src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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_ElecCoul_VdwLJ_GeomW3W3_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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     int              vdwjidx1A,vdwjidx1B;
  88     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     int              nvdwtype;
 103     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 104     int              *vdwtype;
 105     real             *vdwparam;
 106     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 107     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 108     _fjsp_v2r8       itab_tmp;
 109     _fjsp_v2r8       dummy_mask,cutoff_mask;
 110     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 111     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 112     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 113
 114     x                = xx[0];
 115     f                = ff[0];
 116
 117     nri              = nlist->nri;
 118     iinr             = nlist->iinr;
 119     jindex           = nlist->jindex;
 120     jjnr             = nlist->jjnr;
 121     shiftidx         = nlist->shift;
 122     gid              = nlist->gid;
 123     shiftvec         = fr->shift_vec[0];
 124     fshift           = fr->fshift[0];
 125     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 126     charge           = mdatoms->chargeA;
 127     nvdwtype         = fr->ntype;
 128     vdwparam         = fr->nbfp;
 129     vdwtype          = mdatoms->typeA;
 130
 131     /* Setup water-specific parameters */
 132     inr              = nlist->iinr[0];
 133     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 134     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 135     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 136     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 137
 138     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 139     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 140     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 141     vdwjidx0A        = 2*vdwtype[inr+0];
 142     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 143     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 144     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 145     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 146     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 147     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 148     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 149     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 150     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 151     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 152     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 153
 154     /* Avoid stupid compiler warnings */
 155     jnrA = jnrB = 0;
 156     j_coord_offsetA = 0;
 157     j_coord_offsetB = 0;
 158
 159     outeriter        = 0;
 160     inneriter        = 0;
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 178                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 179
 180         fix0             = _fjsp_setzero_v2r8();
 181         fiy0             = _fjsp_setzero_v2r8();
 182         fiz0             = _fjsp_setzero_v2r8();
 183         fix1             = _fjsp_setzero_v2r8();
 184         fiy1             = _fjsp_setzero_v2r8();
 185         fiz1             = _fjsp_setzero_v2r8();
 186         fix2             = _fjsp_setzero_v2r8();
 187         fiy2             = _fjsp_setzero_v2r8();
 188         fiz2             = _fjsp_setzero_v2r8();
 189
 190         /* Reset potential sums */
 191         velecsum         = _fjsp_setzero_v2r8();
 192         vvdwsum          = _fjsp_setzero_v2r8();
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 196         {
 197
 198             /* Get j neighbor index, and coordinate index */
 199             jnrA             = jjnr[jidx];
 200             jnrB             = jjnr[jidx+1];
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203
 204             /* load j atom coordinates */
 205             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 206                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 207
 208             /* Calculate displacement vector */
 209             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 210             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 211             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 212             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 213             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 214             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 215             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 216             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 217             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 218             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 219             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 220             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 221             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 222             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 223             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 224             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 225             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 226             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 227             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 228             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 229             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 230             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 231             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 232             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 233             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 234             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 235             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 236
 237             /* Calculate squared distance and things based on it */
 238             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 239             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 240             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 241             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 242             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 243             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 244             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 245             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 246             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 247
 248             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 249             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 250             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 251             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 252             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 253             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 254             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 255             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 256             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 257
 258             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 259             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 260             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 261             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 262             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 263             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 264             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 265             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 266             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 267
 268             fjx0             = _fjsp_setzero_v2r8();
 269             fjy0             = _fjsp_setzero_v2r8();
 270             fjz0             = _fjsp_setzero_v2r8();
 271             fjx1             = _fjsp_setzero_v2r8();
 272             fjy1             = _fjsp_setzero_v2r8();
 273             fjz1             = _fjsp_setzero_v2r8();
 274             fjx2             = _fjsp_setzero_v2r8();
 275             fjy2             = _fjsp_setzero_v2r8();
 276             fjz2             = _fjsp_setzero_v2r8();
 277
 278             /**************************
 279              * CALCULATE INTERACTIONS *
 280              **************************/
 281
 282             /* COULOMB ELECTROSTATICS */
 283             velec            = _fjsp_mul_v2r8(qq00,rinv00);
 284             felec            = _fjsp_mul_v2r8(velec,rinvsq00);
 285
 286             /* LENNARD-JONES DISPERSION/REPULSION */
 287
 288             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 289             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 290             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 291             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 292             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 293
 294             /* Update potential sum for this i atom from the interaction with this j atom. */
 295             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 296             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 297
 298             fscal            = _fjsp_add_v2r8(felec,fvdw);
 299
 300             /* Update vectorial force */
 301             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 302             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 303             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 304
 305             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 306             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 307             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             /* COULOMB ELECTROSTATICS */
 314             velec            = _fjsp_mul_v2r8(qq01,rinv01);
 315             felec            = _fjsp_mul_v2r8(velec,rinvsq01);
 316
 317             /* Update potential sum for this i atom from the interaction with this j atom. */
 318             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 319
 320             fscal            = felec;
 321
 322             /* Update vectorial force */
 323             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 324             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 325             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 326
 327             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 328             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 329             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 330
 331             /**************************
 332              * CALCULATE INTERACTIONS *
 333              **************************/
 334
 335             /* COULOMB ELECTROSTATICS */
 336             velec            = _fjsp_mul_v2r8(qq02,rinv02);
 337             felec            = _fjsp_mul_v2r8(velec,rinvsq02);
 338
 339             /* Update potential sum for this i atom from the interaction with this j atom. */
 340             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 341
 342             fscal            = felec;
 343
 344             /* Update vectorial force */
 345             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 346             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 347             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 348
 349             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 350             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 351             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             /* COULOMB ELECTROSTATICS */
 358             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 359             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 360
 361             /* Update potential sum for this i atom from the interaction with this j atom. */
 362             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 363
 364             fscal            = felec;
 365
 366             /* Update vectorial force */
 367             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 368             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 369             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 370
 371             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 372             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 373             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 374
 375             /**************************
 376              * CALCULATE INTERACTIONS *
 377              **************************/
 378
 379             /* COULOMB ELECTROSTATICS */
 380             velec            = _fjsp_mul_v2r8(qq11,rinv11);
 381             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             /* Update vectorial force */
 389             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 390             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 391             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 392
 393             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 394             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 395             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 396
 397             /**************************
 398              * CALCULATE INTERACTIONS *
 399              **************************/
 400
 401             /* COULOMB ELECTROSTATICS */
 402             velec            = _fjsp_mul_v2r8(qq12,rinv12);
 403             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
 404
 405             /* Update potential sum for this i atom from the interaction with this j atom. */
 406             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             /* Update vectorial force */
 411             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 412             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 413             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 414
 415             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 416             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 417             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             /* COULOMB ELECTROSTATICS */
 424             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 425             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 426
 427             /* Update potential sum for this i atom from the interaction with this j atom. */
 428             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 429
 430             fscal            = felec;
 431
 432             /* Update vectorial force */
 433             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 434             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 435             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 436
 437             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 438             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 439             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 440
 441             /**************************
 442              * CALCULATE INTERACTIONS *
 443              **************************/
 444
 445             /* COULOMB ELECTROSTATICS */
 446             velec            = _fjsp_mul_v2r8(qq21,rinv21);
 447             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
 448
 449             /* Update potential sum for this i atom from the interaction with this j atom. */
 450             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 451
 452             fscal            = felec;
 453
 454             /* Update vectorial force */
 455             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 456             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 457             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 458
 459             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 460             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 461             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 462
 463             /**************************
 464              * CALCULATE INTERACTIONS *
 465              **************************/
 466
 467             /* COULOMB ELECTROSTATICS */
 468             velec            = _fjsp_mul_v2r8(qq22,rinv22);
 469             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
 470
 471             /* Update potential sum for this i atom from the interaction with this j atom. */
 472             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 473
 474             fscal            = felec;
 475
 476             /* Update vectorial force */
 477             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 478             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 479             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 480
 481             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 482             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 483             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 484
 485             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 486
 487             /* Inner loop uses 291 flops */
 488         }
 489
 490         if(jidx<j_index_end)
 491         {
 492
 493             jnrA             = jjnr[jidx];
 494             j_coord_offsetA  = DIM*jnrA;
 495
 496             /* load j atom coordinates */
 497             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 498                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 499
 500             /* Calculate displacement vector */
 501             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 502             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 503             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 504             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 505             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 506             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 507             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 508             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 509             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 510             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 511             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 512             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 513             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 514             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 515             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 516             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 517             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 518             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 519             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 520             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 521             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 522             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 523             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 524             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 525             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 526             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 527             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 528
 529             /* Calculate squared distance and things based on it */
 530             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 531             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 532             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 533             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 534             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 535             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 536             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 537             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 538             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 539
 540             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 541             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 542             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 543             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 544             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 545             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 546             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 547             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 548             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 549
 550             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 551             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 552             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 553             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 554             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 555             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 556             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 557             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 558             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 559
 560             fjx0             = _fjsp_setzero_v2r8();
 561             fjy0             = _fjsp_setzero_v2r8();
 562             fjz0             = _fjsp_setzero_v2r8();
 563             fjx1             = _fjsp_setzero_v2r8();
 564             fjy1             = _fjsp_setzero_v2r8();
 565             fjz1             = _fjsp_setzero_v2r8();
 566             fjx2             = _fjsp_setzero_v2r8();
 567             fjy2             = _fjsp_setzero_v2r8();
 568             fjz2             = _fjsp_setzero_v2r8();
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             /* COULOMB ELECTROSTATICS */
 575             velec            = _fjsp_mul_v2r8(qq00,rinv00);
 576             felec            = _fjsp_mul_v2r8(velec,rinvsq00);
 577
 578             /* LENNARD-JONES DISPERSION/REPULSION */
 579
 580             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 581             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 582             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 583             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 584             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 585
 586             /* Update potential sum for this i atom from the interaction with this j atom. */
 587             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 588             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 589             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 590             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 591
 592             fscal            = _fjsp_add_v2r8(felec,fvdw);
 593
 594             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 595
 596             /* Update vectorial force */
 597             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 598             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 599             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 600
 601             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 602             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 603             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 604
 605             /**************************
 606              * CALCULATE INTERACTIONS *
 607              **************************/
 608
 609             /* COULOMB ELECTROSTATICS */
 610             velec            = _fjsp_mul_v2r8(qq01,rinv01);
 611             felec            = _fjsp_mul_v2r8(velec,rinvsq01);
 612
 613             /* Update potential sum for this i atom from the interaction with this j atom. */
 614             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 615             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 616
 617             fscal            = felec;
 618
 619             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 620
 621             /* Update vectorial force */
 622             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 623             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 624             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 625
 626             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 627             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 628             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 629
 630             /**************************
 631              * CALCULATE INTERACTIONS *
 632              **************************/
 633
 634             /* COULOMB ELECTROSTATICS */
 635             velec            = _fjsp_mul_v2r8(qq02,rinv02);
 636             felec            = _fjsp_mul_v2r8(velec,rinvsq02);
 637
 638             /* Update potential sum for this i atom from the interaction with this j atom. */
 639             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 640             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 641
 642             fscal            = felec;
 643
 644             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 645
 646             /* Update vectorial force */
 647             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 648             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 649             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 650
 651             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 652             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 653             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 654
 655             /**************************
 656              * CALCULATE INTERACTIONS *
 657              **************************/
 658
 659             /* COULOMB ELECTROSTATICS */
 660             velec            = _fjsp_mul_v2r8(qq10,rinv10);
 661             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
 662
 663             /* Update potential sum for this i atom from the interaction with this j atom. */
 664             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 665             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 666
 667             fscal            = felec;
 668
 669             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 670
 671             /* Update vectorial force */
 672             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 673             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 674             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 675
 676             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 677             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 678             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 679
 680             /**************************
 681              * CALCULATE INTERACTIONS *
 682              **************************/
 683
 684             /* COULOMB ELECTROSTATICS */
 685             velec            = _fjsp_mul_v2r8(qq11,rinv11);
 686             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
 687
 688             /* Update potential sum for this i atom from the interaction with this j atom. */
 689             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 690             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 691
 692             fscal            = felec;
 693
 694             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 695
 696             /* Update vectorial force */
 697             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 698             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 699             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 700
 701             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 702             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 703             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 704
 705             /**************************
 706              * CALCULATE INTERACTIONS *
 707              **************************/
 708
 709             /* COULOMB ELECTROSTATICS */
 710             velec            = _fjsp_mul_v2r8(qq12,rinv12);
 711             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
 712
 713             /* Update potential sum for this i atom from the interaction with this j atom. */
 714             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 715             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 716
 717             fscal            = felec;
 718
 719             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 720
 721             /* Update vectorial force */
 722             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 723             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 724             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 725
 726             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 727             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 728             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 729
 730             /**************************
 731              * CALCULATE INTERACTIONS *
 732              **************************/
 733
 734             /* COULOMB ELECTROSTATICS */
 735             velec            = _fjsp_mul_v2r8(qq20,rinv20);
 736             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
 737
 738             /* Update potential sum for this i atom from the interaction with this j atom. */
 739             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 740             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 741
 742             fscal            = felec;
 743
 744             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 745
 746             /* Update vectorial force */
 747             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 748             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 749             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 750
 751             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 752             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 753             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             /* COULOMB ELECTROSTATICS */
 760             velec            = _fjsp_mul_v2r8(qq21,rinv21);
 761             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 765             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 770
 771             /* Update vectorial force */
 772             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 773             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 774             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 775
 776             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 777             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 778             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 779
 780             /**************************
 781              * CALCULATE INTERACTIONS *
 782              **************************/
 783
 784             /* COULOMB ELECTROSTATICS */
 785             velec            = _fjsp_mul_v2r8(qq22,rinv22);
 786             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
 787
 788             /* Update potential sum for this i atom from the interaction with this j atom. */
 789             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 790             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 791
 792             fscal            = felec;
 793
 794             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 795
 796             /* Update vectorial force */
 797             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 798             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 799             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 800
 801             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 802             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 803             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 804
 805             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 806
 807             /* Inner loop uses 291 flops */
 808         }
 809
 810         /* End of innermost loop */
 811
 812         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 813                                               f+i_coord_offset,fshift+i_shift_offset);
 814
 815         ggid                        = gid[iidx];
 816         /* Update potential energies */
 817         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 818         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 819
 820         /* Increment number of inner iterations */
 821         inneriter                  += j_index_end - j_index_start;
 822
 823         /* Outer loop uses 20 flops */
 824     }
 825
 826     /* Increment number of outer iterations */
 827     outeriter        += nri;
 828
 829     /* Update outer/inner flops */
 830
 831     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*291);
 832 }
 833 /*
 834  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
 835  * Electrostatics interaction: Coulomb
 836  * VdW interaction:            LennardJones
 837  * Geometry:                   Water3-Water3
 838  * Calculate force/pot:        Force
 839  */
 840 void
 841 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
 842                     (t_nblist                    * gmx_restrict       nlist,
 843                      rvec                        * gmx_restrict          xx,
 844                      rvec                        * gmx_restrict          ff,
 845                      t_forcerec                  * gmx_restrict          fr,
 846                      t_mdatoms                   * gmx_restrict     mdatoms,
 847                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 848                      t_nrnb                      * gmx_restrict        nrnb)
 849 {
 850     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 851      * just 0 for non-waters.
 852      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 853      * jnr indices corresponding to data put in the four positions in the SIMD register.
 854      */
 855     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 856     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 857     int              jnrA,jnrB;
 858     int              j_coord_offsetA,j_coord_offsetB;
 859     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 860     real             rcutoff_scalar;
 861     real             *shiftvec,*fshift,*x,*f;
 862     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 863     int              vdwioffset0;
 864     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 865     int              vdwioffset1;
 866     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 867     int              vdwioffset2;
 868     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 869     int              vdwjidx0A,vdwjidx0B;
 870     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 871     int              vdwjidx1A,vdwjidx1B;
 872     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 873     int              vdwjidx2A,vdwjidx2B;
 874     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 875     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 876     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 877     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 878     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 879     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 880     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 881     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 882     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 883     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 884     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 885     real             *charge;
 886     int              nvdwtype;
 887     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 888     int              *vdwtype;
 889     real             *vdwparam;
 890     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 891     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 892     _fjsp_v2r8       itab_tmp;
 893     _fjsp_v2r8       dummy_mask,cutoff_mask;
 894     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 895     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 896     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 897
 898     x                = xx[0];
 899     f                = ff[0];
 900
 901     nri              = nlist->nri;
 902     iinr             = nlist->iinr;
 903     jindex           = nlist->jindex;
 904     jjnr             = nlist->jjnr;
 905     shiftidx         = nlist->shift;
 906     gid              = nlist->gid;
 907     shiftvec         = fr->shift_vec[0];
 908     fshift           = fr->fshift[0];
 909     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 910     charge           = mdatoms->chargeA;
 911     nvdwtype         = fr->ntype;
 912     vdwparam         = fr->nbfp;
 913     vdwtype          = mdatoms->typeA;
 914
 915     /* Setup water-specific parameters */
 916     inr              = nlist->iinr[0];
 917     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 918     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 919     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 920     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 921
 922     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 923     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 924     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 925     vdwjidx0A        = 2*vdwtype[inr+0];
 926     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 927     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 928     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 929     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 930     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 931     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 932     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 933     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 934     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 935     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 936     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 937
 938     /* Avoid stupid compiler warnings */
 939     jnrA = jnrB = 0;
 940     j_coord_offsetA = 0;
 941     j_coord_offsetB = 0;
 942
 943     outeriter        = 0;
 944     inneriter        = 0;
 945
 946     /* Start outer loop over neighborlists */
 947     for(iidx=0; iidx<nri; iidx++)
 948     {
 949         /* Load shift vector for this list */
 950         i_shift_offset   = DIM*shiftidx[iidx];
 951
 952         /* Load limits for loop over neighbors */
 953         j_index_start    = jindex[iidx];
 954         j_index_end      = jindex[iidx+1];
 955
 956         /* Get outer coordinate index */
 957         inr              = iinr[iidx];
 958         i_coord_offset   = DIM*inr;
 959
 960         /* Load i particle coords and add shift vector */
 961         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 962                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 963
 964         fix0             = _fjsp_setzero_v2r8();
 965         fiy0             = _fjsp_setzero_v2r8();
 966         fiz0             = _fjsp_setzero_v2r8();
 967         fix1             = _fjsp_setzero_v2r8();
 968         fiy1             = _fjsp_setzero_v2r8();
 969         fiz1             = _fjsp_setzero_v2r8();
 970         fix2             = _fjsp_setzero_v2r8();
 971         fiy2             = _fjsp_setzero_v2r8();
 972         fiz2             = _fjsp_setzero_v2r8();
 973
 974         /* Start inner kernel loop */
 975         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 976         {
 977
 978             /* Get j neighbor index, and coordinate index */
 979             jnrA             = jjnr[jidx];
 980             jnrB             = jjnr[jidx+1];
 981             j_coord_offsetA  = DIM*jnrA;
 982             j_coord_offsetB  = DIM*jnrB;
 983
 984             /* load j atom coordinates */
 985             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 986                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 987
 988             /* Calculate displacement vector */
 989             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 990             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 991             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 992             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 993             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 994             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 995             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 996             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 997             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 998             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 999             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1000             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1001             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1002             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1003             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1004             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1005             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1006             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1007             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1008             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1009             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1010             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1011             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1012             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1013             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1014             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1015             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1016
1017             /* Calculate squared distance and things based on it */
1018             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1019             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1020             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1021             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1022             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1023             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1024             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1025             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1026             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1027
1028             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1029             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1030             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1031             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1032             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1033             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1034             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1035             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1036             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1037
1038             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1039             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1040             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1041             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1042             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1043             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1044             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1045             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1046             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1047
1048             fjx0             = _fjsp_setzero_v2r8();
1049             fjy0             = _fjsp_setzero_v2r8();
1050             fjz0             = _fjsp_setzero_v2r8();
1051             fjx1             = _fjsp_setzero_v2r8();
1052             fjy1             = _fjsp_setzero_v2r8();
1053             fjz1             = _fjsp_setzero_v2r8();
1054             fjx2             = _fjsp_setzero_v2r8();
1055             fjy2             = _fjsp_setzero_v2r8();
1056             fjz2             = _fjsp_setzero_v2r8();
1057
1058             /**************************
1059              * CALCULATE INTERACTIONS *
1060              **************************/
1061
1062             /* COULOMB ELECTROSTATICS */
1063             velec            = _fjsp_mul_v2r8(qq00,rinv00);
1064             felec            = _fjsp_mul_v2r8(velec,rinvsq00);
1065
1066             /* LENNARD-JONES DISPERSION/REPULSION */
1067
1068             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1069             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1070
1071             fscal            = _fjsp_add_v2r8(felec,fvdw);
1072
1073             /* Update vectorial force */
1074             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1075             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1076             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1077
1078             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1079             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1080             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1081
1082             /**************************
1083              * CALCULATE INTERACTIONS *
1084              **************************/
1085
1086             /* COULOMB ELECTROSTATICS */
1087             velec            = _fjsp_mul_v2r8(qq01,rinv01);
1088             felec            = _fjsp_mul_v2r8(velec,rinvsq01);
1089
1090             fscal            = felec;
1091
1092             /* Update vectorial force */
1093             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1094             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1095             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1096
1097             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1098             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1099             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1100
1101             /**************************
1102              * CALCULATE INTERACTIONS *
1103              **************************/
1104
1105             /* COULOMB ELECTROSTATICS */
1106             velec            = _fjsp_mul_v2r8(qq02,rinv02);
1107             felec            = _fjsp_mul_v2r8(velec,rinvsq02);
1108
1109             fscal            = felec;
1110
1111             /* Update vectorial force */
1112             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1113             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1114             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1115
1116             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1117             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1118             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1119
1120             /**************************
1121              * CALCULATE INTERACTIONS *
1122              **************************/
1123
1124             /* COULOMB ELECTROSTATICS */
1125             velec            = _fjsp_mul_v2r8(qq10,rinv10);
1126             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
1127
1128             fscal            = felec;
1129
1130             /* Update vectorial force */
1131             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1132             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1133             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1134
1135             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1136             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1137             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1138
1139             /**************************
1140              * CALCULATE INTERACTIONS *
1141              **************************/
1142
1143             /* COULOMB ELECTROSTATICS */
1144             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1145             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1146
1147             fscal            = felec;
1148
1149             /* Update vectorial force */
1150             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1151             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1152             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1153
1154             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1155             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1156             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1157
1158             /**************************
1159              * CALCULATE INTERACTIONS *
1160              **************************/
1161
1162             /* COULOMB ELECTROSTATICS */
1163             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1164             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1165
1166             fscal            = felec;
1167
1168             /* Update vectorial force */
1169             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1170             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1171             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1172
1173             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1174             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1175             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1176
1177             /**************************
1178              * CALCULATE INTERACTIONS *
1179              **************************/
1180
1181             /* COULOMB ELECTROSTATICS */
1182             velec            = _fjsp_mul_v2r8(qq20,rinv20);
1183             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
1184
1185             fscal            = felec;
1186
1187             /* Update vectorial force */
1188             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1189             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1190             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1191
1192             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1193             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1194             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1195
1196             /**************************
1197              * CALCULATE INTERACTIONS *
1198              **************************/
1199
1200             /* COULOMB ELECTROSTATICS */
1201             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1202             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1203
1204             fscal            = felec;
1205
1206             /* Update vectorial force */
1207             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1208             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1209             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1210
1211             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1212             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1213             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1214
1215             /**************************
1216              * CALCULATE INTERACTIONS *
1217              **************************/
1218
1219             /* COULOMB ELECTROSTATICS */
1220             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1221             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1222
1223             fscal            = felec;
1224
1225             /* Update vectorial force */
1226             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1227             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1228             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1229
1230             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1231             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1232             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1233
1234             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1235
1236             /* Inner loop uses 277 flops */
1237         }
1238
1239         if(jidx<j_index_end)
1240         {
1241
1242             jnrA             = jjnr[jidx];
1243             j_coord_offsetA  = DIM*jnrA;
1244
1245             /* load j atom coordinates */
1246             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1247                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1248
1249             /* Calculate displacement vector */
1250             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1251             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1252             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1253             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1254             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1255             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1256             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1257             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1258             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1259             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1260             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1261             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1262             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1263             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1264             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1265             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1266             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1267             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1268             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1269             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1270             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1271             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1272             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1273             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1274             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1275             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1276             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1277
1278             /* Calculate squared distance and things based on it */
1279             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1280             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1281             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1282             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1283             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1284             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1285             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1286             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1287             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1288
1289             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1290             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1291             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1292             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1293             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1294             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1295             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1296             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1297             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1298
1299             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1300             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1301             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1302             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1303             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1304             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1305             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1306             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1307             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1308
1309             fjx0             = _fjsp_setzero_v2r8();
1310             fjy0             = _fjsp_setzero_v2r8();
1311             fjz0             = _fjsp_setzero_v2r8();
1312             fjx1             = _fjsp_setzero_v2r8();
1313             fjy1             = _fjsp_setzero_v2r8();
1314             fjz1             = _fjsp_setzero_v2r8();
1315             fjx2             = _fjsp_setzero_v2r8();
1316             fjy2             = _fjsp_setzero_v2r8();
1317             fjz2             = _fjsp_setzero_v2r8();
1318
1319             /**************************
1320              * CALCULATE INTERACTIONS *
1321              **************************/
1322
1323             /* COULOMB ELECTROSTATICS */
1324             velec            = _fjsp_mul_v2r8(qq00,rinv00);
1325             felec            = _fjsp_mul_v2r8(velec,rinvsq00);
1326
1327             /* LENNARD-JONES DISPERSION/REPULSION */
1328
1329             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1330             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1331
1332             fscal            = _fjsp_add_v2r8(felec,fvdw);
1333
1334             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1335
1336             /* Update vectorial force */
1337             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1338             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1339             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1340
1341             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1342             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1343             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1344
1345             /**************************
1346              * CALCULATE INTERACTIONS *
1347              **************************/
1348
1349             /* COULOMB ELECTROSTATICS */
1350             velec            = _fjsp_mul_v2r8(qq01,rinv01);
1351             felec            = _fjsp_mul_v2r8(velec,rinvsq01);
1352
1353             fscal            = felec;
1354
1355             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1356
1357             /* Update vectorial force */
1358             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1359             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1360             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1361
1362             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1363             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1364             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* COULOMB ELECTROSTATICS */
1371             velec            = _fjsp_mul_v2r8(qq02,rinv02);
1372             felec            = _fjsp_mul_v2r8(velec,rinvsq02);
1373
1374             fscal            = felec;
1375
1376             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1377
1378             /* Update vectorial force */
1379             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1380             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1381             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1382
1383             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1384             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1385             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             /* COULOMB ELECTROSTATICS */
1392             velec            = _fjsp_mul_v2r8(qq10,rinv10);
1393             felec            = _fjsp_mul_v2r8(velec,rinvsq10);
1394
1395             fscal            = felec;
1396
1397             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1398
1399             /* Update vectorial force */
1400             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1401             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1402             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1403
1404             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1405             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1406             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1407
1408             /**************************
1409              * CALCULATE INTERACTIONS *
1410              **************************/
1411
1412             /* COULOMB ELECTROSTATICS */
1413             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1414             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1415
1416             fscal            = felec;
1417
1418             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1419
1420             /* Update vectorial force */
1421             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1422             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1423             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1424
1425             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1426             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1427             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1428
1429             /**************************
1430              * CALCULATE INTERACTIONS *
1431              **************************/
1432
1433             /* COULOMB ELECTROSTATICS */
1434             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1435             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1436
1437             fscal            = felec;
1438
1439             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1440
1441             /* Update vectorial force */
1442             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1443             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1444             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1445
1446             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1447             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1448             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             /* COULOMB ELECTROSTATICS */
1455             velec            = _fjsp_mul_v2r8(qq20,rinv20);
1456             felec            = _fjsp_mul_v2r8(velec,rinvsq20);
1457
1458             fscal            = felec;
1459
1460             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1461
1462             /* Update vectorial force */
1463             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1464             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1465             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1466
1467             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1468             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1469             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1470
1471             /**************************
1472              * CALCULATE INTERACTIONS *
1473              **************************/
1474
1475             /* COULOMB ELECTROSTATICS */
1476             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1477             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1478
1479             fscal            = felec;
1480
1481             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1482
1483             /* Update vectorial force */
1484             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1485             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1486             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1487
1488             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1489             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1490             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1491
1492             /**************************
1493              * CALCULATE INTERACTIONS *
1494              **************************/
1495
1496             /* COULOMB ELECTROSTATICS */
1497             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1498             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1499
1500             fscal            = felec;
1501
1502             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1503
1504             /* Update vectorial force */
1505             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1506             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1507             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1508
1509             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1510             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1511             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1512
1513             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1514
1515             /* Inner loop uses 277 flops */
1516         }
1517
1518         /* End of innermost loop */
1519
1520         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1521                                               f+i_coord_offset,fshift+i_shift_offset);
1522
1523         /* Increment number of inner iterations */
1524         inneriter                  += j_index_end - j_index_start;
1525
1526         /* Outer loop uses 18 flops */
1527     }
1528
1529     /* Increment number of outer iterations */
1530     outeriter        += nri;
1531
1532     /* Update outer/inner flops */
1533
1534     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1535 }