src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRFCut_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, 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 "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_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_ElecRFCut_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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 147     rcutoff_scalar   = fr->rcoulomb;
 148     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
 149     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
 150
 151     /* Avoid stupid compiler warnings */
 152     jnrA = jnrB = 0;
 153     j_coord_offsetA = 0;
 154     j_coord_offsetB = 0;
 155
 156     outeriter        = 0;
 157     inneriter        = 0;
 158
 159     /* Start outer loop over neighborlists */
 160     for(iidx=0; iidx<nri; iidx++)
 161     {
 162         /* Load shift vector for this list */
 163         i_shift_offset   = DIM*shiftidx[iidx];
 164
 165         /* Load limits for loop over neighbors */
 166         j_index_start    = jindex[iidx];
 167         j_index_end      = jindex[iidx+1];
 168
 169         /* Get outer coordinate index */
 170         inr              = iinr[iidx];
 171         i_coord_offset   = DIM*inr;
 172
 173         /* Load i particle coords and add shift vector */
 174         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 175                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 176
 177         fix0             = _fjsp_setzero_v2r8();
 178         fiy0             = _fjsp_setzero_v2r8();
 179         fiz0             = _fjsp_setzero_v2r8();
 180         fix1             = _fjsp_setzero_v2r8();
 181         fiy1             = _fjsp_setzero_v2r8();
 182         fiz1             = _fjsp_setzero_v2r8();
 183         fix2             = _fjsp_setzero_v2r8();
 184         fiy2             = _fjsp_setzero_v2r8();
 185         fiz2             = _fjsp_setzero_v2r8();
 186
 187         /* Reset potential sums */
 188         velecsum         = _fjsp_setzero_v2r8();
 189
 190         /* Start inner kernel loop */
 191         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 192         {
 193
 194             /* Get j neighbor index, and coordinate index */
 195             jnrA             = jjnr[jidx];
 196             jnrB             = jjnr[jidx+1];
 197             j_coord_offsetA  = DIM*jnrA;
 198             j_coord_offsetB  = DIM*jnrB;
 199
 200             /* load j atom coordinates */
 201             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 202                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 203
 204             /* Calculate displacement vector */
 205             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 206             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 207             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 208             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 209             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 210             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 211             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 212             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 213             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 214             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 215             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 216             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 217             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 218             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 219             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 220             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 221             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 222             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 223             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 224             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 225             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 226             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 227             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 228             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 229             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 230             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 231             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 232
 233             /* Calculate squared distance and things based on it */
 234             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 235             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 236             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 237             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 238             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 239             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 240             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 241             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 242             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 243
 244             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 245             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 246             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 247             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 248             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 249             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 250             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 251             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 252             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 253
 254             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 255             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 256             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 257             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 258             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 259             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 260             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 261             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 262             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 263
 264             fjx0             = _fjsp_setzero_v2r8();
 265             fjy0             = _fjsp_setzero_v2r8();
 266             fjz0             = _fjsp_setzero_v2r8();
 267             fjx1             = _fjsp_setzero_v2r8();
 268             fjy1             = _fjsp_setzero_v2r8();
 269             fjz1             = _fjsp_setzero_v2r8();
 270             fjx2             = _fjsp_setzero_v2r8();
 271             fjy2             = _fjsp_setzero_v2r8();
 272             fjz2             = _fjsp_setzero_v2r8();
 273
 274             /**************************
 275              * CALCULATE INTERACTIONS *
 276              **************************/
 277
 278             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 279             {
 280
 281             /* REACTION-FIELD ELECTROSTATICS */
 282             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 283             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 284
 285             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 286
 287             /* Update potential sum for this i atom from the interaction with this j atom. */
 288             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 289             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 290
 291             fscal            = felec;
 292
 293             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 294
 295             /* Update vectorial force */
 296             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 297             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 298             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 299
 300             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 301             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 302             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 303
 304             }
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 311             {
 312
 313             /* REACTION-FIELD ELECTROSTATICS */
 314             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 315             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 316
 317             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 321             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 322
 323             fscal            = felec;
 324
 325             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 326
 327             /* Update vectorial force */
 328             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 329             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 330             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 331
 332             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 333             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 334             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 335
 336             }
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 343             {
 344
 345             /* REACTION-FIELD ELECTROSTATICS */
 346             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 347             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 348
 349             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 350
 351             /* Update potential sum for this i atom from the interaction with this j atom. */
 352             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 353             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 354
 355             fscal            = felec;
 356
 357             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 358
 359             /* Update vectorial force */
 360             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 361             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 362             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 363
 364             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 365             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 366             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 367
 368             }
 369
 370             /**************************
 371              * CALCULATE INTERACTIONS *
 372              **************************/
 373
 374             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 375             {
 376
 377             /* REACTION-FIELD ELECTROSTATICS */
 378             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 379             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 380
 381             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 385             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 386
 387             fscal            = felec;
 388
 389             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 390
 391             /* Update vectorial force */
 392             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 393             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 394             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 395
 396             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 397             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 398             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 399
 400             }
 401
 402             /**************************
 403              * CALCULATE INTERACTIONS *
 404              **************************/
 405
 406             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 407             {
 408
 409             /* REACTION-FIELD ELECTROSTATICS */
 410             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 411             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 412
 413             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 414
 415             /* Update potential sum for this i atom from the interaction with this j atom. */
 416             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 417             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 418
 419             fscal            = felec;
 420
 421             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 422
 423             /* Update vectorial force */
 424             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 425             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 426             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 427
 428             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 429             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 430             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 431
 432             }
 433
 434             /**************************
 435              * CALCULATE INTERACTIONS *
 436              **************************/
 437
 438             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 439             {
 440
 441             /* REACTION-FIELD ELECTROSTATICS */
 442             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 443             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 444
 445             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 446
 447             /* Update potential sum for this i atom from the interaction with this j atom. */
 448             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 449             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 450
 451             fscal            = felec;
 452
 453             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 454
 455             /* Update vectorial force */
 456             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 457             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 458             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 459
 460             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 461             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 462             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 463
 464             }
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 471             {
 472
 473             /* REACTION-FIELD ELECTROSTATICS */
 474             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 475             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 476
 477             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 478
 479             /* Update potential sum for this i atom from the interaction with this j atom. */
 480             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 481             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 482
 483             fscal            = felec;
 484
 485             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 486
 487             /* Update vectorial force */
 488             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 489             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 490             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 491
 492             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 493             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 494             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 495
 496             }
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 503             {
 504
 505             /* REACTION-FIELD ELECTROSTATICS */
 506             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 507             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 508
 509             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 510
 511             /* Update potential sum for this i atom from the interaction with this j atom. */
 512             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 513             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 514
 515             fscal            = felec;
 516
 517             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 518
 519             /* Update vectorial force */
 520             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 521             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 522             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 523
 524             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 525             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 526             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 527
 528             }
 529
 530             /**************************
 531              * CALCULATE INTERACTIONS *
 532              **************************/
 533
 534             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 535             {
 536
 537             /* REACTION-FIELD ELECTROSTATICS */
 538             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 539             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 540
 541             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 542
 543             /* Update potential sum for this i atom from the interaction with this j atom. */
 544             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 545             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 546
 547             fscal            = felec;
 548
 549             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 550
 551             /* Update vectorial force */
 552             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 553             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 554             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 555
 556             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 557             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 558             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 559
 560             }
 561
 562             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 563
 564             /* Inner loop uses 351 flops */
 565         }
 566
 567         if(jidx<j_index_end)
 568         {
 569
 570             jnrA             = jjnr[jidx];
 571             j_coord_offsetA  = DIM*jnrA;
 572
 573             /* load j atom coordinates */
 574             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 575                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 576
 577             /* Calculate displacement vector */
 578             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 579             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 580             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 581             dx01             = _fjsp_sub_v2r8(ix0,jx1);
 582             dy01             = _fjsp_sub_v2r8(iy0,jy1);
 583             dz01             = _fjsp_sub_v2r8(iz0,jz1);
 584             dx02             = _fjsp_sub_v2r8(ix0,jx2);
 585             dy02             = _fjsp_sub_v2r8(iy0,jy2);
 586             dz02             = _fjsp_sub_v2r8(iz0,jz2);
 587             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 588             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 589             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 590             dx11             = _fjsp_sub_v2r8(ix1,jx1);
 591             dy11             = _fjsp_sub_v2r8(iy1,jy1);
 592             dz11             = _fjsp_sub_v2r8(iz1,jz1);
 593             dx12             = _fjsp_sub_v2r8(ix1,jx2);
 594             dy12             = _fjsp_sub_v2r8(iy1,jy2);
 595             dz12             = _fjsp_sub_v2r8(iz1,jz2);
 596             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 597             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 598             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 599             dx21             = _fjsp_sub_v2r8(ix2,jx1);
 600             dy21             = _fjsp_sub_v2r8(iy2,jy1);
 601             dz21             = _fjsp_sub_v2r8(iz2,jz1);
 602             dx22             = _fjsp_sub_v2r8(ix2,jx2);
 603             dy22             = _fjsp_sub_v2r8(iy2,jy2);
 604             dz22             = _fjsp_sub_v2r8(iz2,jz2);
 605
 606             /* Calculate squared distance and things based on it */
 607             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 608             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
 609             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
 610             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 611             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
 612             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
 613             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 614             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
 615             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
 616
 617             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 618             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
 619             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
 620             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 621             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
 622             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
 623             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 624             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
 625             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
 626
 627             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 628             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
 629             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
 630             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 631             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
 632             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
 633             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 634             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
 635             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
 636
 637             fjx0             = _fjsp_setzero_v2r8();
 638             fjy0             = _fjsp_setzero_v2r8();
 639             fjz0             = _fjsp_setzero_v2r8();
 640             fjx1             = _fjsp_setzero_v2r8();
 641             fjy1             = _fjsp_setzero_v2r8();
 642             fjz1             = _fjsp_setzero_v2r8();
 643             fjx2             = _fjsp_setzero_v2r8();
 644             fjy2             = _fjsp_setzero_v2r8();
 645             fjz2             = _fjsp_setzero_v2r8();
 646
 647             /**************************
 648              * CALCULATE INTERACTIONS *
 649              **************************/
 650
 651             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
 652             {
 653
 654             /* REACTION-FIELD ELECTROSTATICS */
 655             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 656             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 657
 658             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
 659
 660             /* Update potential sum for this i atom from the interaction with this j atom. */
 661             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 662             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 663             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 664
 665             fscal            = felec;
 666
 667             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 668
 669             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 670
 671             /* Update vectorial force */
 672             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 673             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 674             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 675
 676             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 677             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 678             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 679
 680             }
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
 687             {
 688
 689             /* REACTION-FIELD ELECTROSTATICS */
 690             velec            = _fjsp_mul_v2r8(qq01,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq01,rinv01),crf));
 691             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
 692
 693             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
 694
 695             /* Update potential sum for this i atom from the interaction with this j atom. */
 696             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 697             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 698             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 699
 700             fscal            = felec;
 701
 702             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 703
 704             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 705
 706             /* Update vectorial force */
 707             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
 708             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
 709             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
 710
 711             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
 712             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
 713             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
 714
 715             }
 716
 717             /**************************
 718              * CALCULATE INTERACTIONS *
 719              **************************/
 720
 721             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
 722             {
 723
 724             /* REACTION-FIELD ELECTROSTATICS */
 725             velec            = _fjsp_mul_v2r8(qq02,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq02,rinv02),crf));
 726             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
 727
 728             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
 729
 730             /* Update potential sum for this i atom from the interaction with this j atom. */
 731             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 732             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 733             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 734
 735             fscal            = felec;
 736
 737             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 738
 739             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 740
 741             /* Update vectorial force */
 742             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
 743             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
 744             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
 745
 746             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
 747             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
 748             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
 749
 750             }
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
 757             {
 758
 759             /* REACTION-FIELD ELECTROSTATICS */
 760             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 761             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 762
 763             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
 764
 765             /* Update potential sum for this i atom from the interaction with this j atom. */
 766             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 767             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 768             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 769
 770             fscal            = felec;
 771
 772             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 773
 774             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 775
 776             /* Update vectorial force */
 777             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 778             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 779             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 780
 781             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 782             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 783             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 784
 785             }
 786
 787             /**************************
 788              * CALCULATE INTERACTIONS *
 789              **************************/
 790
 791             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
 792             {
 793
 794             /* REACTION-FIELD ELECTROSTATICS */
 795             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
 796             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
 797
 798             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
 799
 800             /* Update potential sum for this i atom from the interaction with this j atom. */
 801             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 802             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 803             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 804
 805             fscal            = felec;
 806
 807             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 808
 809             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 810
 811             /* Update vectorial force */
 812             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
 813             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
 814             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
 815
 816             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
 817             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
 818             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
 819
 820             }
 821
 822             /**************************
 823              * CALCULATE INTERACTIONS *
 824              **************************/
 825
 826             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
 827             {
 828
 829             /* REACTION-FIELD ELECTROSTATICS */
 830             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
 831             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
 832
 833             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
 834
 835             /* Update potential sum for this i atom from the interaction with this j atom. */
 836             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 837             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 838             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 839
 840             fscal            = felec;
 841
 842             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 843
 844             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 845
 846             /* Update vectorial force */
 847             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
 848             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
 849             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
 850
 851             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
 852             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
 853             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
 854
 855             }
 856
 857             /**************************
 858              * CALCULATE INTERACTIONS *
 859              **************************/
 860
 861             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
 862             {
 863
 864             /* REACTION-FIELD ELECTROSTATICS */
 865             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 866             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 867
 868             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
 869
 870             /* Update potential sum for this i atom from the interaction with this j atom. */
 871             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 872             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 873             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 874
 875             fscal            = felec;
 876
 877             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 878
 879             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 880
 881             /* Update vectorial force */
 882             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 883             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 884             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 885
 886             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 887             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 888             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 889
 890             }
 891
 892             /**************************
 893              * CALCULATE INTERACTIONS *
 894              **************************/
 895
 896             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
 897             {
 898
 899             /* REACTION-FIELD ELECTROSTATICS */
 900             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
 901             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
 902
 903             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
 904
 905             /* Update potential sum for this i atom from the interaction with this j atom. */
 906             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 907             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 908             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 909
 910             fscal            = felec;
 911
 912             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 913
 914             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 915
 916             /* Update vectorial force */
 917             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
 918             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
 919             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
 920
 921             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
 922             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
 923             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
 924
 925             }
 926
 927             /**************************
 928              * CALCULATE INTERACTIONS *
 929              **************************/
 930
 931             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
 932             {
 933
 934             /* REACTION-FIELD ELECTROSTATICS */
 935             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
 936             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
 937
 938             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
 939
 940             /* Update potential sum for this i atom from the interaction with this j atom. */
 941             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
 942             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 943             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 944
 945             fscal            = felec;
 946
 947             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
 948
 949             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 950
 951             /* Update vectorial force */
 952             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
 953             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
 954             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
 955
 956             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
 957             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
 958             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
 959
 960             }
 961
 962             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 963
 964             /* Inner loop uses 351 flops */
 965         }
 966
 967         /* End of innermost loop */
 968
 969         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 970                                               f+i_coord_offset,fshift+i_shift_offset);
 971
 972         ggid                        = gid[iidx];
 973         /* Update potential energies */
 974         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 975
 976         /* Increment number of inner iterations */
 977         inneriter                  += j_index_end - j_index_start;
 978
 979         /* Outer loop uses 19 flops */
 980     }
 981
 982     /* Increment number of outer iterations */
 983     outeriter        += nri;
 984
 985     /* Update outer/inner flops */
 986
 987     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*351);
 988 }
 989 /*
 990  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_sparc64_hpc_ace_double
 991  * Electrostatics interaction: ReactionField
 992  * VdW interaction:            None
 993  * Geometry:                   Water3-Water3
 994  * Calculate force/pot:        Force
 995  */
 996 void
 997 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_sparc64_hpc_ace_double
 998                     (t_nblist                    * gmx_restrict       nlist,
 999                      rvec                        * gmx_restrict          xx,
1000                      rvec                        * gmx_restrict          ff,
1001                      t_forcerec                  * gmx_restrict          fr,
1002                      t_mdatoms                   * gmx_restrict     mdatoms,
1003                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1004                      t_nrnb                      * gmx_restrict        nrnb)
1005 {
1006     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1007      * just 0 for non-waters.
1008      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1009      * jnr indices corresponding to data put in the four positions in the SIMD register.
1010      */
1011     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1012     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1013     int              jnrA,jnrB;
1014     int              j_coord_offsetA,j_coord_offsetB;
1015     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1016     real             rcutoff_scalar;
1017     real             *shiftvec,*fshift,*x,*f;
1018     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1019     int              vdwioffset0;
1020     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1021     int              vdwioffset1;
1022     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1023     int              vdwioffset2;
1024     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1025     int              vdwjidx0A,vdwjidx0B;
1026     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1027     int              vdwjidx1A,vdwjidx1B;
1028     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1029     int              vdwjidx2A,vdwjidx2B;
1030     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1031     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1032     _fjsp_v2r8       dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1033     _fjsp_v2r8       dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1034     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1035     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1036     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1037     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1038     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1039     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1040     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1041     real             *charge;
1042     _fjsp_v2r8       itab_tmp;
1043     _fjsp_v2r8       dummy_mask,cutoff_mask;
1044     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1045     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1046     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1047
1048     x                = xx[0];
1049     f                = ff[0];
1050
1051     nri              = nlist->nri;
1052     iinr             = nlist->iinr;
1053     jindex           = nlist->jindex;
1054     jjnr             = nlist->jjnr;
1055     shiftidx         = nlist->shift;
1056     gid              = nlist->gid;
1057     shiftvec         = fr->shift_vec[0];
1058     fshift           = fr->fshift[0];
1059     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1060     charge           = mdatoms->chargeA;
1061     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
1062     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
1063     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
1064
1065     /* Setup water-specific parameters */
1066     inr              = nlist->iinr[0];
1067     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
1068     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1069     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1070
1071     jq0              = gmx_fjsp_set1_v2r8(charge[inr+0]);
1072     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1073     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1074     qq00             = _fjsp_mul_v2r8(iq0,jq0);
1075     qq01             = _fjsp_mul_v2r8(iq0,jq1);
1076     qq02             = _fjsp_mul_v2r8(iq0,jq2);
1077     qq10             = _fjsp_mul_v2r8(iq1,jq0);
1078     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1079     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1080     qq20             = _fjsp_mul_v2r8(iq2,jq0);
1081     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1082     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1083
1084     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1085     rcutoff_scalar   = fr->rcoulomb;
1086     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1087     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1088
1089     /* Avoid stupid compiler warnings */
1090     jnrA = jnrB = 0;
1091     j_coord_offsetA = 0;
1092     j_coord_offsetB = 0;
1093
1094     outeriter        = 0;
1095     inneriter        = 0;
1096
1097     /* Start outer loop over neighborlists */
1098     for(iidx=0; iidx<nri; iidx++)
1099     {
1100         /* Load shift vector for this list */
1101         i_shift_offset   = DIM*shiftidx[iidx];
1102
1103         /* Load limits for loop over neighbors */
1104         j_index_start    = jindex[iidx];
1105         j_index_end      = jindex[iidx+1];
1106
1107         /* Get outer coordinate index */
1108         inr              = iinr[iidx];
1109         i_coord_offset   = DIM*inr;
1110
1111         /* Load i particle coords and add shift vector */
1112         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1113                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1114
1115         fix0             = _fjsp_setzero_v2r8();
1116         fiy0             = _fjsp_setzero_v2r8();
1117         fiz0             = _fjsp_setzero_v2r8();
1118         fix1             = _fjsp_setzero_v2r8();
1119         fiy1             = _fjsp_setzero_v2r8();
1120         fiz1             = _fjsp_setzero_v2r8();
1121         fix2             = _fjsp_setzero_v2r8();
1122         fiy2             = _fjsp_setzero_v2r8();
1123         fiz2             = _fjsp_setzero_v2r8();
1124
1125         /* Start inner kernel loop */
1126         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1127         {
1128
1129             /* Get j neighbor index, and coordinate index */
1130             jnrA             = jjnr[jidx];
1131             jnrB             = jjnr[jidx+1];
1132             j_coord_offsetA  = DIM*jnrA;
1133             j_coord_offsetB  = DIM*jnrB;
1134
1135             /* load j atom coordinates */
1136             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1137                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1138
1139             /* Calculate displacement vector */
1140             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1141             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1142             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1143             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1144             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1145             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1146             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1147             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1148             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1149             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1150             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1151             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1152             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1153             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1154             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1155             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1156             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1157             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1158             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1159             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1160             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1161             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1162             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1163             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1164             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1165             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1166             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1167
1168             /* Calculate squared distance and things based on it */
1169             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1170             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1171             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1172             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1173             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1174             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1175             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1176             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1177             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1178
1179             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1180             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1181             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1182             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1183             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1184             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1185             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1186             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1187             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1188
1189             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1190             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1191             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1192             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1193             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1194             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1195             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1196             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1197             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1198
1199             fjx0             = _fjsp_setzero_v2r8();
1200             fjy0             = _fjsp_setzero_v2r8();
1201             fjz0             = _fjsp_setzero_v2r8();
1202             fjx1             = _fjsp_setzero_v2r8();
1203             fjy1             = _fjsp_setzero_v2r8();
1204             fjz1             = _fjsp_setzero_v2r8();
1205             fjx2             = _fjsp_setzero_v2r8();
1206             fjy2             = _fjsp_setzero_v2r8();
1207             fjz2             = _fjsp_setzero_v2r8();
1208
1209             /**************************
1210              * CALCULATE INTERACTIONS *
1211              **************************/
1212
1213             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1214             {
1215
1216             /* REACTION-FIELD ELECTROSTATICS */
1217             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1218
1219             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1220
1221             fscal            = felec;
1222
1223             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1224
1225             /* Update vectorial force */
1226             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1227             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1228             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1229
1230             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1231             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1232             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1233
1234             }
1235
1236             /**************************
1237              * CALCULATE INTERACTIONS *
1238              **************************/
1239
1240             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1241             {
1242
1243             /* REACTION-FIELD ELECTROSTATICS */
1244             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1245
1246             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1247
1248             fscal            = felec;
1249
1250             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1251
1252             /* Update vectorial force */
1253             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1254             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1255             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1256
1257             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1258             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1259             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1260
1261             }
1262
1263             /**************************
1264              * CALCULATE INTERACTIONS *
1265              **************************/
1266
1267             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1268             {
1269
1270             /* REACTION-FIELD ELECTROSTATICS */
1271             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1272
1273             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1274
1275             fscal            = felec;
1276
1277             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1278
1279             /* Update vectorial force */
1280             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1281             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1282             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1283
1284             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1285             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1286             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1287
1288             }
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1295             {
1296
1297             /* REACTION-FIELD ELECTROSTATICS */
1298             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1299
1300             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1301
1302             fscal            = felec;
1303
1304             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1305
1306             /* Update vectorial force */
1307             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1308             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1309             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1310
1311             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1312             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1313             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1314
1315             }
1316
1317             /**************************
1318              * CALCULATE INTERACTIONS *
1319              **************************/
1320
1321             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1322             {
1323
1324             /* REACTION-FIELD ELECTROSTATICS */
1325             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1326
1327             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1328
1329             fscal            = felec;
1330
1331             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1332
1333             /* Update vectorial force */
1334             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1335             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1336             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1337
1338             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1339             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1340             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1341
1342             }
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1349             {
1350
1351             /* REACTION-FIELD ELECTROSTATICS */
1352             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1353
1354             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1355
1356             fscal            = felec;
1357
1358             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1359
1360             /* Update vectorial force */
1361             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1362             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1363             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1364
1365             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1366             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1367             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1368
1369             }
1370
1371             /**************************
1372              * CALCULATE INTERACTIONS *
1373              **************************/
1374
1375             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1376             {
1377
1378             /* REACTION-FIELD ELECTROSTATICS */
1379             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1380
1381             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1382
1383             fscal            = felec;
1384
1385             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1386
1387             /* Update vectorial force */
1388             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1389             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1390             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1391
1392             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1393             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1394             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1395
1396             }
1397
1398             /**************************
1399              * CALCULATE INTERACTIONS *
1400              **************************/
1401
1402             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1403             {
1404
1405             /* REACTION-FIELD ELECTROSTATICS */
1406             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1407
1408             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1409
1410             fscal            = felec;
1411
1412             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1413
1414             /* Update vectorial force */
1415             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1416             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1417             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1418
1419             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1420             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1421             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1422
1423             }
1424
1425             /**************************
1426              * CALCULATE INTERACTIONS *
1427              **************************/
1428
1429             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1430             {
1431
1432             /* REACTION-FIELD ELECTROSTATICS */
1433             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1434
1435             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1436
1437             fscal            = felec;
1438
1439             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1440
1441             /* Update vectorial force */
1442             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1443             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1444             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1445
1446             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1447             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1448             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1449
1450             }
1451
1452             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1453
1454             /* Inner loop uses 297 flops */
1455         }
1456
1457         if(jidx<j_index_end)
1458         {
1459
1460             jnrA             = jjnr[jidx];
1461             j_coord_offsetA  = DIM*jnrA;
1462
1463             /* load j atom coordinates */
1464             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1465                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1466
1467             /* Calculate displacement vector */
1468             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1469             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1470             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1471             dx01             = _fjsp_sub_v2r8(ix0,jx1);
1472             dy01             = _fjsp_sub_v2r8(iy0,jy1);
1473             dz01             = _fjsp_sub_v2r8(iz0,jz1);
1474             dx02             = _fjsp_sub_v2r8(ix0,jx2);
1475             dy02             = _fjsp_sub_v2r8(iy0,jy2);
1476             dz02             = _fjsp_sub_v2r8(iz0,jz2);
1477             dx10             = _fjsp_sub_v2r8(ix1,jx0);
1478             dy10             = _fjsp_sub_v2r8(iy1,jy0);
1479             dz10             = _fjsp_sub_v2r8(iz1,jz0);
1480             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1481             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1482             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1483             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1484             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1485             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1486             dx20             = _fjsp_sub_v2r8(ix2,jx0);
1487             dy20             = _fjsp_sub_v2r8(iy2,jy0);
1488             dz20             = _fjsp_sub_v2r8(iz2,jz0);
1489             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1490             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1491             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1492             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1493             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1494             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1495
1496             /* Calculate squared distance and things based on it */
1497             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1498             rsq01            = gmx_fjsp_calc_rsq_v2r8(dx01,dy01,dz01);
1499             rsq02            = gmx_fjsp_calc_rsq_v2r8(dx02,dy02,dz02);
1500             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
1501             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1502             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1503             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
1504             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1505             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1506
1507             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1508             rinv01           = gmx_fjsp_invsqrt_v2r8(rsq01);
1509             rinv02           = gmx_fjsp_invsqrt_v2r8(rsq02);
1510             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
1511             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1512             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1513             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
1514             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1515             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1516
1517             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1518             rinvsq01         = _fjsp_mul_v2r8(rinv01,rinv01);
1519             rinvsq02         = _fjsp_mul_v2r8(rinv02,rinv02);
1520             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
1521             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1522             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1523             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
1524             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1525             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1526
1527             fjx0             = _fjsp_setzero_v2r8();
1528             fjy0             = _fjsp_setzero_v2r8();
1529             fjz0             = _fjsp_setzero_v2r8();
1530             fjx1             = _fjsp_setzero_v2r8();
1531             fjy1             = _fjsp_setzero_v2r8();
1532             fjz1             = _fjsp_setzero_v2r8();
1533             fjx2             = _fjsp_setzero_v2r8();
1534             fjy2             = _fjsp_setzero_v2r8();
1535             fjz2             = _fjsp_setzero_v2r8();
1536
1537             /**************************
1538              * CALCULATE INTERACTIONS *
1539              **************************/
1540
1541             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1542             {
1543
1544             /* REACTION-FIELD ELECTROSTATICS */
1545             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
1546
1547             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1548
1549             fscal            = felec;
1550
1551             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1552
1553             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1554
1555             /* Update vectorial force */
1556             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1557             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1558             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1559
1560             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1561             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1562             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1563
1564             }
1565
1566             /**************************
1567              * CALCULATE INTERACTIONS *
1568              **************************/
1569
1570             if (gmx_fjsp_any_lt_v2r8(rsq01,rcutoff2))
1571             {
1572
1573             /* REACTION-FIELD ELECTROSTATICS */
1574             felec            = _fjsp_mul_v2r8(qq01,_fjsp_msub_v2r8(rinv01,rinvsq01,krf2));
1575
1576             cutoff_mask      = _fjsp_cmplt_v2r8(rsq01,rcutoff2);
1577
1578             fscal            = felec;
1579
1580             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1581
1582             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1583
1584             /* Update vectorial force */
1585             fix0             = _fjsp_madd_v2r8(dx01,fscal,fix0);
1586             fiy0             = _fjsp_madd_v2r8(dy01,fscal,fiy0);
1587             fiz0             = _fjsp_madd_v2r8(dz01,fscal,fiz0);
1588
1589             fjx1             = _fjsp_madd_v2r8(dx01,fscal,fjx1);
1590             fjy1             = _fjsp_madd_v2r8(dy01,fscal,fjy1);
1591             fjz1             = _fjsp_madd_v2r8(dz01,fscal,fjz1);
1592
1593             }
1594
1595             /**************************
1596              * CALCULATE INTERACTIONS *
1597              **************************/
1598
1599             if (gmx_fjsp_any_lt_v2r8(rsq02,rcutoff2))
1600             {
1601
1602             /* REACTION-FIELD ELECTROSTATICS */
1603             felec            = _fjsp_mul_v2r8(qq02,_fjsp_msub_v2r8(rinv02,rinvsq02,krf2));
1604
1605             cutoff_mask      = _fjsp_cmplt_v2r8(rsq02,rcutoff2);
1606
1607             fscal            = felec;
1608
1609             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1610
1611             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1612
1613             /* Update vectorial force */
1614             fix0             = _fjsp_madd_v2r8(dx02,fscal,fix0);
1615             fiy0             = _fjsp_madd_v2r8(dy02,fscal,fiy0);
1616             fiz0             = _fjsp_madd_v2r8(dz02,fscal,fiz0);
1617
1618             fjx2             = _fjsp_madd_v2r8(dx02,fscal,fjx2);
1619             fjy2             = _fjsp_madd_v2r8(dy02,fscal,fjy2);
1620             fjz2             = _fjsp_madd_v2r8(dz02,fscal,fjz2);
1621
1622             }
1623
1624             /**************************
1625              * CALCULATE INTERACTIONS *
1626              **************************/
1627
1628             if (gmx_fjsp_any_lt_v2r8(rsq10,rcutoff2))
1629             {
1630
1631             /* REACTION-FIELD ELECTROSTATICS */
1632             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
1633
1634             cutoff_mask      = _fjsp_cmplt_v2r8(rsq10,rcutoff2);
1635
1636             fscal            = felec;
1637
1638             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1639
1640             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1641
1642             /* Update vectorial force */
1643             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
1644             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
1645             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
1646
1647             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
1648             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
1649             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
1650
1651             }
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1658             {
1659
1660             /* REACTION-FIELD ELECTROSTATICS */
1661             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1662
1663             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1664
1665             fscal            = felec;
1666
1667             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1668
1669             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1670
1671             /* Update vectorial force */
1672             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1673             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1674             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1675
1676             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1677             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1678             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1679
1680             }
1681
1682             /**************************
1683              * CALCULATE INTERACTIONS *
1684              **************************/
1685
1686             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1687             {
1688
1689             /* REACTION-FIELD ELECTROSTATICS */
1690             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1691
1692             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1693
1694             fscal            = felec;
1695
1696             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1697
1698             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1699
1700             /* Update vectorial force */
1701             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1702             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1703             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1704
1705             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1706             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1707             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1708
1709             }
1710
1711             /**************************
1712              * CALCULATE INTERACTIONS *
1713              **************************/
1714
1715             if (gmx_fjsp_any_lt_v2r8(rsq20,rcutoff2))
1716             {
1717
1718             /* REACTION-FIELD ELECTROSTATICS */
1719             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
1720
1721             cutoff_mask      = _fjsp_cmplt_v2r8(rsq20,rcutoff2);
1722
1723             fscal            = felec;
1724
1725             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1726
1727             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1728
1729             /* Update vectorial force */
1730             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
1731             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
1732             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
1733
1734             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
1735             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
1736             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
1737
1738             }
1739
1740             /**************************
1741              * CALCULATE INTERACTIONS *
1742              **************************/
1743
1744             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1745             {
1746
1747             /* REACTION-FIELD ELECTROSTATICS */
1748             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1749
1750             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1751
1752             fscal            = felec;
1753
1754             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1755
1756             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1757
1758             /* Update vectorial force */
1759             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1760             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1761             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1762
1763             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1764             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1765             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1766
1767             }
1768
1769             /**************************
1770              * CALCULATE INTERACTIONS *
1771              **************************/
1772
1773             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1774             {
1775
1776             /* REACTION-FIELD ELECTROSTATICS */
1777             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1778
1779             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1780
1781             fscal            = felec;
1782
1783             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1784
1785             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1786
1787             /* Update vectorial force */
1788             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1789             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1790             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1791
1792             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1793             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1794             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1795
1796             }
1797
1798             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1799
1800             /* Inner loop uses 297 flops */
1801         }
1802
1803         /* End of innermost loop */
1804
1805         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1806                                               f+i_coord_offset,fshift+i_shift_offset);
1807
1808         /* Increment number of inner iterations */
1809         inneriter                  += j_index_end - j_index_start;
1810
1811         /* Outer loop uses 18 flops */
1812     }
1813
1814     /* Increment number of outer iterations */
1815     outeriter        += nri;
1816
1817     /* Update outer/inner flops */
1818
1819     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*297);
1820 }