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 "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_ElecRF_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: ReactionField
  54  * VdW interaction:            None
  55  * Geometry:                   Water4-Water4
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecRF_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     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 124     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 125     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 130     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 131     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 132
 133     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 134     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 135     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 136     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 137     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 138     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 139     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 140     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 141     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 142     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 143     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 144     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 145
 146     /* Avoid stupid compiler warnings */
 147     jnrA = jnrB = 0;
 148     j_coord_offsetA = 0;
 149     j_coord_offsetB = 0;
 150
 151     outeriter        = 0;
 152     inneriter        = 0;
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 170                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 171
 172         fix1             = _fjsp_setzero_v2r8();
 173         fiy1             = _fjsp_setzero_v2r8();
 174         fiz1             = _fjsp_setzero_v2r8();
 175         fix2             = _fjsp_setzero_v2r8();
 176         fiy2             = _fjsp_setzero_v2r8();
 177         fiz2             = _fjsp_setzero_v2r8();
 178         fix3             = _fjsp_setzero_v2r8();
 179         fiy3             = _fjsp_setzero_v2r8();
 180         fiz3             = _fjsp_setzero_v2r8();
 181
 182         /* Reset potential sums */
 183         velecsum         = _fjsp_setzero_v2r8();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194
 195             /* load j atom coordinates */
 196             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 197                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 198
 199             /* Calculate displacement vector */
 200             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 201             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 202             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 203             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 204             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 205             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 206             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 207             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 208             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 209             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 210             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 211             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 212             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 213             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 214             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 215             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 216             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 217             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 218             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 219             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 220             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 221             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 222             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 223             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 224             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 225             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 226             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 227
 228             /* Calculate squared distance and things based on it */
 229             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 230             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 231             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 232             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 233             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 234             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 235             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 236             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 237             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 238
 239             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 240             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 241             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 242             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 243             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 244             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 245             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 246             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 247             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 248
 249             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 250             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 251             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 252             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 253             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 254             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 255             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 256             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 257             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 258
 259             fjx1             = _fjsp_setzero_v2r8();
 260             fjy1             = _fjsp_setzero_v2r8();
 261             fjz1             = _fjsp_setzero_v2r8();
 262             fjx2             = _fjsp_setzero_v2r8();
 263             fjy2             = _fjsp_setzero_v2r8();
 264             fjz2             = _fjsp_setzero_v2r8();
 265             fjx3             = _fjsp_setzero_v2r8();
 266             fjy3             = _fjsp_setzero_v2r8();
 267             fjz3             = _fjsp_setzero_v2r8();
 268
 269             /**************************
 270              * CALCULATE INTERACTIONS *
 271              **************************/
 272
 273             /* REACTION-FIELD ELECTROSTATICS */
 274             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 275             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 276
 277             /* Update potential sum for this i atom from the interaction with this j atom. */
 278             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 279
 280             fscal            = felec;
 281
 282             /* Update vectorial force */
 283             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 284             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 285             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 286
 287             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 288             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 289             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 290
 291             /**************************
 292              * CALCULATE INTERACTIONS *
 293              **************************/
 294
 295             /* REACTION-FIELD ELECTROSTATICS */
 296             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 297             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 298
 299             /* Update potential sum for this i atom from the interaction with this j atom. */
 300             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 301
 302             fscal            = felec;
 303
 304             /* Update vectorial force */
 305             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 306             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 307             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 308
 309             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 310             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 311             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             /* REACTION-FIELD ELECTROSTATICS */
 318             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq13,rinv13),crf));
 319             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
 320
 321             /* Update potential sum for this i atom from the interaction with this j atom. */
 322             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 323
 324             fscal            = felec;
 325
 326             /* Update vectorial force */
 327             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 328             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 329             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 330
 331             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 332             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 333             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             /* REACTION-FIELD ELECTROSTATICS */
 340             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 341             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 345
 346             fscal            = felec;
 347
 348             /* Update vectorial force */
 349             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 350             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 351             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 352
 353             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 354             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 355             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             /* REACTION-FIELD ELECTROSTATICS */
 362             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 363             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370             /* Update vectorial force */
 371             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 372             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 373             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 374
 375             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 376             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 377             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* REACTION-FIELD ELECTROSTATICS */
 384             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq23,rinv23),crf));
 385             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392             /* Update vectorial force */
 393             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 394             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 395             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 396
 397             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 398             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 399             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             /* REACTION-FIELD ELECTROSTATICS */
 406             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq31,rinv31),crf));
 407             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
 408
 409             /* Update potential sum for this i atom from the interaction with this j atom. */
 410             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414             /* Update vectorial force */
 415             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 416             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 417             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 418
 419             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 420             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 421             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             /* REACTION-FIELD ELECTROSTATICS */
 428             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq32,rinv32),crf));
 429             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
 430
 431             /* Update potential sum for this i atom from the interaction with this j atom. */
 432             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436             /* Update vectorial force */
 437             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 438             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 439             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 440
 441             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 442             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 443             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* REACTION-FIELD ELECTROSTATICS */
 450             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq33,rinv33),crf));
 451             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
 452
 453             /* Update potential sum for this i atom from the interaction with this j atom. */
 454             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 455
 456             fscal            = felec;
 457
 458             /* Update vectorial force */
 459             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 460             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 461             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 462
 463             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 464             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 465             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 466
 467             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);
 468
 469             /* Inner loop uses 315 flops */
 470         }
 471
 472         if(jidx<j_index_end)
 473         {
 474
 475             jnrA             = jjnr[jidx];
 476             j_coord_offsetA  = DIM*jnrA;
 477
 478             /* load j atom coordinates */
 479             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
 480                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 481
 482             /* Calculate displacement vector */
 483             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 484             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 485             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 486             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 487             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 488             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 489             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 490             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 491             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 492             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 493             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 494             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 495             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 496             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 497             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 498             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 499             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 500             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 501             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 502             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 503             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 504             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 505             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 506             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 507             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 508             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 509             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 510
 511             /* Calculate squared distance and things based on it */
 512             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 513             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 514             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 515             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 516             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 517             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 518             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 519             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 520             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 521
 522             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 523             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 524             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 525             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 526             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 527             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 528             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 529             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 530             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 531
 532             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 533             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 534             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
 535             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 536             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 537             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
 538             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
 539             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
 540             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
 541
 542             fjx1             = _fjsp_setzero_v2r8();
 543             fjy1             = _fjsp_setzero_v2r8();
 544             fjz1             = _fjsp_setzero_v2r8();
 545             fjx2             = _fjsp_setzero_v2r8();
 546             fjy2             = _fjsp_setzero_v2r8();
 547             fjz2             = _fjsp_setzero_v2r8();
 548             fjx3             = _fjsp_setzero_v2r8();
 549             fjy3             = _fjsp_setzero_v2r8();
 550             fjz3             = _fjsp_setzero_v2r8();
 551
 552             /**************************
 553              * CALCULATE INTERACTIONS *
 554              **************************/
 555
 556             /* REACTION-FIELD ELECTROSTATICS */
 557             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 558             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 559
 560             /* Update potential sum for this i atom from the interaction with this j atom. */
 561             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 562             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 563
 564             fscal            = felec;
 565
 566             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 567
 568             /* Update vectorial force */
 569             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 570             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 571             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 572
 573             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 574             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 575             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 576
 577             /**************************
 578              * CALCULATE INTERACTIONS *
 579              **************************/
 580
 581             /* REACTION-FIELD ELECTROSTATICS */
 582             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 583             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 584
 585             /* Update potential sum for this i atom from the interaction with this j atom. */
 586             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 587             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 588
 589             fscal            = felec;
 590
 591             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 592
 593             /* Update vectorial force */
 594             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 595             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 596             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 597
 598             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 599             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 600             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 601
 602             /**************************
 603              * CALCULATE INTERACTIONS *
 604              **************************/
 605
 606             /* REACTION-FIELD ELECTROSTATICS */
 607             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq13,rinv13),crf));
 608             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
 609
 610             /* Update potential sum for this i atom from the interaction with this j atom. */
 611             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 612             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 613
 614             fscal            = felec;
 615
 616             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 617
 618             /* Update vectorial force */
 619             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
 620             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
 621             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
 622
 623             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
 624             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
 625             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
 626
 627             /**************************
 628              * CALCULATE INTERACTIONS *
 629              **************************/
 630
 631             /* REACTION-FIELD ELECTROSTATICS */
 632             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 633             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 634
 635             /* Update potential sum for this i atom from the interaction with this j atom. */
 636             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 637             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 638
 639             fscal            = felec;
 640
 641             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 642
 643             /* Update vectorial force */
 644             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 645             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 646             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 647
 648             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 649             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 650             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 651
 652             /**************************
 653              * CALCULATE INTERACTIONS *
 654              **************************/
 655
 656             /* REACTION-FIELD ELECTROSTATICS */
 657             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 658             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 659
 660             /* Update potential sum for this i atom from the interaction with this j atom. */
 661             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 662             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 663
 664             fscal            = felec;
 665
 666             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 667
 668             /* Update vectorial force */
 669             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 670             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 671             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 672
 673             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 674             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 675             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             /* REACTION-FIELD ELECTROSTATICS */
 682             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq23,rinv23),crf));
 683             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
 684
 685             /* Update potential sum for this i atom from the interaction with this j atom. */
 686             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 687             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 688
 689             fscal            = felec;
 690
 691             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 692
 693             /* Update vectorial force */
 694             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
 695             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
 696             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
 697
 698             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
 699             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
 700             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
 701
 702             /**************************
 703              * CALCULATE INTERACTIONS *
 704              **************************/
 705
 706             /* REACTION-FIELD ELECTROSTATICS */
 707             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq31,rinv31),crf));
 708             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
 709
 710             /* Update potential sum for this i atom from the interaction with this j atom. */
 711             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 712             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 713
 714             fscal            = felec;
 715
 716             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 717
 718             /* Update vectorial force */
 719             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
 720             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
 721             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
 722
 723             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
 724             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
 725             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             /* REACTION-FIELD ELECTROSTATICS */
 732             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq32,rinv32),crf));
 733             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
 734
 735             /* Update potential sum for this i atom from the interaction with this j atom. */
 736             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 737             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 738
 739             fscal            = felec;
 740
 741             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 742
 743             /* Update vectorial force */
 744             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
 745             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
 746             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
 747
 748             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
 749             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
 750             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             /* REACTION-FIELD ELECTROSTATICS */
 757             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq33,rinv33),crf));
 758             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
 759
 760             /* Update potential sum for this i atom from the interaction with this j atom. */
 761             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 762             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 763
 764             fscal            = felec;
 765
 766             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 767
 768             /* Update vectorial force */
 769             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
 770             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
 771             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
 772
 773             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
 774             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
 775             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
 776
 777             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 778
 779             /* Inner loop uses 315 flops */
 780         }
 781
 782         /* End of innermost loop */
 783
 784         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 785                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 786
 787         ggid                        = gid[iidx];
 788         /* Update potential energies */
 789         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 790
 791         /* Increment number of inner iterations */
 792         inneriter                  += j_index_end - j_index_start;
 793
 794         /* Outer loop uses 19 flops */
 795     }
 796
 797     /* Increment number of outer iterations */
 798     outeriter        += nri;
 799
 800     /* Update outer/inner flops */
 801
 802     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*315);
 803 }
 804 /*
 805  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
 806  * Electrostatics interaction: ReactionField
 807  * VdW interaction:            None
 808  * Geometry:                   Water4-Water4
 809  * Calculate force/pot:        Force
 810  */
 811 void
 812 nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
 813                     (t_nblist                    * gmx_restrict       nlist,
 814                      rvec                        * gmx_restrict          xx,
 815                      rvec                        * gmx_restrict          ff,
 816                      t_forcerec                  * gmx_restrict          fr,
 817                      t_mdatoms                   * gmx_restrict     mdatoms,
 818                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 819                      t_nrnb                      * gmx_restrict        nrnb)
 820 {
 821     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 822      * just 0 for non-waters.
 823      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 824      * jnr indices corresponding to data put in the four positions in the SIMD register.
 825      */
 826     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 827     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 828     int              jnrA,jnrB;
 829     int              j_coord_offsetA,j_coord_offsetB;
 830     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 831     real             rcutoff_scalar;
 832     real             *shiftvec,*fshift,*x,*f;
 833     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 834     int              vdwioffset1;
 835     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 836     int              vdwioffset2;
 837     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 838     int              vdwioffset3;
 839     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 840     int              vdwjidx1A,vdwjidx1B;
 841     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 842     int              vdwjidx2A,vdwjidx2B;
 843     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 844     int              vdwjidx3A,vdwjidx3B;
 845     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 846     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 847     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 848     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 849     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 850     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 851     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 852     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 853     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 854     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 855     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 856     real             *charge;
 857     _fjsp_v2r8       itab_tmp;
 858     _fjsp_v2r8       dummy_mask,cutoff_mask;
 859     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 860     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 861     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 862
 863     x                = xx[0];
 864     f                = ff[0];
 865
 866     nri              = nlist->nri;
 867     iinr             = nlist->iinr;
 868     jindex           = nlist->jindex;
 869     jjnr             = nlist->jjnr;
 870     shiftidx         = nlist->shift;
 871     gid              = nlist->gid;
 872     shiftvec         = fr->shift_vec[0];
 873     fshift           = fr->fshift[0];
 874     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 875     charge           = mdatoms->chargeA;
 876     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 877     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 878     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 879
 880     /* Setup water-specific parameters */
 881     inr              = nlist->iinr[0];
 882     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 883     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 884     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
 885
 886     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 887     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 888     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
 889     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 890     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 891     qq13             = _fjsp_mul_v2r8(iq1,jq3);
 892     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 893     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 894     qq23             = _fjsp_mul_v2r8(iq2,jq3);
 895     qq31             = _fjsp_mul_v2r8(iq3,jq1);
 896     qq32             = _fjsp_mul_v2r8(iq3,jq2);
 897     qq33             = _fjsp_mul_v2r8(iq3,jq3);
 898
 899     /* Avoid stupid compiler warnings */
 900     jnrA = jnrB = 0;
 901     j_coord_offsetA = 0;
 902     j_coord_offsetB = 0;
 903
 904     outeriter        = 0;
 905     inneriter        = 0;
 906
 907     /* Start outer loop over neighborlists */
 908     for(iidx=0; iidx<nri; iidx++)
 909     {
 910         /* Load shift vector for this list */
 911         i_shift_offset   = DIM*shiftidx[iidx];
 912
 913         /* Load limits for loop over neighbors */
 914         j_index_start    = jindex[iidx];
 915         j_index_end      = jindex[iidx+1];
 916
 917         /* Get outer coordinate index */
 918         inr              = iinr[iidx];
 919         i_coord_offset   = DIM*inr;
 920
 921         /* Load i particle coords and add shift vector */
 922         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 923                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 924
 925         fix1             = _fjsp_setzero_v2r8();
 926         fiy1             = _fjsp_setzero_v2r8();
 927         fiz1             = _fjsp_setzero_v2r8();
 928         fix2             = _fjsp_setzero_v2r8();
 929         fiy2             = _fjsp_setzero_v2r8();
 930         fiz2             = _fjsp_setzero_v2r8();
 931         fix3             = _fjsp_setzero_v2r8();
 932         fiy3             = _fjsp_setzero_v2r8();
 933         fiz3             = _fjsp_setzero_v2r8();
 934
 935         /* Start inner kernel loop */
 936         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 937         {
 938
 939             /* Get j neighbor index, and coordinate index */
 940             jnrA             = jjnr[jidx];
 941             jnrB             = jjnr[jidx+1];
 942             j_coord_offsetA  = DIM*jnrA;
 943             j_coord_offsetB  = DIM*jnrB;
 944
 945             /* load j atom coordinates */
 946             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 947                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 948
 949             /* Calculate displacement vector */
 950             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 951             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 952             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 953             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 954             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 955             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 956             dx13             = _fjsp_sub_v2r8(ix1,jx3);
 957             dy13             = _fjsp_sub_v2r8(iy1,jy3);
 958             dz13             = _fjsp_sub_v2r8(iz1,jz3);
 959             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 960             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 961             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 962             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 963             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 964             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 965             dx23             = _fjsp_sub_v2r8(ix2,jx3);
 966             dy23             = _fjsp_sub_v2r8(iy2,jy3);
 967             dz23             = _fjsp_sub_v2r8(iz2,jz3);
 968             dx31             = _fjsp_sub_v2r8(ix3,jx1);
 969             dy31             = _fjsp_sub_v2r8(iy3,jy1);
 970             dz31             = _fjsp_sub_v2r8(iz3,jz1);
 971             dx32             = _fjsp_sub_v2r8(ix3,jx2);
 972             dy32             = _fjsp_sub_v2r8(iy3,jy2);
 973             dz32             = _fjsp_sub_v2r8(iz3,jz2);
 974             dx33             = _fjsp_sub_v2r8(ix3,jx3);
 975             dy33             = _fjsp_sub_v2r8(iy3,jy3);
 976             dz33             = _fjsp_sub_v2r8(iz3,jz3);
 977
 978             /* Calculate squared distance and things based on it */
 979             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 980             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 981             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
 982             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 983             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 984             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
 985             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
 986             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
 987             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
 988
 989             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 990             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 991             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
 992             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 993             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 994             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
 995             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
 996             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
 997             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
 998
 999             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1000             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1001             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1002             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1003             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1004             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1005             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1006             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1007             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1008
1009             fjx1             = _fjsp_setzero_v2r8();
1010             fjy1             = _fjsp_setzero_v2r8();
1011             fjz1             = _fjsp_setzero_v2r8();
1012             fjx2             = _fjsp_setzero_v2r8();
1013             fjy2             = _fjsp_setzero_v2r8();
1014             fjz2             = _fjsp_setzero_v2r8();
1015             fjx3             = _fjsp_setzero_v2r8();
1016             fjy3             = _fjsp_setzero_v2r8();
1017             fjz3             = _fjsp_setzero_v2r8();
1018
1019             /**************************
1020              * CALCULATE INTERACTIONS *
1021              **************************/
1022
1023             /* REACTION-FIELD ELECTROSTATICS */
1024             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1025
1026             fscal            = felec;
1027
1028             /* Update vectorial force */
1029             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1030             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1031             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1032
1033             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1034             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1035             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1036
1037             /**************************
1038              * CALCULATE INTERACTIONS *
1039              **************************/
1040
1041             /* REACTION-FIELD ELECTROSTATICS */
1042             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1043
1044             fscal            = felec;
1045
1046             /* Update vectorial force */
1047             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1048             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1049             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1050
1051             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1052             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1053             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1054
1055             /**************************
1056              * CALCULATE INTERACTIONS *
1057              **************************/
1058
1059             /* REACTION-FIELD ELECTROSTATICS */
1060             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
1061
1062             fscal            = felec;
1063
1064             /* Update vectorial force */
1065             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1066             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1067             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1068
1069             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1070             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1071             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1072
1073             /**************************
1074              * CALCULATE INTERACTIONS *
1075              **************************/
1076
1077             /* REACTION-FIELD ELECTROSTATICS */
1078             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1079
1080             fscal            = felec;
1081
1082             /* Update vectorial force */
1083             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1084             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1085             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1086
1087             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1088             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1089             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1090
1091             /**************************
1092              * CALCULATE INTERACTIONS *
1093              **************************/
1094
1095             /* REACTION-FIELD ELECTROSTATICS */
1096             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1097
1098             fscal            = felec;
1099
1100             /* Update vectorial force */
1101             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1102             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1103             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1104
1105             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1106             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1107             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1108
1109             /**************************
1110              * CALCULATE INTERACTIONS *
1111              **************************/
1112
1113             /* REACTION-FIELD ELECTROSTATICS */
1114             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
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             /* REACTION-FIELD ELECTROSTATICS */
1132             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
1133
1134             fscal            = felec;
1135
1136             /* Update vectorial force */
1137             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1138             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1139             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1140
1141             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1142             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1143             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1144
1145             /**************************
1146              * CALCULATE INTERACTIONS *
1147              **************************/
1148
1149             /* REACTION-FIELD ELECTROSTATICS */
1150             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
1151
1152             fscal            = felec;
1153
1154             /* Update vectorial force */
1155             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1156             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1157             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1158
1159             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1160             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1161             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1162
1163             /**************************
1164              * CALCULATE INTERACTIONS *
1165              **************************/
1166
1167             /* REACTION-FIELD ELECTROSTATICS */
1168             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
1169
1170             fscal            = felec;
1171
1172             /* Update vectorial force */
1173             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1174             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1175             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1176
1177             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1178             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1179             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1180
1181             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);
1182
1183             /* Inner loop uses 270 flops */
1184         }
1185
1186         if(jidx<j_index_end)
1187         {
1188
1189             jnrA             = jjnr[jidx];
1190             j_coord_offsetA  = DIM*jnrA;
1191
1192             /* load j atom coordinates */
1193             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1194                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1195
1196             /* Calculate displacement vector */
1197             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1198             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1199             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1200             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1201             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1202             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1203             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1204             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1205             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1206             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1207             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1208             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1209             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1210             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1211             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1212             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1213             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1214             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1215             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1216             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1217             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1218             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1219             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1220             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1221             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1222             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1223             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1224
1225             /* Calculate squared distance and things based on it */
1226             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1227             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1228             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1229             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1230             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1231             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1232             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1233             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1234             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1235
1236             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1237             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1238             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1239             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1240             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1241             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1242             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1243             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1244             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1245
1246             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1247             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1248             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1249             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1250             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1251             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1252             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1253             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1254             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1255
1256             fjx1             = _fjsp_setzero_v2r8();
1257             fjy1             = _fjsp_setzero_v2r8();
1258             fjz1             = _fjsp_setzero_v2r8();
1259             fjx2             = _fjsp_setzero_v2r8();
1260             fjy2             = _fjsp_setzero_v2r8();
1261             fjz2             = _fjsp_setzero_v2r8();
1262             fjx3             = _fjsp_setzero_v2r8();
1263             fjy3             = _fjsp_setzero_v2r8();
1264             fjz3             = _fjsp_setzero_v2r8();
1265
1266             /**************************
1267              * CALCULATE INTERACTIONS *
1268              **************************/
1269
1270             /* REACTION-FIELD ELECTROSTATICS */
1271             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1272
1273             fscal            = felec;
1274
1275             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1276
1277             /* Update vectorial force */
1278             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1279             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1280             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1281
1282             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1283             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1284             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1285
1286             /**************************
1287              * CALCULATE INTERACTIONS *
1288              **************************/
1289
1290             /* REACTION-FIELD ELECTROSTATICS */
1291             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1292
1293             fscal            = felec;
1294
1295             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1296
1297             /* Update vectorial force */
1298             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1299             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1300             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1301
1302             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1303             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1304             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* REACTION-FIELD ELECTROSTATICS */
1311             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
1312
1313             fscal            = felec;
1314
1315             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1316
1317             /* Update vectorial force */
1318             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1319             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1320             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1321
1322             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1323             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1324             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1325
1326             /**************************
1327              * CALCULATE INTERACTIONS *
1328              **************************/
1329
1330             /* REACTION-FIELD ELECTROSTATICS */
1331             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1332
1333             fscal            = felec;
1334
1335             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1336
1337             /* Update vectorial force */
1338             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1339             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1340             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1341
1342             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1343             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1344             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1345
1346             /**************************
1347              * CALCULATE INTERACTIONS *
1348              **************************/
1349
1350             /* REACTION-FIELD ELECTROSTATICS */
1351             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1352
1353             fscal            = felec;
1354
1355             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1356
1357             /* Update vectorial force */
1358             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1359             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1360             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1361
1362             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1363             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1364             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* REACTION-FIELD ELECTROSTATICS */
1371             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
1372
1373             fscal            = felec;
1374
1375             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1376
1377             /* Update vectorial force */
1378             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1379             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1380             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1381
1382             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1383             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1384             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             /* REACTION-FIELD ELECTROSTATICS */
1391             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
1392
1393             fscal            = felec;
1394
1395             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1396
1397             /* Update vectorial force */
1398             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1399             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1400             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1401
1402             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1403             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1404             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1405
1406             /**************************
1407              * CALCULATE INTERACTIONS *
1408              **************************/
1409
1410             /* REACTION-FIELD ELECTROSTATICS */
1411             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
1412
1413             fscal            = felec;
1414
1415             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1416
1417             /* Update vectorial force */
1418             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1419             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1420             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1421
1422             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1423             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1424             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1425
1426             /**************************
1427              * CALCULATE INTERACTIONS *
1428              **************************/
1429
1430             /* REACTION-FIELD ELECTROSTATICS */
1431             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
1432
1433             fscal            = felec;
1434
1435             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1436
1437             /* Update vectorial force */
1438             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1439             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1440             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1441
1442             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1443             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1444             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1445
1446             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1447
1448             /* Inner loop uses 270 flops */
1449         }
1450
1451         /* End of innermost loop */
1452
1453         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1454                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1455
1456         /* Increment number of inner iterations */
1457         inneriter                  += j_index_end - j_index_start;
1458
1459         /* Outer loop uses 18 flops */
1460     }
1461
1462     /* Increment number of outer iterations */
1463     outeriter        += nri;
1464
1465     /* Update outer/inner flops */
1466
1467     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1468 }