src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRFCut_VdwCSTab_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_ElecRFCut_VdwCSTab_GeomW3W3_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_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       rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 109     real             *vftab;
 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     vftab            = kernel_data->table_vdw->data;
 137     vftabscale       = gmx_fjsp_set1_v2r8(kernel_data->table_vdw->scale);
 138
 139     /* Setup water-specific parameters */
 140     inr              = nlist->iinr[0];
 141     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 142     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 143     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 144     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 145
 146     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 147     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 148     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 149     vdwjidx0A        = 2*vdwtype[inr+0];
 150     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 151     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 152     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 153     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 154     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 155     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 156     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 157     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 158     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 159     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 160     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 161
 162     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 163     rcutoff_scalar   = fr->rcoulomb;
 164     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 165     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 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             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 299
 300             /* Calculate table index by multiplying r with table scale and truncate to integer */
 301             rt               = _fjsp_mul_v2r8(r00,vftabscale);
 302             itab_tmp         = _fjsp_dtox_v2r8(rt);
 303             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 304             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 305             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 306
 307             vfconv.i[0]     *= 8;
 308             vfconv.i[1]     *= 8;
 309
 310             /* REACTION-FIELD ELECTROSTATICS */
 311             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 312             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 313
 314             /* CUBIC SPLINE TABLE DISPERSION */
 315             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 316             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
 317             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 318             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 2 );
 319             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] + 2 );
 320             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 321             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
 322             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 323             vvdw6            = _fjsp_mul_v2r8(c6_00,VV);
 324             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
 325             fvdw6            = _fjsp_mul_v2r8(c6_00,FF);
 326
 327             /* CUBIC SPLINE TABLE REPULSION */
 328             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 4 );
 329             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] + 4 );
 330             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 331             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 6 );
 332             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] + 6 );
 333             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 334             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
 335             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 336             vvdw12           = _fjsp_mul_v2r8(c12_00,VV);
 337             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
 338             fvdw12           = _fjsp_mul_v2r8(c12_00,FF);
 339             vvdw             = _fjsp_add_v2r8(vvdw12,vvdw6);
 340             fvdw             = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_add_v2r8(fvdw6,fvdw12),_fjsp_mul_v2r8(vftabscale,rinv00)));
 341
 342             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 346             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 347             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 348             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 349
 350             fscal            = _fjsp_add_v2r8(felec,fvdw);
 351
 352             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 353
 354             /* Update vectorial force */
 355             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 356             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 357             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 358
 359             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 360             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 361             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 362
 363             }
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 370             {
 371
 372             /* REACTION-FIELD ELECTROSTATICS */
 373             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 374             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 375
 376             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 377
 378             /* Update potential sum for this i atom from the interaction with this j atom. */
 379             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 380             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 385
 386             /* Update vectorial force */
 387             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 388             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 389             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 390
 391             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 392             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 393             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 394
 395             }
 396
 397             /**************************
 398              * CALCULATE INTERACTIONS *
 399              **************************/
 400
 401             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 402             {
 403
 404             /* REACTION-FIELD ELECTROSTATICS */
 405             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 406             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 407
 408             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 409
 410             /* Update potential sum for this i atom from the interaction with this j atom. */
 411             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 412             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 417
 418             /* Update vectorial force */
 419             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 420             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 421             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 422
 423             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 424             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 425             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 426
 427             }
 428
 429             /**************************
 430              * CALCULATE INTERACTIONS *
 431              **************************/
 432
 433             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 434             {
 435
 436             /* REACTION-FIELD ELECTROSTATICS */
 437             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 438             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 439
 440             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 441
 442             /* Update potential sum for this i atom from the interaction with this j atom. */
 443             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 444             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 449
 450             /* Update vectorial force */
 451             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 452             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 453             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 454
 455             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 456             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 457             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 458
 459             }
 460
 461             /**************************
 462              * CALCULATE INTERACTIONS *
 463              **************************/
 464
 465             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 466             {
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 470             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 471
 472             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 473
 474             /* Update potential sum for this i atom from the interaction with this j atom. */
 475             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 476             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 481
 482             /* Update vectorial force */
 483             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 484             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 485             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 486
 487             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 488             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 489             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 490
 491             }
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 498             {
 499
 500             /* REACTION-FIELD ELECTROSTATICS */
 501             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 502             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 503
 504             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 508             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 513
 514             /* Update vectorial force */
 515             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 516             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 517             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 518
 519             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 520             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 521             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 522
 523             }
 524
 525             /**************************
 526              * CALCULATE INTERACTIONS *
 527              **************************/
 528
 529             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 530             {
 531
 532             /* REACTION-FIELD ELECTROSTATICS */
 533             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 534             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 535
 536             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 540             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 541
 542             fscal            = felec;
 543
 544             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 545
 546             /* Update vectorial force */
 547             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 548             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 549             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 550
 551             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 552             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 553             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 554
 555             }
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 562             {
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 566             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 567
 568             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 572             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 573
 574             fscal            = felec;
 575
 576             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 577
 578             /* Update vectorial force */
 579             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 580             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 581             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 582
 583             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 584             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 585             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 586
 587             }
 588
 589             /**************************
 590              * CALCULATE INTERACTIONS *
 591              **************************/
 592
 593             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 594             {
 595
 596             /* REACTION-FIELD ELECTROSTATICS */
 597             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 598             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 599
 600             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 601
 602             /* Update potential sum for this i atom from the interaction with this j atom. */
 603             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 604             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 605
 606             fscal            = felec;
 607
 608             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 609
 610             /* Update vectorial force */
 611             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 612             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 613             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 614
 615             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 616             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 617             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 618
 619             }
 620
 621             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 622
 623             /* Inner loop uses 387 flops */
 624         }
 625
 626         if(jidx<j_index_end)
 627         {
 628
 629             jnrA             = jjnr[jidx];
 630             j_coord_offsetA  = DIM*jnrA;
 631
 632             /* load j atom coordinates */
 633             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 634                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 635
 636             /* Calculate displacement vector */
 637             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 638             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 639             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 640             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 641             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 642             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 643             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 644             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 645             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 646             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 647             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 648             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 649             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 650             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 651             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 652             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 653             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 654             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 655             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 656             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 657             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 658             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 659             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 660             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 661             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 662             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 663             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 664
 665             /* Calculate squared distance and things based on it */
 666             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 667             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 668             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 669             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 670             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 671             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 672             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 673             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 674             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 675
 676             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 677             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 678             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 679             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 680             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 681             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 682             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 683             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 684             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 685
 686             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 687             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 688             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 689             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 690             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 691             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 692             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 693             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 694             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 695
 696             fjx0             = _fjsp_setzero_v2r8();
 697             fjy0             = _fjsp_setzero_v2r8();
 698             fjz0             = _fjsp_setzero_v2r8();
 699             fjx1             = _fjsp_setzero_v2r8();
 700             fjy1             = _fjsp_setzero_v2r8();
 701             fjz1             = _fjsp_setzero_v2r8();
 702             fjx2             = _fjsp_setzero_v2r8();
 703             fjy2             = _fjsp_setzero_v2r8();
 704             fjz2             = _fjsp_setzero_v2r8();
 705
 706             /**************************
 707              * CALCULATE INTERACTIONS *
 708              **************************/
 709
 710             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 711             {
 712
 713             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
 714
 715             /* Calculate table index by multiplying r with table scale and truncate to integer */
 716             rt               = _fjsp_mul_v2r8(r00,vftabscale);
 717             itab_tmp         = _fjsp_dtox_v2r8(rt);
 718             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
 719             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
 720             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
 721
 722             vfconv.i[0]     *= 8;
 723             vfconv.i[1]     *= 8;
 724
 725             /* REACTION-FIELD ELECTROSTATICS */
 726             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 727             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 728
 729             /* CUBIC SPLINE TABLE DISPERSION */
 730             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
 731             F                = _fjsp_setzero_v2r8();
 732             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 733             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 2 );
 734             H                = _fjsp_setzero_v2r8();
 735             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 736             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
 737             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 738             vvdw6            = _fjsp_mul_v2r8(c6_00,VV);
 739             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
 740             fvdw6            = _fjsp_mul_v2r8(c6_00,FF);
 741
 742             /* CUBIC SPLINE TABLE REPULSION */
 743             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 4 );
 744             F                = _fjsp_setzero_v2r8();
 745             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
 746             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 6 );
 747             H                = _fjsp_setzero_v2r8();
 748             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
 749             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
 750             VV               = _fjsp_madd_v2r8(vfeps,Fp,Y);
 751             vvdw12           = _fjsp_mul_v2r8(c12_00,VV);
 752             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
 753             fvdw12           = _fjsp_mul_v2r8(c12_00,FF);
 754             vvdw             = _fjsp_add_v2r8(vvdw12,vvdw6);
 755             fvdw             = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_add_v2r8(fvdw6,fvdw12),_fjsp_mul_v2r8(vftabscale,rinv00)));
 756
 757             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 758
 759             /* Update potential sum for this i atom from the interaction with this j atom. */
 760             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 761             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 762             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 763             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
 764             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 765             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 766
 767             fscal            = _fjsp_add_v2r8(felec,fvdw);
 768
 769             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 770
 771             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 772
 773             /* Update vectorial force */
 774             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 775             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 776             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 777
 778             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 779             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 780             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 781
 782             }
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 789             {
 790
 791             /* REACTION-FIELD ELECTROSTATICS */
 792             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 793             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 794
 795             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 796
 797             /* Update potential sum for this i atom from the interaction with this j atom. */
 798             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 799             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 800             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 801
 802             fscal            = felec;
 803
 804             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 805
 806             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 807
 808             /* Update vectorial force */
 809             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 810             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 811             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 812
 813             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 814             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 815             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 816
 817             }
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 824             {
 825
 826             /* REACTION-FIELD ELECTROSTATICS */
 827             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 828             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 829
 830             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 831
 832             /* Update potential sum for this i atom from the interaction with this j atom. */
 833             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 834             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 835             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 836
 837             fscal            = felec;
 838
 839             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 840
 841             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 842
 843             /* Update vectorial force */
 844             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 845             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 846             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 847
 848             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 849             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 850             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 851
 852             }
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 859             {
 860
 861             /* REACTION-FIELD ELECTROSTATICS */
 862             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 863             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 864
 865             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 866
 867             /* Update potential sum for this i atom from the interaction with this j atom. */
 868             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 869             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 870             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 871
 872             fscal            = felec;
 873
 874             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 875
 876             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 877
 878             /* Update vectorial force */
 879             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 880             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 881             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 882
 883             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 884             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 885             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 886
 887             }
 888
 889             /**************************
 890              * CALCULATE INTERACTIONS *
 891              **************************/
 892
 893             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 894             {
 895
 896             /* REACTION-FIELD ELECTROSTATICS */
 897             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 898             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 899
 900             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 901
 902             /* Update potential sum for this i atom from the interaction with this j atom. */
 903             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 904             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 905             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 906
 907             fscal            = felec;
 908
 909             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 910
 911             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 912
 913             /* Update vectorial force */
 914             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 915             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 916             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 917
 918             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 919             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 920             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 921
 922             }
 923
 924             /**************************
 925              * CALCULATE INTERACTIONS *
 926              **************************/
 927
 928             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 929             {
 930
 931             /* REACTION-FIELD ELECTROSTATICS */
 932             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 933             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 934
 935             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 936
 937             /* Update potential sum for this i atom from the interaction with this j atom. */
 938             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 939             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 940             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 941
 942             fscal            = felec;
 943
 944             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 945
 946             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 947
 948             /* Update vectorial force */
 949             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 950             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 951             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 952
 953             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 954             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 955             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 956
 957             }
 958
 959             /**************************
 960              * CALCULATE INTERACTIONS *
 961              **************************/
 962
 963             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 964             {
 965
 966             /* REACTION-FIELD ELECTROSTATICS */
 967             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 968             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 969
 970             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 971
 972             /* Update potential sum for this i atom from the interaction with this j atom. */
 973             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 974             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 975             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 976
 977             fscal            = felec;
 978
 979             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 980
 981             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 982
 983             /* Update vectorial force */
 984             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 985             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 986             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 987
 988             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 989             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 990             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 991
 992             }
 993
 994             /**************************
 995              * CALCULATE INTERACTIONS *
 996              **************************/
 997
 998             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 999             {
1000
1001             /* REACTION-FIELD ELECTROSTATICS */
1002             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
1003             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1004
1005             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1006
1007             /* Update potential sum for this i atom from the interaction with this j atom. */
1008             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1009             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1010             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1011
1012             fscal            = felec;
1013
1014             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1015
1016             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1017
1018             /* Update vectorial force */
1019             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1020             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1021             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1022
1023             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1024             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1025             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1026
1027             }
1028
1029             /**************************
1030              * CALCULATE INTERACTIONS *
1031              **************************/
1032
1033             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1034             {
1035
1036             /* REACTION-FIELD ELECTROSTATICS */
1037             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
1038             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1039
1040             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1041
1042             /* Update potential sum for this i atom from the interaction with this j atom. */
1043             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1044             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1045             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1046
1047             fscal            = felec;
1048
1049             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1050
1051             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1052
1053             /* Update vectorial force */
1054             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1055             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1056             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1057
1058             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1059             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1060             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1061
1062             }
1063
1064             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1065
1066             /* Inner loop uses 387 flops */
1067         }
1068
1069         /* End of innermost loop */
1070
1071         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1072                                               f+i_coord_offset,fshift+i_shift_offset);
1073
1074         ggid                        = gid[iidx];
1075         /* Update potential energies */
1076         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1077         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
1078
1079         /* Increment number of inner iterations */
1080         inneriter                  += j_index_end - j_index_start;
1081
1082         /* Outer loop uses 20 flops */
1083     }
1084
1085     /* Increment number of outer iterations */
1086     outeriter        += nri;
1087
1088     /* Update outer/inner flops */
1089
1090     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*387);
1091 }
1092 /*
1093  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sparc64_hpc_ace_double
1094  * Electrostatics interaction: ReactionField
1095  * VdW interaction:            CubicSplineTable
1096  * Geometry:                   Water3-Water3
1097  * Calculate force/pot:        Force
1098  */
1099 void
1100 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sparc64_hpc_ace_double
1101                     (t_nblist                    * gmx_restrict       nlist,
1102                      rvec                        * gmx_restrict          xx,
1103                      rvec                        * gmx_restrict          ff,
1104                      t_forcerec                  * gmx_restrict          fr,
1105                      t_mdatoms                   * gmx_restrict     mdatoms,
1106                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1107                      t_nrnb                      * gmx_restrict        nrnb)
1108 {
1109     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1110      * just 0 for non-waters.
1111      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1112      * jnr indices corresponding to data put in the four positions in the SIMD register.
1113      */
1114     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1115     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1116     int              jnrA,jnrB;
1117     int              j_coord_offsetA,j_coord_offsetB;
1118     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1119     real             rcutoff_scalar;
1120     real             *shiftvec,*fshift,*x,*f;
1121     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1122     int              vdwioffset0;
1123     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1124     int              vdwioffset1;
1125     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1126     int              vdwioffset2;
1127     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1128     int              vdwjidx0A,vdwjidx0B;
1129     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1130     int              vdwjidx1A,vdwjidx1B;
1131     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1132     int              vdwjidx2A,vdwjidx2B;
1133     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1134     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1135     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1136     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1137     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1138     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1139     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1140     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1141     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1142     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1143     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1144     real             *charge;
1145     int              nvdwtype;
1146     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1147     int              *vdwtype;
1148     real             *vdwparam;
1149     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
1150     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
1151     _fjsp_v2r8       rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1152     real             *vftab;
1153     _fjsp_v2r8       itab_tmp;
1154     _fjsp_v2r8       dummy_mask,cutoff_mask;
1155     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1156     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1157     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1158
1159     x                = xx[0];
1160     f                = ff[0];
1161
1162     nri              = nlist->nri;
1163     iinr             = nlist->iinr;
1164     jindex           = nlist->jindex;
1165     jjnr             = nlist->jjnr;
1166     shiftidx         = nlist->shift;
1167     gid              = nlist->gid;
1168     shiftvec         = fr->shift_vec[0];
1169     fshift           = fr->fshift[0];
1170     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1171     charge           = mdatoms->chargeA;
1172     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
1173     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
1174     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
1175     nvdwtype         = fr->ntype;
1176     vdwparam         = fr->nbfp;
1177     vdwtype          = mdatoms->typeA;
1178
1179     vftab            = kernel_data->table_vdw->data;
1180     vftabscale       = gmx_fjsp_set1_v2r8(kernel_data->table_vdw->scale);
1181
1182     /* Setup water-specific parameters */
1183     inr              = nlist->iinr[0];
1184     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1185     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1186     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1187     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1188
1189     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1190     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1191     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1192     vdwjidx0A        = 2*vdwtype[inr+0];
1193     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1194     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
1195     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
1196     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1197     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1198     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1199     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1200     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1201     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1202     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1203     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1204
1205     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1206     rcutoff_scalar   = fr->rcoulomb;
1207     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1208     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1209
1210     /* Avoid stupid compiler warnings */
1211     jnrA = jnrB = 0;
1212     j_coord_offsetA = 0;
1213     j_coord_offsetB = 0;
1214
1215     outeriter        = 0;
1216     inneriter        = 0;
1217
1218     /* Start outer loop over neighborlists */
1219     for(iidx=0; iidx<nri; iidx++)
1220     {
1221         /* Load shift vector for this list */
1222         i_shift_offset   = DIM*shiftidx[iidx];
1223
1224         /* Load limits for loop over neighbors */
1225         j_index_start    = jindex[iidx];
1226         j_index_end      = jindex[iidx+1];
1227
1228         /* Get outer coordinate index */
1229         inr              = iinr[iidx];
1230         i_coord_offset   = DIM*inr;
1231
1232         /* Load i particle coords and add shift vector */
1233         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1234                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1235
1236         fix0             = _fjsp_setzero_v2r8();
1237         fiy0             = _fjsp_setzero_v2r8();
1238         fiz0             = _fjsp_setzero_v2r8();
1239         fix1             = _fjsp_setzero_v2r8();
1240         fiy1             = _fjsp_setzero_v2r8();
1241         fiz1             = _fjsp_setzero_v2r8();
1242         fix2             = _fjsp_setzero_v2r8();
1243         fiy2             = _fjsp_setzero_v2r8();
1244         fiz2             = _fjsp_setzero_v2r8();
1245
1246         /* Start inner kernel loop */
1247         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1248         {
1249
1250             /* Get j neighbor index, and coordinate index */
1251             jnrA             = jjnr[jidx];
1252             jnrB             = jjnr[jidx+1];
1253             j_coord_offsetA  = DIM*jnrA;
1254             j_coord_offsetB  = DIM*jnrB;
1255
1256             /* load j atom coordinates */
1257             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1258                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1259
1260             /* Calculate displacement vector */
1261             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1262             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1263             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1264             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1265             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1266             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1267             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1268             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1269             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1270             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1271             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1272             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1273             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1274             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1275             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1276             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1277             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1278             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1279             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1280             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1281             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1282             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1283             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1284             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1285             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1286             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1287             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1288
1289             /* Calculate squared distance and things based on it */
1290             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1291             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1292             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1293             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1294             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1295             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1296             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1297             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1298             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1299
1300             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1301             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1302             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1303             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1304             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1305             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1306             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1307             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1308             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1309
1310             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1311             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1312             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1313             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1314             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1315             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1316             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1317             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1318             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1319
1320             fjx0             = _fjsp_setzero_v2r8();
1321             fjy0             = _fjsp_setzero_v2r8();
1322             fjz0             = _fjsp_setzero_v2r8();
1323             fjx1             = _fjsp_setzero_v2r8();
1324             fjy1             = _fjsp_setzero_v2r8();
1325             fjz1             = _fjsp_setzero_v2r8();
1326             fjx2             = _fjsp_setzero_v2r8();
1327             fjy2             = _fjsp_setzero_v2r8();
1328             fjz2             = _fjsp_setzero_v2r8();
1329
1330             /**************************
1331              * CALCULATE INTERACTIONS *
1332              **************************/
1333
1334             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1335             {
1336
1337             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1338
1339             /* Calculate table index by multiplying r with table scale and truncate to integer */
1340             rt               = _fjsp_mul_v2r8(r00,vftabscale);
1341             itab_tmp         = _fjsp_dtox_v2r8(rt);
1342             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1343             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1344             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1345
1346             vfconv.i[0]     *= 8;
1347             vfconv.i[1]     *= 8;
1348
1349             /* REACTION-FIELD ELECTROSTATICS */
1350             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1351
1352             /* CUBIC SPLINE TABLE DISPERSION */
1353             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1354             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] );
1355             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1356             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 2 );
1357             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] + 2 );
1358             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1359             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
1360             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
1361             fvdw6            = _fjsp_mul_v2r8(c6_00,FF);
1362
1363             /* CUBIC SPLINE TABLE REPULSION */
1364             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 4 );
1365             F                = _fjsp_load_v2r8( vftab + vfconv.i[1] + 4 );
1366             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1367             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 6 );
1368             H                = _fjsp_load_v2r8( vftab + vfconv.i[1] + 6 );
1369             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1370             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
1371             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
1372             fvdw12           = _fjsp_mul_v2r8(c12_00,FF);
1373             fvdw             = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_add_v2r8(fvdw6,fvdw12),_fjsp_mul_v2r8(vftabscale,rinv00)));
1374
1375             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1376
1377             fscal            = _fjsp_add_v2r8(felec,fvdw);
1378
1379             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1380
1381             /* Update vectorial force */
1382             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1383             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1384             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1385
1386             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1387             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1388             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1389
1390             }
1391
1392             /**************************
1393              * CALCULATE INTERACTIONS *
1394              **************************/
1395
1396             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1397             {
1398
1399             /* REACTION-FIELD ELECTROSTATICS */
1400             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1401
1402             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1403
1404             fscal            = felec;
1405
1406             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1407
1408             /* Update vectorial force */
1409             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1410             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1411             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1412
1413             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1414             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1415             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1416
1417             }
1418
1419             /**************************
1420              * CALCULATE INTERACTIONS *
1421              **************************/
1422
1423             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1424             {
1425
1426             /* REACTION-FIELD ELECTROSTATICS */
1427             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1428
1429             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1430
1431             fscal            = felec;
1432
1433             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1434
1435             /* Update vectorial force */
1436             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1437             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1438             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1439
1440             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1441             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1442             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1443
1444             }
1445
1446             /**************************
1447              * CALCULATE INTERACTIONS *
1448              **************************/
1449
1450             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1451             {
1452
1453             /* REACTION-FIELD ELECTROSTATICS */
1454             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1455
1456             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1457
1458             fscal            = felec;
1459
1460             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1461
1462             /* Update vectorial force */
1463             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1464             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1465             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1466
1467             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1468             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1469             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1470
1471             }
1472
1473             /**************************
1474              * CALCULATE INTERACTIONS *
1475              **************************/
1476
1477             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1478             {
1479
1480             /* REACTION-FIELD ELECTROSTATICS */
1481             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1482
1483             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1484
1485             fscal            = felec;
1486
1487             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1488
1489             /* Update vectorial force */
1490             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1491             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1492             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1493
1494             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1495             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1496             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1497
1498             }
1499
1500             /**************************
1501              * CALCULATE INTERACTIONS *
1502              **************************/
1503
1504             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1505             {
1506
1507             /* REACTION-FIELD ELECTROSTATICS */
1508             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1509
1510             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1511
1512             fscal            = felec;
1513
1514             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1515
1516             /* Update vectorial force */
1517             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1518             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1519             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1520
1521             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1522             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1523             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1524
1525             }
1526
1527             /**************************
1528              * CALCULATE INTERACTIONS *
1529              **************************/
1530
1531             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1532             {
1533
1534             /* REACTION-FIELD ELECTROSTATICS */
1535             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1536
1537             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1538
1539             fscal            = felec;
1540
1541             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1542
1543             /* Update vectorial force */
1544             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1545             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1546             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1547
1548             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1549             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1550             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1551
1552             }
1553
1554             /**************************
1555              * CALCULATE INTERACTIONS *
1556              **************************/
1557
1558             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1559             {
1560
1561             /* REACTION-FIELD ELECTROSTATICS */
1562             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1563
1564             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1565
1566             fscal            = felec;
1567
1568             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1569
1570             /* Update vectorial force */
1571             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1572             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1573             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1574
1575             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1576             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1577             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1578
1579             }
1580
1581             /**************************
1582              * CALCULATE INTERACTIONS *
1583              **************************/
1584
1585             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1586             {
1587
1588             /* REACTION-FIELD ELECTROSTATICS */
1589             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1590
1591             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1592
1593             fscal            = felec;
1594
1595             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1596
1597             /* Update vectorial force */
1598             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1599             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1600             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1601
1602             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1603             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1604             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1605
1606             }
1607
1608             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1609
1610             /* Inner loop uses 324 flops */
1611         }
1612
1613         if(jidx<j_index_end)
1614         {
1615
1616             jnrA             = jjnr[jidx];
1617             j_coord_offsetA  = DIM*jnrA;
1618
1619             /* load j atom coordinates */
1620             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1621                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1622
1623             /* Calculate displacement vector */
1624             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1625             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1626             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1627             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1628             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1629             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1630             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1631             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1632             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1633             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1634             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1635             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1636             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1637             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1638             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1639             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1640             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1641             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1642             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1643             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1644             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1645             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1646             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1647             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1648             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1649             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1650             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1651
1652             /* Calculate squared distance and things based on it */
1653             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1654             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1655             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1656             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1657             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1658             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1659             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1660             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1661             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1662
1663             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1664             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1665             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1666             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1667             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1668             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1669             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1670             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1671             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1672
1673             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1674             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1675             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1676             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1677             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1678             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1679             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1680             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1681             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1682
1683             fjx0             = _fjsp_setzero_v2r8();
1684             fjy0             = _fjsp_setzero_v2r8();
1685             fjz0             = _fjsp_setzero_v2r8();
1686             fjx1             = _fjsp_setzero_v2r8();
1687             fjy1             = _fjsp_setzero_v2r8();
1688             fjz1             = _fjsp_setzero_v2r8();
1689             fjx2             = _fjsp_setzero_v2r8();
1690             fjy2             = _fjsp_setzero_v2r8();
1691             fjz2             = _fjsp_setzero_v2r8();
1692
1693             /**************************
1694              * CALCULATE INTERACTIONS *
1695              **************************/
1696
1697             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1698             {
1699
1700             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1701
1702             /* Calculate table index by multiplying r with table scale and truncate to integer */
1703             rt               = _fjsp_mul_v2r8(r00,vftabscale);
1704             itab_tmp         = _fjsp_dtox_v2r8(rt);
1705             vfeps            = _fjsp_sub_v2r8(rt, _fjsp_xtod_v2r8(itab_tmp));
1706             twovfeps         = _fjsp_add_v2r8(vfeps,vfeps);
1707             _fjsp_store_v2r8(&vfconv.simd,itab_tmp);
1708
1709             vfconv.i[0]     *= 8;
1710             vfconv.i[1]     *= 8;
1711
1712             /* REACTION-FIELD ELECTROSTATICS */
1713             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1714
1715             /* CUBIC SPLINE TABLE DISPERSION */
1716             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] );
1717             F                = _fjsp_setzero_v2r8();
1718             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1719             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 2 );
1720             H                = _fjsp_setzero_v2r8();
1721             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1722             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
1723             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
1724             fvdw6            = _fjsp_mul_v2r8(c6_00,FF);
1725
1726             /* CUBIC SPLINE TABLE REPULSION */
1727             Y                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 4 );
1728             F                = _fjsp_setzero_v2r8();
1729             GMX_FJSP_TRANSPOSE2_V2R8(Y,F);
1730             G                = _fjsp_load_v2r8( vftab + vfconv.i[0] + 6 );
1731             H                = _fjsp_setzero_v2r8();
1732             GMX_FJSP_TRANSPOSE2_V2R8(G,H);
1733             Fp               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(H,vfeps,G),F);
1734             FF               = _fjsp_madd_v2r8(vfeps,_fjsp_madd_v2r8(twovfeps,H,G),Fp);
1735             fvdw12           = _fjsp_mul_v2r8(c12_00,FF);
1736             fvdw             = _fjsp_neg_v2r8(_fjsp_mul_v2r8(_fjsp_add_v2r8(fvdw6,fvdw12),_fjsp_mul_v2r8(vftabscale,rinv00)));
1737
1738             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1739
1740             fscal            = _fjsp_add_v2r8(felec,fvdw);
1741
1742             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1743
1744             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1745
1746             /* Update vectorial force */
1747             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1748             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1749             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1750
1751             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1752             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1753             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1754
1755             }
1756
1757             /**************************
1758              * CALCULATE INTERACTIONS *
1759              **************************/
1760
1761             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1762             {
1763
1764             /* REACTION-FIELD ELECTROSTATICS */
1765             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1766
1767             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1768
1769             fscal            = felec;
1770
1771             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1772
1773             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1774
1775             /* Update vectorial force */
1776             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1777             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1778             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1779
1780             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1781             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1782             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1783
1784             }
1785
1786             /**************************
1787              * CALCULATE INTERACTIONS *
1788              **************************/
1789
1790             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1791             {
1792
1793             /* REACTION-FIELD ELECTROSTATICS */
1794             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1795
1796             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1797
1798             fscal            = felec;
1799
1800             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1801
1802             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1803
1804             /* Update vectorial force */
1805             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1806             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1807             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1808
1809             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1810             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1811             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1812
1813             }
1814
1815             /**************************
1816              * CALCULATE INTERACTIONS *
1817              **************************/
1818
1819             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1820             {
1821
1822             /* REACTION-FIELD ELECTROSTATICS */
1823             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1824
1825             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1826
1827             fscal            = felec;
1828
1829             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1830
1831             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1832
1833             /* Update vectorial force */
1834             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1835             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1836             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1837
1838             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1839             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1840             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1841
1842             }
1843
1844             /**************************
1845              * CALCULATE INTERACTIONS *
1846              **************************/
1847
1848             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1849             {
1850
1851             /* REACTION-FIELD ELECTROSTATICS */
1852             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1853
1854             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1855
1856             fscal            = felec;
1857
1858             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1859
1860             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1861
1862             /* Update vectorial force */
1863             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1864             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1865             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1866
1867             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1868             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1869             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1870
1871             }
1872
1873             /**************************
1874              * CALCULATE INTERACTIONS *
1875              **************************/
1876
1877             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1878             {
1879
1880             /* REACTION-FIELD ELECTROSTATICS */
1881             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1882
1883             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1884
1885             fscal            = felec;
1886
1887             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1888
1889             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1890
1891             /* Update vectorial force */
1892             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1893             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1894             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1895
1896             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1897             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1898             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1899
1900             }
1901
1902             /**************************
1903              * CALCULATE INTERACTIONS *
1904              **************************/
1905
1906             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1907             {
1908
1909             /* REACTION-FIELD ELECTROSTATICS */
1910             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1911
1912             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1913
1914             fscal            = felec;
1915
1916             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1917
1918             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1919
1920             /* Update vectorial force */
1921             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1922             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1923             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1924
1925             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1926             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1927             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1928
1929             }
1930
1931             /**************************
1932              * CALCULATE INTERACTIONS *
1933              **************************/
1934
1935             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1936             {
1937
1938             /* REACTION-FIELD ELECTROSTATICS */
1939             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1940
1941             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1942
1943             fscal            = felec;
1944
1945             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1946
1947             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1948
1949             /* Update vectorial force */
1950             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1951             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1952             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1953
1954             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1955             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1956             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1957
1958             }
1959
1960             /**************************
1961              * CALCULATE INTERACTIONS *
1962              **************************/
1963
1964             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1965             {
1966
1967             /* REACTION-FIELD ELECTROSTATICS */
1968             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1969
1970             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1971
1972             fscal            = felec;
1973
1974             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1975
1976             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1977
1978             /* Update vectorial force */
1979             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1980             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1981             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1982
1983             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1984             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1985             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1986
1987             }
1988
1989             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1990
1991             /* Inner loop uses 324 flops */
1992         }
1993
1994         /* End of innermost loop */
1995
1996         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1997                                               f+i_coord_offset,fshift+i_shift_offset);
1998
1999         /* Increment number of inner iterations */
2000         inneriter                  += j_index_end - j_index_start;
2001
2002         /* Outer loop uses 18 flops */
2003     }
2004
2005     /* Increment number of outer iterations */
2006     outeriter        += nri;
2007
2008     /* Update outer/inner flops */
2009
2010     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*324);
2011 }