src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRF_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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "kernelutil_sparc64_hpc_ace_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            None
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      t_forcerec                  * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset1;
  80     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  81     int              vdwioffset2;
  82     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  83     int              vdwioffset3;
  84     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  85     int              vdwjidx1A,vdwjidx1B;
  86     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  87     int              vdwjidx2A,vdwjidx2B;
  88     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  89     int              vdwjidx3A,vdwjidx3B;
  90     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  91     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  92     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  93     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  94     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  95     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  96     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  97     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  98     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  99     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 100     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     _fjsp_v2r8       itab_tmp;
 103     _fjsp_v2r8       dummy_mask,cutoff_mask;
 104     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 105     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 106     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 107
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 122     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 123     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 124
 125     /* Setup water-specific parameters */
 126     inr              = nlist->iinr[0];
 127     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 128     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 129     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 130
 131     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 132     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 133     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 134     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 135     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 136     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 137     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 138     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 139     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 140     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 141     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 142     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 143
 144     /* Avoid stupid compiler warnings */
 145     jnrA = jnrB = 0;
 146     j_coord_offsetA = 0;
 147     j_coord_offsetB = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157
 158         /* Load limits for loop over neighbors */
 159         j_index_start    = jindex[iidx];
 160         j_index_end      = jindex[iidx+1];
 161
 162         /* Get outer coordinate index */
 163         inr              = iinr[iidx];
 164         i_coord_offset   = DIM*inr;
 165
 166         /* Load i particle coords and add shift vector */
 167         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 168                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 169
 170         fix1             = _fjsp_setzero_v2r8();
 171         fiy1             = _fjsp_setzero_v2r8();
 172         fiz1             = _fjsp_setzero_v2r8();
 173         fix2             = _fjsp_setzero_v2r8();
 174         fiy2             = _fjsp_setzero_v2r8();
 175         fiz2             = _fjsp_setzero_v2r8();
 176         fix3             = _fjsp_setzero_v2r8();
 177         fiy3             = _fjsp_setzero_v2r8();
 178         fiz3             = _fjsp_setzero_v2r8();
 179
 180         /* Reset potential sums */
 181         velecsum         = _fjsp_setzero_v2r8();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             j_coord_offsetA  = DIM*jnrA;
 191             j_coord_offsetB  = DIM*jnrB;
 192
 193             /* load j atom coordinates */
 194             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 195                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 196
 197             /* Calculate displacement vector */
 198             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 199             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 200             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 201             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 202             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 203             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 204             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 205             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 206             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 207             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 208             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 209             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 210             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 211             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 212             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 213             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 214             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 215             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 216             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 217             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 218             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 219             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 220             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 221             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 222             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 223             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 224             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 225
 226             /* Calculate squared distance and things based on it */
 227             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 228             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 229             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 230             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 231             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 232             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 233             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 234             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 235             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 236
 237             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 238             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 239             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 240             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 241             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 242             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 243             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 244             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 245             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 246
 247             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 248             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 249             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 250             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 251             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 252             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 253             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 254             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 255             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 256
 257             fjx1             = _fjsp_setzero_v2r8();
 258             fjy1             = _fjsp_setzero_v2r8();
 259             fjz1             = _fjsp_setzero_v2r8();
 260             fjx2             = _fjsp_setzero_v2r8();
 261             fjy2             = _fjsp_setzero_v2r8();
 262             fjz2             = _fjsp_setzero_v2r8();
 263             fjx3             = _fjsp_setzero_v2r8();
 264             fjy3             = _fjsp_setzero_v2r8();
 265             fjz3             = _fjsp_setzero_v2r8();
 266
 267             /**************************
 268              * CALCULATE INTERACTIONS *
 269              **************************/
 270
 271             /* REACTION-FIELD ELECTROSTATICS */
 272             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 273             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 274
 275             /* Update potential sum for this i atom from the interaction with this j atom. */
 276             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 277
 278             fscal            = felec;
 279
 280             /* Update vectorial force */
 281             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 282             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 283             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 284
 285             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 286             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 287             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 288
 289             /**************************
 290              * CALCULATE INTERACTIONS *
 291              **************************/
 292
 293             /* REACTION-FIELD ELECTROSTATICS */
 294             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 295             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 299
 300             fscal            = felec;
 301
 302             /* Update vectorial force */
 303             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 304             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 305             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 306
 307             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 308             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 309             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             /* REACTION-FIELD ELECTROSTATICS */
 316             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq13,rinv13),crf));
 317             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324             /* Update vectorial force */
 325             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 326             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 327             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 328
 329             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 330             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 331             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             /* REACTION-FIELD ELECTROSTATICS */
 338             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 339             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 340
 341             /* Update potential sum for this i atom from the interaction with this j atom. */
 342             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 343
 344             fscal            = felec;
 345
 346             /* Update vectorial force */
 347             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 348             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 349             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 350
 351             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 352             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 353             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             /* REACTION-FIELD ELECTROSTATICS */
 360             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 361             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368             /* Update vectorial force */
 369             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 370             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 371             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 372
 373             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 374             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 375             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 376
 377             /**************************
 378              * CALCULATE INTERACTIONS *
 379              **************************/
 380
 381             /* REACTION-FIELD ELECTROSTATICS */
 382             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq23,rinv23),crf));
 383             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 387
 388             fscal            = felec;
 389
 390             /* Update vectorial force */
 391             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 392             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 393             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 394
 395             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 396             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 397             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             /* REACTION-FIELD ELECTROSTATICS */
 404             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq31,rinv31),crf));
 405             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412             /* Update vectorial force */
 413             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 414             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 415             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 416
 417             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 418             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 419             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             /* REACTION-FIELD ELECTROSTATICS */
 426             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq32,rinv32),crf));
 427             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
 428
 429             /* Update potential sum for this i atom from the interaction with this j atom. */
 430             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434             /* Update vectorial force */
 435             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 436             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 437             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 438
 439             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 440             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 441             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             /* REACTION-FIELD ELECTROSTATICS */
 448             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq33,rinv33),crf));
 449             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
 450
 451             /* Update potential sum for this i atom from the interaction with this j atom. */
 452             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 453
 454             fscal            = felec;
 455
 456             /* Update vectorial force */
 457             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 458             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 459             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 460
 461             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 462             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 463             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 464
 465             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);
 466
 467             /* Inner loop uses 315 flops */
 468         }
 469
 470         if(jidx<j_index_end)
 471         {
 472
 473             jnrA             = jjnr[jidx];
 474             j_coord_offsetA  = DIM*jnrA;
 475
 476             /* load j atom coordinates */
 477             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
 478                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 479
 480             /* Calculate displacement vector */
 481             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 482             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 483             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 484             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 485             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 486             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 487             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 488             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 489             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 490             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 491             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 492             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 493             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 494             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 495             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 496             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 497             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 498             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 499             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 500             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 501             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 502             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 503             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 504             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 505             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 506             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 507             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 508
 509             /* Calculate squared distance and things based on it */
 510             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 511             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 512             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 513             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 514             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 515             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 516             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 517             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 518             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 519
 520             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 521             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 522             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 523             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 524             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 525             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 526             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 527             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 528             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 529
 530             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 531             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 532             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 533             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 534             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 535             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 536             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 537             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 538             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 539
 540             fjx1             = _fjsp_setzero_v2r8();
 541             fjy1             = _fjsp_setzero_v2r8();
 542             fjz1             = _fjsp_setzero_v2r8();
 543             fjx2             = _fjsp_setzero_v2r8();
 544             fjy2             = _fjsp_setzero_v2r8();
 545             fjz2             = _fjsp_setzero_v2r8();
 546             fjx3             = _fjsp_setzero_v2r8();
 547             fjy3             = _fjsp_setzero_v2r8();
 548             fjz3             = _fjsp_setzero_v2r8();
 549
 550             /**************************
 551              * CALCULATE INTERACTIONS *
 552              **************************/
 553
 554             /* REACTION-FIELD ELECTROSTATICS */
 555             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 556             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 557
 558             /* Update potential sum for this i atom from the interaction with this j atom. */
 559             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 560             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 561
 562             fscal            = felec;
 563
 564             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 565
 566             /* Update vectorial force */
 567             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 568             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 569             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 570
 571             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 572             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 573             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             /* REACTION-FIELD ELECTROSTATICS */
 580             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 581             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 582
 583             /* Update potential sum for this i atom from the interaction with this j atom. */
 584             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 585             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 586
 587             fscal            = felec;
 588
 589             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 590
 591             /* Update vectorial force */
 592             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 593             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 594             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 595
 596             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 597             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 598             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 599
 600             /**************************
 601              * CALCULATE INTERACTIONS *
 602              **************************/
 603
 604             /* REACTION-FIELD ELECTROSTATICS */
 605             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq13,rinv13),crf));
 606             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
 607
 608             /* Update potential sum for this i atom from the interaction with this j atom. */
 609             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 610             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 611
 612             fscal            = felec;
 613
 614             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 615
 616             /* Update vectorial force */
 617             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 618             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 619             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 620
 621             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 622             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 623             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 624
 625             /**************************
 626              * CALCULATE INTERACTIONS *
 627              **************************/
 628
 629             /* REACTION-FIELD ELECTROSTATICS */
 630             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 631             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 632
 633             /* Update potential sum for this i atom from the interaction with this j atom. */
 634             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 635             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 636
 637             fscal            = felec;
 638
 639             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 640
 641             /* Update vectorial force */
 642             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 643             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 644             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 645
 646             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 647             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 648             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             /* REACTION-FIELD ELECTROSTATICS */
 655             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 656             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 657
 658             /* Update potential sum for this i atom from the interaction with this j atom. */
 659             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 660             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 661
 662             fscal            = felec;
 663
 664             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 665
 666             /* Update vectorial force */
 667             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 668             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 669             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 670
 671             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 672             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 673             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             /* REACTION-FIELD ELECTROSTATICS */
 680             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq23,rinv23),crf));
 681             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
 682
 683             /* Update potential sum for this i atom from the interaction with this j atom. */
 684             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 685             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 686
 687             fscal            = felec;
 688
 689             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 690
 691             /* Update vectorial force */
 692             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 693             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 694             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 695
 696             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 697             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 698             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 699
 700             /**************************
 701              * CALCULATE INTERACTIONS *
 702              **************************/
 703
 704             /* REACTION-FIELD ELECTROSTATICS */
 705             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq31,rinv31),crf));
 706             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
 707
 708             /* Update potential sum for this i atom from the interaction with this j atom. */
 709             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 710             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 711
 712             fscal            = felec;
 713
 714             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 715
 716             /* Update vectorial force */
 717             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 718             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 719             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 720
 721             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 722             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 723             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 724
 725             /**************************
 726              * CALCULATE INTERACTIONS *
 727              **************************/
 728
 729             /* REACTION-FIELD ELECTROSTATICS */
 730             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq32,rinv32),crf));
 731             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
 732
 733             /* Update potential sum for this i atom from the interaction with this j atom. */
 734             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 735             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 736
 737             fscal            = felec;
 738
 739             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 740
 741             /* Update vectorial force */
 742             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 743             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 744             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 745
 746             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 747             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 748             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 749
 750             /**************************
 751              * CALCULATE INTERACTIONS *
 752              **************************/
 753
 754             /* REACTION-FIELD ELECTROSTATICS */
 755             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq33,rinv33),crf));
 756             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
 757
 758             /* Update potential sum for this i atom from the interaction with this j atom. */
 759             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 760             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 761
 762             fscal            = felec;
 763
 764             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 765
 766             /* Update vectorial force */
 767             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 768             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 769             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 770
 771             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 772             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 773             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 774
 775             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 776
 777             /* Inner loop uses 315 flops */
 778         }
 779
 780         /* End of innermost loop */
 781
 782         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 783                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 784
 785         ggid                        = gid[iidx];
 786         /* Update potential energies */
 787         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 788
 789         /* Increment number of inner iterations */
 790         inneriter                  += j_index_end - j_index_start;
 791
 792         /* Outer loop uses 19 flops */
 793     }
 794
 795     /* Increment number of outer iterations */
 796     outeriter        += nri;
 797
 798     /* Update outer/inner flops */
 799
 800     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*315);
 801 }
 802 /*
 803  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
 804  * Electrostatics interaction: ReactionField
 805  * VdW interaction:            None
 806  * Geometry:                   Water4-Water4
 807  * Calculate force/pot:        Force
 808  */
 809 void
 810 nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
 811                     (t_nblist                    * gmx_restrict       nlist,
 812                      rvec                        * gmx_restrict          xx,
 813                      rvec                        * gmx_restrict          ff,
 814                      t_forcerec                  * gmx_restrict          fr,
 815                      t_mdatoms                   * gmx_restrict     mdatoms,
 816                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 817                      t_nrnb                      * gmx_restrict        nrnb)
 818 {
 819     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 820      * just 0 for non-waters.
 821      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 822      * jnr indices corresponding to data put in the four positions in the SIMD register.
 823      */
 824     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 825     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 826     int              jnrA,jnrB;
 827     int              j_coord_offsetA,j_coord_offsetB;
 828     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 829     real             rcutoff_scalar;
 830     real             *shiftvec,*fshift,*x,*f;
 831     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 832     int              vdwioffset1;
 833     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 834     int              vdwioffset2;
 835     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 836     int              vdwioffset3;
 837     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 838     int              vdwjidx1A,vdwjidx1B;
 839     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 840     int              vdwjidx2A,vdwjidx2B;
 841     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 842     int              vdwjidx3A,vdwjidx3B;
 843     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 844     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 845     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 846     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 847     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 848     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 849     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 850     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 851     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 852     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 853     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 854     real             *charge;
 855     _fjsp_v2r8       itab_tmp;
 856     _fjsp_v2r8       dummy_mask,cutoff_mask;
 857     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 858     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 859     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 860
 861     x                = xx[0];
 862     f                = ff[0];
 863
 864     nri              = nlist->nri;
 865     iinr             = nlist->iinr;
 866     jindex           = nlist->jindex;
 867     jjnr             = nlist->jjnr;
 868     shiftidx         = nlist->shift;
 869     gid              = nlist->gid;
 870     shiftvec         = fr->shift_vec[0];
 871     fshift           = fr->fshift[0];
 872     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 873     charge           = mdatoms->chargeA;
 874     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 875     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 876     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 877
 878     /* Setup water-specific parameters */
 879     inr              = nlist->iinr[0];
 880     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 881     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 882     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 883
 884     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 885     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 886     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 887     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 888     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 889     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 890     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 891     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 892     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 893     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 894     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 895     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 896
 897     /* Avoid stupid compiler warnings */
 898     jnrA = jnrB = 0;
 899     j_coord_offsetA = 0;
 900     j_coord_offsetB = 0;
 901
 902     outeriter        = 0;
 903     inneriter        = 0;
 904
 905     /* Start outer loop over neighborlists */
 906     for(iidx=0; iidx<nri; iidx++)
 907     {
 908         /* Load shift vector for this list */
 909         i_shift_offset   = DIM*shiftidx[iidx];
 910
 911         /* Load limits for loop over neighbors */
 912         j_index_start    = jindex[iidx];
 913         j_index_end      = jindex[iidx+1];
 914
 915         /* Get outer coordinate index */
 916         inr              = iinr[iidx];
 917         i_coord_offset   = DIM*inr;
 918
 919         /* Load i particle coords and add shift vector */
 920         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 921                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 922
 923         fix1             = _fjsp_setzero_v2r8();
 924         fiy1             = _fjsp_setzero_v2r8();
 925         fiz1             = _fjsp_setzero_v2r8();
 926         fix2             = _fjsp_setzero_v2r8();
 927         fiy2             = _fjsp_setzero_v2r8();
 928         fiz2             = _fjsp_setzero_v2r8();
 929         fix3             = _fjsp_setzero_v2r8();
 930         fiy3             = _fjsp_setzero_v2r8();
 931         fiz3             = _fjsp_setzero_v2r8();
 932
 933         /* Start inner kernel loop */
 934         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 935         {
 936
 937             /* Get j neighbor index, and coordinate index */
 938             jnrA             = jjnr[jidx];
 939             jnrB             = jjnr[jidx+1];
 940             j_coord_offsetA  = DIM*jnrA;
 941             j_coord_offsetB  = DIM*jnrB;
 942
 943             /* load j atom coordinates */
 944             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 945                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 946
 947             /* Calculate displacement vector */
 948             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 949             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 950             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 951             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 952             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 953             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 954             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 955             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 956             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 957             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 958             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 959             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 960             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 961             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 962             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 963             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 964             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 965             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 966             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 967             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 968             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 969             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 970             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 971             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 972             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 973             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 974             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 975
 976             /* Calculate squared distance and things based on it */
 977             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 978             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 979             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 980             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 981             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 982             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 983             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 984             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 985             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 986
 987             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 988             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 989             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 990             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 991             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 992             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 993             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 994             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 995             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 996
 997             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 998             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 999             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1000             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1001             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1002             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1003             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1004             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1005             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1006
1007             fjx1             = _fjsp_setzero_v2r8();
1008             fjy1             = _fjsp_setzero_v2r8();
1009             fjz1             = _fjsp_setzero_v2r8();
1010             fjx2             = _fjsp_setzero_v2r8();
1011             fjy2             = _fjsp_setzero_v2r8();
1012             fjz2             = _fjsp_setzero_v2r8();
1013             fjx3             = _fjsp_setzero_v2r8();
1014             fjy3             = _fjsp_setzero_v2r8();
1015             fjz3             = _fjsp_setzero_v2r8();
1016
1017             /**************************
1018              * CALCULATE INTERACTIONS *
1019              **************************/
1020
1021             /* REACTION-FIELD ELECTROSTATICS */
1022             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1023
1024             fscal            = felec;
1025
1026             /* Update vectorial force */
1027             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1028             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1029             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1030
1031             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1032             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1033             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1034
1035             /**************************
1036              * CALCULATE INTERACTIONS *
1037              **************************/
1038
1039             /* REACTION-FIELD ELECTROSTATICS */
1040             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1041
1042             fscal            = felec;
1043
1044             /* Update vectorial force */
1045             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1046             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1047             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1048
1049             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1050             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1051             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1052
1053             /**************************
1054              * CALCULATE INTERACTIONS *
1055              **************************/
1056
1057             /* REACTION-FIELD ELECTROSTATICS */
1058             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
1059
1060             fscal            = felec;
1061
1062             /* Update vectorial force */
1063             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1064             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1065             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1066
1067             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1068             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1069             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1070
1071             /**************************
1072              * CALCULATE INTERACTIONS *
1073              **************************/
1074
1075             /* REACTION-FIELD ELECTROSTATICS */
1076             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
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             /* REACTION-FIELD ELECTROSTATICS */
1094             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1095
1096             fscal            = felec;
1097
1098             /* Update vectorial force */
1099             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1100             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1101             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1102
1103             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1104             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1105             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1106
1107             /**************************
1108              * CALCULATE INTERACTIONS *
1109              **************************/
1110
1111             /* REACTION-FIELD ELECTROSTATICS */
1112             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
1113
1114             fscal            = felec;
1115
1116             /* Update vectorial force */
1117             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1118             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1119             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1120
1121             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1122             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1123             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1124
1125             /**************************
1126              * CALCULATE INTERACTIONS *
1127              **************************/
1128
1129             /* REACTION-FIELD ELECTROSTATICS */
1130             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
1131
1132             fscal            = felec;
1133
1134             /* Update vectorial force */
1135             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1136             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1137             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1138
1139             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1140             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1141             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             /* REACTION-FIELD ELECTROSTATICS */
1148             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
1149
1150             fscal            = felec;
1151
1152             /* Update vectorial force */
1153             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1154             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1155             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1156
1157             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1158             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1159             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1160
1161             /**************************
1162              * CALCULATE INTERACTIONS *
1163              **************************/
1164
1165             /* REACTION-FIELD ELECTROSTATICS */
1166             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
1167
1168             fscal            = felec;
1169
1170             /* Update vectorial force */
1171             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1172             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1173             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1174
1175             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1176             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1177             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1178
1179             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);
1180
1181             /* Inner loop uses 270 flops */
1182         }
1183
1184         if(jidx<j_index_end)
1185         {
1186
1187             jnrA             = jjnr[jidx];
1188             j_coord_offsetA  = DIM*jnrA;
1189
1190             /* load j atom coordinates */
1191             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1192                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1193
1194             /* Calculate displacement vector */
1195             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1196             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1197             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1198             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1199             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1200             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1201             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1202             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1203             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1204             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1205             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1206             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1207             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1208             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1209             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1210             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1211             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1212             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1213             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1214             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1215             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1216             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1217             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1218             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1219             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1220             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1221             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1222
1223             /* Calculate squared distance and things based on it */
1224             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1225             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1226             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1227             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1228             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1229             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1230             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1231             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1232             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1233
1234             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1235             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1236             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1237             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1238             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1239             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1240             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1241             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1242             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1243
1244             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1245             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1246             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1247             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1248             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1249             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1250             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1251             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1252             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1253
1254             fjx1             = _fjsp_setzero_v2r8();
1255             fjy1             = _fjsp_setzero_v2r8();
1256             fjz1             = _fjsp_setzero_v2r8();
1257             fjx2             = _fjsp_setzero_v2r8();
1258             fjy2             = _fjsp_setzero_v2r8();
1259             fjz2             = _fjsp_setzero_v2r8();
1260             fjx3             = _fjsp_setzero_v2r8();
1261             fjy3             = _fjsp_setzero_v2r8();
1262             fjz3             = _fjsp_setzero_v2r8();
1263
1264             /**************************
1265              * CALCULATE INTERACTIONS *
1266              **************************/
1267
1268             /* REACTION-FIELD ELECTROSTATICS */
1269             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1270
1271             fscal            = felec;
1272
1273             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1274
1275             /* Update vectorial force */
1276             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1277             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1278             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1279
1280             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1281             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1282             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1283
1284             /**************************
1285              * CALCULATE INTERACTIONS *
1286              **************************/
1287
1288             /* REACTION-FIELD ELECTROSTATICS */
1289             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1290
1291             fscal            = felec;
1292
1293             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1294
1295             /* Update vectorial force */
1296             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1297             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1298             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1299
1300             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1301             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1302             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1303
1304             /**************************
1305              * CALCULATE INTERACTIONS *
1306              **************************/
1307
1308             /* REACTION-FIELD ELECTROSTATICS */
1309             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
1310
1311             fscal            = felec;
1312
1313             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1314
1315             /* Update vectorial force */
1316             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1317             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1318             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1319
1320             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1321             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1322             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1323
1324             /**************************
1325              * CALCULATE INTERACTIONS *
1326              **************************/
1327
1328             /* REACTION-FIELD ELECTROSTATICS */
1329             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1330
1331             fscal            = felec;
1332
1333             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1334
1335             /* Update vectorial force */
1336             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1337             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1338             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1339
1340             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1341             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1342             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             /* REACTION-FIELD ELECTROSTATICS */
1349             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1350
1351             fscal            = felec;
1352
1353             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1354
1355             /* Update vectorial force */
1356             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1357             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1358             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1359
1360             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1361             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1362             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1363
1364             /**************************
1365              * CALCULATE INTERACTIONS *
1366              **************************/
1367
1368             /* REACTION-FIELD ELECTROSTATICS */
1369             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
1370
1371             fscal            = felec;
1372
1373             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1374
1375             /* Update vectorial force */
1376             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1377             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1378             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1379
1380             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1381             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1382             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             /* REACTION-FIELD ELECTROSTATICS */
1389             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
1390
1391             fscal            = felec;
1392
1393             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1394
1395             /* Update vectorial force */
1396             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1397             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1398             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1399
1400             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1401             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1402             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1403
1404             /**************************
1405              * CALCULATE INTERACTIONS *
1406              **************************/
1407
1408             /* REACTION-FIELD ELECTROSTATICS */
1409             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
1410
1411             fscal            = felec;
1412
1413             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1414
1415             /* Update vectorial force */
1416             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1417             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1418             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1419
1420             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1421             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1422             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1423
1424             /**************************
1425              * CALCULATE INTERACTIONS *
1426              **************************/
1427
1428             /* REACTION-FIELD ELECTROSTATICS */
1429             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
1430
1431             fscal            = felec;
1432
1433             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1434
1435             /* Update vectorial force */
1436             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1437             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1438             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1439
1440             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1441             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1442             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1443
1444             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1445
1446             /* Inner loop uses 270 flops */
1447         }
1448
1449         /* End of innermost loop */
1450
1451         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1452                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1453
1454         /* Increment number of inner iterations */
1455         inneriter                  += j_index_end - j_index_start;
1456
1457         /* Outer loop uses 18 flops */
1458     }
1459
1460     /* Increment number of outer iterations */
1461     outeriter        += nri;
1462
1463     /* Update outer/inner flops */
1464
1465     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1466 }