src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRF_VdwNone_GeomW3W3_sparc64_hpc_ace_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sparc64_hpc_ace_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "gromacs/legacyheaders/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: ReactionField
  54  * VdW interaction:            None
  55  * Geometry:                   Water3-Water3
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  94     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  95     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  96     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  97     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 100     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 103     real             *charge;
 104     _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     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 130     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 131     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 132
 133     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 134     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 135     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 136     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 137     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 138     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 139     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 140     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 141     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 142     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 143     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 144     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 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,
 170                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 171
 172         fix0             = _fjsp_setzero_v2r8();
 173         fiy0             = _fjsp_setzero_v2r8();
 174         fiz0             = _fjsp_setzero_v2r8();
 175         fix1             = _fjsp_setzero_v2r8();
 176         fiy1             = _fjsp_setzero_v2r8();
 177         fiz1             = _fjsp_setzero_v2r8();
 178         fix2             = _fjsp_setzero_v2r8();
 179         fiy2             = _fjsp_setzero_v2r8();
 180         fiz2             = _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,x+j_coord_offsetB,
 197                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 198
 199             /* Calculate displacement vector */
 200             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 201             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 202             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 203             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 204             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 205             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 206             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 207             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 208             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 209             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 210             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 211             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 212             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 213             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 214             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 215             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 216             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 217             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 218             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 219             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 220             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 221             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 222             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 223             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 224             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 225             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 226             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 227
 228             /* Calculate squared distance and things based on it */
 229             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 230             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 231             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 232             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 233             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 234             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 235             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 236             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 237             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 238
 239             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 240             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 241             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 242             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 243             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 244             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 245             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 246             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 247             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 248
 249             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 250             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 251             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 252             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 253             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 254             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 255             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 256             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 257             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 258
 259             fjx0             = _fjsp_setzero_v2r8();
 260             fjy0             = _fjsp_setzero_v2r8();
 261             fjz0             = _fjsp_setzero_v2r8();
 262             fjx1             = _fjsp_setzero_v2r8();
 263             fjy1             = _fjsp_setzero_v2r8();
 264             fjz1             = _fjsp_setzero_v2r8();
 265             fjx2             = _fjsp_setzero_v2r8();
 266             fjy2             = _fjsp_setzero_v2r8();
 267             fjz2             = _fjsp_setzero_v2r8();
 268
 269             /**************************
 270              * CALCULATE INTERACTIONS *
 271              **************************/
 272
 273             /* REACTION-FIELD ELECTROSTATICS */
 274             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 275             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,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             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 284             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 285             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 286
 287             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 288             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 289             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 290
 291             /**************************
 292              * CALCULATE INTERACTIONS *
 293              **************************/
 294
 295             /* REACTION-FIELD ELECTROSTATICS */
 296             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 297             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,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             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 306             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 307             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 308
 309             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 310             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 311             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             /* REACTION-FIELD ELECTROSTATICS */
 318             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 319             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,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             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 328             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 329             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 330
 331             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 332             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 333             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             /* REACTION-FIELD ELECTROSTATICS */
 340             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 341             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,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             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 350             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 351             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 352
 353             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 354             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 355             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             /* REACTION-FIELD ELECTROSTATICS */
 362             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 363             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,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             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 372             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 373             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 374
 375             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 376             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 377             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* REACTION-FIELD ELECTROSTATICS */
 384             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 385             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,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             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 394             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 395             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 396
 397             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 398             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 399             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             /* REACTION-FIELD ELECTROSTATICS */
 406             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 407             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,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             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 416             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 417             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 418
 419             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 420             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 421             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             /* REACTION-FIELD ELECTROSTATICS */
 428             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 429             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,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             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 438             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 439             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 440
 441             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 442             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 443             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* REACTION-FIELD ELECTROSTATICS */
 450             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 451             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,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             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 460             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 461             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 462
 463             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 464             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 465             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 466
 467             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 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,
 480                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 481
 482             /* Calculate displacement vector */
 483             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 484             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 485             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 486             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 487             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 488             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 489             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 490             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 491             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 492             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 493             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 494             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 495             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 496             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 497             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 498             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 499             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 500             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 501             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 502             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 503             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 504             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 505             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 506             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 507             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 508             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 509             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 510
 511             /* Calculate squared distance and things based on it */
 512             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 513             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 514             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 515             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 516             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 517             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 518             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 519             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 520             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 521
 522             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 523             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 524             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 525             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 526             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 527             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 528             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 529             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 530             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 531
 532             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 533             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 534             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 535             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 536             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 537             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 538             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 539             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 540             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 541
 542             fjx0             = _fjsp_setzero_v2r8();
 543             fjy0             = _fjsp_setzero_v2r8();
 544             fjz0             = _fjsp_setzero_v2r8();
 545             fjx1             = _fjsp_setzero_v2r8();
 546             fjy1             = _fjsp_setzero_v2r8();
 547             fjz1             = _fjsp_setzero_v2r8();
 548             fjx2             = _fjsp_setzero_v2r8();
 549             fjy2             = _fjsp_setzero_v2r8();
 550             fjz2             = _fjsp_setzero_v2r8();
 551
 552             /**************************
 553              * CALCULATE INTERACTIONS *
 554              **************************/
 555
 556             /* REACTION-FIELD ELECTROSTATICS */
 557             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 558             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,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             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 570             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 571             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 572
 573             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 574             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 575             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 576
 577             /**************************
 578              * CALCULATE INTERACTIONS *
 579              **************************/
 580
 581             /* REACTION-FIELD ELECTROSTATICS */
 582             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 583             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,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             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 595             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 596             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 597
 598             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 599             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 600             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 601
 602             /**************************
 603              * CALCULATE INTERACTIONS *
 604              **************************/
 605
 606             /* REACTION-FIELD ELECTROSTATICS */
 607             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 608             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,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             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 620             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 621             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 622
 623             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 624             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 625             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 626
 627             /**************************
 628              * CALCULATE INTERACTIONS *
 629              **************************/
 630
 631             /* REACTION-FIELD ELECTROSTATICS */
 632             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 633             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,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             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 645             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 646             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 647
 648             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 649             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 650             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 651
 652             /**************************
 653              * CALCULATE INTERACTIONS *
 654              **************************/
 655
 656             /* REACTION-FIELD ELECTROSTATICS */
 657             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 658             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,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             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 670             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 671             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 672
 673             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 674             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 675             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             /* REACTION-FIELD ELECTROSTATICS */
 682             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 683             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,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             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 695             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 696             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 697
 698             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 699             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 700             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 701
 702             /**************************
 703              * CALCULATE INTERACTIONS *
 704              **************************/
 705
 706             /* REACTION-FIELD ELECTROSTATICS */
 707             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 708             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,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             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 720             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 721             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 722
 723             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 724             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 725             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             /* REACTION-FIELD ELECTROSTATICS */
 732             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 733             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,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             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 745             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 746             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 747
 748             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 749             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 750             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             /* REACTION-FIELD ELECTROSTATICS */
 757             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 758             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,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             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 770             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 771             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 772
 773             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 774             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 775             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 776
 777             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 778
 779             /* Inner loop uses 315 flops */
 780         }
 781
 782         /* End of innermost loop */
 783
 784         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 785                                               f+i_coord_offset,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_W3W3_VF,outeriter*19 + inneriter*315);
 803 }
 804 /*
 805  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sparc64_hpc_ace_double
 806  * Electrostatics interaction: ReactionField
 807  * VdW interaction:            None
 808  * Geometry:                   Water3-Water3
 809  * Calculate force/pot:        Force
 810  */
 811 void
 812 nb_kernel_ElecRF_VdwNone_GeomW3W3_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              vdwioffset0;
 835     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 836     int              vdwioffset1;
 837     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 838     int              vdwioffset2;
 839     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 840     int              vdwjidx0A,vdwjidx0B;
 841     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 842     int              vdwjidx1A,vdwjidx1B;
 843     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 844     int              vdwjidx2A,vdwjidx2B;
 845     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 846     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 847     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 848     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 849     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 850     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 851     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 852     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 853     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 854     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 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     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 883     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 884     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 885
 886     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
 887     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
 888     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
 889     qq00             = _fjsp_mul_v2r8(iq0,jq0);
 890     qq01             = _fjsp_mul_v2r8(iq0,jq1);
 891     qq02             = _fjsp_mul_v2r8(iq0,jq2);
 892     qq10             = _fjsp_mul_v2r8(iq1,jq0);
 893     qq11             = _fjsp_mul_v2r8(iq1,jq1);
 894     qq12             = _fjsp_mul_v2r8(iq1,jq2);
 895     qq20             = _fjsp_mul_v2r8(iq2,jq0);
 896     qq21             = _fjsp_mul_v2r8(iq2,jq1);
 897     qq22             = _fjsp_mul_v2r8(iq2,jq2);
 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,
 923                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 924
 925         fix0             = _fjsp_setzero_v2r8();
 926         fiy0             = _fjsp_setzero_v2r8();
 927         fiz0             = _fjsp_setzero_v2r8();
 928         fix1             = _fjsp_setzero_v2r8();
 929         fiy1             = _fjsp_setzero_v2r8();
 930         fiz1             = _fjsp_setzero_v2r8();
 931         fix2             = _fjsp_setzero_v2r8();
 932         fiy2             = _fjsp_setzero_v2r8();
 933         fiz2             = _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,x+j_coord_offsetB,
 947                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 948
 949             /* Calculate displacement vector */
 950             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 951             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 952             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 953             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 954             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 955             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 956             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 957             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 958             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 959             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 960             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 961             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 962             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 963             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 964             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 965             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 966             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 967             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 968             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 969             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 970             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 971             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 972             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 973             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 974             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 975             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 976             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 977
 978             /* Calculate squared distance and things based on it */
 979             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 980             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 981             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 982             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 983             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 984             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 985             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 986             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 987             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 988
 989             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 990             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 991             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 992             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 993             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 994             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 995             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 996             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 997             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 998
 999             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1000             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1001             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1002             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1003             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1004             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1005             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1006             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1007             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1008
1009             fjx0             = _fjsp_setzero_v2r8();
1010             fjy0             = _fjsp_setzero_v2r8();
1011             fjz0             = _fjsp_setzero_v2r8();
1012             fjx1             = _fjsp_setzero_v2r8();
1013             fjy1             = _fjsp_setzero_v2r8();
1014             fjz1             = _fjsp_setzero_v2r8();
1015             fjx2             = _fjsp_setzero_v2r8();
1016             fjy2             = _fjsp_setzero_v2r8();
1017             fjz2             = _fjsp_setzero_v2r8();
1018
1019             /**************************
1020              * CALCULATE INTERACTIONS *
1021              **************************/
1022
1023             /* REACTION-FIELD ELECTROSTATICS */
1024             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1025
1026             fscal            = felec;
1027
1028             /* Update vectorial force */
1029             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1030             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1031             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1032
1033             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1034             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1035             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1036
1037             /**************************
1038              * CALCULATE INTERACTIONS *
1039              **************************/
1040
1041             /* REACTION-FIELD ELECTROSTATICS */
1042             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1043
1044             fscal            = felec;
1045
1046             /* Update vectorial force */
1047             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1048             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1049             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1050
1051             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1052             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1053             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1054
1055             /**************************
1056              * CALCULATE INTERACTIONS *
1057              **************************/
1058
1059             /* REACTION-FIELD ELECTROSTATICS */
1060             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1061
1062             fscal            = felec;
1063
1064             /* Update vectorial force */
1065             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1066             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1067             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1068
1069             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1070             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1071             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1072
1073             /**************************
1074              * CALCULATE INTERACTIONS *
1075              **************************/
1076
1077             /* REACTION-FIELD ELECTROSTATICS */
1078             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1079
1080             fscal            = felec;
1081
1082             /* Update vectorial force */
1083             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1084             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1085             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1086
1087             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1088             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1089             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1090
1091             /**************************
1092              * CALCULATE INTERACTIONS *
1093              **************************/
1094
1095             /* REACTION-FIELD ELECTROSTATICS */
1096             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1097
1098             fscal            = felec;
1099
1100             /* Update vectorial force */
1101             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1102             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1103             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1104
1105             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1106             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1107             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1108
1109             /**************************
1110              * CALCULATE INTERACTIONS *
1111              **************************/
1112
1113             /* REACTION-FIELD ELECTROSTATICS */
1114             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1115
1116             fscal            = felec;
1117
1118             /* Update vectorial force */
1119             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1120             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1121             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1122
1123             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1124             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1125             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1126
1127             /**************************
1128              * CALCULATE INTERACTIONS *
1129              **************************/
1130
1131             /* REACTION-FIELD ELECTROSTATICS */
1132             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1133
1134             fscal            = felec;
1135
1136             /* Update vectorial force */
1137             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1138             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1139             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1140
1141             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1142             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1143             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1144
1145             /**************************
1146              * CALCULATE INTERACTIONS *
1147              **************************/
1148
1149             /* REACTION-FIELD ELECTROSTATICS */
1150             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1151
1152             fscal            = felec;
1153
1154             /* Update vectorial force */
1155             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1156             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1157             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1158
1159             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1160             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1161             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1162
1163             /**************************
1164              * CALCULATE INTERACTIONS *
1165              **************************/
1166
1167             /* REACTION-FIELD ELECTROSTATICS */
1168             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1169
1170             fscal            = felec;
1171
1172             /* Update vectorial force */
1173             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1174             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1175             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1176
1177             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1178             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1179             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1180
1181             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
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,
1194                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1195
1196             /* Calculate displacement vector */
1197             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1198             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1199             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1200             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1201             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1202             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1203             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1204             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1205             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1206             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1207             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1208             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1209             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1210             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1211             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1212             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1213             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1214             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1215             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1216             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1217             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1218             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1219             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1220             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1221             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1222             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1223             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1224
1225             /* Calculate squared distance and things based on it */
1226             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1227             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1228             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1229             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1230             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1231             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1232             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1233             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1234             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1235
1236             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1237             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1238             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1239             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1240             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1241             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1242             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1243             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1244             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1245
1246             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1247             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1248             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1249             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1250             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1251             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1252             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1253             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1254             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1255
1256             fjx0             = _fjsp_setzero_v2r8();
1257             fjy0             = _fjsp_setzero_v2r8();
1258             fjz0             = _fjsp_setzero_v2r8();
1259             fjx1             = _fjsp_setzero_v2r8();
1260             fjy1             = _fjsp_setzero_v2r8();
1261             fjz1             = _fjsp_setzero_v2r8();
1262             fjx2             = _fjsp_setzero_v2r8();
1263             fjy2             = _fjsp_setzero_v2r8();
1264             fjz2             = _fjsp_setzero_v2r8();
1265
1266             /**************************
1267              * CALCULATE INTERACTIONS *
1268              **************************/
1269
1270             /* REACTION-FIELD ELECTROSTATICS */
1271             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1272
1273             fscal            = felec;
1274
1275             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1276
1277             /* Update vectorial force */
1278             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1279             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1280             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1281
1282             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1283             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1284             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1285
1286             /**************************
1287              * CALCULATE INTERACTIONS *
1288              **************************/
1289
1290             /* REACTION-FIELD ELECTROSTATICS */
1291             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1292
1293             fscal            = felec;
1294
1295             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1296
1297             /* Update vectorial force */
1298             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1299             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1300             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1301
1302             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1303             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1304             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* REACTION-FIELD ELECTROSTATICS */
1311             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1312
1313             fscal            = felec;
1314
1315             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1316
1317             /* Update vectorial force */
1318             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1319             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1320             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1321
1322             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1323             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1324             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1325
1326             /**************************
1327              * CALCULATE INTERACTIONS *
1328              **************************/
1329
1330             /* REACTION-FIELD ELECTROSTATICS */
1331             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1332
1333             fscal            = felec;
1334
1335             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1336
1337             /* Update vectorial force */
1338             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1339             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1340             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1341
1342             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1343             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1344             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1345
1346             /**************************
1347              * CALCULATE INTERACTIONS *
1348              **************************/
1349
1350             /* REACTION-FIELD ELECTROSTATICS */
1351             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1352
1353             fscal            = felec;
1354
1355             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1356
1357             /* Update vectorial force */
1358             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1359             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1360             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1361
1362             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1363             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1364             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* REACTION-FIELD ELECTROSTATICS */
1371             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1372
1373             fscal            = felec;
1374
1375             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1376
1377             /* Update vectorial force */
1378             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1379             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1380             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1381
1382             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1383             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1384             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             /* REACTION-FIELD ELECTROSTATICS */
1391             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1392
1393             fscal            = felec;
1394
1395             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1396
1397             /* Update vectorial force */
1398             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1399             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1400             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1401
1402             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1403             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1404             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1405
1406             /**************************
1407              * CALCULATE INTERACTIONS *
1408              **************************/
1409
1410             /* REACTION-FIELD ELECTROSTATICS */
1411             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1412
1413             fscal            = felec;
1414
1415             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1416
1417             /* Update vectorial force */
1418             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1419             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1420             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1421
1422             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1423             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1424             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1425
1426             /**************************
1427              * CALCULATE INTERACTIONS *
1428              **************************/
1429
1430             /* REACTION-FIELD ELECTROSTATICS */
1431             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1432
1433             fscal            = felec;
1434
1435             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1436
1437             /* Update vectorial force */
1438             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1439             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1440             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1441
1442             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1443             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1444             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1445
1446             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1447
1448             /* Inner loop uses 270 flops */
1449         }
1450
1451         /* End of innermost loop */
1452
1453         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1454                                               f+i_coord_offset,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_W3W3_F,outeriter*18 + inneriter*270);
1468 }