src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecCoul_VdwNone_GeomW4W4_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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: Coulomb
  54  * VdW interaction:            None
  55  * Geometry:                   Water4-Water4
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecCoul_VdwNone_GeomW4W4_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              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              vdwioffset3;
  86     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  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     int              vdwjidx3A,vdwjidx3B;
  92     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  93     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  94     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  95     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  96     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  97     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  98     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  99     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 100     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 101     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 102     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 103     real             *charge;
 104     _fjsp_v2r8       itab_tmp;
 105     _fjsp_v2r8       dummy_mask,cutoff_mask;
 106     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 107     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 108     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 109
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 127     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 128     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 129
 130     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 131     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 132     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 133     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 134     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 135     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 136     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 137     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 138     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 139     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 140     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 141     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 142
 143     /* Avoid stupid compiler warnings */
 144     jnrA = jnrB = 0;
 145     j_coord_offsetA = 0;
 146     j_coord_offsetB = 0;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 167                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 168
 169         fix1             = _fjsp_setzero_v2r8();
 170         fiy1             = _fjsp_setzero_v2r8();
 171         fiz1             = _fjsp_setzero_v2r8();
 172         fix2             = _fjsp_setzero_v2r8();
 173         fiy2             = _fjsp_setzero_v2r8();
 174         fiz2             = _fjsp_setzero_v2r8();
 175         fix3             = _fjsp_setzero_v2r8();
 176         fiy3             = _fjsp_setzero_v2r8();
 177         fiz3             = _fjsp_setzero_v2r8();
 178
 179         /* Reset potential sums */
 180         velecsum         = _fjsp_setzero_v2r8();
 181
 182         /* Start inner kernel loop */
 183         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 184         {
 185
 186             /* Get j neighbor index, and coordinate index */
 187             jnrA             = jjnr[jidx];
 188             jnrB             = jjnr[jidx+1];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191
 192             /* load j atom coordinates */
 193             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 194                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 195
 196             /* Calculate displacement vector */
 197             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 198             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 199             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 200             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 201             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 202             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 203             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 204             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 205             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 206             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 207             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 208             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 209             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 210             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 211             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 212             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 213             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 214             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 215             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 216             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 217             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 218             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 219             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 220             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 221             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 222             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 223             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 224
 225             /* Calculate squared distance and things based on it */
 226             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 227             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 228             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 229             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 230             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 231             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 232             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 233             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 234             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 235
 236             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 237             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 238             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 239             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 240             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 241             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 242             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 243             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 244             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 245
 246             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 247             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 248             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 249             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 250             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 251             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 252             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 253             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 254             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 255
 256             fjx1             = _fjsp_setzero_v2r8();
 257             fjy1             = _fjsp_setzero_v2r8();
 258             fjz1             = _fjsp_setzero_v2r8();
 259             fjx2             = _fjsp_setzero_v2r8();
 260             fjy2             = _fjsp_setzero_v2r8();
 261             fjz2             = _fjsp_setzero_v2r8();
 262             fjx3             = _fjsp_setzero_v2r8();
 263             fjy3             = _fjsp_setzero_v2r8();
 264             fjz3             = _fjsp_setzero_v2r8();
 265
 266             /**************************
 267              * CALCULATE INTERACTIONS *
 268              **************************/
 269
 270             /* COULOMB ELECTROSTATICS */
 271             velec            = _fjsp_mul_v2r8(qq11,rinv11);
 272             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
 273
 274             /* Update potential sum for this i atom from the interaction with this j atom. */
 275             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 276
 277             fscal            = felec;
 278
 279             /* Update vectorial force */
 280             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 281             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 282             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 283
 284             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 285             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 286             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 287
 288             /**************************
 289              * CALCULATE INTERACTIONS *
 290              **************************/
 291
 292             /* COULOMB ELECTROSTATICS */
 293             velec            = _fjsp_mul_v2r8(qq12,rinv12);
 294             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
 295
 296             /* Update potential sum for this i atom from the interaction with this j atom. */
 297             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 298
 299             fscal            = felec;
 300
 301             /* Update vectorial force */
 302             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 303             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 304             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 305
 306             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 307             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 308             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 309
 310             /**************************
 311              * CALCULATE INTERACTIONS *
 312              **************************/
 313
 314             /* COULOMB ELECTROSTATICS */
 315             velec            = _fjsp_mul_v2r8(qq13,rinv13);
 316             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 320
 321             fscal            = felec;
 322
 323             /* Update vectorial force */
 324             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 325             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 326             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 327
 328             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 329             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 330             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 331
 332             /**************************
 333              * CALCULATE INTERACTIONS *
 334              **************************/
 335
 336             /* COULOMB ELECTROSTATICS */
 337             velec            = _fjsp_mul_v2r8(qq21,rinv21);
 338             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
 339
 340             /* Update potential sum for this i atom from the interaction with this j atom. */
 341             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 342
 343             fscal            = felec;
 344
 345             /* Update vectorial force */
 346             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 347             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 348             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 349
 350             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 351             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 352             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             /* COULOMB ELECTROSTATICS */
 359             velec            = _fjsp_mul_v2r8(qq22,rinv22);
 360             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 364
 365             fscal            = felec;
 366
 367             /* Update vectorial force */
 368             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 369             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 370             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 371
 372             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 373             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 374             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 375
 376             /**************************
 377              * CALCULATE INTERACTIONS *
 378              **************************/
 379
 380             /* COULOMB ELECTROSTATICS */
 381             velec            = _fjsp_mul_v2r8(qq23,rinv23);
 382             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
 383
 384             /* Update potential sum for this i atom from the interaction with this j atom. */
 385             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 386
 387             fscal            = felec;
 388
 389             /* Update vectorial force */
 390             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 391             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 392             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 393
 394             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 395             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 396             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             /* COULOMB ELECTROSTATICS */
 403             velec            = _fjsp_mul_v2r8(qq31,rinv31);
 404             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411             /* Update vectorial force */
 412             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 413             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 414             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 415
 416             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 417             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 418             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             /* COULOMB ELECTROSTATICS */
 425             velec            = _fjsp_mul_v2r8(qq32,rinv32);
 426             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
 427
 428             /* Update potential sum for this i atom from the interaction with this j atom. */
 429             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 430
 431             fscal            = felec;
 432
 433             /* Update vectorial force */
 434             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 435             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 436             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 437
 438             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 439             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 440             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 441
 442             /**************************
 443              * CALCULATE INTERACTIONS *
 444              **************************/
 445
 446             /* COULOMB ELECTROSTATICS */
 447             velec            = _fjsp_mul_v2r8(qq33,rinv33);
 448             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
 449
 450             /* Update potential sum for this i atom from the interaction with this j atom. */
 451             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 452
 453             fscal            = felec;
 454
 455             /* Update vectorial force */
 456             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 457             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 458             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 459
 460             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 461             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 462             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 463
 464             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 465
 466             /* Inner loop uses 279 flops */
 467         }
 468
 469         if(jidx<j_index_end)
 470         {
 471
 472             jnrA             = jjnr[jidx];
 473             j_coord_offsetA  = DIM*jnrA;
 474
 475             /* load j atom coordinates */
 476             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
 477                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 478
 479             /* Calculate displacement vector */
 480             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 481             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 482             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 483             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 484             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 485             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 486             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 487             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 488             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 489             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 490             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 491             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 492             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 493             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 494             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 495             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 496             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 497             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 498             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 499             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 500             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 501             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 502             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 503             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 504             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 505             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 506             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 507
 508             /* Calculate squared distance and things based on it */
 509             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 510             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 511             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 512             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 513             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 514             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 515             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 516             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 517             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 518
 519             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 520             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 521             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 522             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 523             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 524             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 525             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 526             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 527             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 528
 529             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 530             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 531             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 532             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 533             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 534             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 535             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 536             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 537             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 538
 539             fjx1             = _fjsp_setzero_v2r8();
 540             fjy1             = _fjsp_setzero_v2r8();
 541             fjz1             = _fjsp_setzero_v2r8();
 542             fjx2             = _fjsp_setzero_v2r8();
 543             fjy2             = _fjsp_setzero_v2r8();
 544             fjz2             = _fjsp_setzero_v2r8();
 545             fjx3             = _fjsp_setzero_v2r8();
 546             fjy3             = _fjsp_setzero_v2r8();
 547             fjz3             = _fjsp_setzero_v2r8();
 548
 549             /**************************
 550              * CALCULATE INTERACTIONS *
 551              **************************/
 552
 553             /* COULOMB ELECTROSTATICS */
 554             velec            = _fjsp_mul_v2r8(qq11,rinv11);
 555             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
 556
 557             /* Update potential sum for this i atom from the interaction with this j atom. */
 558             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 559             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 560
 561             fscal            = felec;
 562
 563             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 564
 565             /* Update vectorial force */
 566             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 567             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 568             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 569
 570             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 571             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 572             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 573
 574             /**************************
 575              * CALCULATE INTERACTIONS *
 576              **************************/
 577
 578             /* COULOMB ELECTROSTATICS */
 579             velec            = _fjsp_mul_v2r8(qq12,rinv12);
 580             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
 581
 582             /* Update potential sum for this i atom from the interaction with this j atom. */
 583             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 584             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 585
 586             fscal            = felec;
 587
 588             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 589
 590             /* Update vectorial force */
 591             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 592             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 593             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 594
 595             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 596             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 597             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             /* COULOMB ELECTROSTATICS */
 604             velec            = _fjsp_mul_v2r8(qq13,rinv13);
 605             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
 606
 607             /* Update potential sum for this i atom from the interaction with this j atom. */
 608             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 609             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 610
 611             fscal            = felec;
 612
 613             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 614
 615             /* Update vectorial force */
 616             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 617             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 618             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 619
 620             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 621             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 622             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 623
 624             /**************************
 625              * CALCULATE INTERACTIONS *
 626              **************************/
 627
 628             /* COULOMB ELECTROSTATICS */
 629             velec            = _fjsp_mul_v2r8(qq21,rinv21);
 630             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
 631
 632             /* Update potential sum for this i atom from the interaction with this j atom. */
 633             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 634             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 635
 636             fscal            = felec;
 637
 638             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 639
 640             /* Update vectorial force */
 641             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 642             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 643             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 644
 645             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 646             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 647             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 648
 649             /**************************
 650              * CALCULATE INTERACTIONS *
 651              **************************/
 652
 653             /* COULOMB ELECTROSTATICS */
 654             velec            = _fjsp_mul_v2r8(qq22,rinv22);
 655             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
 656
 657             /* Update potential sum for this i atom from the interaction with this j atom. */
 658             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 659             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 660
 661             fscal            = felec;
 662
 663             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 664
 665             /* Update vectorial force */
 666             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 667             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 668             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 669
 670             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 671             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 672             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 673
 674             /**************************
 675              * CALCULATE INTERACTIONS *
 676              **************************/
 677
 678             /* COULOMB ELECTROSTATICS */
 679             velec            = _fjsp_mul_v2r8(qq23,rinv23);
 680             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
 681
 682             /* Update potential sum for this i atom from the interaction with this j atom. */
 683             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 684             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 685
 686             fscal            = felec;
 687
 688             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 689
 690             /* Update vectorial force */
 691             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 692             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 693             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 694
 695             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 696             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 697             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 698
 699             /**************************
 700              * CALCULATE INTERACTIONS *
 701              **************************/
 702
 703             /* COULOMB ELECTROSTATICS */
 704             velec            = _fjsp_mul_v2r8(qq31,rinv31);
 705             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
 706
 707             /* Update potential sum for this i atom from the interaction with this j atom. */
 708             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 709             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 710
 711             fscal            = felec;
 712
 713             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 714
 715             /* Update vectorial force */
 716             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 717             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 718             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 719
 720             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 721             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 722             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             /* COULOMB ELECTROSTATICS */
 729             velec            = _fjsp_mul_v2r8(qq32,rinv32);
 730             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
 731
 732             /* Update potential sum for this i atom from the interaction with this j atom. */
 733             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 734             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 735
 736             fscal            = felec;
 737
 738             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 739
 740             /* Update vectorial force */
 741             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 742             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 743             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 744
 745             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 746             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 747             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 748
 749             /**************************
 750              * CALCULATE INTERACTIONS *
 751              **************************/
 752
 753             /* COULOMB ELECTROSTATICS */
 754             velec            = _fjsp_mul_v2r8(qq33,rinv33);
 755             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
 756
 757             /* Update potential sum for this i atom from the interaction with this j atom. */
 758             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 759             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 760
 761             fscal            = felec;
 762
 763             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 764
 765             /* Update vectorial force */
 766             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 767             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 768             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 769
 770             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 771             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 772             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 773
 774             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 775
 776             /* Inner loop uses 279 flops */
 777         }
 778
 779         /* End of innermost loop */
 780
 781         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 782                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 783
 784         ggid                        = gid[iidx];
 785         /* Update potential energies */
 786         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 787
 788         /* Increment number of inner iterations */
 789         inneriter                  += j_index_end - j_index_start;
 790
 791         /* Outer loop uses 19 flops */
 792     }
 793
 794     /* Increment number of outer iterations */
 795     outeriter        += nri;
 796
 797     /* Update outer/inner flops */
 798
 799     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*279);
 800 }
 801 /*
 802  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
 803  * Electrostatics interaction: Coulomb
 804  * VdW interaction:            None
 805  * Geometry:                   Water4-Water4
 806  * Calculate force/pot:        Force
 807  */
 808 void
 809 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
 810                     (t_nblist                    * gmx_restrict       nlist,
 811                      rvec                        * gmx_restrict          xx,
 812                      rvec                        * gmx_restrict          ff,
 813                      t_forcerec                  * gmx_restrict          fr,
 814                      t_mdatoms                   * gmx_restrict     mdatoms,
 815                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 816                      t_nrnb                      * gmx_restrict        nrnb)
 817 {
 818     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 819      * just 0 for non-waters.
 820      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 821      * jnr indices corresponding to data put in the four positions in the SIMD register.
 822      */
 823     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 824     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 825     int              jnrA,jnrB;
 826     int              j_coord_offsetA,j_coord_offsetB;
 827     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 828     real             rcutoff_scalar;
 829     real             *shiftvec,*fshift,*x,*f;
 830     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 831     int              vdwioffset1;
 832     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 833     int              vdwioffset2;
 834     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 835     int              vdwioffset3;
 836     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 837     int              vdwjidx1A,vdwjidx1B;
 838     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 839     int              vdwjidx2A,vdwjidx2B;
 840     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 841     int              vdwjidx3A,vdwjidx3B;
 842     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 843     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 844     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 845     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 846     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 847     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 848     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 849     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 850     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 851     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 852     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 853     real             *charge;
 854     _fjsp_v2r8       itab_tmp;
 855     _fjsp_v2r8       dummy_mask,cutoff_mask;
 856     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 857     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 858     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 859
 860     x                = xx[0];
 861     f                = ff[0];
 862
 863     nri              = nlist->nri;
 864     iinr             = nlist->iinr;
 865     jindex           = nlist->jindex;
 866     jjnr             = nlist->jjnr;
 867     shiftidx         = nlist->shift;
 868     gid              = nlist->gid;
 869     shiftvec         = fr->shift_vec[0];
 870     fshift           = fr->fshift[0];
 871     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 872     charge           = mdatoms->chargeA;
 873
 874     /* Setup water-specific parameters */
 875     inr              = nlist->iinr[0];
 876     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 877     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 878     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 879
 880     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 881     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 882     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 883     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 884     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 885     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 886     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 887     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 888     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 889     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 890     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 891     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 892
 893     /* Avoid stupid compiler warnings */
 894     jnrA = jnrB = 0;
 895     j_coord_offsetA = 0;
 896     j_coord_offsetB = 0;
 897
 898     outeriter        = 0;
 899     inneriter        = 0;
 900
 901     /* Start outer loop over neighborlists */
 902     for(iidx=0; iidx<nri; iidx++)
 903     {
 904         /* Load shift vector for this list */
 905         i_shift_offset   = DIM*shiftidx[iidx];
 906
 907         /* Load limits for loop over neighbors */
 908         j_index_start    = jindex[iidx];
 909         j_index_end      = jindex[iidx+1];
 910
 911         /* Get outer coordinate index */
 912         inr              = iinr[iidx];
 913         i_coord_offset   = DIM*inr;
 914
 915         /* Load i particle coords and add shift vector */
 916         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 917                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 918
 919         fix1             = _fjsp_setzero_v2r8();
 920         fiy1             = _fjsp_setzero_v2r8();
 921         fiz1             = _fjsp_setzero_v2r8();
 922         fix2             = _fjsp_setzero_v2r8();
 923         fiy2             = _fjsp_setzero_v2r8();
 924         fiz2             = _fjsp_setzero_v2r8();
 925         fix3             = _fjsp_setzero_v2r8();
 926         fiy3             = _fjsp_setzero_v2r8();
 927         fiz3             = _fjsp_setzero_v2r8();
 928
 929         /* Start inner kernel loop */
 930         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 931         {
 932
 933             /* Get j neighbor index, and coordinate index */
 934             jnrA             = jjnr[jidx];
 935             jnrB             = jjnr[jidx+1];
 936             j_coord_offsetA  = DIM*jnrA;
 937             j_coord_offsetB  = DIM*jnrB;
 938
 939             /* load j atom coordinates */
 940             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 941                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 942
 943             /* Calculate displacement vector */
 944             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 945             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 946             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 947             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 948             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 949             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 950             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 951             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 952             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 953             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 954             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 955             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 956             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 957             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 958             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 959             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 960             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 961             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 962             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 963             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 964             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 965             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 966             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 967             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 968             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 969             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 970             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 971
 972             /* Calculate squared distance and things based on it */
 973             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 974             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 975             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 976             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 977             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 978             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 979             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 980             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 981             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 982
 983             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 984             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 985             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 986             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 987             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 988             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 989             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 990             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 991             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 992
 993             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 994             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 995             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 996             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 997             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 998             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 999             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1000             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1001             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1002
1003             fjx1             = _fjsp_setzero_v2r8();
1004             fjy1             = _fjsp_setzero_v2r8();
1005             fjz1             = _fjsp_setzero_v2r8();
1006             fjx2             = _fjsp_setzero_v2r8();
1007             fjy2             = _fjsp_setzero_v2r8();
1008             fjz2             = _fjsp_setzero_v2r8();
1009             fjx3             = _fjsp_setzero_v2r8();
1010             fjy3             = _fjsp_setzero_v2r8();
1011             fjz3             = _fjsp_setzero_v2r8();
1012
1013             /**************************
1014              * CALCULATE INTERACTIONS *
1015              **************************/
1016
1017             /* COULOMB ELECTROSTATICS */
1018             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1019             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1020
1021             fscal            = felec;
1022
1023             /* Update vectorial force */
1024             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1025             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1026             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1027
1028             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1029             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1030             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1031
1032             /**************************
1033              * CALCULATE INTERACTIONS *
1034              **************************/
1035
1036             /* COULOMB ELECTROSTATICS */
1037             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1038             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1039
1040             fscal            = felec;
1041
1042             /* Update vectorial force */
1043             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1044             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1045             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1046
1047             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1048             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1049             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1050
1051             /**************************
1052              * CALCULATE INTERACTIONS *
1053              **************************/
1054
1055             /* COULOMB ELECTROSTATICS */
1056             velec            = _fjsp_mul_v2r8(qq13,rinv13);
1057             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
1058
1059             fscal            = felec;
1060
1061             /* Update vectorial force */
1062             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1063             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1064             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1065
1066             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1067             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1068             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1069
1070             /**************************
1071              * CALCULATE INTERACTIONS *
1072              **************************/
1073
1074             /* COULOMB ELECTROSTATICS */
1075             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1076             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1077
1078             fscal            = felec;
1079
1080             /* Update vectorial force */
1081             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1082             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1083             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1084
1085             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1086             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1087             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1088
1089             /**************************
1090              * CALCULATE INTERACTIONS *
1091              **************************/
1092
1093             /* COULOMB ELECTROSTATICS */
1094             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1095             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1096
1097             fscal            = felec;
1098
1099             /* Update vectorial force */
1100             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1101             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1102             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1103
1104             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1105             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1106             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1107
1108             /**************************
1109              * CALCULATE INTERACTIONS *
1110              **************************/
1111
1112             /* COULOMB ELECTROSTATICS */
1113             velec            = _fjsp_mul_v2r8(qq23,rinv23);
1114             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
1115
1116             fscal            = felec;
1117
1118             /* Update vectorial force */
1119             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1120             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1121             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1122
1123             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1124             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1125             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1126
1127             /**************************
1128              * CALCULATE INTERACTIONS *
1129              **************************/
1130
1131             /* COULOMB ELECTROSTATICS */
1132             velec            = _fjsp_mul_v2r8(qq31,rinv31);
1133             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
1134
1135             fscal            = felec;
1136
1137             /* Update vectorial force */
1138             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1139             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1140             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1141
1142             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1143             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1144             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1145
1146             /**************************
1147              * CALCULATE INTERACTIONS *
1148              **************************/
1149
1150             /* COULOMB ELECTROSTATICS */
1151             velec            = _fjsp_mul_v2r8(qq32,rinv32);
1152             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
1153
1154             fscal            = felec;
1155
1156             /* Update vectorial force */
1157             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1158             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1159             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1160
1161             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1162             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1163             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1164
1165             /**************************
1166              * CALCULATE INTERACTIONS *
1167              **************************/
1168
1169             /* COULOMB ELECTROSTATICS */
1170             velec            = _fjsp_mul_v2r8(qq33,rinv33);
1171             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
1172
1173             fscal            = felec;
1174
1175             /* Update vectorial force */
1176             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1177             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1178             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1179
1180             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1181             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1182             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1183
1184             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1185
1186             /* Inner loop uses 270 flops */
1187         }
1188
1189         if(jidx<j_index_end)
1190         {
1191
1192             jnrA             = jjnr[jidx];
1193             j_coord_offsetA  = DIM*jnrA;
1194
1195             /* load j atom coordinates */
1196             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1197                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1198
1199             /* Calculate displacement vector */
1200             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1201             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1202             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1203             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1204             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1205             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1206             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1207             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1208             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1209             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1210             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1211             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1212             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1213             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1214             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1215             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1216             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1217             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1218             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1219             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1220             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1221             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1222             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1223             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1224             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1225             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1226             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1227
1228             /* Calculate squared distance and things based on it */
1229             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1230             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1231             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1232             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1233             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1234             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1235             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1236             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1237             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1238
1239             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1240             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1241             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1242             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1243             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1244             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1245             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1246             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1247             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1248
1249             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1250             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1251             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1252             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1253             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1254             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1255             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1256             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1257             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1258
1259             fjx1             = _fjsp_setzero_v2r8();
1260             fjy1             = _fjsp_setzero_v2r8();
1261             fjz1             = _fjsp_setzero_v2r8();
1262             fjx2             = _fjsp_setzero_v2r8();
1263             fjy2             = _fjsp_setzero_v2r8();
1264             fjz2             = _fjsp_setzero_v2r8();
1265             fjx3             = _fjsp_setzero_v2r8();
1266             fjy3             = _fjsp_setzero_v2r8();
1267             fjz3             = _fjsp_setzero_v2r8();
1268
1269             /**************************
1270              * CALCULATE INTERACTIONS *
1271              **************************/
1272
1273             /* COULOMB ELECTROSTATICS */
1274             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1275             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1276
1277             fscal            = felec;
1278
1279             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1280
1281             /* Update vectorial force */
1282             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1283             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1284             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1285
1286             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1287             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1288             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             /* COULOMB ELECTROSTATICS */
1295             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1296             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1297
1298             fscal            = felec;
1299
1300             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1301
1302             /* Update vectorial force */
1303             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1304             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1305             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1306
1307             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1308             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1309             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1310
1311             /**************************
1312              * CALCULATE INTERACTIONS *
1313              **************************/
1314
1315             /* COULOMB ELECTROSTATICS */
1316             velec            = _fjsp_mul_v2r8(qq13,rinv13);
1317             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
1318
1319             fscal            = felec;
1320
1321             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1322
1323             /* Update vectorial force */
1324             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1325             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1326             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1327
1328             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1329             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1330             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1331
1332             /**************************
1333              * CALCULATE INTERACTIONS *
1334              **************************/
1335
1336             /* COULOMB ELECTROSTATICS */
1337             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1338             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1339
1340             fscal            = felec;
1341
1342             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1343
1344             /* Update vectorial force */
1345             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1346             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1347             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1348
1349             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1350             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1351             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1352
1353             /**************************
1354              * CALCULATE INTERACTIONS *
1355              **************************/
1356
1357             /* COULOMB ELECTROSTATICS */
1358             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1359             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1360
1361             fscal            = felec;
1362
1363             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1364
1365             /* Update vectorial force */
1366             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1367             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1368             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1369
1370             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1371             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1372             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1373
1374             /**************************
1375              * CALCULATE INTERACTIONS *
1376              **************************/
1377
1378             /* COULOMB ELECTROSTATICS */
1379             velec            = _fjsp_mul_v2r8(qq23,rinv23);
1380             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
1381
1382             fscal            = felec;
1383
1384             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1385
1386             /* Update vectorial force */
1387             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1388             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1389             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1390
1391             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1392             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1393             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1394
1395             /**************************
1396              * CALCULATE INTERACTIONS *
1397              **************************/
1398
1399             /* COULOMB ELECTROSTATICS */
1400             velec            = _fjsp_mul_v2r8(qq31,rinv31);
1401             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
1402
1403             fscal            = felec;
1404
1405             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1406
1407             /* Update vectorial force */
1408             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1409             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1410             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1411
1412             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1413             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1414             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             /* COULOMB ELECTROSTATICS */
1421             velec            = _fjsp_mul_v2r8(qq32,rinv32);
1422             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
1423
1424             fscal            = felec;
1425
1426             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1427
1428             /* Update vectorial force */
1429             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1430             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1431             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1432
1433             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1434             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1435             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1436
1437             /**************************
1438              * CALCULATE INTERACTIONS *
1439              **************************/
1440
1441             /* COULOMB ELECTROSTATICS */
1442             velec            = _fjsp_mul_v2r8(qq33,rinv33);
1443             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
1444
1445             fscal            = felec;
1446
1447             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1448
1449             /* Update vectorial force */
1450             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1451             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1452             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1453
1454             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1455             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1456             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1457
1458             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1459
1460             /* Inner loop uses 270 flops */
1461         }
1462
1463         /* End of innermost loop */
1464
1465         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1466                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1467
1468         /* Increment number of inner iterations */
1469         inneriter                  += j_index_end - j_index_start;
1470
1471         /* Outer loop uses 18 flops */
1472     }
1473
1474     /* Increment number of outer iterations */
1475     outeriter        += nri;
1476
1477     /* Update outer/inner flops */
1478
1479     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1480 }