src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEw_VdwLJ_GeomW4P1_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 "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_GeomW4P1_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: Ewald
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water4-Particle
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecEw_VdwLJ_GeomW4P1_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwioffset3;
  88     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  89     int              vdwjidx0A,vdwjidx0B;
  90     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  93     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  94     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  95     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 102     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 103     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 104     real             *ewtab;
 105     _fjsp_v2r8       itab_tmp;
 106     _fjsp_v2r8       dummy_mask,cutoff_mask;
 107     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 108     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 109     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 110
 111     x                = xx[0];
 112     f                = ff[0];
 113
 114     nri              = nlist->nri;
 115     iinr             = nlist->iinr;
 116     jindex           = nlist->jindex;
 117     jjnr             = nlist->jjnr;
 118     shiftidx         = nlist->shift;
 119     gid              = nlist->gid;
 120     shiftvec         = fr->shift_vec[0];
 121     fshift           = fr->fshift[0];
 122     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 123     charge           = mdatoms->chargeA;
 124     nvdwtype         = fr->ntype;
 125     vdwparam         = fr->nbfp;
 126     vdwtype          = mdatoms->typeA;
 127
 128     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 129     ewtab            = fr->ic->tabq_coul_FDV0;
 130     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 131     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 132
 133     /* Setup water-specific parameters */
 134     inr              = nlist->iinr[0];
 135     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 136     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 137     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 138     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 139
 140     /* Avoid stupid compiler warnings */
 141     jnrA = jnrB = 0;
 142     j_coord_offsetA = 0;
 143     j_coord_offsetB = 0;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 164                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 165
 166         fix0             = _fjsp_setzero_v2r8();
 167         fiy0             = _fjsp_setzero_v2r8();
 168         fiz0             = _fjsp_setzero_v2r8();
 169         fix1             = _fjsp_setzero_v2r8();
 170         fiy1             = _fjsp_setzero_v2r8();
 171         fiz1             = _fjsp_setzero_v2r8();
 172         fix2             = _fjsp_setzero_v2r8();
 173         fiy2             = _fjsp_setzero_v2r8();
 174         fiz2             = _fjsp_setzero_v2r8();
 175         fix3             = _fjsp_setzero_v2r8();
 176         fiy3             = _fjsp_setzero_v2r8();
 177         fiz3             = _fjsp_setzero_v2r8();
 178
 179         /* Reset potential sums */
 180         velecsum         = _fjsp_setzero_v2r8();
 181         vvdwsum          = _fjsp_setzero_v2r8();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             j_coord_offsetA  = DIM*jnrA;
 191             j_coord_offsetB  = DIM*jnrB;
 192
 193             /* load j atom coordinates */
 194             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 195                                               &jx0,&jy0,&jz0);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 199             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 200             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 201             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 202             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 203             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 204             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 205             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 206             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 207             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 208             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 209             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 210
 211             /* Calculate squared distance and things based on it */
 212             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 213             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 214             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 215             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 216
 217             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 218             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 219             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 220
 221             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 222             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 223             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 224             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 225
 226             /* Load parameters for j particles */
 227             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 228             vdwjidx0A        = 2*vdwtype[jnrA+0];
 229             vdwjidx0B        = 2*vdwtype[jnrB+0];
 230
 231             fjx0             = _fjsp_setzero_v2r8();
 232             fjy0             = _fjsp_setzero_v2r8();
 233             fjz0             = _fjsp_setzero_v2r8();
 234
 235             /**************************
 236              * CALCULATE INTERACTIONS *
 237              **************************/
 238
 239             /* Compute parameters for interactions between i and j atoms */
 240             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 241                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 242
 243             /* LENNARD-JONES DISPERSION/REPULSION */
 244
 245             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 246             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 247             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 248             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 249             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 250
 251             /* Update potential sum for this i atom from the interaction with this j atom. */
 252             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 253
 254             fscal            = fvdw;
 255
 256             /* Update vectorial force */
 257             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 258             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 259             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 260
 261             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 262             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 263             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 264
 265             /**************************
 266              * CALCULATE INTERACTIONS *
 267              **************************/
 268
 269             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 270
 271             /* Compute parameters for interactions between i and j atoms */
 272             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 273
 274             /* EWALD ELECTROSTATICS */
 275
 276             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 277             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 278             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 279             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 280             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 281
 282             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 283             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 284             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 285             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 286             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 287             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 288             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 289             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 290             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 291             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 292
 293             /* Update potential sum for this i atom from the interaction with this j atom. */
 294             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 295
 296             fscal            = felec;
 297
 298             /* Update vectorial force */
 299             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 300             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 301             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 302
 303             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 304             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 305             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 306
 307             /**************************
 308              * CALCULATE INTERACTIONS *
 309              **************************/
 310
 311             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 312
 313             /* Compute parameters for interactions between i and j atoms */
 314             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 315
 316             /* EWALD ELECTROSTATICS */
 317
 318             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 319             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 320             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 321             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 322             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 323
 324             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 325             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 326             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 327             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 328             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 329             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 330             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 331             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 332             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 333             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 334
 335             /* Update potential sum for this i atom from the interaction with this j atom. */
 336             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 337
 338             fscal            = felec;
 339
 340             /* Update vectorial force */
 341             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 342             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 343             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 344
 345             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 346             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 347             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 348
 349             /**************************
 350              * CALCULATE INTERACTIONS *
 351              **************************/
 352
 353             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
 354
 355             /* Compute parameters for interactions between i and j atoms */
 356             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 357
 358             /* EWALD ELECTROSTATICS */
 359
 360             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 361             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
 362             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 363             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 364             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 365
 366             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 367             ewtabD           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[1] );
 368             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 369             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 370             ewtabFn          = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[1] +2);
 371             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 372             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 373             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 374             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(rinv30,velec));
 375             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
 376
 377             /* Update potential sum for this i atom from the interaction with this j atom. */
 378             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 379
 380             fscal            = felec;
 381
 382             /* Update vectorial force */
 383             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 384             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 385             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 386
 387             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 388             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 389             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 390
 391             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 392
 393             /* Inner loop uses 170 flops */
 394         }
 395
 396         if(jidx<j_index_end)
 397         {
 398
 399             jnrA             = jjnr[jidx];
 400             j_coord_offsetA  = DIM*jnrA;
 401
 402             /* load j atom coordinates */
 403             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 404                                               &jx0,&jy0,&jz0);
 405
 406             /* Calculate displacement vector */
 407             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 408             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 409             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 410             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 411             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 412             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 413             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 414             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 415             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 416             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 417             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 418             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 419
 420             /* Calculate squared distance and things based on it */
 421             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 422             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 423             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 424             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 425
 426             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 427             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 428             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 429
 430             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 431             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 432             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 433             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 434
 435             /* Load parameters for j particles */
 436             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 437             vdwjidx0A        = 2*vdwtype[jnrA+0];
 438
 439             fjx0             = _fjsp_setzero_v2r8();
 440             fjy0             = _fjsp_setzero_v2r8();
 441             fjz0             = _fjsp_setzero_v2r8();
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             /* Compute parameters for interactions between i and j atoms */
 448             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 449
 450             /* LENNARD-JONES DISPERSION/REPULSION */
 451
 452             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 453             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 454             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 455             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 456             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 457
 458             /* Update potential sum for this i atom from the interaction with this j atom. */
 459             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 460             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 461
 462             fscal            = fvdw;
 463
 464             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 465
 466             /* Update vectorial force */
 467             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 468             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 469             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 470
 471             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 472             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 473             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 474
 475             /**************************
 476              * CALCULATE INTERACTIONS *
 477              **************************/
 478
 479             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 480
 481             /* Compute parameters for interactions between i and j atoms */
 482             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 483
 484             /* EWALD ELECTROSTATICS */
 485
 486             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 487             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 488             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 489             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 490             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 491
 492             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 493             ewtabD           = _fjsp_setzero_v2r8();
 494             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 495             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 496             ewtabFn          = _fjsp_setzero_v2r8();
 497             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 498             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 499             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 500             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(rinv10,velec));
 501             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 502
 503             /* Update potential sum for this i atom from the interaction with this j atom. */
 504             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 505             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 506
 507             fscal            = felec;
 508
 509             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 510
 511             /* Update vectorial force */
 512             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 513             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 514             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 515
 516             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 517             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 518             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 519
 520             /**************************
 521              * CALCULATE INTERACTIONS *
 522              **************************/
 523
 524             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 525
 526             /* Compute parameters for interactions between i and j atoms */
 527             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 528
 529             /* EWALD ELECTROSTATICS */
 530
 531             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 532             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 533             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 534             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 535             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 536
 537             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 538             ewtabD           = _fjsp_setzero_v2r8();
 539             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 540             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 541             ewtabFn          = _fjsp_setzero_v2r8();
 542             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 543             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 544             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 545             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(rinv20,velec));
 546             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 547
 548             /* Update potential sum for this i atom from the interaction with this j atom. */
 549             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 550             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 551
 552             fscal            = felec;
 553
 554             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 555
 556             /* Update vectorial force */
 557             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 558             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 559             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 560
 561             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 562             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 563             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 564
 565             /**************************
 566              * CALCULATE INTERACTIONS *
 567              **************************/
 568
 569             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
 570
 571             /* Compute parameters for interactions between i and j atoms */
 572             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 573
 574             /* EWALD ELECTROSTATICS */
 575
 576             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 577             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
 578             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 579             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 580             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 581
 582             ewtabF           = _fjsp_load_v2r8( ewtab + 4*ewconv.i[0] );
 583             ewtabD           = _fjsp_setzero_v2r8();
 584             GMX_FJSP_TRANSPOSE2_V2R8(ewtabF,ewtabD);
 585             ewtabV           = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(), ewtab + 4*ewconv.i[0] +2);
 586             ewtabFn          = _fjsp_setzero_v2r8();
 587             GMX_FJSP_TRANSPOSE2_V2R8(ewtabV,ewtabFn);
 588             felec            = _fjsp_madd_v2r8(eweps,ewtabD,ewtabF);
 589             velec            = _fjsp_nmsub_v2r8(_fjsp_mul_v2r8(ewtabhalfspace,eweps) ,_fjsp_add_v2r8(ewtabF,felec), ewtabV);
 590             velec            = _fjsp_mul_v2r8(qq30,_fjsp_sub_v2r8(rinv30,velec));
 591             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
 592
 593             /* Update potential sum for this i atom from the interaction with this j atom. */
 594             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 595             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 596
 597             fscal            = felec;
 598
 599             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 600
 601             /* Update vectorial force */
 602             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 603             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 604             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 605
 606             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 607             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 608             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 609
 610             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 611
 612             /* Inner loop uses 170 flops */
 613         }
 614
 615         /* End of innermost loop */
 616
 617         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 618                                               f+i_coord_offset,fshift+i_shift_offset);
 619
 620         ggid                        = gid[iidx];
 621         /* Update potential energies */
 622         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 623         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 624
 625         /* Increment number of inner iterations */
 626         inneriter                  += j_index_end - j_index_start;
 627
 628         /* Outer loop uses 26 flops */
 629     }
 630
 631     /* Increment number of outer iterations */
 632     outeriter        += nri;
 633
 634     /* Update outer/inner flops */
 635
 636     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*170);
 637 }
 638 /*
 639  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 640  * Electrostatics interaction: Ewald
 641  * VdW interaction:            LennardJones
 642  * Geometry:                   Water4-Particle
 643  * Calculate force/pot:        Force
 644  */
 645 void
 646 nb_kernel_ElecEw_VdwLJ_GeomW4P1_F_sparc64_hpc_ace_double
 647                     (t_nblist                    * gmx_restrict       nlist,
 648                      rvec                        * gmx_restrict          xx,
 649                      rvec                        * gmx_restrict          ff,
 650                      t_forcerec                  * gmx_restrict          fr,
 651                      t_mdatoms                   * gmx_restrict     mdatoms,
 652                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 653                      t_nrnb                      * gmx_restrict        nrnb)
 654 {
 655     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 656      * just 0 for non-waters.
 657      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 658      * jnr indices corresponding to data put in the four positions in the SIMD register.
 659      */
 660     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 661     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 662     int              jnrA,jnrB;
 663     int              j_coord_offsetA,j_coord_offsetB;
 664     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 665     real             rcutoff_scalar;
 666     real             *shiftvec,*fshift,*x,*f;
 667     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 668     int              vdwioffset0;
 669     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 670     int              vdwioffset1;
 671     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 672     int              vdwioffset2;
 673     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 674     int              vdwioffset3;
 675     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 676     int              vdwjidx0A,vdwjidx0B;
 677     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 678     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 679     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 680     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 681     _fjsp_v2r8       dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 682     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 683     real             *charge;
 684     int              nvdwtype;
 685     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 686     int              *vdwtype;
 687     real             *vdwparam;
 688     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 689     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 690     _fjsp_v2r8       ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 691     real             *ewtab;
 692     _fjsp_v2r8       itab_tmp;
 693     _fjsp_v2r8       dummy_mask,cutoff_mask;
 694     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 695     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 696     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 697
 698     x                = xx[0];
 699     f                = ff[0];
 700
 701     nri              = nlist->nri;
 702     iinr             = nlist->iinr;
 703     jindex           = nlist->jindex;
 704     jjnr             = nlist->jjnr;
 705     shiftidx         = nlist->shift;
 706     gid              = nlist->gid;
 707     shiftvec         = fr->shift_vec[0];
 708     fshift           = fr->fshift[0];
 709     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 710     charge           = mdatoms->chargeA;
 711     nvdwtype         = fr->ntype;
 712     vdwparam         = fr->nbfp;
 713     vdwtype          = mdatoms->typeA;
 714
 715     sh_ewald         = gmx_fjsp_set1_v2r8(fr->ic->sh_ewald);
 716     ewtab            = fr->ic->tabq_coul_F;
 717     ewtabscale       = gmx_fjsp_set1_v2r8(fr->ic->tabq_scale);
 718     ewtabhalfspace   = gmx_fjsp_set1_v2r8(0.5/fr->ic->tabq_scale);
 719
 720     /* Setup water-specific parameters */
 721     inr              = nlist->iinr[0];
 722     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 723     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 724     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 725     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 726
 727     /* Avoid stupid compiler warnings */
 728     jnrA = jnrB = 0;
 729     j_coord_offsetA = 0;
 730     j_coord_offsetB = 0;
 731
 732     outeriter        = 0;
 733     inneriter        = 0;
 734
 735     /* Start outer loop over neighborlists */
 736     for(iidx=0; iidx<nri; iidx++)
 737     {
 738         /* Load shift vector for this list */
 739         i_shift_offset   = DIM*shiftidx[iidx];
 740
 741         /* Load limits for loop over neighbors */
 742         j_index_start    = jindex[iidx];
 743         j_index_end      = jindex[iidx+1];
 744
 745         /* Get outer coordinate index */
 746         inr              = iinr[iidx];
 747         i_coord_offset   = DIM*inr;
 748
 749         /* Load i particle coords and add shift vector */
 750         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 751                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 752
 753         fix0             = _fjsp_setzero_v2r8();
 754         fiy0             = _fjsp_setzero_v2r8();
 755         fiz0             = _fjsp_setzero_v2r8();
 756         fix1             = _fjsp_setzero_v2r8();
 757         fiy1             = _fjsp_setzero_v2r8();
 758         fiz1             = _fjsp_setzero_v2r8();
 759         fix2             = _fjsp_setzero_v2r8();
 760         fiy2             = _fjsp_setzero_v2r8();
 761         fiz2             = _fjsp_setzero_v2r8();
 762         fix3             = _fjsp_setzero_v2r8();
 763         fiy3             = _fjsp_setzero_v2r8();
 764         fiz3             = _fjsp_setzero_v2r8();
 765
 766         /* Start inner kernel loop */
 767         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 768         {
 769
 770             /* Get j neighbor index, and coordinate index */
 771             jnrA             = jjnr[jidx];
 772             jnrB             = jjnr[jidx+1];
 773             j_coord_offsetA  = DIM*jnrA;
 774             j_coord_offsetB  = DIM*jnrB;
 775
 776             /* load j atom coordinates */
 777             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 778                                               &jx0,&jy0,&jz0);
 779
 780             /* Calculate displacement vector */
 781             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 782             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 783             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 784             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 785             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 786             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 787             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 788             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 789             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 790             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 791             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 792             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 793
 794             /* Calculate squared distance and things based on it */
 795             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 796             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 797             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 798             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 799
 800             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 801             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 802             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 803
 804             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 805             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 806             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 807             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 808
 809             /* Load parameters for j particles */
 810             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 811             vdwjidx0A        = 2*vdwtype[jnrA+0];
 812             vdwjidx0B        = 2*vdwtype[jnrB+0];
 813
 814             fjx0             = _fjsp_setzero_v2r8();
 815             fjy0             = _fjsp_setzero_v2r8();
 816             fjz0             = _fjsp_setzero_v2r8();
 817
 818             /**************************
 819              * CALCULATE INTERACTIONS *
 820              **************************/
 821
 822             /* Compute parameters for interactions between i and j atoms */
 823             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 824                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 825
 826             /* LENNARD-JONES DISPERSION/REPULSION */
 827
 828             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 829             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 830
 831             fscal            = fvdw;
 832
 833             /* Update vectorial force */
 834             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 835             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 836             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 837
 838             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 839             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 840             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 841
 842             /**************************
 843              * CALCULATE INTERACTIONS *
 844              **************************/
 845
 846             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
 847
 848             /* Compute parameters for interactions between i and j atoms */
 849             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 850
 851             /* EWALD ELECTROSTATICS */
 852
 853             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 854             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
 855             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 856             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 857             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 858
 859             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 860                                          &ewtabF,&ewtabFn);
 861             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 862             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
 863
 864             fscal            = felec;
 865
 866             /* Update vectorial force */
 867             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 868             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 869             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 870
 871             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 872             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 873             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
 880
 881             /* Compute parameters for interactions between i and j atoms */
 882             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 883
 884             /* EWALD ELECTROSTATICS */
 885
 886             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 887             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
 888             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 889             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 890             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 891
 892             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 893                                          &ewtabF,&ewtabFn);
 894             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 895             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
 896
 897             fscal            = felec;
 898
 899             /* Update vectorial force */
 900             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 901             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 902             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 903
 904             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 905             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 906             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 907
 908             /**************************
 909              * CALCULATE INTERACTIONS *
 910              **************************/
 911
 912             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
 913
 914             /* Compute parameters for interactions between i and j atoms */
 915             qq30             = _fjsp_mul_v2r8(iq3,jq0);
 916
 917             /* EWALD ELECTROSTATICS */
 918
 919             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 920             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
 921             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
 922             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
 923             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
 924
 925             gmx_fjsp_load_2pair_swizzle_v2r8(ewtab+ewconv.i[0],ewtab+ewconv.i[1],
 926                                          &ewtabF,&ewtabFn);
 927             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
 928             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
 929
 930             fscal            = felec;
 931
 932             /* Update vectorial force */
 933             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
 934             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
 935             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
 936
 937             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
 938             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
 939             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
 940
 941             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 942
 943             /* Inner loop uses 150 flops */
 944         }
 945
 946         if(jidx<j_index_end)
 947         {
 948
 949             jnrA             = jjnr[jidx];
 950             j_coord_offsetA  = DIM*jnrA;
 951
 952             /* load j atom coordinates */
 953             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 954                                               &jx0,&jy0,&jz0);
 955
 956             /* Calculate displacement vector */
 957             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 958             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 959             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 960             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 961             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 962             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 963             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 964             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 965             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 966             dx30             = _fjsp_sub_v2r8(ix3,jx0);
 967             dy30             = _fjsp_sub_v2r8(iy3,jy0);
 968             dz30             = _fjsp_sub_v2r8(iz3,jz0);
 969
 970             /* Calculate squared distance and things based on it */
 971             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 972             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 973             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 974             rsq30            = gmx_fjsp_calc_rsq_v2r8(dx30,dy30,dz30);
 975
 976             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 977             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 978             rinv30           = gmx_fjsp_invsqrt_v2r8(rsq30);
 979
 980             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
 981             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 982             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 983             rinvsq30         = _fjsp_mul_v2r8(rinv30,rinv30);
 984
 985             /* Load parameters for j particles */
 986             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 987             vdwjidx0A        = 2*vdwtype[jnrA+0];
 988
 989             fjx0             = _fjsp_setzero_v2r8();
 990             fjy0             = _fjsp_setzero_v2r8();
 991             fjz0             = _fjsp_setzero_v2r8();
 992
 993             /**************************
 994              * CALCULATE INTERACTIONS *
 995              **************************/
 996
 997             /* Compute parameters for interactions between i and j atoms */
 998             gmx_fjsp_load_1pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 999
1000             /* LENNARD-JONES DISPERSION/REPULSION */
1001
1002             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1003             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1004
1005             fscal            = fvdw;
1006
1007             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1008
1009             /* Update vectorial force */
1010             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1011             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1012             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1013
1014             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1015             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1016             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1017
1018             /**************************
1019              * CALCULATE INTERACTIONS *
1020              **************************/
1021
1022             r10              = _fjsp_mul_v2r8(rsq10,rinv10);
1023
1024             /* Compute parameters for interactions between i and j atoms */
1025             qq10             = _fjsp_mul_v2r8(iq1,jq0);
1026
1027             /* EWALD ELECTROSTATICS */
1028
1029             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1030             ewrt             = _fjsp_mul_v2r8(r10,ewtabscale);
1031             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1032             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1033             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1034
1035             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1036             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1037             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq10,rinv10),_fjsp_sub_v2r8(rinvsq10,felec));
1038
1039             fscal            = felec;
1040
1041             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1042
1043             /* Update vectorial force */
1044             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1045             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1046             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1047
1048             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1049             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1050             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1051
1052             /**************************
1053              * CALCULATE INTERACTIONS *
1054              **************************/
1055
1056             r20              = _fjsp_mul_v2r8(rsq20,rinv20);
1057
1058             /* Compute parameters for interactions between i and j atoms */
1059             qq20             = _fjsp_mul_v2r8(iq2,jq0);
1060
1061             /* EWALD ELECTROSTATICS */
1062
1063             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1064             ewrt             = _fjsp_mul_v2r8(r20,ewtabscale);
1065             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1066             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1067             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1068
1069             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1070             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1071             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq20,rinv20),_fjsp_sub_v2r8(rinvsq20,felec));
1072
1073             fscal            = felec;
1074
1075             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1076
1077             /* Update vectorial force */
1078             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1079             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1080             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1081
1082             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1083             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1084             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1085
1086             /**************************
1087              * CALCULATE INTERACTIONS *
1088              **************************/
1089
1090             r30              = _fjsp_mul_v2r8(rsq30,rinv30);
1091
1092             /* Compute parameters for interactions between i and j atoms */
1093             qq30             = _fjsp_mul_v2r8(iq3,jq0);
1094
1095             /* EWALD ELECTROSTATICS */
1096
1097             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1098             ewrt             = _fjsp_mul_v2r8(r30,ewtabscale);
1099             itab_tmp         = _fjsp_dtox_v2r8(ewrt);
1100             eweps            = _fjsp_sub_v2r8(ewrt,_fjsp_xtod_v2r8(itab_tmp));
1101             _fjsp_store_v2r8(&ewconv.simd,itab_tmp);
1102
1103             gmx_fjsp_load_1pair_swizzle_v2r8(ewtab+ewconv.i[0],&ewtabF,&ewtabFn);
1104             felec            = _fjsp_madd_v2r8(eweps,ewtabFn,_fjsp_nmsub_v2r8(eweps,ewtabF,ewtabF));
1105             felec            = _fjsp_mul_v2r8(_fjsp_mul_v2r8(qq30,rinv30),_fjsp_sub_v2r8(rinvsq30,felec));
1106
1107             fscal            = felec;
1108
1109             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1110
1111             /* Update vectorial force */
1112             fix3             = _fjsp_madd_v2r8(dx30,fscal,fix3);
1113             fiy3             = _fjsp_madd_v2r8(dy30,fscal,fiy3);
1114             fiz3             = _fjsp_madd_v2r8(dz30,fscal,fiz3);
1115
1116             fjx0             = _fjsp_madd_v2r8(dx30,fscal,fjx0);
1117             fjy0             = _fjsp_madd_v2r8(dy30,fscal,fjy0);
1118             fjz0             = _fjsp_madd_v2r8(dz30,fscal,fjz0);
1119
1120             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1121
1122             /* Inner loop uses 150 flops */
1123         }
1124
1125         /* End of innermost loop */
1126
1127         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1128                                               f+i_coord_offset,fshift+i_shift_offset);
1129
1130         /* Increment number of inner iterations */
1131         inneriter                  += j_index_end - j_index_start;
1132
1133         /* Outer loop uses 24 flops */
1134     }
1135
1136     /* Increment number of outer iterations */
1137     outeriter        += nri;
1138
1139     /* Update outer/inner flops */
1140
1141     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*150);
1142 }