupid compiler warnings */
 670     jnrA = jnrB = 0;
 671     j_coord_offsetA = 0;
 672     j_coord_offsetB = 0;
 673
 674     outeriter        = 0;
 675     inneriter        = 0;
 676
 677     /* Start outer loop over neighborlists */
 678     for(iidx=0; iidx<nri; iidx++)
 679     {
 680         /* Load shift vector for this list */
 681         i_shift_offset   = DIM*shiftidx[iidx];
 682
 683         /* Load limits for loop over neighbors */
 684         j_index_start    = jindex[iidx];
 685         j_index_end      = jindex[iidx+1];
 686
 687         /* Get outer coordinate index */
 688         inr              = iinr[iidx];
 689         i_coord_offset   = DIM*inr;
 690
 691         /* Load i particle coords and add shift vector */
 692         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 693                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 694
 695         fix0             = _fjsp_setzero_v2r8();
 696         fiy0             = _fjsp_setzero_v2r8();
 697         fiz0             = _fjsp_setzero_v2r8();
 698         fix1             = _fjsp_setzero_v2r8();
 699         fiy1             = _fjsp_setzero_v2r8();
 700         fiz1             = _fjsp_setzero_v2r8();
 701         fix2             = _fjsp_setzero_v2r8();
 702         fiy2             = _fjsp_setzero_v2r8();
 703         fiz2             = _fjsp_setzero_v2r8();
 704
 705         /* Start inner kernel loop */
 706         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 707         {
 708
 709             /* Get j neighbor index, and coordinate index */
 710             jnrA             = jjnr[jidx];
 711             jnrB             = jjnr[jidx+1];
 712             j_coord_offsetA  = DIM*jnrA;
 713             j_coord_offsetB  = DIM*jnrB;
 714
 715             /* load j atom coordinates */
 716             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 717                                               &jx0,&jy0,&jz0);
 718
 719             /* Calculate displacement vector */
 720             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 721             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 722             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 723             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 724             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 725             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 726             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 727             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 728             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 729
 730             /* Calculate squared distance and things based on it */
 731             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 732             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 733             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 734
 735             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 736             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 737             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 738
 739             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 740             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 741             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 742
 743             /* Load parameters for j particles */
 744             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 745             vdwjidx0A        = 2*vdwtype[jnrA+0];
 746             vdwjidx0B        = 2*vdwtype[jnrB+0];
 747
 748             fjx0             = _fjsp_setzero_v2r8();
 749             fjy0             = _fjsp_setzero_v2r8();
 750             fjz0             = _fjsp_setzero_v2r8();
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 757
 758             /* Compute parameters for interactions between i and j atoms */
 759             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 760             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 761                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 762
 763             /* EWALD ELECTROSTATICS */
 764
 765             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 766             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 767             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 768             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 769             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 770
 771             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 772                                          &ewtabF,&ewtabFn);
 773             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 774             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 775
 776             /* LENNARD-JONES DISPERSION/REPULSION */
 777
 778             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 779             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 780
 781             fscal            = _fjsp_add_v2r8(felec,fvdw);
 782
 783             /* Update vectorial force */
 784             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 785             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 786             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 787
 788             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 789             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 790             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 791
 792             /**************************
 793              * CALCULATE INTERACTIONS *
 794              **************************/
 795
 796             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 797
 798             /* Compute parameters for interactions between i and j atoms */
 799             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 800
 801             /* EWALD ELECTROSTATICS */
 802
 803             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 804             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 805             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 806             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 807             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 808
 809             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 810                                          &ewtabF,&ewtabFn);
 811             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 812             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 813
 814             fscal            = felec;
 815
 816             /* Update vectorial force */
 817             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 818             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 819             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 820
 821             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 822             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 823             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 824
 825             /**************************
 826              * CALCULATE INTERACTIONS *
 827              **************************/
 828
 829             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 830
 831             /* Compute parameters for interactions between i and j atoms */
 832             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 833
 834             /* EWALD ELECTROSTATICS */
 835
 836             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 837             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 838             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 839             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 840             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 841
 842             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 843                                          &ewtabF,&ewtabFn);
 844             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 845             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 846
 847             fscal            = felec;
 848
 849             /* Update vectorial force */
 850             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 851             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 852             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 853
 854             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 855             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 856             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 857
 858             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 859
 860             /* Inner loop uses 127 flops */
 861         }
 862
 863         if(jidx<j_index_end)
 864         {
 865
 866             jnrA             = jjnr[jidx];
 867             j_coord_offsetA  = DIM*jnrA;
 868
 869             /* load j atom coordinates */
 870             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 871                                               &jx0,&jy0,&jz0);
 872
 873             /* Calculate displacement vector */
 874             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 875             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 876             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 877             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 878             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 879             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 880             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 881             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 882             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 883
 884             /* Calculate squared distance and things based on it */
 885             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 886             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 887             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 888
 889             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 890             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 891             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 892
 893             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 894             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 895             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 896
 897             /* Load parameters for j particles */
 898             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 899             vdwjidx0A        = 2*vdwtype[jnrA+0];
 900
 901             fjx0             = _fjsp_setzero_v2r8();
 902             fjy0             = _fjsp_setzero_v2r8();
 903             fjz0             = _fjsp_setzero_v2r8();
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 910
 911             /* Compute parameters for interactions between i and j atoms */
 912             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 913             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 914                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 915
 916             /* EWALD ELECTROSTATICS */
 917
 918             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 919             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 920             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 921             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 922             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 923
 924             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 925             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 926             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 927
 928             /* LENNARD-JONES DISPERSION/REPULSION */
 929
 930             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 931             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 932
 933             fscal            = _fjsp_add_v2r8(felec,fvdw);
 934
 935             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 936
 937             /* Update vectorial force */
 938             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 939             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 940             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 941
 942             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 943             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 944             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 945
 946             /**************************
 947              * CALCULATE INTERACTIONS *
 948              **************************/
 949
 950             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 951
 952             /* Compute parameters for interactions between i and j atoms */
 953             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 954
 955             /* EWALD ELECTROSTATICS */
 956
 957             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 958             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 959             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 960             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 961             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 962
 963             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 964             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 965             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 966
 967             fscal            = felec;
 968
 969             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 970
 971             /* Update vectorial force */
 972             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 973             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 974             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 975
 976             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 977             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 978             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 979
 980             /**************************
 981              * CALCULATE INTERACTIONS *
 982              **************************/
 983
 984             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 985
 986             /* Compute parameters for interactions between i and j atoms */
 987             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 988
 989             /* EWALD ELECTROSTATICS */
 990
 991             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 992             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 993             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 994             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 995             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 996
 997             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 998             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 999             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1000
1001             fscal            = felec;
1002
1003             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1004
1005             /* Update vectorial force */
1006             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1007             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1008             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1009
1010             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1011             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1012             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1013
1014             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1015
1016             /* Inner loop uses 127 flops */
1017         }
1018
1019         /* End of innermost loop */
1020
1021         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1022                                               f+i_coord_offset,fshift+i_shift_offset);
1023
1024         /* Increment number of inner iterations */
1025         inneriter                  += j_index_end - j_index_start;
1026
1027         /* Outer loop uses 18 flops */
1028     }
1029
1030     /* Increment number of outer iterations */
1031     outeriter        += nri;
1032
1033     /* Update outer/inner flops */
1034
1035     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*127);
1036 }