src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRFCut_VdwLJSh_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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "gromacs/legacyheaders/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: ReactionField
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water3-Water3
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  94     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  95     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  96     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  97     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 100     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 103     real             *charge;
 104     int              nvdwtype;
 105     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 106     int              *vdwtype;
 107     real             *vdwparam;
 108     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 109     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 110     _fjsp_v2r8       itab_tmp;
 111     _fjsp_v2r8       dummy_mask,cutoff_mask;
 112     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 113     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 114     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 115
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 128     charge           = mdatoms->chargeA;
 129     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 130     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 131     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 132     nvdwtype         = fr->ntype;
 133     vdwparam         = fr->nbfp;
 134     vdwtype          = mdatoms->typeA;
 135
 136     /* Setup water-specific parameters */
 137     inr              = nlist->iinr[0];
 138     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 139     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 140     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 141     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 142
 143     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 144     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 145     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 146     vdwjidx0A        = 2*vdwtype[inr+0];
 147     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 148     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 149     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 150     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 151     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 152     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 153     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 154     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 155     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 156     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 157     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 158
 159     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 160     rcutoff_scalar   = fr->rcoulomb;
 161     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 162     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 163
 164     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
 165     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
 166
 167     /* Avoid stupid compiler warnings */
 168     jnrA = jnrB = 0;
 169     j_coord_offsetA = 0;
 170     j_coord_offsetB = 0;
 171
 172     outeriter        = 0;
 173     inneriter        = 0;
 174
 175     /* Start outer loop over neighborlists */
 176     for(iidx=0; iidx<nri; iidx++)
 177     {
 178         /* Load shift vector for this list */
 179         i_shift_offset   = DIM*shiftidx[iidx];
 180
 181         /* Load limits for loop over neighbors */
 182         j_index_start    = jindex[iidx];
 183         j_index_end      = jindex[iidx+1];
 184
 185         /* Get outer coordinate index */
 186         inr              = iinr[iidx];
 187         i_coord_offset   = DIM*inr;
 188
 189         /* Load i particle coords and add shift vector */
 190         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 191                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 192
 193         fix0             = _fjsp_setzero_v2r8();
 194         fiy0             = _fjsp_setzero_v2r8();
 195         fiz0             = _fjsp_setzero_v2r8();
 196         fix1             = _fjsp_setzero_v2r8();
 197         fiy1             = _fjsp_setzero_v2r8();
 198         fiz1             = _fjsp_setzero_v2r8();
 199         fix2             = _fjsp_setzero_v2r8();
 200         fiy2             = _fjsp_setzero_v2r8();
 201         fiz2             = _fjsp_setzero_v2r8();
 202
 203         /* Reset potential sums */
 204         velecsum         = _fjsp_setzero_v2r8();
 205         vvdwsum          = _fjsp_setzero_v2r8();
 206
 207         /* Start inner kernel loop */
 208         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 209         {
 210
 211             /* Get j neighbor index, and coordinate index */
 212             jnrA             = jjnr[jidx];
 213             jnrB             = jjnr[jidx+1];
 214             j_coord_offsetA  = DIM*jnrA;
 215             j_coord_offsetB  = DIM*jnrB;
 216
 217             /* load j atom coordinates */
 218             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 219                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 220
 221             /* Calculate displacement vector */
 222             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 223             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 224             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 225             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 226             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 227             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 228             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 229             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 230             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 231             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 232             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 233             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 234             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 235             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 236             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 237             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 238             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 239             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 240             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 241             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 242             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 243             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 244             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 245             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 246             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 247             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 248             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 249
 250             /* Calculate squared distance and things based on it */
 251             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 252             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 253             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 254             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 255             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 256             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 257             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 258             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 259             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 260
 261             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 262             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 263             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 264             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 265             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 266             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 267             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 268             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 269             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 270
 271             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 272             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 273             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 274             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 275             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 276             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 277             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 278             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 279             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 280
 281             fjx0             = _fjsp_setzero_v2r8();
 282             fjy0             = _fjsp_setzero_v2r8();
 283             fjz0             = _fjsp_setzero_v2r8();
 284             fjx1             = _fjsp_setzero_v2r8();
 285             fjy1             = _fjsp_setzero_v2r8();
 286             fjz1             = _fjsp_setzero_v2r8();
 287             fjx2             = _fjsp_setzero_v2r8();
 288             fjy2             = _fjsp_setzero_v2r8();
 289             fjz2             = _fjsp_setzero_v2r8();
 290
 291             /**************************
 292              * CALCULATE INTERACTIONS *
 293              **************************/
 294
 295             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 296             {
 297
 298             /* REACTION-FIELD ELECTROSTATICS */
 299             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 300             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 301
 302             /* LENNARD-JONES DISPERSION/REPULSION */
 303
 304             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 305             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 306             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 307             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 308                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 309             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 310
 311             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 312
 313             /* Update potential sum for this i atom from the interaction with this j atom. */
 314             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 315             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 316             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 317             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 318
 319             fscal            = _fjsp_add_v2r8(felec,fvdw);
 320
 321             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 322
 323             /* Update vectorial force */
 324             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 325             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 326             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 327
 328             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 329             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 330             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 331
 332             }
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 339             {
 340
 341             /* REACTION-FIELD ELECTROSTATICS */
 342             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 343             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 344
 345             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 349             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 350
 351             fscal            = felec;
 352
 353             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 354
 355             /* Update vectorial force */
 356             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 357             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 358             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 359
 360             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 361             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 362             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 363
 364             }
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 371             {
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 375             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 376
 377             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 381             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 386
 387             /* Update vectorial force */
 388             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 389             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 390             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 391
 392             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 393             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 394             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 395
 396             }
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 403             {
 404
 405             /* REACTION-FIELD ELECTROSTATICS */
 406             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 407             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 408
 409             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 410
 411             /* Update potential sum for this i atom from the interaction with this j atom. */
 412             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 413             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 414
 415             fscal            = felec;
 416
 417             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 418
 419             /* Update vectorial force */
 420             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 421             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 422             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 423
 424             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 425             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 426             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 427
 428             }
 429
 430             /**************************
 431              * CALCULATE INTERACTIONS *
 432              **************************/
 433
 434             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 435             {
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 439             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 440
 441             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 445             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 446
 447             fscal            = felec;
 448
 449             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 450
 451             /* Update vectorial force */
 452             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 453             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 454             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 455
 456             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 457             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 458             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 459
 460             }
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 467             {
 468
 469             /* REACTION-FIELD ELECTROSTATICS */
 470             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 471             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 472
 473             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 477             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 478
 479             fscal            = felec;
 480
 481             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 482
 483             /* Update vectorial force */
 484             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 485             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 486             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 487
 488             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 489             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 490             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 491
 492             }
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 499             {
 500
 501             /* REACTION-FIELD ELECTROSTATICS */
 502             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 503             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 504
 505             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 509             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 510
 511             fscal            = felec;
 512
 513             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 514
 515             /* Update vectorial force */
 516             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 517             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 518             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 519
 520             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 521             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 522             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 523
 524             }
 525
 526             /**************************
 527              * CALCULATE INTERACTIONS *
 528              **************************/
 529
 530             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 531             {
 532
 533             /* REACTION-FIELD ELECTROSTATICS */
 534             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 535             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 536
 537             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 538
 539             /* Update potential sum for this i atom from the interaction with this j atom. */
 540             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 541             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 542
 543             fscal            = felec;
 544
 545             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 546
 547             /* Update vectorial force */
 548             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 549             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 550             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 551
 552             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 553             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 554             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 555
 556             }
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 563             {
 564
 565             /* REACTION-FIELD ELECTROSTATICS */
 566             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 567             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 568
 569             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 570
 571             /* Update potential sum for this i atom from the interaction with this j atom. */
 572             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 573             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 574
 575             fscal            = felec;
 576
 577             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 578
 579             /* Update vectorial force */
 580             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 581             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 582             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 583
 584             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 585             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 586             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 587
 588             }
 589
 590             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 591
 592             /* Inner loop uses 369 flops */
 593         }
 594
 595         if(jidx<j_index_end)
 596         {
 597
 598             jnrA             = jjnr[jidx];
 599             j_coord_offsetA  = DIM*jnrA;
 600
 601             /* load j atom coordinates */
 602             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 603                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 604
 605             /* Calculate displacement vector */
 606             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 607             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 608             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 609             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 610             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 611             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 612             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 613             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 614             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 615             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 616             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 617             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 618             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 619             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 620             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 621             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 622             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 623             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 624             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 625             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 626             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 627             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 628             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 629             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 630             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 631             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 632             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 633
 634             /* Calculate squared distance and things based on it */
 635             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 636             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 637             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 638             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 639             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 640             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 641             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 642             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 643             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 644
 645             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 646             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 647             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 648             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 649             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 650             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 651             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 652             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 653             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 654
 655             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 656             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 657             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 658             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 659             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 660             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 661             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 662             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 663             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 664
 665             fjx0             = _fjsp_setzero_v2r8();
 666             fjy0             = _fjsp_setzero_v2r8();
 667             fjz0             = _fjsp_setzero_v2r8();
 668             fjx1             = _fjsp_setzero_v2r8();
 669             fjy1             = _fjsp_setzero_v2r8();
 670             fjz1             = _fjsp_setzero_v2r8();
 671             fjx2             = _fjsp_setzero_v2r8();
 672             fjy2             = _fjsp_setzero_v2r8();
 673             fjz2             = _fjsp_setzero_v2r8();
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 680             {
 681
 682             /* REACTION-FIELD ELECTROSTATICS */
 683             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 684             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 685
 686             /* LENNARD-JONES DISPERSION/REPULSION */
 687
 688             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 689             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 690             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 691             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 692                                            _fjsp_mul_v2r8(_fjsp_nmsub_v2r8( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 693             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 694
 695             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 696
 697             /* Update potential sum for this i atom from the interaction with this j atom. */
 698             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 699             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 700             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 701             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 702             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 703             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 704
 705             fscal            = _fjsp_add_v2r8(felec,fvdw);
 706
 707             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 708
 709             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 710
 711             /* Update vectorial force */
 712             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 713             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 714             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 715
 716             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 717             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 718             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 719
 720             }
 721
 722             /**************************
 723              * CALCULATE INTERACTIONS *
 724              **************************/
 725
 726             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 727             {
 728
 729             /* REACTION-FIELD ELECTROSTATICS */
 730             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 731             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 732
 733             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 734
 735             /* Update potential sum for this i atom from the interaction with this j atom. */
 736             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 737             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 738             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 739
 740             fscal            = felec;
 741
 742             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 743
 744             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 745
 746             /* Update vectorial force */
 747             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 748             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 749             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 750
 751             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 752             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 753             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 754
 755             }
 756
 757             /**************************
 758              * CALCULATE INTERACTIONS *
 759              **************************/
 760
 761             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 762             {
 763
 764             /* REACTION-FIELD ELECTROSTATICS */
 765             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 766             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 767
 768             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 769
 770             /* Update potential sum for this i atom from the interaction with this j atom. */
 771             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 772             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 773             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 774
 775             fscal            = felec;
 776
 777             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 778
 779             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 780
 781             /* Update vectorial force */
 782             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 783             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 784             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 785
 786             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 787             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 788             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 789
 790             }
 791
 792             /**************************
 793              * CALCULATE INTERACTIONS *
 794              **************************/
 795
 796             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 797             {
 798
 799             /* REACTION-FIELD ELECTROSTATICS */
 800             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 801             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 802
 803             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 804
 805             /* Update potential sum for this i atom from the interaction with this j atom. */
 806             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 807             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 808             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 809
 810             fscal            = felec;
 811
 812             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 813
 814             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 815
 816             /* Update vectorial force */
 817             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 818             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 819             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 820
 821             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 822             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 823             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 824
 825             }
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 832             {
 833
 834             /* REACTION-FIELD ELECTROSTATICS */
 835             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 836             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 837
 838             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 839
 840             /* Update potential sum for this i atom from the interaction with this j atom. */
 841             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 842             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 843             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 844
 845             fscal            = felec;
 846
 847             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 848
 849             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 850
 851             /* Update vectorial force */
 852             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 853             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 854             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 855
 856             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 857             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 858             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 859
 860             }
 861
 862             /**************************
 863              * CALCULATE INTERACTIONS *
 864              **************************/
 865
 866             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 867             {
 868
 869             /* REACTION-FIELD ELECTROSTATICS */
 870             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 871             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 872
 873             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 874
 875             /* Update potential sum for this i atom from the interaction with this j atom. */
 876             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 877             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 878             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 879
 880             fscal            = felec;
 881
 882             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 883
 884             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 885
 886             /* Update vectorial force */
 887             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 888             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 889             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 890
 891             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 892             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 893             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 894
 895             }
 896
 897             /**************************
 898              * CALCULATE INTERACTIONS *
 899              **************************/
 900
 901             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 902             {
 903
 904             /* REACTION-FIELD ELECTROSTATICS */
 905             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 906             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 907
 908             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 909
 910             /* Update potential sum for this i atom from the interaction with this j atom. */
 911             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 912             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 913             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 914
 915             fscal            = felec;
 916
 917             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 918
 919             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 920
 921             /* Update vectorial force */
 922             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 923             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 924             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 925
 926             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 927             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 928             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 929
 930             }
 931
 932             /**************************
 933              * CALCULATE INTERACTIONS *
 934              **************************/
 935
 936             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 937             {
 938
 939             /* REACTION-FIELD ELECTROSTATICS */
 940             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 941             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 942
 943             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 944
 945             /* Update potential sum for this i atom from the interaction with this j atom. */
 946             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 947             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 948             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 949
 950             fscal            = felec;
 951
 952             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 953
 954             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 955
 956             /* Update vectorial force */
 957             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 958             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 959             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 960
 961             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 962             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 963             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 964
 965             }
 966
 967             /**************************
 968              * CALCULATE INTERACTIONS *
 969              **************************/
 970
 971             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 972             {
 973
 974             /* REACTION-FIELD ELECTROSTATICS */
 975             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 976             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 977
 978             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 979
 980             /* Update potential sum for this i atom from the interaction with this j atom. */
 981             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 982             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 983             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 984
 985             fscal            = felec;
 986
 987             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 988
 989             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 990
 991             /* Update vectorial force */
 992             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 993             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 994             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 995
 996             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 997             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 998             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 999
1000             }
1001
1002             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1003
1004             /* Inner loop uses 369 flops */
1005         }
1006
1007         /* End of innermost loop */
1008
1009         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1010                                               f+i_coord_offset,fshift+i_shift_offset);
1011
1012         ggid                        = gid[iidx];
1013         /* Update potential energies */
1014         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1015         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
1016
1017         /* Increment number of inner iterations */
1018         inneriter                  += j_index_end - j_index_start;
1019
1020         /* Outer loop uses 20 flops */
1021     }
1022
1023     /* Increment number of outer iterations */
1024     outeriter        += nri;
1025
1026     /* Update outer/inner flops */
1027
1028     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*369);
1029 }
1030 /*
1031  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sparc64_hpc_ace_double
1032  * Electrostatics interaction: ReactionField
1033  * VdW interaction:            LennardJones
1034  * Geometry:                   Water3-Water3
1035  * Calculate force/pot:        Force
1036  */
1037 void
1038 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sparc64_hpc_ace_double
1039                     (t_nblist                    * gmx_restrict       nlist,
1040                      rvec                        * gmx_restrict          xx,
1041                      rvec                        * gmx_restrict          ff,
1042                      t_forcerec                  * gmx_restrict          fr,
1043                      t_mdatoms                   * gmx_restrict     mdatoms,
1044                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1045                      t_nrnb                      * gmx_restrict        nrnb)
1046 {
1047     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1048      * just 0 for non-waters.
1049      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1050      * jnr indices corresponding to data put in the four positions in the SIMD register.
1051      */
1052     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1053     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1054     int              jnrA,jnrB;
1055     int              j_coord_offsetA,j_coord_offsetB;
1056     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1057     real             rcutoff_scalar;
1058     real             *shiftvec,*fshift,*x,*f;
1059     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1060     int              vdwioffset0;
1061     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1062     int              vdwioffset1;
1063     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1064     int              vdwioffset2;
1065     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1066     int              vdwjidx0A,vdwjidx0B;
1067     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1068     int              vdwjidx1A,vdwjidx1B;
1069     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1070     int              vdwjidx2A,vdwjidx2B;
1071     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1072     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1073     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1074     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1075     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1076     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1077     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1078     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1079     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1080     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1081     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1082     real             *charge;
1083     int              nvdwtype;
1084     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1085     int              *vdwtype;
1086     real             *vdwparam;
1087     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
1088     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
1089     _fjsp_v2r8       itab_tmp;
1090     _fjsp_v2r8       dummy_mask,cutoff_mask;
1091     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1092     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1093     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1094
1095     x                = xx[0];
1096     f                = ff[0];
1097
1098     nri              = nlist->nri;
1099     iinr             = nlist->iinr;
1100     jindex           = nlist->jindex;
1101     jjnr             = nlist->jjnr;
1102     shiftidx         = nlist->shift;
1103     gid              = nlist->gid;
1104     shiftvec         = fr->shift_vec[0];
1105     fshift           = fr->fshift[0];
1106     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1107     charge           = mdatoms->chargeA;
1108     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
1109     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
1110     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
1111     nvdwtype         = fr->ntype;
1112     vdwparam         = fr->nbfp;
1113     vdwtype          = mdatoms->typeA;
1114
1115     /* Setup water-specific parameters */
1116     inr              = nlist->iinr[0];
1117     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1118     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1119     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1120     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1121
1122     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1123     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1124     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1125     vdwjidx0A        = 2*vdwtype[inr+0];
1126     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1127     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
1128     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
1129     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1130     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1131     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1132     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1133     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1134     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1135     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1136     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1137
1138     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1139     rcutoff_scalar   = fr->rcoulomb;
1140     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1141     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1142
1143     sh_vdw_invrcut6  = gmx_fjsp_set1_v2r8(fr->ic->sh_invrc6);
1144     rvdw             = gmx_fjsp_set1_v2r8(fr->rvdw);
1145
1146     /* Avoid stupid compiler warnings */
1147     jnrA = jnrB = 0;
1148     j_coord_offsetA = 0;
1149     j_coord_offsetB = 0;
1150
1151     outeriter        = 0;
1152     inneriter        = 0;
1153
1154     /* Start outer loop over neighborlists */
1155     for(iidx=0; iidx<nri; iidx++)
1156     {
1157         /* Load shift vector for this list */
1158         i_shift_offset   = DIM*shiftidx[iidx];
1159
1160         /* Load limits for loop over neighbors */
1161         j_index_start    = jindex[iidx];
1162         j_index_end      = jindex[iidx+1];
1163
1164         /* Get outer coordinate index */
1165         inr              = iinr[iidx];
1166         i_coord_offset   = DIM*inr;
1167
1168         /* Load i particle coords and add shift vector */
1169         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1170                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1171
1172         fix0             = _fjsp_setzero_v2r8();
1173         fiy0             = _fjsp_setzero_v2r8();
1174         fiz0             = _fjsp_setzero_v2r8();
1175         fix1             = _fjsp_setzero_v2r8();
1176         fiy1             = _fjsp_setzero_v2r8();
1177         fiz1             = _fjsp_setzero_v2r8();
1178         fix2             = _fjsp_setzero_v2r8();
1179         fiy2             = _fjsp_setzero_v2r8();
1180         fiz2             = _fjsp_setzero_v2r8();
1181
1182         /* Start inner kernel loop */
1183         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1184         {
1185
1186             /* Get j neighbor index, and coordinate index */
1187             jnrA             = jjnr[jidx];
1188             jnrB             = jjnr[jidx+1];
1189             j_coord_offsetA  = DIM*jnrA;
1190             j_coord_offsetB  = DIM*jnrB;
1191
1192             /* load j atom coordinates */
1193             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1194                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1195
1196             /* Calculate displacement vector */
1197             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1198             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1199             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1200             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1201             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1202             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1203             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1204             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1205             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1206             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1207             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1208             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1209             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1210             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1211             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1212             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1213             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1214             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1215             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1216             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1217             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1218             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1219             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1220             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1221             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1222             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1223             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1224
1225             /* Calculate squared distance and things based on it */
1226             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1227             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1228             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1229             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1230             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1231             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1232             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1233             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1234             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1235
1236             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1237             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1238             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1239             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1240             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1241             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1242             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1243             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1244             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1245
1246             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1247             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1248             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1249             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1250             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1251             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1252             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1253             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1254             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1255
1256             fjx0             = _fjsp_setzero_v2r8();
1257             fjy0             = _fjsp_setzero_v2r8();
1258             fjz0             = _fjsp_setzero_v2r8();
1259             fjx1             = _fjsp_setzero_v2r8();
1260             fjy1             = _fjsp_setzero_v2r8();
1261             fjz1             = _fjsp_setzero_v2r8();
1262             fjx2             = _fjsp_setzero_v2r8();
1263             fjy2             = _fjsp_setzero_v2r8();
1264             fjz2             = _fjsp_setzero_v2r8();
1265
1266             /**************************
1267              * CALCULATE INTERACTIONS *
1268              **************************/
1269
1270             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1271             {
1272
1273             /* REACTION-FIELD ELECTROSTATICS */
1274             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1275
1276             /* LENNARD-JONES DISPERSION/REPULSION */
1277
1278             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1279             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1280
1281             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1282
1283             fscal            = _fjsp_add_v2r8(felec,fvdw);
1284
1285             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1286
1287             /* Update vectorial force */
1288             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1289             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1290             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1291
1292             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1293             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1294             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1295
1296             }
1297
1298             /**************************
1299              * CALCULATE INTERACTIONS *
1300              **************************/
1301
1302             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1303             {
1304
1305             /* REACTION-FIELD ELECTROSTATICS */
1306             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1307
1308             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1309
1310             fscal            = felec;
1311
1312             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1313
1314             /* Update vectorial force */
1315             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1316             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1317             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1318
1319             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1320             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1321             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1322
1323             }
1324
1325             /**************************
1326              * CALCULATE INTERACTIONS *
1327              **************************/
1328
1329             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1330             {
1331
1332             /* REACTION-FIELD ELECTROSTATICS */
1333             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1334
1335             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1336
1337             fscal            = felec;
1338
1339             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1340
1341             /* Update vectorial force */
1342             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1343             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1344             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1345
1346             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1347             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1348             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1349
1350             }
1351
1352             /**************************
1353              * CALCULATE INTERACTIONS *
1354              **************************/
1355
1356             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1357             {
1358
1359             /* REACTION-FIELD ELECTROSTATICS */
1360             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1361
1362             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1363
1364             fscal            = felec;
1365
1366             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1367
1368             /* Update vectorial force */
1369             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1370             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1371             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1372
1373             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1374             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1375             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1376
1377             }
1378
1379             /**************************
1380              * CALCULATE INTERACTIONS *
1381              **************************/
1382
1383             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1384             {
1385
1386             /* REACTION-FIELD ELECTROSTATICS */
1387             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1388
1389             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1390
1391             fscal            = felec;
1392
1393             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1394
1395             /* Update vectorial force */
1396             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1397             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1398             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1399
1400             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1401             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1402             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1403
1404             }
1405
1406             /**************************
1407              * CALCULATE INTERACTIONS *
1408              **************************/
1409
1410             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1411             {
1412
1413             /* REACTION-FIELD ELECTROSTATICS */
1414             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1415
1416             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1417
1418             fscal            = felec;
1419
1420             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1421
1422             /* Update vectorial force */
1423             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1424             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1425             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1426
1427             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1428             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1429             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1430
1431             }
1432
1433             /**************************
1434              * CALCULATE INTERACTIONS *
1435              **************************/
1436
1437             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1438             {
1439
1440             /* REACTION-FIELD ELECTROSTATICS */
1441             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1442
1443             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1444
1445             fscal            = felec;
1446
1447             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1448
1449             /* Update vectorial force */
1450             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1451             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1452             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1453
1454             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1455             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1456             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1457
1458             }
1459
1460             /**************************
1461              * CALCULATE INTERACTIONS *
1462              **************************/
1463
1464             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1465             {
1466
1467             /* REACTION-FIELD ELECTROSTATICS */
1468             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1469
1470             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1471
1472             fscal            = felec;
1473
1474             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1475
1476             /* Update vectorial force */
1477             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1478             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1479             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1480
1481             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1482             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1483             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1484
1485             }
1486
1487             /**************************
1488              * CALCULATE INTERACTIONS *
1489              **************************/
1490
1491             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1492             {
1493
1494             /* REACTION-FIELD ELECTROSTATICS */
1495             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1496
1497             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1498
1499             fscal            = felec;
1500
1501             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1502
1503             /* Update vectorial force */
1504             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1505             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1506             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1507
1508             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1509             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1510             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1511
1512             }
1513
1514             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1515
1516             /* Inner loop uses 304 flops */
1517         }
1518
1519         if(jidx<j_index_end)
1520         {
1521
1522             jnrA             = jjnr[jidx];
1523             j_coord_offsetA  = DIM*jnrA;
1524
1525             /* load j atom coordinates */
1526             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1527                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1528
1529             /* Calculate displacement vector */
1530             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1531             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1532             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1533             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1534             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1535             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1536             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1537             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1538             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1539             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1540             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1541             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1542             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1543             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1544             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1545             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1546             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1547             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1548             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1549             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1550             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1551             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1552             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1553             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1554             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1555             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1556             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1557
1558             /* Calculate squared distance and things based on it */
1559             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1560             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1561             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1562             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1563             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1564             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1565             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1566             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1567             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1568
1569             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1570             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1571             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1572             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1573             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1574             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1575             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1576             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1577             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1578
1579             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1580             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1581             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1582             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1583             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1584             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1585             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1586             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1587             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1588
1589             fjx0             = _fjsp_setzero_v2r8();
1590             fjy0             = _fjsp_setzero_v2r8();
1591             fjz0             = _fjsp_setzero_v2r8();
1592             fjx1             = _fjsp_setzero_v2r8();
1593             fjy1             = _fjsp_setzero_v2r8();
1594             fjz1             = _fjsp_setzero_v2r8();
1595             fjx2             = _fjsp_setzero_v2r8();
1596             fjy2             = _fjsp_setzero_v2r8();
1597             fjz2             = _fjsp_setzero_v2r8();
1598
1599             /**************************
1600              * CALCULATE INTERACTIONS *
1601              **************************/
1602
1603             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1604             {
1605
1606             /* REACTION-FIELD ELECTROSTATICS */
1607             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1608
1609             /* LENNARD-JONES DISPERSION/REPULSION */
1610
1611             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1612             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1613
1614             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1615
1616             fscal            = _fjsp_add_v2r8(felec,fvdw);
1617
1618             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1619
1620             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1621
1622             /* Update vectorial force */
1623             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1624             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1625             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1626
1627             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1628             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1629             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1630
1631             }
1632
1633             /**************************
1634              * CALCULATE INTERACTIONS *
1635              **************************/
1636
1637             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1638             {
1639
1640             /* REACTION-FIELD ELECTROSTATICS */
1641             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1642
1643             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1644
1645             fscal            = felec;
1646
1647             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1648
1649             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1650
1651             /* Update vectorial force */
1652             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1653             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1654             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1655
1656             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1657             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1658             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1659
1660             }
1661
1662             /**************************
1663              * CALCULATE INTERACTIONS *
1664              **************************/
1665
1666             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1667             {
1668
1669             /* REACTION-FIELD ELECTROSTATICS */
1670             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1671
1672             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1673
1674             fscal            = felec;
1675
1676             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1677
1678             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1679
1680             /* Update vectorial force */
1681             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1682             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1683             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1684
1685             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1686             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1687             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1688
1689             }
1690
1691             /**************************
1692              * CALCULATE INTERACTIONS *
1693              **************************/
1694
1695             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1696             {
1697
1698             /* REACTION-FIELD ELECTROSTATICS */
1699             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1700
1701             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1702
1703             fscal            = felec;
1704
1705             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1706
1707             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1708
1709             /* Update vectorial force */
1710             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1711             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1712             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1713
1714             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1715             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1716             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1717
1718             }
1719
1720             /**************************
1721              * CALCULATE INTERACTIONS *
1722              **************************/
1723
1724             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1725             {
1726
1727             /* REACTION-FIELD ELECTROSTATICS */
1728             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1729
1730             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1731
1732             fscal            = felec;
1733
1734             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1735
1736             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1737
1738             /* Update vectorial force */
1739             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1740             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1741             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1742
1743             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1744             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1745             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1746
1747             }
1748
1749             /**************************
1750              * CALCULATE INTERACTIONS *
1751              **************************/
1752
1753             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1754             {
1755
1756             /* REACTION-FIELD ELECTROSTATICS */
1757             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1758
1759             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1760
1761             fscal            = felec;
1762
1763             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1764
1765             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1766
1767             /* Update vectorial force */
1768             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1769             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1770             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1771
1772             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1773             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1774             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1775
1776             }
1777
1778             /**************************
1779              * CALCULATE INTERACTIONS *
1780              **************************/
1781
1782             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1783             {
1784
1785             /* REACTION-FIELD ELECTROSTATICS */
1786             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1787
1788             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1789
1790             fscal            = felec;
1791
1792             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1793
1794             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1795
1796             /* Update vectorial force */
1797             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1798             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1799             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1800
1801             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1802             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1803             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1804
1805             }
1806
1807             /**************************
1808              * CALCULATE INTERACTIONS *
1809              **************************/
1810
1811             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1812             {
1813
1814             /* REACTION-FIELD ELECTROSTATICS */
1815             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1816
1817             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1818
1819             fscal            = felec;
1820
1821             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1822
1823             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1824
1825             /* Update vectorial force */
1826             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1827             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1828             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1829
1830             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1831             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1832             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1833
1834             }
1835
1836             /**************************
1837              * CALCULATE INTERACTIONS *
1838              **************************/
1839
1840             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1841             {
1842
1843             /* REACTION-FIELD ELECTROSTATICS */
1844             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1845
1846             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1847
1848             fscal            = felec;
1849
1850             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1851
1852             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1853
1854             /* Update vectorial force */
1855             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1856             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1857             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1858
1859             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1860             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1861             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1862
1863             }
1864
1865             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1866
1867             /* Inner loop uses 304 flops */
1868         }
1869
1870         /* End of innermost loop */
1871
1872         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1873                                               f+i_coord_offset,fshift+i_shift_offset);
1874
1875         /* Increment number of inner iterations */
1876         inneriter                  += j_index_end - j_index_start;
1877
1878         /* Outer loop uses 18 flops */
1879     }
1880
1881     /* Increment number of outer iterations */
1882     outeriter        += nri;
1883
1884     /* Update outer/inner flops */
1885
1886     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*304);
1887 }