src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRF_VdwLJ_GeomW3W3_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sparc64_hpc_ace_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "gromacs/legacyheaders/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: ReactionField
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water3-Water3
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  94     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  95     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  96     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  97     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 100     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 103     real             *charge;
 104     int              nvdwtype;
 105     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 106     int              *vdwtype;
 107     real             *vdwparam;
 108     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 109     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 110     _fjsp_v2r8       itab_tmp;
 111     _fjsp_v2r8       dummy_mask,cutoff_mask;
 112     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 113     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 114     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 115
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 128     charge           = mdatoms->chargeA;
 129     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 130     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 131     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 132     nvdwtype         = fr->ntype;
 133     vdwparam         = fr->nbfp;
 134     vdwtype          = mdatoms->typeA;
 135
 136     /* Setup water-specific parameters */
 137     inr              = nlist->iinr[0];
 138     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 139     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 140     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 141     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 142
 143     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 144     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 145     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 146     vdwjidx0A        = 2*vdwtype[inr+0];
 147     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 148     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 149     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 150     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 151     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 152     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 153     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 154     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 155     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 156     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 157     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 158
 159     /* Avoid stupid compiler warnings */
 160     jnrA = jnrB = 0;
 161     j_coord_offsetA = 0;
 162     j_coord_offsetB = 0;
 163
 164     outeriter        = 0;
 165     inneriter        = 0;
 166
 167     /* Start outer loop over neighborlists */
 168     for(iidx=0; iidx<nri; iidx++)
 169     {
 170         /* Load shift vector for this list */
 171         i_shift_offset   = DIM*shiftidx[iidx];
 172
 173         /* Load limits for loop over neighbors */
 174         j_index_start    = jindex[iidx];
 175         j_index_end      = jindex[iidx+1];
 176
 177         /* Get outer coordinate index */
 178         inr              = iinr[iidx];
 179         i_coord_offset   = DIM*inr;
 180
 181         /* Load i particle coords and add shift vector */
 182         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 183                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 184
 185         fix0             = _fjsp_setzero_v2r8();
 186         fiy0             = _fjsp_setzero_v2r8();
 187         fiz0             = _fjsp_setzero_v2r8();
 188         fix1             = _fjsp_setzero_v2r8();
 189         fiy1             = _fjsp_setzero_v2r8();
 190         fiz1             = _fjsp_setzero_v2r8();
 191         fix2             = _fjsp_setzero_v2r8();
 192         fiy2             = _fjsp_setzero_v2r8();
 193         fiz2             = _fjsp_setzero_v2r8();
 194
 195         /* Reset potential sums */
 196         velecsum         = _fjsp_setzero_v2r8();
 197         vvdwsum          = _fjsp_setzero_v2r8();
 198
 199         /* Start inner kernel loop */
 200         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 201         {
 202
 203             /* Get j neighbor index, and coordinate index */
 204             jnrA             = jjnr[jidx];
 205             jnrB             = jjnr[jidx+1];
 206             j_coord_offsetA  = DIM*jnrA;
 207             j_coord_offsetB  = DIM*jnrB;
 208
 209             /* load j atom coordinates */
 210             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 211                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 212
 213             /* Calculate displacement vector */
 214             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 215             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 216             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 217             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 218             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 219             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 220             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 221             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 222             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 223             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 224             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 225             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 226             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 227             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 228             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 229             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 230             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 231             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 232             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 233             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 234             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 235             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 236             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 237             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 238             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 239             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 240             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 241
 242             /* Calculate squared distance and things based on it */
 243             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 244             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 245             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 246             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 247             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 248             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 249             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 250             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 251             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 252
 253             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 254             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 255             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 256             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 257             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 258             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 259             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 260             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 261             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 262
 263             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 264             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 265             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 266             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 267             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 268             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 269             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 270             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 271             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 272
 273             fjx0             = _fjsp_setzero_v2r8();
 274             fjy0             = _fjsp_setzero_v2r8();
 275             fjz0             = _fjsp_setzero_v2r8();
 276             fjx1             = _fjsp_setzero_v2r8();
 277             fjy1             = _fjsp_setzero_v2r8();
 278             fjz1             = _fjsp_setzero_v2r8();
 279             fjx2             = _fjsp_setzero_v2r8();
 280             fjy2             = _fjsp_setzero_v2r8();
 281             fjz2             = _fjsp_setzero_v2r8();
 282
 283             /**************************
 284              * CALCULATE INTERACTIONS *
 285              **************************/
 286
 287             /* REACTION-FIELD ELECTROSTATICS */
 288             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 289             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 290
 291             /* LENNARD-JONES DISPERSION/REPULSION */
 292
 293             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 294             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 295             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 296             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 297             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 298
 299             /* Update potential sum for this i atom from the interaction with this j atom. */
 300             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 301             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 302
 303             fscal            = _fjsp_add_v2r8(felec,fvdw);
 304
 305             /* Update vectorial force */
 306             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 307             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 308             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 309
 310             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 311             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 312             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 313
 314             /**************************
 315              * CALCULATE INTERACTIONS *
 316              **************************/
 317
 318             /* REACTION-FIELD ELECTROSTATICS */
 319             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 320             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 321
 322             /* Update potential sum for this i atom from the interaction with this j atom. */
 323             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 324
 325             fscal            = felec;
 326
 327             /* Update vectorial force */
 328             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 329             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 330             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 331
 332             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 333             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 334             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 335
 336             /**************************
 337              * CALCULATE INTERACTIONS *
 338              **************************/
 339
 340             /* REACTION-FIELD ELECTROSTATICS */
 341             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 342             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 346
 347             fscal            = felec;
 348
 349             /* Update vectorial force */
 350             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 351             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 352             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 353
 354             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 355             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 356             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 357
 358             /**************************
 359              * CALCULATE INTERACTIONS *
 360              **************************/
 361
 362             /* REACTION-FIELD ELECTROSTATICS */
 363             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 364             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 365
 366             /* Update potential sum for this i atom from the interaction with this j atom. */
 367             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 368
 369             fscal            = felec;
 370
 371             /* Update vectorial force */
 372             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 373             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 374             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 375
 376             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 377             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 378             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 379
 380             /**************************
 381              * CALCULATE INTERACTIONS *
 382              **************************/
 383
 384             /* REACTION-FIELD ELECTROSTATICS */
 385             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 386             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 387
 388             /* Update potential sum for this i atom from the interaction with this j atom. */
 389             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 390
 391             fscal            = felec;
 392
 393             /* Update vectorial force */
 394             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 395             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 396             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 397
 398             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 399             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 400             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 401
 402             /**************************
 403              * CALCULATE INTERACTIONS *
 404              **************************/
 405
 406             /* REACTION-FIELD ELECTROSTATICS */
 407             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 408             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 409
 410             /* Update potential sum for this i atom from the interaction with this j atom. */
 411             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 412
 413             fscal            = felec;
 414
 415             /* Update vectorial force */
 416             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 417             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 418             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 419
 420             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 421             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 422             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 423
 424             /**************************
 425              * CALCULATE INTERACTIONS *
 426              **************************/
 427
 428             /* REACTION-FIELD ELECTROSTATICS */
 429             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 430             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 434
 435             fscal            = felec;
 436
 437             /* Update vectorial force */
 438             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 439             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 440             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 441
 442             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 443             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 444             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 445
 446             /**************************
 447              * CALCULATE INTERACTIONS *
 448              **************************/
 449
 450             /* REACTION-FIELD ELECTROSTATICS */
 451             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 452             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 453
 454             /* Update potential sum for this i atom from the interaction with this j atom. */
 455             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 456
 457             fscal            = felec;
 458
 459             /* Update vectorial force */
 460             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 461             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 462             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 463
 464             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 465             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 466             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 467
 468             /**************************
 469              * CALCULATE INTERACTIONS *
 470              **************************/
 471
 472             /* REACTION-FIELD ELECTROSTATICS */
 473             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 474             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 475
 476             /* Update potential sum for this i atom from the interaction with this j atom. */
 477             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 478
 479             fscal            = felec;
 480
 481             /* Update vectorial force */
 482             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 483             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 484             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 485
 486             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 487             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 488             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 489
 490             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 491
 492             /* Inner loop uses 327 flops */
 493         }
 494
 495         if(jidx<j_index_end)
 496         {
 497
 498             jnrA             = jjnr[jidx];
 499             j_coord_offsetA  = DIM*jnrA;
 500
 501             /* load j atom coordinates */
 502             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 503                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 504
 505             /* Calculate displacement vector */
 506             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 507             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 508             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 509             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 510             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 511             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 512             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 513             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 514             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 515             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 516             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 517             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 518             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 519             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 520             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 521             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 522             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 523             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 524             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 525             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 526             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 527             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 528             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 529             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 530             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 531             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 532             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 533
 534             /* Calculate squared distance and things based on it */
 535             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 536             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 537             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 538             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 539             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 540             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 541             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 542             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 543             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 544
 545             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 546             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 547             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 548             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 549             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 550             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 551             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 552             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 553             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 554
 555             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 556             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 557             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 558             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 559             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 560             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 561             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 562             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 563             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 564
 565             fjx0             = _fjsp_setzero_v2r8();
 566             fjy0             = _fjsp_setzero_v2r8();
 567             fjz0             = _fjsp_setzero_v2r8();
 568             fjx1             = _fjsp_setzero_v2r8();
 569             fjy1             = _fjsp_setzero_v2r8();
 570             fjz1             = _fjsp_setzero_v2r8();
 571             fjx2             = _fjsp_setzero_v2r8();
 572             fjy2             = _fjsp_setzero_v2r8();
 573             fjz2             = _fjsp_setzero_v2r8();
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             /* REACTION-FIELD ELECTROSTATICS */
 580             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 581             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 582
 583             /* LENNARD-JONES DISPERSION/REPULSION */
 584
 585             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 586             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 587             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 588             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 589             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 590
 591             /* Update potential sum for this i atom from the interaction with this j atom. */
 592             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 593             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 594             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 595             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 596
 597             fscal            = _fjsp_add_v2r8(felec,fvdw);
 598
 599             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 600
 601             /* Update vectorial force */
 602             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 603             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 604             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 605
 606             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 607             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 608             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 609
 610             /**************************
 611              * CALCULATE INTERACTIONS *
 612              **************************/
 613
 614             /* REACTION-FIELD ELECTROSTATICS */
 615             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 616             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 617
 618             /* Update potential sum for this i atom from the interaction with this j atom. */
 619             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 620             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 621
 622             fscal            = felec;
 623
 624             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 625
 626             /* Update vectorial force */
 627             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 628             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 629             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 630
 631             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 632             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 633             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 634
 635             /**************************
 636              * CALCULATE INTERACTIONS *
 637              **************************/
 638
 639             /* REACTION-FIELD ELECTROSTATICS */
 640             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 641             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 642
 643             /* Update potential sum for this i atom from the interaction with this j atom. */
 644             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 645             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 646
 647             fscal            = felec;
 648
 649             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 650
 651             /* Update vectorial force */
 652             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 653             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 654             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 655
 656             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 657             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 658             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 659
 660             /**************************
 661              * CALCULATE INTERACTIONS *
 662              **************************/
 663
 664             /* REACTION-FIELD ELECTROSTATICS */
 665             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 666             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 667
 668             /* Update potential sum for this i atom from the interaction with this j atom. */
 669             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 670             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 671
 672             fscal            = felec;
 673
 674             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 675
 676             /* Update vectorial force */
 677             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 678             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 679             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 680
 681             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 682             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 683             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 684
 685             /**************************
 686              * CALCULATE INTERACTIONS *
 687              **************************/
 688
 689             /* REACTION-FIELD ELECTROSTATICS */
 690             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 691             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 692
 693             /* Update potential sum for this i atom from the interaction with this j atom. */
 694             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 695             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 696
 697             fscal            = felec;
 698
 699             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 700
 701             /* Update vectorial force */
 702             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 703             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 704             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 705
 706             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 707             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 708             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 709
 710             /**************************
 711              * CALCULATE INTERACTIONS *
 712              **************************/
 713
 714             /* REACTION-FIELD ELECTROSTATICS */
 715             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 716             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 717
 718             /* Update potential sum for this i atom from the interaction with this j atom. */
 719             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 720             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 721
 722             fscal            = felec;
 723
 724             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 725
 726             /* Update vectorial force */
 727             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 728             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 729             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 730
 731             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 732             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 733             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 734
 735             /**************************
 736              * CALCULATE INTERACTIONS *
 737              **************************/
 738
 739             /* REACTION-FIELD ELECTROSTATICS */
 740             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 741             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 742
 743             /* Update potential sum for this i atom from the interaction with this j atom. */
 744             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 745             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 746
 747             fscal            = felec;
 748
 749             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 750
 751             /* Update vectorial force */
 752             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 753             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 754             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 755
 756             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 757             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 758             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 759
 760             /**************************
 761              * CALCULATE INTERACTIONS *
 762              **************************/
 763
 764             /* REACTION-FIELD ELECTROSTATICS */
 765             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 766             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 767
 768             /* Update potential sum for this i atom from the interaction with this j atom. */
 769             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 770             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 771
 772             fscal            = felec;
 773
 774             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 775
 776             /* Update vectorial force */
 777             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 778             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 779             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 780
 781             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 782             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 783             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 784
 785             /**************************
 786              * CALCULATE INTERACTIONS *
 787              **************************/
 788
 789             /* REACTION-FIELD ELECTROSTATICS */
 790             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 791             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 792
 793             /* Update potential sum for this i atom from the interaction with this j atom. */
 794             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 795             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 796
 797             fscal            = felec;
 798
 799             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 800
 801             /* Update vectorial force */
 802             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 803             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 804             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 805
 806             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 807             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 808             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 809
 810             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 811
 812             /* Inner loop uses 327 flops */
 813         }
 814
 815         /* End of innermost loop */
 816
 817         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 818                                               f+i_coord_offset,fshift+i_shift_offset);
 819
 820         ggid                        = gid[iidx];
 821         /* Update potential energies */
 822         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 823         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 824
 825         /* Increment number of inner iterations */
 826         inneriter                  += j_index_end - j_index_start;
 827
 828         /* Outer loop uses 20 flops */
 829     }
 830
 831     /* Increment number of outer iterations */
 832     outeriter        += nri;
 833
 834     /* Update outer/inner flops */
 835
 836     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*327);
 837 }
 838 /*
 839  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
 840  * Electrostatics interaction: ReactionField
 841  * VdW interaction:            LennardJones
 842  * Geometry:                   Water3-Water3
 843  * Calculate force/pot:        Force
 844  */
 845 void
 846 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sparc64_hpc_ace_double
 847                     (t_nblist                    * gmx_restrict       nlist,
 848                      rvec                        * gmx_restrict          xx,
 849                      rvec                        * gmx_restrict          ff,
 850                      t_forcerec                  * gmx_restrict          fr,
 851                      t_mdatoms                   * gmx_restrict     mdatoms,
 852                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 853                      t_nrnb                      * gmx_restrict        nrnb)
 854 {
 855     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 856      * just 0 for non-waters.
 857      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 858      * jnr indices corresponding to data put in the four positions in the SIMD register.
 859      */
 860     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 861     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 862     int              jnrA,jnrB;
 863     int              j_coord_offsetA,j_coord_offsetB;
 864     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 865     real             rcutoff_scalar;
 866     real             *shiftvec,*fshift,*x,*f;
 867     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 868     int              vdwioffset0;
 869     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 870     int              vdwioffset1;
 871     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 872     int              vdwioffset2;
 873     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 874     int              vdwjidx0A,vdwjidx0B;
 875     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 876     int              vdwjidx1A,vdwjidx1B;
 877     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 878     int              vdwjidx2A,vdwjidx2B;
 879     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 880     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 881     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 882     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 883     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 884     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 885     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 886     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 887     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 888     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 889     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 890     real             *charge;
 891     int              nvdwtype;
 892     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 893     int              *vdwtype;
 894     real             *vdwparam;
 895     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 896     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 897     _fjsp_v2r8       itab_tmp;
 898     _fjsp_v2r8       dummy_mask,cutoff_mask;
 899     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 900     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 901     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 902
 903     x                = xx[0];
 904     f                = ff[0];
 905
 906     nri              = nlist->nri;
 907     iinr             = nlist->iinr;
 908     jindex           = nlist->jindex;
 909     jjnr             = nlist->jjnr;
 910     shiftidx         = nlist->shift;
 911     gid              = nlist->gid;
 912     shiftvec         = fr->shift_vec[0];
 913     fshift           = fr->fshift[0];
 914     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 915     charge           = mdatoms->chargeA;
 916     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 917     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 918     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 919     nvdwtype         = fr->ntype;
 920     vdwparam         = fr->nbfp;
 921     vdwtype          = mdatoms->typeA;
 922
 923     /* Setup water-specific parameters */
 924     inr              = nlist->iinr[0];
 925     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 926     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 927     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 928     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 929
 930     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 931     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 932     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 933     vdwjidx0A        = 2*vdwtype[inr+0];
 934     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 935     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
 936     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
 937     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 938     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 939     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 940     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 941     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 942     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 943     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 944     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 945
 946     /* Avoid stupid compiler warnings */
 947     jnrA = jnrB = 0;
 948     j_coord_offsetA = 0;
 949     j_coord_offsetB = 0;
 950
 951     outeriter        = 0;
 952     inneriter        = 0;
 953
 954     /* Start outer loop over neighborlists */
 955     for(iidx=0; iidx<nri; iidx++)
 956     {
 957         /* Load shift vector for this list */
 958         i_shift_offset   = DIM*shiftidx[iidx];
 959
 960         /* Load limits for loop over neighbors */
 961         j_index_start    = jindex[iidx];
 962         j_index_end      = jindex[iidx+1];
 963
 964         /* Get outer coordinate index */
 965         inr              = iinr[iidx];
 966         i_coord_offset   = DIM*inr;
 967
 968         /* Load i particle coords and add shift vector */
 969         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 970                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 971
 972         fix0             = _fjsp_setzero_v2r8();
 973         fiy0             = _fjsp_setzero_v2r8();
 974         fiz0             = _fjsp_setzero_v2r8();
 975         fix1             = _fjsp_setzero_v2r8();
 976         fiy1             = _fjsp_setzero_v2r8();
 977         fiz1             = _fjsp_setzero_v2r8();
 978         fix2             = _fjsp_setzero_v2r8();
 979         fiy2             = _fjsp_setzero_v2r8();
 980         fiz2             = _fjsp_setzero_v2r8();
 981
 982         /* Start inner kernel loop */
 983         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 984         {
 985
 986             /* Get j neighbor index, and coordinate index */
 987             jnrA             = jjnr[jidx];
 988             jnrB             = jjnr[jidx+1];
 989             j_coord_offsetA  = DIM*jnrA;
 990             j_coord_offsetB  = DIM*jnrB;
 991
 992             /* load j atom coordinates */
 993             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 994                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 995
 996             /* Calculate displacement vector */
 997             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 998             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 999             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1000             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1001             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1002             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1003             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1004             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1005             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1006             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1007             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1008             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1009             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1010             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1011             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1012             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1013             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1014             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1015             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1016             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1017             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1018             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1019             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1020             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1021             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1022             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1023             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1024
1025             /* Calculate squared distance and things based on it */
1026             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1027             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1028             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1029             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1030             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1031             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1032             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1033             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1034             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1035
1036             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1037             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1038             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1039             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1040             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1041             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1042             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1043             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1044             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1045
1046             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1047             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1048             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1049             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1050             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1051             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1052             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1053             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1054             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1055
1056             fjx0             = _fjsp_setzero_v2r8();
1057             fjy0             = _fjsp_setzero_v2r8();
1058             fjz0             = _fjsp_setzero_v2r8();
1059             fjx1             = _fjsp_setzero_v2r8();
1060             fjy1             = _fjsp_setzero_v2r8();
1061             fjz1             = _fjsp_setzero_v2r8();
1062             fjx2             = _fjsp_setzero_v2r8();
1063             fjy2             = _fjsp_setzero_v2r8();
1064             fjz2             = _fjsp_setzero_v2r8();
1065
1066             /**************************
1067              * CALCULATE INTERACTIONS *
1068              **************************/
1069
1070             /* REACTION-FIELD ELECTROSTATICS */
1071             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1072
1073             /* LENNARD-JONES DISPERSION/REPULSION */
1074
1075             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1076             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1077
1078             fscal            = _fjsp_add_v2r8(felec,fvdw);
1079
1080             /* Update vectorial force */
1081             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1082             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1083             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1084
1085             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1086             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1087             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1088
1089             /**************************
1090              * CALCULATE INTERACTIONS *
1091              **************************/
1092
1093             /* REACTION-FIELD ELECTROSTATICS */
1094             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1095
1096             fscal            = felec;
1097
1098             /* Update vectorial force */
1099             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1100             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1101             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1102
1103             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1104             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1105             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1106
1107             /**************************
1108              * CALCULATE INTERACTIONS *
1109              **************************/
1110
1111             /* REACTION-FIELD ELECTROSTATICS */
1112             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1113
1114             fscal            = felec;
1115
1116             /* Update vectorial force */
1117             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1118             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1119             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1120
1121             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1122             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1123             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1124
1125             /**************************
1126              * CALCULATE INTERACTIONS *
1127              **************************/
1128
1129             /* REACTION-FIELD ELECTROSTATICS */
1130             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1131
1132             fscal            = felec;
1133
1134             /* Update vectorial force */
1135             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1136             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1137             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1138
1139             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1140             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1141             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             /* REACTION-FIELD ELECTROSTATICS */
1148             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1149
1150             fscal            = felec;
1151
1152             /* Update vectorial force */
1153             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1154             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1155             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1156
1157             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1158             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1159             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1160
1161             /**************************
1162              * CALCULATE INTERACTIONS *
1163              **************************/
1164
1165             /* REACTION-FIELD ELECTROSTATICS */
1166             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1167
1168             fscal            = felec;
1169
1170             /* Update vectorial force */
1171             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1172             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1173             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1174
1175             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1176             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1177             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1178
1179             /**************************
1180              * CALCULATE INTERACTIONS *
1181              **************************/
1182
1183             /* REACTION-FIELD ELECTROSTATICS */
1184             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1185
1186             fscal            = felec;
1187
1188             /* Update vectorial force */
1189             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1190             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1191             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1192
1193             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1194             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1195             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1196
1197             /**************************
1198              * CALCULATE INTERACTIONS *
1199              **************************/
1200
1201             /* REACTION-FIELD ELECTROSTATICS */
1202             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1203
1204             fscal            = felec;
1205
1206             /* Update vectorial force */
1207             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1208             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1209             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1210
1211             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1212             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1213             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1214
1215             /**************************
1216              * CALCULATE INTERACTIONS *
1217              **************************/
1218
1219             /* REACTION-FIELD ELECTROSTATICS */
1220             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1221
1222             fscal            = felec;
1223
1224             /* Update vectorial force */
1225             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1226             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1227             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1228
1229             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1230             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1231             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1232
1233             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1234
1235             /* Inner loop uses 277 flops */
1236         }
1237
1238         if(jidx<j_index_end)
1239         {
1240
1241             jnrA             = jjnr[jidx];
1242             j_coord_offsetA  = DIM*jnrA;
1243
1244             /* load j atom coordinates */
1245             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1246                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1247
1248             /* Calculate displacement vector */
1249             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1250             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1251             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1252             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1253             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1254             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1255             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1256             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1257             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1258             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1259             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1260             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1261             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1262             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1263             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1264             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1265             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1266             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1267             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1268             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1269             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1270             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1271             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1272             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1273             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1274             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1275             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1276
1277             /* Calculate squared distance and things based on it */
1278             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1279             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1280             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1281             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1282             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1283             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1284             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1285             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1286             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1287
1288             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1289             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1290             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1291             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1292             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1293             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1294             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1295             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1296             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1297
1298             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1299             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1300             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1301             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1302             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1303             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1304             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1305             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1306             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1307
1308             fjx0             = _fjsp_setzero_v2r8();
1309             fjy0             = _fjsp_setzero_v2r8();
1310             fjz0             = _fjsp_setzero_v2r8();
1311             fjx1             = _fjsp_setzero_v2r8();
1312             fjy1             = _fjsp_setzero_v2r8();
1313             fjz1             = _fjsp_setzero_v2r8();
1314             fjx2             = _fjsp_setzero_v2r8();
1315             fjy2             = _fjsp_setzero_v2r8();
1316             fjz2             = _fjsp_setzero_v2r8();
1317
1318             /**************************
1319              * CALCULATE INTERACTIONS *
1320              **************************/
1321
1322             /* REACTION-FIELD ELECTROSTATICS */
1323             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1324
1325             /* LENNARD-JONES DISPERSION/REPULSION */
1326
1327             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1328             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1329
1330             fscal            = _fjsp_add_v2r8(felec,fvdw);
1331
1332             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1333
1334             /* Update vectorial force */
1335             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1336             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1337             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1338
1339             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1340             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1341             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1342
1343             /**************************
1344              * CALCULATE INTERACTIONS *
1345              **************************/
1346
1347             /* REACTION-FIELD ELECTROSTATICS */
1348             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1349
1350             fscal            = felec;
1351
1352             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1353
1354             /* Update vectorial force */
1355             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1356             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1357             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1358
1359             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1360             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1361             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1362
1363             /**************************
1364              * CALCULATE INTERACTIONS *
1365              **************************/
1366
1367             /* REACTION-FIELD ELECTROSTATICS */
1368             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1369
1370             fscal            = felec;
1371
1372             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1373
1374             /* Update vectorial force */
1375             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1376             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1377             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1378
1379             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1380             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1381             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1382
1383             /**************************
1384              * CALCULATE INTERACTIONS *
1385              **************************/
1386
1387             /* REACTION-FIELD ELECTROSTATICS */
1388             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1389
1390             fscal            = felec;
1391
1392             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1393
1394             /* Update vectorial force */
1395             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1396             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1397             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1398
1399             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1400             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1401             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1402
1403             /**************************
1404              * CALCULATE INTERACTIONS *
1405              **************************/
1406
1407             /* REACTION-FIELD ELECTROSTATICS */
1408             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1409
1410             fscal            = felec;
1411
1412             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1413
1414             /* Update vectorial force */
1415             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1416             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1417             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1418
1419             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1420             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1421             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1422
1423             /**************************
1424              * CALCULATE INTERACTIONS *
1425              **************************/
1426
1427             /* REACTION-FIELD ELECTROSTATICS */
1428             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1429
1430             fscal            = felec;
1431
1432             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1433
1434             /* Update vectorial force */
1435             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1436             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1437             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1438
1439             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1440             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1441             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1442
1443             /**************************
1444              * CALCULATE INTERACTIONS *
1445              **************************/
1446
1447             /* REACTION-FIELD ELECTROSTATICS */
1448             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1449
1450             fscal            = felec;
1451
1452             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1453
1454             /* Update vectorial force */
1455             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1456             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1457             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1458
1459             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1460             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1461             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1462
1463             /**************************
1464              * CALCULATE INTERACTIONS *
1465              **************************/
1466
1467             /* REACTION-FIELD ELECTROSTATICS */
1468             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1469
1470             fscal            = felec;
1471
1472             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1473
1474             /* Update vectorial force */
1475             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1476             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1477             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1478
1479             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1480             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1481             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1482
1483             /**************************
1484              * CALCULATE INTERACTIONS *
1485              **************************/
1486
1487             /* REACTION-FIELD ELECTROSTATICS */
1488             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1489
1490             fscal            = felec;
1491
1492             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1493
1494             /* Update vectorial force */
1495             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1496             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1497             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1498
1499             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1500             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1501             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1502
1503             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1504
1505             /* Inner loop uses 277 flops */
1506         }
1507
1508         /* End of innermost loop */
1509
1510         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1511                                               f+i_coord_offset,fshift+i_shift_offset);
1512
1513         /* Increment number of inner iterations */
1514         inneriter                  += j_index_end - j_index_start;
1515
1516         /* Outer loop uses 18 flops */
1517     }
1518
1519     /* Increment number of outer iterations */
1520     outeriter        += nri;
1521
1522     /* Update outer/inner flops */
1523
1524     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1525 }