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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3P1_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJ_GeomW3P1_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwjidx0A,vdwjidx0B;
  86     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  89     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  90     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
  97     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
  98     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  99     real             *ewtab;
 100     _fjsp_v2r8       itab_tmp;
 101     _fjsp_v2r8       dummy_mask,cutoff_mask;
 102     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 103     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 104     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 105
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 124     ewtab            = fr->ic->tabq_coul_FDV0;
 125     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 126     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 127
 128     /* Setup water-specific parameters */
 129     inr              = nlist->iinr[0];
 130     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 131     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 132     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 133     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 134
 135     /* Avoid stupid compiler warnings */
 136     jnrA = jnrB = 0;
 137     j_coord_offsetA = 0;
 138     j_coord_offsetB = 0;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     /* Start outer loop over neighborlists */
 144     for(iidx=0; iidx<nri; iidx++)
 145     {
 146         /* Load shift vector for this list */
 147         i_shift_offset   = DIM*shiftidx[iidx];
 148
 149         /* Load limits for loop over neighbors */
 150         j_index_start    = jindex[iidx];
 151         j_index_end      = jindex[iidx+1];
 152
 153         /* Get outer coordinate index */
 154         inr              = iinr[iidx];
 155         i_coord_offset   = DIM*inr;
 156
 157         /* Load i particle coords and add shift vector */
 158         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 159                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 160
 161         fix0             = _fjsp_setzero_v2r8();
 162         fiy0             = _fjsp_setzero_v2r8();
 163         fiz0             = _fjsp_setzero_v2r8();
 164         fix1             = _fjsp_setzero_v2r8();
 165         fiy1             = _fjsp_setzero_v2r8();
 166         fiz1             = _fjsp_setzero_v2r8();
 167         fix2             = _fjsp_setzero_v2r8();
 168         fiy2             = _fjsp_setzero_v2r8();
 169         fiz2             = _fjsp_setzero_v2r8();
 170
 171         /* Reset potential sums */
 172         velecsum         = _fjsp_setzero_v2r8();
 173         vvdwsum          = _fjsp_setzero_v2r8();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             j_coord_offsetA  = DIM*jnrA;
 183             j_coord_offsetB  = DIM*jnrB;
 184
 185             /* load j atom coordinates */
 186             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 187                                               &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 191             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 192             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 193             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 194             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 195             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 196             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 197             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 198             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 199
 200             /* Calculate squared distance and things based on it */
 201             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 202             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 203             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 204
 205             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 206             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 207             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 208
 209             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 210             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 211             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 212
 213             /* Load parameters for j particles */
 214             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 215             vdwjidx0A        = 2*vdwtype[jnrA+0];
 216             vdwjidx0B        = 2*vdwtype[jnrB+0];
 217
 218             fjx0             = _fjsp_setzero_v2r8();
 219             fjy0             = _fjsp_setzero_v2r8();
 220             fjz0             = _fjsp_setzero_v2r8();
 221
 222             /**************************
 223              * CALCULATE INTERACTIONS *
 224              **************************/
 225
 226             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 227
 228             /* Compute parameters for interactions between i and j atoms */
 229             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 230             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 231                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 232
 233             /* EWALD ELECTROSTATICS */
 234
 235             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 236             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 237             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 238             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 239             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 240
 241             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 242             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 243             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 244             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 245             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 246             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 247             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 248             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 249             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 250             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 251
 252             /* LENNARD-JONES DISPERSION/REPULSION */
 253
 254             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 255             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 256             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 257             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 258             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 259
 260             /* Update potential sum for this i atom from the interaction with this j atom. */
 261             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 262             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 263
 264             fscal            = _fjsp_add_v2r8(felec,fvdw);
 265
 266             /* Update vectorial force */
 267             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 268             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 269             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 270
 271             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 272             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 273             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 274
 275             /**************************
 276              * CALCULATE INTERACTIONS *
 277              **************************/
 278
 279             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 280
 281             /* Compute parameters for interactions between i and j atoms */
 282             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 283
 284             /* EWALD ELECTROSTATICS */
 285
 286             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 287             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 288             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 289             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 290             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 291
 292             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 293             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 294             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 295             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 296             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 297             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 298             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 299             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 300             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 301             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 302
 303             /* Update potential sum for this i atom from the interaction with this j atom. */
 304             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 305
 306             fscal            = felec;
 307
 308             /* Update vectorial force */
 309             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 310             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 311             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 312
 313             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 314             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 315             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 316
 317             /**************************
 318              * CALCULATE INTERACTIONS *
 319              **************************/
 320
 321             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 322
 323             /* Compute parameters for interactions between i and j atoms */
 324             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 325
 326             /* EWALD ELECTROSTATICS */
 327
 328             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 329             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 330             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 331             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 332             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 333
 334             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 335             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 336             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 337             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 338             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 339             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 340             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 341             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 342             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 343             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 344
 345             /* Update potential sum for this i atom from the interaction with this j atom. */
 346             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             /* Update vectorial force */
 351             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 352             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 353             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 354
 355             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 356             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 357             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 358
 359             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 360
 361             /* Inner loop uses 147 flops */
 362         }
 363
 364         if(jidx<j_index_end)
 365         {
 366
 367             jnrA             = jjnr[jidx];
 368             j_coord_offsetA  = DIM*jnrA;
 369
 370             /* load j atom coordinates */
 371             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 372                                               &jx0,&jy0,&jz0);
 373
 374             /* Calculate displacement vector */
 375             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 376             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 377             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 378             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 379             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 380             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 381             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 382             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 383             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 384
 385             /* Calculate squared distance and things based on it */
 386             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 387             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 388             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 389
 390             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 391             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 392             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 393
 394             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 395             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 396             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 397
 398             /* Load parameters for j particles */
 399             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 400             vdwjidx0A        = 2*vdwtype[jnrA+0];
 401
 402             fjx0             = _fjsp_setzero_v2r8();
 403             fjy0             = _fjsp_setzero_v2r8();
 404             fjz0             = _fjsp_setzero_v2r8();
 405
 406             /**************************
 407              * CALCULATE INTERACTIONS *
 408              **************************/
 409
 410             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 411
 412             /* Compute parameters for interactions between i and j atoms */
 413             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 414             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 415
 416             /* EWALD ELECTROSTATICS */
 417
 418             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 419             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 420             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 421             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 422             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 423
 424             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 425             ewtabD           = _fjsp_setzero_v2r8();
 426             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 427             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 428             ewtabFn          = _fjsp_setzero_v2r8();
 429             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 430             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 431             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 432             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(rinv00,velec));
 433             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 434
 435             /* LENNARD-JONES DISPERSION/REPULSION */
 436
 437             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 438             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 439             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 440             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 441             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 445             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 446             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 447             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 448
 449             fscal            = _fjsp_add_v2r8(felec,fvdw);
 450
 451             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 452
 453             /* Update vectorial force */
 454             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 455             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 456             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 457
 458             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 459             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 460             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 467
 468             /* Compute parameters for interactions between i and j atoms */
 469             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 470
 471             /* EWALD ELECTROSTATICS */
 472
 473             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 474             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 475             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 476             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 477             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 478
 479             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 480             ewtabD           = _fjsp_setzero_v2r8();
 481             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 482             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 483             ewtabFn          = _fjsp_setzero_v2r8();
 484             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 485             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 486             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 487             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 488             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 489
 490             /* Update potential sum for this i atom from the interaction with this j atom. */
 491             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 492             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 493
 494             fscal            = felec;
 495
 496             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 497
 498             /* Update vectorial force */
 499             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 500             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 501             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 502
 503             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 504             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 505             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 506
 507             /**************************
 508              * CALCULATE INTERACTIONS *
 509              **************************/
 510
 511             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 512
 513             /* Compute parameters for interactions between i and j atoms */
 514             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 515
 516             /* EWALD ELECTROSTATICS */
 517
 518             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 519             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 520             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 521             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 522             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 523
 524             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 525             ewtabD           = _fjsp_setzero_v2r8();
 526             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 527             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 528             ewtabFn          = _fjsp_setzero_v2r8();
 529             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 530             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 531             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 532             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 533             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 534
 535             /* Update potential sum for this i atom from the interaction with this j atom. */
 536             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 537             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 538
 539             fscal            = felec;
 540
 541             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 542
 543             /* Update vectorial force */
 544             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 545             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 546             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 547
 548             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 549             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 550             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 551
 552             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 553
 554             /* Inner loop uses 147 flops */
 555         }
 556
 557         /* End of innermost loop */
 558
 559         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 560                                               f+i_coord_offset,fshift+i_shift_offset);
 561
 562         ggid                        = gid[iidx];
 563         /* Update potential energies */
 564         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 565         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 566
 567         /* Increment number of inner iterations */
 568         inneriter                  += j_index_end - j_index_start;
 569
 570         /* Outer loop uses 20 flops */
 571     }
 572
 573     /* Increment number of outer iterations */
 574     outeriter        += nri;
 575
 576     /* Update outer/inner flops */
 577
 578     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*147);
 579 }
 580 /*
 581  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 582  * Electrostatics interaction: Ewald
 583  * VdW interaction:            LennardJones
 584  * Geometry:                   Water3-Particle
 585  * Calculate force/pot:        Force
 586  */
 587 void
 588 nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 589                     (t_nblist                    * gmx_restrict       nlist,
 590                      rvec                        * gmx_restrict          xx,
 591                      rvec                        * gmx_restrict          ff,
 592                      t_forcerec                  * gmx_restrict          fr,
 593                      t_mdatoms                   * gmx_restrict     mdatoms,
 594                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 595                      t_nrnb                      * gmx_restrict        nrnb)
 596 {
 597     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 598      * just 0 for non-waters.
 599      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 600      * jnr indices corresponding to data put in the four positions in the SIMD register.
 601      */
 602     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 603     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 604     int              jnrA,jnrB;
 605     int              j_coord_offsetA,j_coord_offsetB;
 606     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 607     real             rcutoff_scalar;
 608     real             *shiftvec,*fshift,*x,*f;
 609     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 610     int              vdwioffset0;
 611     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 612     int              vdwioffset1;
 613     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 614     int              vdwioffset2;
 615     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 616     int              vdwjidx0A,vdwjidx0B;
 617     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 618     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 619     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 620     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 621     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 622     real             *charge;
 623     int              nvdwtype;
 624     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 625     int              *vdwtype;
 626     real             *vdwparam;
 627     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 628     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 629     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 630     real             *ewtab;
 631     _fjsp_v2r8       itab_tmp;
 632     _fjsp_v2r8       dummy_mask,cutoff_mask;
 633     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 634     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 635     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 636
 637     x                = xx[0];
 638     f                = ff[0];
 639
 640     nri              = nlist->nri;
 641     iinr             = nlist->iinr;
 642     jindex           = nlist->jindex;
 643     jjnr             = nlist->jjnr;
 644     shiftidx         = nlist->shift;
 645     gid              = nlist->gid;
 646     shiftvec         = fr->shift_vec[0];
 647     fshift           = fr->fshift[0];
 648     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 649     charge           = mdatoms->chargeA;
 650     nvdwtype         = fr->ntype;
 651     vdwparam         = fr->nbfp;
 652     vdwtype          = mdatoms->typeA;
 653
 654     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 655     ewtab            = fr->ic->tabq_coul_F;
 656     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 657     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 658
 659     /* Setup water-specific parameters */
 660     inr              = nlist->iinr[0];
 661     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 662     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 663     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 664     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 665
 666     /* Avoid stupid compiler warnings */
 667     jnrA = jnrB = 0;
 668     j_coord_offsetA = 0;
 669     j_coord_offsetB = 0;
 670
 671     outeriter        = 0;
 672     inneriter        = 0;
 673
 674     /* Start outer loop over neighborlists */
 675     for(iidx=0; iidx<nri; iidx++)
 676     {
 677         /* Load shift vector for this list */
 678         i_shift_offset   = DIM*shiftidx[iidx];
 679
 680         /* Load limits for loop over neighbors */
 681         j_index_start    = jindex[iidx];
 682         j_index_end      = jindex[iidx+1];
 683
 684         /* Get outer coordinate index */
 685         inr              = iinr[iidx];
 686         i_coord_offset   = DIM*inr;
 687
 688         /* Load i particle coords and add shift vector */
 689         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 690                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 691
 692         fix0             = _fjsp_setzero_v2r8();
 693         fiy0             = _fjsp_setzero_v2r8();
 694         fiz0             = _fjsp_setzero_v2r8();
 695         fix1             = _fjsp_setzero_v2r8();
 696         fiy1             = _fjsp_setzero_v2r8();
 697         fiz1             = _fjsp_setzero_v2r8();
 698         fix2             = _fjsp_setzero_v2r8();
 699         fiy2             = _fjsp_setzero_v2r8();
 700         fiz2             = _fjsp_setzero_v2r8();
 701
 702         /* Start inner kernel loop */
 703         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 704         {
 705
 706             /* Get j neighbor index, and coordinate index */
 707             jnrA             = jjnr[jidx];
 708             jnrB             = jjnr[jidx+1];
 709             j_coord_offsetA  = DIM*jnrA;
 710             j_coord_offsetB  = DIM*jnrB;
 711
 712             /* load j atom coordinates */
 713             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 714                                               &jx0,&jy0,&jz0);
 715
 716             /* Calculate displacement vector */
 717             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 718             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 719             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 720             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 721             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 722             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 723             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 724             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 725             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 726
 727             /* Calculate squared distance and things based on it */
 728             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 729             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 730             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 731
 732             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 733             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 734             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 735
 736             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 737             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 738             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 739
 740             /* Load parameters for j particles */
 741             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 742             vdwjidx0A        = 2*vdwtype[jnrA+0];
 743             vdwjidx0B        = 2*vdwtype[jnrB+0];
 744
 745             fjx0             = _fjsp_setzero_v2r8();
 746             fjy0             = _fjsp_setzero_v2r8();
 747             fjz0             = _fjsp_setzero_v2r8();
 748
 749             /**************************
 750              * CALCULATE INTERACTIONS *
 751              **************************/
 752
 753             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 754
 755             /* Compute parameters for interactions between i and j atoms */
 756             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 757             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 758                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 759
 760             /* EWALD ELECTROSTATICS */
 761
 762             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 763             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 764             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 765             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 766             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 767
 768             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 769                                          &ewtabF,&ewtabFn);
 770             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 771             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 772
 773             /* LENNARD-JONES DISPERSION/REPULSION */
 774
 775             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 776             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 777
 778             fscal            = _fjsp_add_v2r8(felec,fvdw);
 779
 780             /* Update vectorial force */
 781             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 782             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 783             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 784
 785             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 786             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 787             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 794
 795             /* Compute parameters for interactions between i and j atoms */
 796             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 797
 798             /* EWALD ELECTROSTATICS */
 799
 800             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 801             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 802             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 803             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 804             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 805
 806             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 807                                          &ewtabF,&ewtabFn);
 808             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 809             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 810
 811             fscal            = felec;
 812
 813             /* Update vectorial force */
 814             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 815             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 816             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 817
 818             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 819             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 820             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 821
 822             /**************************
 823              * CALCULATE INTERACTIONS *
 824              **************************/
 825
 826             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 827
 828             /* Compute parameters for interactions between i and j atoms */
 829             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 830
 831             /* EWALD ELECTROSTATICS */
 832
 833             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 834             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 835             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 836             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 837             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 838
 839             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 840                                          &ewtabF,&ewtabFn);
 841             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 842             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 843
 844             fscal            = felec;
 845
 846             /* Update vectorial force */
 847             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 848             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 849             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 850
 851             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 852             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 853             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 854
 855             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 856
 857             /* Inner loop uses 127 flops */
 858         }
 859
 860         if(jidx<j_index_end)
 861         {
 862
 863             jnrA             = jjnr[jidx];
 864             j_coord_offsetA  = DIM*jnrA;
 865
 866             /* load j atom coordinates */
 867             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 868                                               &jx0,&jy0,&jz0);
 869
 870             /* Calculate displacement vector */
 871             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 872             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 873             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 874             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 875             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 876             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 877             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 878             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 879             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 880
 881             /* Calculate squared distance and things based on it */
 882             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 883             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 884             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 885
 886             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 887             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 888             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 889
 890             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 891             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 892             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 893
 894             /* Load parameters for j particles */
 895             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 896             vdwjidx0A        = 2*vdwtype[jnrA+0];
 897
 898             fjx0             = _fjsp_setzero_v2r8();
 899             fjy0             = _fjsp_setzero_v2r8();
 900             fjz0             = _fjsp_setzero_v2r8();
 901
 902             /**************************
 903              * CALCULATE INTERACTIONS *
 904              **************************/
 905
 906             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 907
 908             /* Compute parameters for interactions between i and j atoms */
 909             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 910             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 911
 912             /* EWALD ELECTROSTATICS */
 913
 914             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 915             ewrt             = _fjsp_mul_v2r8(r00,ewtabscale);
 916             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 917             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 918             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 919
 920             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 921             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 922             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq00,rinv00),_fjsp_sub_v2r8(rinvsq00,felec));
 923
 924             /* LENNARD-JONES DISPERSION/REPULSION */
 925
 926             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 927             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 928
 929             fscal            = _fjsp_add_v2r8(felec,fvdw);
 930
 931             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 932
 933             /* Update vectorial force */
 934             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 935             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 936             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 937
 938             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 939             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 940             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 941
 942             /**************************
 943              * CALCULATE INTERACTIONS *
 944              **************************/
 945
 946             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 947
 948             /* Compute parameters for interactions between i and j atoms */
 949             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 950
 951             /* EWALD ELECTROSTATICS */
 952
 953             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 954             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 955             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 956             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 957             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 958
 959             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 960             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 961             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 962
 963             fscal            = felec;
 964
 965             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 966
 967             /* Update vectorial force */
 968             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 969             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 970             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 971
 972             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 973             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 974             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 975
 976             /**************************
 977              * CALCULATE INTERACTIONS *
 978              **************************/
 979
 980             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 981
 982             /* Compute parameters for interactions between i and j atoms */
 983             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 984
 985             /* EWALD ELECTROSTATICS */
 986
 987             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 988             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 989             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 990             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 991             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 992
 993             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
 994             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 995             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 996
 997             fscal            = felec;
 998
 999             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1000
1001             /* Update vectorial force */
1002             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1003             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1004             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1005
1006             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1007             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1008             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1009
1010             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1011
1012             /* Inner loop uses 127 flops */
1013         }
1014
1015         /* End of innermost loop */
1016
1017         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1018                                               f+i_coord_offset,fshift+i_shift_offset);
1019
1020         /* Increment number of inner iterations */
1021         inneriter                  += j_index_end - j_index_start;
1022
1023         /* Outer loop uses 18 flops */
1024     }
1025
1026     /* Increment number of outer iterations */
1027     outeriter        += nri;
1028
1029     /* Update outer/inner flops */
1030
1031     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*127);
1032 }