src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRF_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_ElecRF_VdwLJ_GeomW3W3_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_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     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 128     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 129     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 130     nvdwtype         = fr->ntype;
 131     vdwparam         = fr->nbfp;
 132     vdwtype          = mdatoms->typeA;
 133
 134     /* Setup water-specific parameters */
 135     inr              = nlist->iinr[0];
 136     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 137     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 138     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 139     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 140
 141     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 142     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 143     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 144     vdwjidx0A        = 2*vdwtype[inr+0];
 145     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 146     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 147     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 148     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 149     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 150     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 151     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 152     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 153     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 154     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 155     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 156
 157     /* Avoid stupid compiler warnings */
 158     jnrA = jnrB = 0;
 159     j_coord_offsetA = 0;
 160     j_coord_offsetB = 0;
 161
 162     outeriter        = 0;
 163     inneriter        = 0;
 164
 165     /* Start outer loop over neighborlists */
 166     for(iidx=0; iidx<nri; iidx++)
 167     {
 168         /* Load shift vector for this list */
 169         i_shift_offset   = DIM*shiftidx[iidx];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 181                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 182
 183         fix0             = _fjsp_setzero_v2r8();
 184         fiy0             = _fjsp_setzero_v2r8();
 185         fiz0             = _fjsp_setzero_v2r8();
 186         fix1             = _fjsp_setzero_v2r8();
 187         fiy1             = _fjsp_setzero_v2r8();
 188         fiz1             = _fjsp_setzero_v2r8();
 189         fix2             = _fjsp_setzero_v2r8();
 190         fiy2             = _fjsp_setzero_v2r8();
 191         fiz2             = _fjsp_setzero_v2r8();
 192
 193         /* Reset potential sums */
 194         velecsum         = _fjsp_setzero_v2r8();
 195         vvdwsum          = _fjsp_setzero_v2r8();
 196
 197         /* Start inner kernel loop */
 198         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 199         {
 200
 201             /* Get j neighbor index, and coordinate index */
 202             jnrA             = jjnr[jidx];
 203             jnrB             = jjnr[jidx+1];
 204             j_coord_offsetA  = DIM*jnrA;
 205             j_coord_offsetB  = DIM*jnrB;
 206
 207             /* load j atom coordinates */
 208             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 209                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 210
 211             /* Calculate displacement vector */
 212             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 213             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 214             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 215             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 216             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 217             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 218             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 219             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 220             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 221             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 222             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 223             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 224             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 225             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 226             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 227             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 228             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 229             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 230             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 231             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 232             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 233             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 234             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 235             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 236             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 237             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 238             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 239
 240             /* Calculate squared distance and things based on it */
 241             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 242             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 243             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 244             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 245             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 246             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 247             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 248             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 249             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 250
 251             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 252             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 253             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 254             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 255             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 256             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 257             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 258             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 259             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 260
 261             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 262             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 263             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 264             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 265             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 266             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 267             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 268             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 269             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 270
 271             fjx0             = _fjsp_setzero_v2r8();
 272             fjy0             = _fjsp_setzero_v2r8();
 273             fjz0             = _fjsp_setzero_v2r8();
 274             fjx1             = _fjsp_setzero_v2r8();
 275             fjy1             = _fjsp_setzero_v2r8();
 276             fjz1             = _fjsp_setzero_v2r8();
 277             fjx2             = _fjsp_setzero_v2r8();
 278             fjy2             = _fjsp_setzero_v2r8();
 279             fjz2             = _fjsp_setzero_v2r8();
 280
 281             /**************************
 282              * CALCULATE INTERACTIONS *
 283              **************************/
 284
 285             /* REACTION-FIELD ELECTROSTATICS */
 286             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 287             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 288
 289             /* LENNARD-JONES DISPERSION/REPULSION */
 290
 291             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 292             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 293             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 294             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 295             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 299             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 300
 301             fscal            = _fjsp_add_v2r8(felec,fvdw);
 302
 303             /* Update vectorial force */
 304             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 305             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 306             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 307
 308             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 309             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 310             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 311
 312             /**************************
 313              * CALCULATE INTERACTIONS *
 314              **************************/
 315
 316             /* REACTION-FIELD ELECTROSTATICS */
 317             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 318             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 319
 320             /* Update potential sum for this i atom from the interaction with this j atom. */
 321             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 322
 323             fscal            = felec;
 324
 325             /* Update vectorial force */
 326             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 327             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 328             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 329
 330             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 331             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 332             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             /* REACTION-FIELD ELECTROSTATICS */
 339             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 340             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 341
 342             /* Update potential sum for this i atom from the interaction with this j atom. */
 343             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 344
 345             fscal            = felec;
 346
 347             /* Update vectorial force */
 348             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 349             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 350             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 351
 352             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 353             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 354             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 362             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 363
 364             /* Update potential sum for this i atom from the interaction with this j atom. */
 365             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 366
 367             fscal            = felec;
 368
 369             /* Update vectorial force */
 370             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 371             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 372             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 373
 374             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 375             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 376             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 377
 378             /**************************
 379              * CALCULATE INTERACTIONS *
 380              **************************/
 381
 382             /* REACTION-FIELD ELECTROSTATICS */
 383             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 384             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 385
 386             /* Update potential sum for this i atom from the interaction with this j atom. */
 387             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 388
 389             fscal            = felec;
 390
 391             /* Update vectorial force */
 392             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 393             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 394             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 395
 396             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 397             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 398             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 399
 400             /**************************
 401              * CALCULATE INTERACTIONS *
 402              **************************/
 403
 404             /* REACTION-FIELD ELECTROSTATICS */
 405             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 406             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 407
 408             /* Update potential sum for this i atom from the interaction with this j atom. */
 409             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 410
 411             fscal            = felec;
 412
 413             /* Update vectorial force */
 414             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 415             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 416             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 417
 418             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 419             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 420             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 421
 422             /**************************
 423              * CALCULATE INTERACTIONS *
 424              **************************/
 425
 426             /* REACTION-FIELD ELECTROSTATICS */
 427             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 428             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 432
 433             fscal            = felec;
 434
 435             /* Update vectorial force */
 436             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 437             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 438             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 439
 440             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 441             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 442             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             /* REACTION-FIELD ELECTROSTATICS */
 449             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 450             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 451
 452             /* Update potential sum for this i atom from the interaction with this j atom. */
 453             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 454
 455             fscal            = felec;
 456
 457             /* Update vectorial force */
 458             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 459             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 460             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 461
 462             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 463             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 464             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             /* REACTION-FIELD ELECTROSTATICS */
 471             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 472             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 473
 474             /* Update potential sum for this i atom from the interaction with this j atom. */
 475             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 476
 477             fscal            = felec;
 478
 479             /* Update vectorial force */
 480             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 481             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 482             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 483
 484             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 485             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 486             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 487
 488             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 489
 490             /* Inner loop uses 327 flops */
 491         }
 492
 493         if(jidx<j_index_end)
 494         {
 495
 496             jnrA             = jjnr[jidx];
 497             j_coord_offsetA  = DIM*jnrA;
 498
 499             /* load j atom coordinates */
 500             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 501                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 502
 503             /* Calculate displacement vector */
 504             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 505             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 506             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 507             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 508             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 509             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 510             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 511             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 512             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 513             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 514             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 515             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 516             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 517             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 518             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 519             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 520             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 521             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 522             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 523             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 524             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 525             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 526             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 527             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 528             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 529             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 530             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 531
 532             /* Calculate squared distance and things based on it */
 533             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 534             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 535             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 536             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 537             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 538             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 539             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 540             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 541             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 542
 543             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 544             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 545             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 546             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 547             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 548             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 549             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 550             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 551             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 552
 553             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 554             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 555             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 556             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 557             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 558             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 559             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 560             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 561             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 562
 563             fjx0             = _fjsp_setzero_v2r8();
 564             fjy0             = _fjsp_setzero_v2r8();
 565             fjz0             = _fjsp_setzero_v2r8();
 566             fjx1             = _fjsp_setzero_v2r8();
 567             fjy1             = _fjsp_setzero_v2r8();
 568             fjz1             = _fjsp_setzero_v2r8();
 569             fjx2             = _fjsp_setzero_v2r8();
 570             fjy2             = _fjsp_setzero_v2r8();
 571             fjz2             = _fjsp_setzero_v2r8();
 572
 573             /**************************
 574              * CALCULATE INTERACTIONS *
 575              **************************/
 576
 577             /* REACTION-FIELD ELECTROSTATICS */
 578             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 579             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 580
 581             /* LENNARD-JONES DISPERSION/REPULSION */
 582
 583             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 584             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 585             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 586             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 587             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 588
 589             /* Update potential sum for this i atom from the interaction with this j atom. */
 590             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 591             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 592             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 593             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 594
 595             fscal            = _fjsp_add_v2r8(felec,fvdw);
 596
 597             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 598
 599             /* Update vectorial force */
 600             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 601             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 602             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 603
 604             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 605             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 606             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 607
 608             /**************************
 609              * CALCULATE INTERACTIONS *
 610              **************************/
 611
 612             /* REACTION-FIELD ELECTROSTATICS */
 613             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 614             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 615
 616             /* Update potential sum for this i atom from the interaction with this j atom. */
 617             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 618             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 619
 620             fscal            = felec;
 621
 622             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 623
 624             /* Update vectorial force */
 625             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 626             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 627             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 628
 629             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 630             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 631             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 632
 633             /**************************
 634              * CALCULATE INTERACTIONS *
 635              **************************/
 636
 637             /* REACTION-FIELD ELECTROSTATICS */
 638             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 639             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 640
 641             /* Update potential sum for this i atom from the interaction with this j atom. */
 642             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 643             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 644
 645             fscal            = felec;
 646
 647             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 648
 649             /* Update vectorial force */
 650             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 651             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 652             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 653
 654             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 655             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 656             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 657
 658             /**************************
 659              * CALCULATE INTERACTIONS *
 660              **************************/
 661
 662             /* REACTION-FIELD ELECTROSTATICS */
 663             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 664             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 665
 666             /* Update potential sum for this i atom from the interaction with this j atom. */
 667             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 668             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 669
 670             fscal            = felec;
 671
 672             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 673
 674             /* Update vectorial force */
 675             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 676             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 677             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 678
 679             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 680             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 681             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 682
 683             /**************************
 684              * CALCULATE INTERACTIONS *
 685              **************************/
 686
 687             /* REACTION-FIELD ELECTROSTATICS */
 688             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 689             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 690
 691             /* Update potential sum for this i atom from the interaction with this j atom. */
 692             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 693             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 694
 695             fscal            = felec;
 696
 697             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 698
 699             /* Update vectorial force */
 700             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 701             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 702             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 703
 704             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 705             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 706             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 707
 708             /**************************
 709              * CALCULATE INTERACTIONS *
 710              **************************/
 711
 712             /* REACTION-FIELD ELECTROSTATICS */
 713             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 714             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 715
 716             /* Update potential sum for this i atom from the interaction with this j atom. */
 717             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 718             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 719
 720             fscal            = felec;
 721
 722             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 723
 724             /* Update vectorial force */
 725             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 726             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 727             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 728
 729             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 730             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 731             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 732
 733             /**************************
 734              * CALCULATE INTERACTIONS *
 735              **************************/
 736
 737             /* REACTION-FIELD ELECTROSTATICS */
 738             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 739             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 740
 741             /* Update potential sum for this i atom from the interaction with this j atom. */
 742             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 743             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 744
 745             fscal            = felec;
 746
 747             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 748
 749             /* Update vectorial force */
 750             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 751             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 752             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 753
 754             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 755             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 756             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             /* REACTION-FIELD ELECTROSTATICS */
 763             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 764             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 765
 766             /* Update potential sum for this i atom from the interaction with this j atom. */
 767             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 768             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 769
 770             fscal            = felec;
 771
 772             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 773
 774             /* Update vectorial force */
 775             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 776             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 777             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 778
 779             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 780             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 781             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 782
 783             /**************************
 784              * CALCULATE INTERACTIONS *
 785              **************************/
 786
 787             /* REACTION-FIELD ELECTROSTATICS */
 788             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 789             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 790
 791             /* Update potential sum for this i atom from the interaction with this j atom. */
 792             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 793             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 794
 795             fscal            = felec;
 796
 797             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 798
 799             /* Update vectorial force */
 800             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 801             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 802             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 803
 804             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 805             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 806             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 807
 808             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 809
 810             /* Inner loop uses 327 flops */
 811         }
 812
 813         /* End of innermost loop */
 814
 815         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 816                                               f+i_coord_offset,fshift+i_shift_offset);
 817
 818         ggid                        = gid[iidx];
 819         /* Update potential energies */
 820         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 821         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 822
 823         /* Increment number of inner iterations */
 824         inneriter                  += j_index_end - j_index_start;
 825
 826         /* Outer loop uses 20 flops */
 827     }
 828
 829     /* Increment number of outer iterations */
 830     outeriter        += nri;
 831
 832     /* Update outer/inner flops */
 833
 834     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*327);
 835 }
 836 /*
 837  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
 838  * Electrostatics interaction: ReactionField
 839  * VdW interaction:            LennardJones
 840  * Geometry:                   Water3-Water3
 841  * Calculate force/pot:        Force
 842  */
 843 void
 844 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
 845                     (t_nblist                    * gmx_restrict       nlist,
 846                      rvec                        * gmx_restrict          xx,
 847                      rvec                        * gmx_restrict          ff,
 848                      t_forcerec                  * gmx_restrict          fr,
 849                      t_mdatoms                   * gmx_restrict     mdatoms,
 850                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 851                      t_nrnb                      * gmx_restrict        nrnb)
 852 {
 853     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 854      * just 0 for non-waters.
 855      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 856      * jnr indices corresponding to data put in the four positions in the SIMD register.
 857      */
 858     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 859     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 860     int              jnrA,jnrB;
 861     int              j_coord_offsetA,j_coord_offsetB;
 862     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 863     real             rcutoff_scalar;
 864     real             *shiftvec,*fshift,*x,*f;
 865     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 866     int              vdwioffset0;
 867     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 868     int              vdwioffset1;
 869     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 870     int              vdwioffset2;
 871     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 872     int              vdwjidx0A,vdwjidx0B;
 873     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 874     int              vdwjidx1A,vdwjidx1B;
 875     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 876     int              vdwjidx2A,vdwjidx2B;
 877     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 878     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 879     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 880     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 881     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 882     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 883     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 884     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 885     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 886     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 887     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 888     real             *charge;
 889     int              nvdwtype;
 890     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 891     int              *vdwtype;
 892     real             *vdwparam;
 893     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 894     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 895     _fjsp_v2r8       itab_tmp;
 896     _fjsp_v2r8       dummy_mask,cutoff_mask;
 897     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 898     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 899     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 900
 901     x                = xx[0];
 902     f                = ff[0];
 903
 904     nri              = nlist->nri;
 905     iinr             = nlist->iinr;
 906     jindex           = nlist->jindex;
 907     jjnr             = nlist->jjnr;
 908     shiftidx         = nlist->shift;
 909     gid              = nlist->gid;
 910     shiftvec         = fr->shift_vec[0];
 911     fshift           = fr->fshift[0];
 912     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 913     charge           = mdatoms->chargeA;
 914     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 915     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 916     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 917     nvdwtype         = fr->ntype;
 918     vdwparam         = fr->nbfp;
 919     vdwtype          = mdatoms->typeA;
 920
 921     /* Setup water-specific parameters */
 922     inr              = nlist->iinr[0];
 923     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 924     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 925     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 926     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 927
 928     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 929     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 930     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 931     vdwjidx0A        = 2*vdwtype[inr+0];
 932     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 933     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 934     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 935     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 936     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 937     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 938     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 939     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 940     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 941     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 942     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 943
 944     /* Avoid stupid compiler warnings */
 945     jnrA = jnrB = 0;
 946     j_coord_offsetA = 0;
 947     j_coord_offsetB = 0;
 948
 949     outeriter        = 0;
 950     inneriter        = 0;
 951
 952     /* Start outer loop over neighborlists */
 953     for(iidx=0; iidx<nri; iidx++)
 954     {
 955         /* Load shift vector for this list */
 956         i_shift_offset   = DIM*shiftidx[iidx];
 957
 958         /* Load limits for loop over neighbors */
 959         j_index_start    = jindex[iidx];
 960         j_index_end      = jindex[iidx+1];
 961
 962         /* Get outer coordinate index */
 963         inr              = iinr[iidx];
 964         i_coord_offset   = DIM*inr;
 965
 966         /* Load i particle coords and add shift vector */
 967         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 968                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 969
 970         fix0             = _fjsp_setzero_v2r8();
 971         fiy0             = _fjsp_setzero_v2r8();
 972         fiz0             = _fjsp_setzero_v2r8();
 973         fix1             = _fjsp_setzero_v2r8();
 974         fiy1             = _fjsp_setzero_v2r8();
 975         fiz1             = _fjsp_setzero_v2r8();
 976         fix2             = _fjsp_setzero_v2r8();
 977         fiy2             = _fjsp_setzero_v2r8();
 978         fiz2             = _fjsp_setzero_v2r8();
 979
 980         /* Start inner kernel loop */
 981         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 982         {
 983
 984             /* Get j neighbor index, and coordinate index */
 985             jnrA             = jjnr[jidx];
 986             jnrB             = jjnr[jidx+1];
 987             j_coord_offsetA  = DIM*jnrA;
 988             j_coord_offsetB  = DIM*jnrB;
 989
 990             /* load j atom coordinates */
 991             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 992                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 993
 994             /* Calculate displacement vector */
 995             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 996             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 997             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 998             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 999             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1000             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1001             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1002             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1003             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1004             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1005             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1006             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1007             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1008             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1009             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1010             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1011             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1012             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1013             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1014             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1015             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1016             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1017             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1018             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1019             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1020             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1021             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1022
1023             /* Calculate squared distance and things based on it */
1024             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1025             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1026             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1027             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1028             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1029             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1030             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1031             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1032             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1033
1034             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1035             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1036             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1037             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1038             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1039             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1040             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1041             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1042             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1043
1044             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1045             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1046             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1047             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1048             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1049             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1050             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1051             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1052             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1053
1054             fjx0             = _fjsp_setzero_v2r8();
1055             fjy0             = _fjsp_setzero_v2r8();
1056             fjz0             = _fjsp_setzero_v2r8();
1057             fjx1             = _fjsp_setzero_v2r8();
1058             fjy1             = _fjsp_setzero_v2r8();
1059             fjz1             = _fjsp_setzero_v2r8();
1060             fjx2             = _fjsp_setzero_v2r8();
1061             fjy2             = _fjsp_setzero_v2r8();
1062             fjz2             = _fjsp_setzero_v2r8();
1063
1064             /**************************
1065              * CALCULATE INTERACTIONS *
1066              **************************/
1067
1068             /* REACTION-FIELD ELECTROSTATICS */
1069             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1070
1071             /* LENNARD-JONES DISPERSION/REPULSION */
1072
1073             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1074             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1075
1076             fscal            = _fjsp_add_v2r8(felec,fvdw);
1077
1078             /* Update vectorial force */
1079             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1080             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1081             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1082
1083             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1084             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1085             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1086
1087             /**************************
1088              * CALCULATE INTERACTIONS *
1089              **************************/
1090
1091             /* REACTION-FIELD ELECTROSTATICS */
1092             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1093
1094             fscal            = felec;
1095
1096             /* Update vectorial force */
1097             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1098             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1099             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1100
1101             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1102             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1103             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1104
1105             /**************************
1106              * CALCULATE INTERACTIONS *
1107              **************************/
1108
1109             /* REACTION-FIELD ELECTROSTATICS */
1110             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1111
1112             fscal            = felec;
1113
1114             /* Update vectorial force */
1115             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1116             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1117             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1118
1119             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1120             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1121             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1122
1123             /**************************
1124              * CALCULATE INTERACTIONS *
1125              **************************/
1126
1127             /* REACTION-FIELD ELECTROSTATICS */
1128             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1129
1130             fscal            = felec;
1131
1132             /* Update vectorial force */
1133             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1134             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1135             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1136
1137             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1138             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1139             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1140
1141             /**************************
1142              * CALCULATE INTERACTIONS *
1143              **************************/
1144
1145             /* REACTION-FIELD ELECTROSTATICS */
1146             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1147
1148             fscal            = felec;
1149
1150             /* Update vectorial force */
1151             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1152             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1153             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1154
1155             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1156             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1157             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1158
1159             /**************************
1160              * CALCULATE INTERACTIONS *
1161              **************************/
1162
1163             /* REACTION-FIELD ELECTROSTATICS */
1164             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
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             /* REACTION-FIELD ELECTROSTATICS */
1182             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1183
1184             fscal            = felec;
1185
1186             /* Update vectorial force */
1187             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1188             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1189             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1190
1191             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1192             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1193             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1194
1195             /**************************
1196              * CALCULATE INTERACTIONS *
1197              **************************/
1198
1199             /* REACTION-FIELD ELECTROSTATICS */
1200             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1201
1202             fscal            = felec;
1203
1204             /* Update vectorial force */
1205             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1206             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1207             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1208
1209             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1210             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1211             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1212
1213             /**************************
1214              * CALCULATE INTERACTIONS *
1215              **************************/
1216
1217             /* REACTION-FIELD ELECTROSTATICS */
1218             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1219
1220             fscal            = felec;
1221
1222             /* Update vectorial force */
1223             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1224             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1225             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1226
1227             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1228             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1229             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1230
1231             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1232
1233             /* Inner loop uses 277 flops */
1234         }
1235
1236         if(jidx<j_index_end)
1237         {
1238
1239             jnrA             = jjnr[jidx];
1240             j_coord_offsetA  = DIM*jnrA;
1241
1242             /* load j atom coordinates */
1243             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1244                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1245
1246             /* Calculate displacement vector */
1247             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1248             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1249             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1250             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1251             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1252             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1253             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1254             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1255             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1256             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1257             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1258             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1259             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1260             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1261             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1262             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1263             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1264             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1265             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1266             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1267             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1268             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1269             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1270             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1271             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1272             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1273             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1274
1275             /* Calculate squared distance and things based on it */
1276             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1277             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1278             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1279             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1280             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1281             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1282             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1283             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1284             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1285
1286             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1287             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1288             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1289             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1290             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1291             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1292             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1293             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1294             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1295
1296             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1297             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1298             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1299             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1300             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1301             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1302             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1303             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1304             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1305
1306             fjx0             = _fjsp_setzero_v2r8();
1307             fjy0             = _fjsp_setzero_v2r8();
1308             fjz0             = _fjsp_setzero_v2r8();
1309             fjx1             = _fjsp_setzero_v2r8();
1310             fjy1             = _fjsp_setzero_v2r8();
1311             fjz1             = _fjsp_setzero_v2r8();
1312             fjx2             = _fjsp_setzero_v2r8();
1313             fjy2             = _fjsp_setzero_v2r8();
1314             fjz2             = _fjsp_setzero_v2r8();
1315
1316             /**************************
1317              * CALCULATE INTERACTIONS *
1318              **************************/
1319
1320             /* REACTION-FIELD ELECTROSTATICS */
1321             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1322
1323             /* LENNARD-JONES DISPERSION/REPULSION */
1324
1325             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1326             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1327
1328             fscal            = _fjsp_add_v2r8(felec,fvdw);
1329
1330             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1331
1332             /* Update vectorial force */
1333             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1334             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1335             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1336
1337             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1338             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1339             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1340
1341             /**************************
1342              * CALCULATE INTERACTIONS *
1343              **************************/
1344
1345             /* REACTION-FIELD ELECTROSTATICS */
1346             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1347
1348             fscal            = felec;
1349
1350             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1351
1352             /* Update vectorial force */
1353             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1354             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1355             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1356
1357             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1358             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1359             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1360
1361             /**************************
1362              * CALCULATE INTERACTIONS *
1363              **************************/
1364
1365             /* REACTION-FIELD ELECTROSTATICS */
1366             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1367
1368             fscal            = felec;
1369
1370             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1371
1372             /* Update vectorial force */
1373             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1374             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1375             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1376
1377             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1378             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1379             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1380
1381             /**************************
1382              * CALCULATE INTERACTIONS *
1383              **************************/
1384
1385             /* REACTION-FIELD ELECTROSTATICS */
1386             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1387
1388             fscal            = felec;
1389
1390             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1391
1392             /* Update vectorial force */
1393             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1394             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1395             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1396
1397             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1398             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1399             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1400
1401             /**************************
1402              * CALCULATE INTERACTIONS *
1403              **************************/
1404
1405             /* REACTION-FIELD ELECTROSTATICS */
1406             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1407
1408             fscal            = felec;
1409
1410             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1411
1412             /* Update vectorial force */
1413             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1414             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1415             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1416
1417             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1418             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1419             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1420
1421             /**************************
1422              * CALCULATE INTERACTIONS *
1423              **************************/
1424
1425             /* REACTION-FIELD ELECTROSTATICS */
1426             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1427
1428             fscal            = felec;
1429
1430             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1431
1432             /* Update vectorial force */
1433             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1434             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1435             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1436
1437             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1438             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1439             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1440
1441             /**************************
1442              * CALCULATE INTERACTIONS *
1443              **************************/
1444
1445             /* REACTION-FIELD ELECTROSTATICS */
1446             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1447
1448             fscal            = felec;
1449
1450             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1451
1452             /* Update vectorial force */
1453             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1454             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1455             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1456
1457             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1458             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1459             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1460
1461             /**************************
1462              * CALCULATE INTERACTIONS *
1463              **************************/
1464
1465             /* REACTION-FIELD ELECTROSTATICS */
1466             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1467
1468             fscal            = felec;
1469
1470             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1471
1472             /* Update vectorial force */
1473             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1474             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1475             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1476
1477             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1478             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1479             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1480
1481             /**************************
1482              * CALCULATE INTERACTIONS *
1483              **************************/
1484
1485             /* REACTION-FIELD ELECTROSTATICS */
1486             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1487
1488             fscal            = felec;
1489
1490             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1491
1492             /* Update vectorial force */
1493             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1494             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1495             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1496
1497             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1498             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1499             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1500
1501             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1502
1503             /* Inner loop uses 277 flops */
1504         }
1505
1506         /* End of innermost loop */
1507
1508         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1509                                               f+i_coord_offset,fshift+i_shift_offset);
1510
1511         /* Increment number of inner iterations */
1512         inneriter                  += j_index_end - j_index_start;
1513
1514         /* Outer loop uses 18 flops */
1515     }
1516
1517     /* Increment number of outer iterations */
1518     outeriter        += nri;
1519
1520     /* Update outer/inner flops */
1521
1522     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1523 }