src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecCoul_VdwNone_GeomW4W4_sse4_1_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 sse4_1_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 "gmx_math_x86_sse4_1_double.h"
  50 #include "kernelutil_x86_sse4_1_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse4_1_double
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            None
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse4_1_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     int              vdwjidx3A,vdwjidx3B;
  93     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  94     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  95     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  96     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  97     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  98     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  99     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 100     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 101     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 102     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 103     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     __m128d          dummy_mask,cutoff_mask;
 106     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 107     __m128d          one     = _mm_set1_pd(1.0);
 108     __m128d          two     = _mm_set1_pd(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_pd(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122
 123     /* Setup water-specific parameters */
 124     inr              = nlist->iinr[0];
 125     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 126     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 127     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 128
 129     jq1              = _mm_set1_pd(charge[inr+1]);
 130     jq2              = _mm_set1_pd(charge[inr+2]);
 131     jq3              = _mm_set1_pd(charge[inr+3]);
 132     qq11             = _mm_mul_pd(iq1,jq1);
 133     qq12             = _mm_mul_pd(iq1,jq2);
 134     qq13             = _mm_mul_pd(iq1,jq3);
 135     qq21             = _mm_mul_pd(iq2,jq1);
 136     qq22             = _mm_mul_pd(iq2,jq2);
 137     qq23             = _mm_mul_pd(iq2,jq3);
 138     qq31             = _mm_mul_pd(iq3,jq1);
 139     qq32             = _mm_mul_pd(iq3,jq2);
 140     qq33             = _mm_mul_pd(iq3,jq3);
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146
 147     outeriter        = 0;
 148     inneriter        = 0;
 149
 150     /* Start outer loop over neighborlists */
 151     for(iidx=0; iidx<nri; iidx++)
 152     {
 153         /* Load shift vector for this list */
 154         i_shift_offset   = DIM*shiftidx[iidx];
 155
 156         /* Load limits for loop over neighbors */
 157         j_index_start    = jindex[iidx];
 158         j_index_end      = jindex[iidx+1];
 159
 160         /* Get outer coordinate index */
 161         inr              = iinr[iidx];
 162         i_coord_offset   = DIM*inr;
 163
 164         /* Load i particle coords and add shift vector */
 165         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 166                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 167
 168         fix1             = _mm_setzero_pd();
 169         fiy1             = _mm_setzero_pd();
 170         fiz1             = _mm_setzero_pd();
 171         fix2             = _mm_setzero_pd();
 172         fiy2             = _mm_setzero_pd();
 173         fiz2             = _mm_setzero_pd();
 174         fix3             = _mm_setzero_pd();
 175         fiy3             = _mm_setzero_pd();
 176         fiz3             = _mm_setzero_pd();
 177
 178         /* Reset potential sums */
 179         velecsum         = _mm_setzero_pd();
 180
 181         /* Start inner kernel loop */
 182         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 183         {
 184
 185             /* Get j neighbor index, and coordinate index */
 186             jnrA             = jjnr[jidx];
 187             jnrB             = jjnr[jidx+1];
 188             j_coord_offsetA  = DIM*jnrA;
 189             j_coord_offsetB  = DIM*jnrB;
 190
 191             /* load j atom coordinates */
 192             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 193                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 194
 195             /* Calculate displacement vector */
 196             dx11             = _mm_sub_pd(ix1,jx1);
 197             dy11             = _mm_sub_pd(iy1,jy1);
 198             dz11             = _mm_sub_pd(iz1,jz1);
 199             dx12             = _mm_sub_pd(ix1,jx2);
 200             dy12             = _mm_sub_pd(iy1,jy2);
 201             dz12             = _mm_sub_pd(iz1,jz2);
 202             dx13             = _mm_sub_pd(ix1,jx3);
 203             dy13             = _mm_sub_pd(iy1,jy3);
 204             dz13             = _mm_sub_pd(iz1,jz3);
 205             dx21             = _mm_sub_pd(ix2,jx1);
 206             dy21             = _mm_sub_pd(iy2,jy1);
 207             dz21             = _mm_sub_pd(iz2,jz1);
 208             dx22             = _mm_sub_pd(ix2,jx2);
 209             dy22             = _mm_sub_pd(iy2,jy2);
 210             dz22             = _mm_sub_pd(iz2,jz2);
 211             dx23             = _mm_sub_pd(ix2,jx3);
 212             dy23             = _mm_sub_pd(iy2,jy3);
 213             dz23             = _mm_sub_pd(iz2,jz3);
 214             dx31             = _mm_sub_pd(ix3,jx1);
 215             dy31             = _mm_sub_pd(iy3,jy1);
 216             dz31             = _mm_sub_pd(iz3,jz1);
 217             dx32             = _mm_sub_pd(ix3,jx2);
 218             dy32             = _mm_sub_pd(iy3,jy2);
 219             dz32             = _mm_sub_pd(iz3,jz2);
 220             dx33             = _mm_sub_pd(ix3,jx3);
 221             dy33             = _mm_sub_pd(iy3,jy3);
 222             dz33             = _mm_sub_pd(iz3,jz3);
 223
 224             /* Calculate squared distance and things based on it */
 225             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 226             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 227             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 228             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 229             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 230             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 231             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 232             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 233             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 234
 235             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 236             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 237             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 238             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 239             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 240             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 241             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 242             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 243             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 244
 245             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 246             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 247             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 248             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 249             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 250             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 251             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 252             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 253             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 254
 255             fjx1             = _mm_setzero_pd();
 256             fjy1             = _mm_setzero_pd();
 257             fjz1             = _mm_setzero_pd();
 258             fjx2             = _mm_setzero_pd();
 259             fjy2             = _mm_setzero_pd();
 260             fjz2             = _mm_setzero_pd();
 261             fjx3             = _mm_setzero_pd();
 262             fjy3             = _mm_setzero_pd();
 263             fjz3             = _mm_setzero_pd();
 264
 265             /**************************
 266              * CALCULATE INTERACTIONS *
 267              **************************/
 268
 269             /* COULOMB ELECTROSTATICS */
 270             velec            = _mm_mul_pd(qq11,rinv11);
 271             felec            = _mm_mul_pd(velec,rinvsq11);
 272
 273             /* Update potential sum for this i atom from the interaction with this j atom. */
 274             velecsum         = _mm_add_pd(velecsum,velec);
 275
 276             fscal            = felec;
 277
 278             /* Calculate temporary vectorial force */
 279             tx               = _mm_mul_pd(fscal,dx11);
 280             ty               = _mm_mul_pd(fscal,dy11);
 281             tz               = _mm_mul_pd(fscal,dz11);
 282
 283             /* Update vectorial force */
 284             fix1             = _mm_add_pd(fix1,tx);
 285             fiy1             = _mm_add_pd(fiy1,ty);
 286             fiz1             = _mm_add_pd(fiz1,tz);
 287
 288             fjx1             = _mm_add_pd(fjx1,tx);
 289             fjy1             = _mm_add_pd(fjy1,ty);
 290             fjz1             = _mm_add_pd(fjz1,tz);
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             /* COULOMB ELECTROSTATICS */
 297             velec            = _mm_mul_pd(qq12,rinv12);
 298             felec            = _mm_mul_pd(velec,rinvsq12);
 299
 300             /* Update potential sum for this i atom from the interaction with this j atom. */
 301             velecsum         = _mm_add_pd(velecsum,velec);
 302
 303             fscal            = felec;
 304
 305             /* Calculate temporary vectorial force */
 306             tx               = _mm_mul_pd(fscal,dx12);
 307             ty               = _mm_mul_pd(fscal,dy12);
 308             tz               = _mm_mul_pd(fscal,dz12);
 309
 310             /* Update vectorial force */
 311             fix1             = _mm_add_pd(fix1,tx);
 312             fiy1             = _mm_add_pd(fiy1,ty);
 313             fiz1             = _mm_add_pd(fiz1,tz);
 314
 315             fjx2             = _mm_add_pd(fjx2,tx);
 316             fjy2             = _mm_add_pd(fjy2,ty);
 317             fjz2             = _mm_add_pd(fjz2,tz);
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             /* COULOMB ELECTROSTATICS */
 324             velec            = _mm_mul_pd(qq13,rinv13);
 325             felec            = _mm_mul_pd(velec,rinvsq13);
 326
 327             /* Update potential sum for this i atom from the interaction with this j atom. */
 328             velecsum         = _mm_add_pd(velecsum,velec);
 329
 330             fscal            = felec;
 331
 332             /* Calculate temporary vectorial force */
 333             tx               = _mm_mul_pd(fscal,dx13);
 334             ty               = _mm_mul_pd(fscal,dy13);
 335             tz               = _mm_mul_pd(fscal,dz13);
 336
 337             /* Update vectorial force */
 338             fix1             = _mm_add_pd(fix1,tx);
 339             fiy1             = _mm_add_pd(fiy1,ty);
 340             fiz1             = _mm_add_pd(fiz1,tz);
 341
 342             fjx3             = _mm_add_pd(fjx3,tx);
 343             fjy3             = _mm_add_pd(fjy3,ty);
 344             fjz3             = _mm_add_pd(fjz3,tz);
 345
 346             /**************************
 347              * CALCULATE INTERACTIONS *
 348              **************************/
 349
 350             /* COULOMB ELECTROSTATICS */
 351             velec            = _mm_mul_pd(qq21,rinv21);
 352             felec            = _mm_mul_pd(velec,rinvsq21);
 353
 354             /* Update potential sum for this i atom from the interaction with this j atom. */
 355             velecsum         = _mm_add_pd(velecsum,velec);
 356
 357             fscal            = felec;
 358
 359             /* Calculate temporary vectorial force */
 360             tx               = _mm_mul_pd(fscal,dx21);
 361             ty               = _mm_mul_pd(fscal,dy21);
 362             tz               = _mm_mul_pd(fscal,dz21);
 363
 364             /* Update vectorial force */
 365             fix2             = _mm_add_pd(fix2,tx);
 366             fiy2             = _mm_add_pd(fiy2,ty);
 367             fiz2             = _mm_add_pd(fiz2,tz);
 368
 369             fjx1             = _mm_add_pd(fjx1,tx);
 370             fjy1             = _mm_add_pd(fjy1,ty);
 371             fjz1             = _mm_add_pd(fjz1,tz);
 372
 373             /**************************
 374              * CALCULATE INTERACTIONS *
 375              **************************/
 376
 377             /* COULOMB ELECTROSTATICS */
 378             velec            = _mm_mul_pd(qq22,rinv22);
 379             felec            = _mm_mul_pd(velec,rinvsq22);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velecsum         = _mm_add_pd(velecsum,velec);
 383
 384             fscal            = felec;
 385
 386             /* Calculate temporary vectorial force */
 387             tx               = _mm_mul_pd(fscal,dx22);
 388             ty               = _mm_mul_pd(fscal,dy22);
 389             tz               = _mm_mul_pd(fscal,dz22);
 390
 391             /* Update vectorial force */
 392             fix2             = _mm_add_pd(fix2,tx);
 393             fiy2             = _mm_add_pd(fiy2,ty);
 394             fiz2             = _mm_add_pd(fiz2,tz);
 395
 396             fjx2             = _mm_add_pd(fjx2,tx);
 397             fjy2             = _mm_add_pd(fjy2,ty);
 398             fjz2             = _mm_add_pd(fjz2,tz);
 399
 400             /**************************
 401              * CALCULATE INTERACTIONS *
 402              **************************/
 403
 404             /* COULOMB ELECTROSTATICS */
 405             velec            = _mm_mul_pd(qq23,rinv23);
 406             felec            = _mm_mul_pd(velec,rinvsq23);
 407
 408             /* Update potential sum for this i atom from the interaction with this j atom. */
 409             velecsum         = _mm_add_pd(velecsum,velec);
 410
 411             fscal            = felec;
 412
 413             /* Calculate temporary vectorial force */
 414             tx               = _mm_mul_pd(fscal,dx23);
 415             ty               = _mm_mul_pd(fscal,dy23);
 416             tz               = _mm_mul_pd(fscal,dz23);
 417
 418             /* Update vectorial force */
 419             fix2             = _mm_add_pd(fix2,tx);
 420             fiy2             = _mm_add_pd(fiy2,ty);
 421             fiz2             = _mm_add_pd(fiz2,tz);
 422
 423             fjx3             = _mm_add_pd(fjx3,tx);
 424             fjy3             = _mm_add_pd(fjy3,ty);
 425             fjz3             = _mm_add_pd(fjz3,tz);
 426
 427             /**************************
 428              * CALCULATE INTERACTIONS *
 429              **************************/
 430
 431             /* COULOMB ELECTROSTATICS */
 432             velec            = _mm_mul_pd(qq31,rinv31);
 433             felec            = _mm_mul_pd(velec,rinvsq31);
 434
 435             /* Update potential sum for this i atom from the interaction with this j atom. */
 436             velecsum         = _mm_add_pd(velecsum,velec);
 437
 438             fscal            = felec;
 439
 440             /* Calculate temporary vectorial force */
 441             tx               = _mm_mul_pd(fscal,dx31);
 442             ty               = _mm_mul_pd(fscal,dy31);
 443             tz               = _mm_mul_pd(fscal,dz31);
 444
 445             /* Update vectorial force */
 446             fix3             = _mm_add_pd(fix3,tx);
 447             fiy3             = _mm_add_pd(fiy3,ty);
 448             fiz3             = _mm_add_pd(fiz3,tz);
 449
 450             fjx1             = _mm_add_pd(fjx1,tx);
 451             fjy1             = _mm_add_pd(fjy1,ty);
 452             fjz1             = _mm_add_pd(fjz1,tz);
 453
 454             /**************************
 455              * CALCULATE INTERACTIONS *
 456              **************************/
 457
 458             /* COULOMB ELECTROSTATICS */
 459             velec            = _mm_mul_pd(qq32,rinv32);
 460             felec            = _mm_mul_pd(velec,rinvsq32);
 461
 462             /* Update potential sum for this i atom from the interaction with this j atom. */
 463             velecsum         = _mm_add_pd(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467             /* Calculate temporary vectorial force */
 468             tx               = _mm_mul_pd(fscal,dx32);
 469             ty               = _mm_mul_pd(fscal,dy32);
 470             tz               = _mm_mul_pd(fscal,dz32);
 471
 472             /* Update vectorial force */
 473             fix3             = _mm_add_pd(fix3,tx);
 474             fiy3             = _mm_add_pd(fiy3,ty);
 475             fiz3             = _mm_add_pd(fiz3,tz);
 476
 477             fjx2             = _mm_add_pd(fjx2,tx);
 478             fjy2             = _mm_add_pd(fjy2,ty);
 479             fjz2             = _mm_add_pd(fjz2,tz);
 480
 481             /**************************
 482              * CALCULATE INTERACTIONS *
 483              **************************/
 484
 485             /* COULOMB ELECTROSTATICS */
 486             velec            = _mm_mul_pd(qq33,rinv33);
 487             felec            = _mm_mul_pd(velec,rinvsq33);
 488
 489             /* Update potential sum for this i atom from the interaction with this j atom. */
 490             velecsum         = _mm_add_pd(velecsum,velec);
 491
 492             fscal            = felec;
 493
 494             /* Calculate temporary vectorial force */
 495             tx               = _mm_mul_pd(fscal,dx33);
 496             ty               = _mm_mul_pd(fscal,dy33);
 497             tz               = _mm_mul_pd(fscal,dz33);
 498
 499             /* Update vectorial force */
 500             fix3             = _mm_add_pd(fix3,tx);
 501             fiy3             = _mm_add_pd(fiy3,ty);
 502             fiz3             = _mm_add_pd(fiz3,tz);
 503
 504             fjx3             = _mm_add_pd(fjx3,tx);
 505             fjy3             = _mm_add_pd(fjy3,ty);
 506             fjz3             = _mm_add_pd(fjz3,tz);
 507
 508             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 509
 510             /* Inner loop uses 252 flops */
 511         }
 512
 513         if(jidx<j_index_end)
 514         {
 515
 516             jnrA             = jjnr[jidx];
 517             j_coord_offsetA  = DIM*jnrA;
 518
 519             /* load j atom coordinates */
 520             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 521                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 522
 523             /* Calculate displacement vector */
 524             dx11             = _mm_sub_pd(ix1,jx1);
 525             dy11             = _mm_sub_pd(iy1,jy1);
 526             dz11             = _mm_sub_pd(iz1,jz1);
 527             dx12             = _mm_sub_pd(ix1,jx2);
 528             dy12             = _mm_sub_pd(iy1,jy2);
 529             dz12             = _mm_sub_pd(iz1,jz2);
 530             dx13             = _mm_sub_pd(ix1,jx3);
 531             dy13             = _mm_sub_pd(iy1,jy3);
 532             dz13             = _mm_sub_pd(iz1,jz3);
 533             dx21             = _mm_sub_pd(ix2,jx1);
 534             dy21             = _mm_sub_pd(iy2,jy1);
 535             dz21             = _mm_sub_pd(iz2,jz1);
 536             dx22             = _mm_sub_pd(ix2,jx2);
 537             dy22             = _mm_sub_pd(iy2,jy2);
 538             dz22             = _mm_sub_pd(iz2,jz2);
 539             dx23             = _mm_sub_pd(ix2,jx3);
 540             dy23             = _mm_sub_pd(iy2,jy3);
 541             dz23             = _mm_sub_pd(iz2,jz3);
 542             dx31             = _mm_sub_pd(ix3,jx1);
 543             dy31             = _mm_sub_pd(iy3,jy1);
 544             dz31             = _mm_sub_pd(iz3,jz1);
 545             dx32             = _mm_sub_pd(ix3,jx2);
 546             dy32             = _mm_sub_pd(iy3,jy2);
 547             dz32             = _mm_sub_pd(iz3,jz2);
 548             dx33             = _mm_sub_pd(ix3,jx3);
 549             dy33             = _mm_sub_pd(iy3,jy3);
 550             dz33             = _mm_sub_pd(iz3,jz3);
 551
 552             /* Calculate squared distance and things based on it */
 553             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 554             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 555             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 556             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 557             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 558             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 559             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 560             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 561             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 562
 563             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 564             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 565             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 566             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 567             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 568             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 569             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 570             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 571             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 572
 573             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 574             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 575             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 576             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 577             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 578             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 579             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 580             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 581             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 582
 583             fjx1             = _mm_setzero_pd();
 584             fjy1             = _mm_setzero_pd();
 585             fjz1             = _mm_setzero_pd();
 586             fjx2             = _mm_setzero_pd();
 587             fjy2             = _mm_setzero_pd();
 588             fjz2             = _mm_setzero_pd();
 589             fjx3             = _mm_setzero_pd();
 590             fjy3             = _mm_setzero_pd();
 591             fjz3             = _mm_setzero_pd();
 592
 593             /**************************
 594              * CALCULATE INTERACTIONS *
 595              **************************/
 596
 597             /* COULOMB ELECTROSTATICS */
 598             velec            = _mm_mul_pd(qq11,rinv11);
 599             felec            = _mm_mul_pd(velec,rinvsq11);
 600
 601             /* Update potential sum for this i atom from the interaction with this j atom. */
 602             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 603             velecsum         = _mm_add_pd(velecsum,velec);
 604
 605             fscal            = felec;
 606
 607             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 608
 609             /* Calculate temporary vectorial force */
 610             tx               = _mm_mul_pd(fscal,dx11);
 611             ty               = _mm_mul_pd(fscal,dy11);
 612             tz               = _mm_mul_pd(fscal,dz11);
 613
 614             /* Update vectorial force */
 615             fix1             = _mm_add_pd(fix1,tx);
 616             fiy1             = _mm_add_pd(fiy1,ty);
 617             fiz1             = _mm_add_pd(fiz1,tz);
 618
 619             fjx1             = _mm_add_pd(fjx1,tx);
 620             fjy1             = _mm_add_pd(fjy1,ty);
 621             fjz1             = _mm_add_pd(fjz1,tz);
 622
 623             /**************************
 624              * CALCULATE INTERACTIONS *
 625              **************************/
 626
 627             /* COULOMB ELECTROSTATICS */
 628             velec            = _mm_mul_pd(qq12,rinv12);
 629             felec            = _mm_mul_pd(velec,rinvsq12);
 630
 631             /* Update potential sum for this i atom from the interaction with this j atom. */
 632             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 633             velecsum         = _mm_add_pd(velecsum,velec);
 634
 635             fscal            = felec;
 636
 637             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 638
 639             /* Calculate temporary vectorial force */
 640             tx               = _mm_mul_pd(fscal,dx12);
 641             ty               = _mm_mul_pd(fscal,dy12);
 642             tz               = _mm_mul_pd(fscal,dz12);
 643
 644             /* Update vectorial force */
 645             fix1             = _mm_add_pd(fix1,tx);
 646             fiy1             = _mm_add_pd(fiy1,ty);
 647             fiz1             = _mm_add_pd(fiz1,tz);
 648
 649             fjx2             = _mm_add_pd(fjx2,tx);
 650             fjy2             = _mm_add_pd(fjy2,ty);
 651             fjz2             = _mm_add_pd(fjz2,tz);
 652
 653             /**************************
 654              * CALCULATE INTERACTIONS *
 655              **************************/
 656
 657             /* COULOMB ELECTROSTATICS */
 658             velec            = _mm_mul_pd(qq13,rinv13);
 659             felec            = _mm_mul_pd(velec,rinvsq13);
 660
 661             /* Update potential sum for this i atom from the interaction with this j atom. */
 662             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 663             velecsum         = _mm_add_pd(velecsum,velec);
 664
 665             fscal            = felec;
 666
 667             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 668
 669             /* Calculate temporary vectorial force */
 670             tx               = _mm_mul_pd(fscal,dx13);
 671             ty               = _mm_mul_pd(fscal,dy13);
 672             tz               = _mm_mul_pd(fscal,dz13);
 673
 674             /* Update vectorial force */
 675             fix1             = _mm_add_pd(fix1,tx);
 676             fiy1             = _mm_add_pd(fiy1,ty);
 677             fiz1             = _mm_add_pd(fiz1,tz);
 678
 679             fjx3             = _mm_add_pd(fjx3,tx);
 680             fjy3             = _mm_add_pd(fjy3,ty);
 681             fjz3             = _mm_add_pd(fjz3,tz);
 682
 683             /**************************
 684              * CALCULATE INTERACTIONS *
 685              **************************/
 686
 687             /* COULOMB ELECTROSTATICS */
 688             velec            = _mm_mul_pd(qq21,rinv21);
 689             felec            = _mm_mul_pd(velec,rinvsq21);
 690
 691             /* Update potential sum for this i atom from the interaction with this j atom. */
 692             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 693             velecsum         = _mm_add_pd(velecsum,velec);
 694
 695             fscal            = felec;
 696
 697             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 698
 699             /* Calculate temporary vectorial force */
 700             tx               = _mm_mul_pd(fscal,dx21);
 701             ty               = _mm_mul_pd(fscal,dy21);
 702             tz               = _mm_mul_pd(fscal,dz21);
 703
 704             /* Update vectorial force */
 705             fix2             = _mm_add_pd(fix2,tx);
 706             fiy2             = _mm_add_pd(fiy2,ty);
 707             fiz2             = _mm_add_pd(fiz2,tz);
 708
 709             fjx1             = _mm_add_pd(fjx1,tx);
 710             fjy1             = _mm_add_pd(fjy1,ty);
 711             fjz1             = _mm_add_pd(fjz1,tz);
 712
 713             /**************************
 714              * CALCULATE INTERACTIONS *
 715              **************************/
 716
 717             /* COULOMB ELECTROSTATICS */
 718             velec            = _mm_mul_pd(qq22,rinv22);
 719             felec            = _mm_mul_pd(velec,rinvsq22);
 720
 721             /* Update potential sum for this i atom from the interaction with this j atom. */
 722             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 723             velecsum         = _mm_add_pd(velecsum,velec);
 724
 725             fscal            = felec;
 726
 727             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 728
 729             /* Calculate temporary vectorial force */
 730             tx               = _mm_mul_pd(fscal,dx22);
 731             ty               = _mm_mul_pd(fscal,dy22);
 732             tz               = _mm_mul_pd(fscal,dz22);
 733
 734             /* Update vectorial force */
 735             fix2             = _mm_add_pd(fix2,tx);
 736             fiy2             = _mm_add_pd(fiy2,ty);
 737             fiz2             = _mm_add_pd(fiz2,tz);
 738
 739             fjx2             = _mm_add_pd(fjx2,tx);
 740             fjy2             = _mm_add_pd(fjy2,ty);
 741             fjz2             = _mm_add_pd(fjz2,tz);
 742
 743             /**************************
 744              * CALCULATE INTERACTIONS *
 745              **************************/
 746
 747             /* COULOMB ELECTROSTATICS */
 748             velec            = _mm_mul_pd(qq23,rinv23);
 749             felec            = _mm_mul_pd(velec,rinvsq23);
 750
 751             /* Update potential sum for this i atom from the interaction with this j atom. */
 752             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 753             velecsum         = _mm_add_pd(velecsum,velec);
 754
 755             fscal            = felec;
 756
 757             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 758
 759             /* Calculate temporary vectorial force */
 760             tx               = _mm_mul_pd(fscal,dx23);
 761             ty               = _mm_mul_pd(fscal,dy23);
 762             tz               = _mm_mul_pd(fscal,dz23);
 763
 764             /* Update vectorial force */
 765             fix2             = _mm_add_pd(fix2,tx);
 766             fiy2             = _mm_add_pd(fiy2,ty);
 767             fiz2             = _mm_add_pd(fiz2,tz);
 768
 769             fjx3             = _mm_add_pd(fjx3,tx);
 770             fjy3             = _mm_add_pd(fjy3,ty);
 771             fjz3             = _mm_add_pd(fjz3,tz);
 772
 773             /**************************
 774              * CALCULATE INTERACTIONS *
 775              **************************/
 776
 777             /* COULOMB ELECTROSTATICS */
 778             velec            = _mm_mul_pd(qq31,rinv31);
 779             felec            = _mm_mul_pd(velec,rinvsq31);
 780
 781             /* Update potential sum for this i atom from the interaction with this j atom. */
 782             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 783             velecsum         = _mm_add_pd(velecsum,velec);
 784
 785             fscal            = felec;
 786
 787             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 788
 789             /* Calculate temporary vectorial force */
 790             tx               = _mm_mul_pd(fscal,dx31);
 791             ty               = _mm_mul_pd(fscal,dy31);
 792             tz               = _mm_mul_pd(fscal,dz31);
 793
 794             /* Update vectorial force */
 795             fix3             = _mm_add_pd(fix3,tx);
 796             fiy3             = _mm_add_pd(fiy3,ty);
 797             fiz3             = _mm_add_pd(fiz3,tz);
 798
 799             fjx1             = _mm_add_pd(fjx1,tx);
 800             fjy1             = _mm_add_pd(fjy1,ty);
 801             fjz1             = _mm_add_pd(fjz1,tz);
 802
 803             /**************************
 804              * CALCULATE INTERACTIONS *
 805              **************************/
 806
 807             /* COULOMB ELECTROSTATICS */
 808             velec            = _mm_mul_pd(qq32,rinv32);
 809             felec            = _mm_mul_pd(velec,rinvsq32);
 810
 811             /* Update potential sum for this i atom from the interaction with this j atom. */
 812             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 813             velecsum         = _mm_add_pd(velecsum,velec);
 814
 815             fscal            = felec;
 816
 817             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 818
 819             /* Calculate temporary vectorial force */
 820             tx               = _mm_mul_pd(fscal,dx32);
 821             ty               = _mm_mul_pd(fscal,dy32);
 822             tz               = _mm_mul_pd(fscal,dz32);
 823
 824             /* Update vectorial force */
 825             fix3             = _mm_add_pd(fix3,tx);
 826             fiy3             = _mm_add_pd(fiy3,ty);
 827             fiz3             = _mm_add_pd(fiz3,tz);
 828
 829             fjx2             = _mm_add_pd(fjx2,tx);
 830             fjy2             = _mm_add_pd(fjy2,ty);
 831             fjz2             = _mm_add_pd(fjz2,tz);
 832
 833             /**************************
 834              * CALCULATE INTERACTIONS *
 835              **************************/
 836
 837             /* COULOMB ELECTROSTATICS */
 838             velec            = _mm_mul_pd(qq33,rinv33);
 839             felec            = _mm_mul_pd(velec,rinvsq33);
 840
 841             /* Update potential sum for this i atom from the interaction with this j atom. */
 842             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 843             velecsum         = _mm_add_pd(velecsum,velec);
 844
 845             fscal            = felec;
 846
 847             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 848
 849             /* Calculate temporary vectorial force */
 850             tx               = _mm_mul_pd(fscal,dx33);
 851             ty               = _mm_mul_pd(fscal,dy33);
 852             tz               = _mm_mul_pd(fscal,dz33);
 853
 854             /* Update vectorial force */
 855             fix3             = _mm_add_pd(fix3,tx);
 856             fiy3             = _mm_add_pd(fiy3,ty);
 857             fiz3             = _mm_add_pd(fiz3,tz);
 858
 859             fjx3             = _mm_add_pd(fjx3,tx);
 860             fjy3             = _mm_add_pd(fjy3,ty);
 861             fjz3             = _mm_add_pd(fjz3,tz);
 862
 863             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 864
 865             /* Inner loop uses 252 flops */
 866         }
 867
 868         /* End of innermost loop */
 869
 870         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 871                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 872
 873         ggid                        = gid[iidx];
 874         /* Update potential energies */
 875         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 876
 877         /* Increment number of inner iterations */
 878         inneriter                  += j_index_end - j_index_start;
 879
 880         /* Outer loop uses 19 flops */
 881     }
 882
 883     /* Increment number of outer iterations */
 884     outeriter        += nri;
 885
 886     /* Update outer/inner flops */
 887
 888     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*252);
 889 }
 890 /*
 891  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse4_1_double
 892  * Electrostatics interaction: Coulomb
 893  * VdW interaction:            None
 894  * Geometry:                   Water4-Water4
 895  * Calculate force/pot:        Force
 896  */
 897 void
 898 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse4_1_double
 899                     (t_nblist                    * gmx_restrict       nlist,
 900                      rvec                        * gmx_restrict          xx,
 901                      rvec                        * gmx_restrict          ff,
 902                      t_forcerec                  * gmx_restrict          fr,
 903                      t_mdatoms                   * gmx_restrict     mdatoms,
 904                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 905                      t_nrnb                      * gmx_restrict        nrnb)
 906 {
 907     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 908      * just 0 for non-waters.
 909      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 910      * jnr indices corresponding to data put in the four positions in the SIMD register.
 911      */
 912     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 913     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 914     int              jnrA,jnrB;
 915     int              j_coord_offsetA,j_coord_offsetB;
 916     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 917     real             rcutoff_scalar;
 918     real             *shiftvec,*fshift,*x,*f;
 919     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 920     int              vdwioffset1;
 921     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 922     int              vdwioffset2;
 923     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 924     int              vdwioffset3;
 925     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 926     int              vdwjidx1A,vdwjidx1B;
 927     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 928     int              vdwjidx2A,vdwjidx2B;
 929     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 930     int              vdwjidx3A,vdwjidx3B;
 931     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 932     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 933     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 934     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 935     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 936     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 937     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 938     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 939     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 940     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 941     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 942     real             *charge;
 943     __m128d          dummy_mask,cutoff_mask;
 944     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 945     __m128d          one     = _mm_set1_pd(1.0);
 946     __m128d          two     = _mm_set1_pd(2.0);
 947     x                = xx[0];
 948     f                = ff[0];
 949
 950     nri              = nlist->nri;
 951     iinr             = nlist->iinr;
 952     jindex           = nlist->jindex;
 953     jjnr             = nlist->jjnr;
 954     shiftidx         = nlist->shift;
 955     gid              = nlist->gid;
 956     shiftvec         = fr->shift_vec[0];
 957     fshift           = fr->fshift[0];
 958     facel            = _mm_set1_pd(fr->epsfac);
 959     charge           = mdatoms->chargeA;
 960
 961     /* Setup water-specific parameters */
 962     inr              = nlist->iinr[0];
 963     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 964     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 965     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 966
 967     jq1              = _mm_set1_pd(charge[inr+1]);
 968     jq2              = _mm_set1_pd(charge[inr+2]);
 969     jq3              = _mm_set1_pd(charge[inr+3]);
 970     qq11             = _mm_mul_pd(iq1,jq1);
 971     qq12             = _mm_mul_pd(iq1,jq2);
 972     qq13             = _mm_mul_pd(iq1,jq3);
 973     qq21             = _mm_mul_pd(iq2,jq1);
 974     qq22             = _mm_mul_pd(iq2,jq2);
 975     qq23             = _mm_mul_pd(iq2,jq3);
 976     qq31             = _mm_mul_pd(iq3,jq1);
 977     qq32             = _mm_mul_pd(iq3,jq2);
 978     qq33             = _mm_mul_pd(iq3,jq3);
 979
 980     /* Avoid stupid compiler warnings */
 981     jnrA = jnrB = 0;
 982     j_coord_offsetA = 0;
 983     j_coord_offsetB = 0;
 984
 985     outeriter        = 0;
 986     inneriter        = 0;
 987
 988     /* Start outer loop over neighborlists */
 989     for(iidx=0; iidx<nri; iidx++)
 990     {
 991         /* Load shift vector for this list */
 992         i_shift_offset   = DIM*shiftidx[iidx];
 993
 994         /* Load limits for loop over neighbors */
 995         j_index_start    = jindex[iidx];
 996         j_index_end      = jindex[iidx+1];
 997
 998         /* Get outer coordinate index */
 999         inr              = iinr[iidx];
1000         i_coord_offset   = DIM*inr;
1001
1002         /* Load i particle coords and add shift vector */
1003         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1004                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1005
1006         fix1             = _mm_setzero_pd();
1007         fiy1             = _mm_setzero_pd();
1008         fiz1             = _mm_setzero_pd();
1009         fix2             = _mm_setzero_pd();
1010         fiy2             = _mm_setzero_pd();
1011         fiz2             = _mm_setzero_pd();
1012         fix3             = _mm_setzero_pd();
1013         fiy3             = _mm_setzero_pd();
1014         fiz3             = _mm_setzero_pd();
1015
1016         /* Start inner kernel loop */
1017         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1018         {
1019
1020             /* Get j neighbor index, and coordinate index */
1021             jnrA             = jjnr[jidx];
1022             jnrB             = jjnr[jidx+1];
1023             j_coord_offsetA  = DIM*jnrA;
1024             j_coord_offsetB  = DIM*jnrB;
1025
1026             /* load j atom coordinates */
1027             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1028                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1029
1030             /* Calculate displacement vector */
1031             dx11             = _mm_sub_pd(ix1,jx1);
1032             dy11             = _mm_sub_pd(iy1,jy1);
1033             dz11             = _mm_sub_pd(iz1,jz1);
1034             dx12             = _mm_sub_pd(ix1,jx2);
1035             dy12             = _mm_sub_pd(iy1,jy2);
1036             dz12             = _mm_sub_pd(iz1,jz2);
1037             dx13             = _mm_sub_pd(ix1,jx3);
1038             dy13             = _mm_sub_pd(iy1,jy3);
1039             dz13             = _mm_sub_pd(iz1,jz3);
1040             dx21             = _mm_sub_pd(ix2,jx1);
1041             dy21             = _mm_sub_pd(iy2,jy1);
1042             dz21             = _mm_sub_pd(iz2,jz1);
1043             dx22             = _mm_sub_pd(ix2,jx2);
1044             dy22             = _mm_sub_pd(iy2,jy2);
1045             dz22             = _mm_sub_pd(iz2,jz2);
1046             dx23             = _mm_sub_pd(ix2,jx3);
1047             dy23             = _mm_sub_pd(iy2,jy3);
1048             dz23             = _mm_sub_pd(iz2,jz3);
1049             dx31             = _mm_sub_pd(ix3,jx1);
1050             dy31             = _mm_sub_pd(iy3,jy1);
1051             dz31             = _mm_sub_pd(iz3,jz1);
1052             dx32             = _mm_sub_pd(ix3,jx2);
1053             dy32             = _mm_sub_pd(iy3,jy2);
1054             dz32             = _mm_sub_pd(iz3,jz2);
1055             dx33             = _mm_sub_pd(ix3,jx3);
1056             dy33             = _mm_sub_pd(iy3,jy3);
1057             dz33             = _mm_sub_pd(iz3,jz3);
1058
1059             /* Calculate squared distance and things based on it */
1060             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1061             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1062             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1063             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1064             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1065             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1066             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1067             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1068             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1069
1070             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1071             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1072             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1073             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1074             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1075             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1076             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1077             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1078             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1079
1080             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1081             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1082             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1083             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1084             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1085             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1086             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1087             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1088             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1089
1090             fjx1             = _mm_setzero_pd();
1091             fjy1             = _mm_setzero_pd();
1092             fjz1             = _mm_setzero_pd();
1093             fjx2             = _mm_setzero_pd();
1094             fjy2             = _mm_setzero_pd();
1095             fjz2             = _mm_setzero_pd();
1096             fjx3             = _mm_setzero_pd();
1097             fjy3             = _mm_setzero_pd();
1098             fjz3             = _mm_setzero_pd();
1099
1100             /**************************
1101              * CALCULATE INTERACTIONS *
1102              **************************/
1103
1104             /* COULOMB ELECTROSTATICS */
1105             velec            = _mm_mul_pd(qq11,rinv11);
1106             felec            = _mm_mul_pd(velec,rinvsq11);
1107
1108             fscal            = felec;
1109
1110             /* Calculate temporary vectorial force */
1111             tx               = _mm_mul_pd(fscal,dx11);
1112             ty               = _mm_mul_pd(fscal,dy11);
1113             tz               = _mm_mul_pd(fscal,dz11);
1114
1115             /* Update vectorial force */
1116             fix1             = _mm_add_pd(fix1,tx);
1117             fiy1             = _mm_add_pd(fiy1,ty);
1118             fiz1             = _mm_add_pd(fiz1,tz);
1119
1120             fjx1             = _mm_add_pd(fjx1,tx);
1121             fjy1             = _mm_add_pd(fjy1,ty);
1122             fjz1             = _mm_add_pd(fjz1,tz);
1123
1124             /**************************
1125              * CALCULATE INTERACTIONS *
1126              **************************/
1127
1128             /* COULOMB ELECTROSTATICS */
1129             velec            = _mm_mul_pd(qq12,rinv12);
1130             felec            = _mm_mul_pd(velec,rinvsq12);
1131
1132             fscal            = felec;
1133
1134             /* Calculate temporary vectorial force */
1135             tx               = _mm_mul_pd(fscal,dx12);
1136             ty               = _mm_mul_pd(fscal,dy12);
1137             tz               = _mm_mul_pd(fscal,dz12);
1138
1139             /* Update vectorial force */
1140             fix1             = _mm_add_pd(fix1,tx);
1141             fiy1             = _mm_add_pd(fiy1,ty);
1142             fiz1             = _mm_add_pd(fiz1,tz);
1143
1144             fjx2             = _mm_add_pd(fjx2,tx);
1145             fjy2             = _mm_add_pd(fjy2,ty);
1146             fjz2             = _mm_add_pd(fjz2,tz);
1147
1148             /**************************
1149              * CALCULATE INTERACTIONS *
1150              **************************/
1151
1152             /* COULOMB ELECTROSTATICS */
1153             velec            = _mm_mul_pd(qq13,rinv13);
1154             felec            = _mm_mul_pd(velec,rinvsq13);
1155
1156             fscal            = felec;
1157
1158             /* Calculate temporary vectorial force */
1159             tx               = _mm_mul_pd(fscal,dx13);
1160             ty               = _mm_mul_pd(fscal,dy13);
1161             tz               = _mm_mul_pd(fscal,dz13);
1162
1163             /* Update vectorial force */
1164             fix1             = _mm_add_pd(fix1,tx);
1165             fiy1             = _mm_add_pd(fiy1,ty);
1166             fiz1             = _mm_add_pd(fiz1,tz);
1167
1168             fjx3             = _mm_add_pd(fjx3,tx);
1169             fjy3             = _mm_add_pd(fjy3,ty);
1170             fjz3             = _mm_add_pd(fjz3,tz);
1171
1172             /**************************
1173              * CALCULATE INTERACTIONS *
1174              **************************/
1175
1176             /* COULOMB ELECTROSTATICS */
1177             velec            = _mm_mul_pd(qq21,rinv21);
1178             felec            = _mm_mul_pd(velec,rinvsq21);
1179
1180             fscal            = felec;
1181
1182             /* Calculate temporary vectorial force */
1183             tx               = _mm_mul_pd(fscal,dx21);
1184             ty               = _mm_mul_pd(fscal,dy21);
1185             tz               = _mm_mul_pd(fscal,dz21);
1186
1187             /* Update vectorial force */
1188             fix2             = _mm_add_pd(fix2,tx);
1189             fiy2             = _mm_add_pd(fiy2,ty);
1190             fiz2             = _mm_add_pd(fiz2,tz);
1191
1192             fjx1             = _mm_add_pd(fjx1,tx);
1193             fjy1             = _mm_add_pd(fjy1,ty);
1194             fjz1             = _mm_add_pd(fjz1,tz);
1195
1196             /**************************
1197              * CALCULATE INTERACTIONS *
1198              **************************/
1199
1200             /* COULOMB ELECTROSTATICS */
1201             velec            = _mm_mul_pd(qq22,rinv22);
1202             felec            = _mm_mul_pd(velec,rinvsq22);
1203
1204             fscal            = felec;
1205
1206             /* Calculate temporary vectorial force */
1207             tx               = _mm_mul_pd(fscal,dx22);
1208             ty               = _mm_mul_pd(fscal,dy22);
1209             tz               = _mm_mul_pd(fscal,dz22);
1210
1211             /* Update vectorial force */
1212             fix2             = _mm_add_pd(fix2,tx);
1213             fiy2             = _mm_add_pd(fiy2,ty);
1214             fiz2             = _mm_add_pd(fiz2,tz);
1215
1216             fjx2             = _mm_add_pd(fjx2,tx);
1217             fjy2             = _mm_add_pd(fjy2,ty);
1218             fjz2             = _mm_add_pd(fjz2,tz);
1219
1220             /**************************
1221              * CALCULATE INTERACTIONS *
1222              **************************/
1223
1224             /* COULOMB ELECTROSTATICS */
1225             velec            = _mm_mul_pd(qq23,rinv23);
1226             felec            = _mm_mul_pd(velec,rinvsq23);
1227
1228             fscal            = felec;
1229
1230             /* Calculate temporary vectorial force */
1231             tx               = _mm_mul_pd(fscal,dx23);
1232             ty               = _mm_mul_pd(fscal,dy23);
1233             tz               = _mm_mul_pd(fscal,dz23);
1234
1235             /* Update vectorial force */
1236             fix2             = _mm_add_pd(fix2,tx);
1237             fiy2             = _mm_add_pd(fiy2,ty);
1238             fiz2             = _mm_add_pd(fiz2,tz);
1239
1240             fjx3             = _mm_add_pd(fjx3,tx);
1241             fjy3             = _mm_add_pd(fjy3,ty);
1242             fjz3             = _mm_add_pd(fjz3,tz);
1243
1244             /**************************
1245              * CALCULATE INTERACTIONS *
1246              **************************/
1247
1248             /* COULOMB ELECTROSTATICS */
1249             velec            = _mm_mul_pd(qq31,rinv31);
1250             felec            = _mm_mul_pd(velec,rinvsq31);
1251
1252             fscal            = felec;
1253
1254             /* Calculate temporary vectorial force */
1255             tx               = _mm_mul_pd(fscal,dx31);
1256             ty               = _mm_mul_pd(fscal,dy31);
1257             tz               = _mm_mul_pd(fscal,dz31);
1258
1259             /* Update vectorial force */
1260             fix3             = _mm_add_pd(fix3,tx);
1261             fiy3             = _mm_add_pd(fiy3,ty);
1262             fiz3             = _mm_add_pd(fiz3,tz);
1263
1264             fjx1             = _mm_add_pd(fjx1,tx);
1265             fjy1             = _mm_add_pd(fjy1,ty);
1266             fjz1             = _mm_add_pd(fjz1,tz);
1267
1268             /**************************
1269              * CALCULATE INTERACTIONS *
1270              **************************/
1271
1272             /* COULOMB ELECTROSTATICS */
1273             velec            = _mm_mul_pd(qq32,rinv32);
1274             felec            = _mm_mul_pd(velec,rinvsq32);
1275
1276             fscal            = felec;
1277
1278             /* Calculate temporary vectorial force */
1279             tx               = _mm_mul_pd(fscal,dx32);
1280             ty               = _mm_mul_pd(fscal,dy32);
1281             tz               = _mm_mul_pd(fscal,dz32);
1282
1283             /* Update vectorial force */
1284             fix3             = _mm_add_pd(fix3,tx);
1285             fiy3             = _mm_add_pd(fiy3,ty);
1286             fiz3             = _mm_add_pd(fiz3,tz);
1287
1288             fjx2             = _mm_add_pd(fjx2,tx);
1289             fjy2             = _mm_add_pd(fjy2,ty);
1290             fjz2             = _mm_add_pd(fjz2,tz);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             /* COULOMB ELECTROSTATICS */
1297             velec            = _mm_mul_pd(qq33,rinv33);
1298             felec            = _mm_mul_pd(velec,rinvsq33);
1299
1300             fscal            = felec;
1301
1302             /* Calculate temporary vectorial force */
1303             tx               = _mm_mul_pd(fscal,dx33);
1304             ty               = _mm_mul_pd(fscal,dy33);
1305             tz               = _mm_mul_pd(fscal,dz33);
1306
1307             /* Update vectorial force */
1308             fix3             = _mm_add_pd(fix3,tx);
1309             fiy3             = _mm_add_pd(fiy3,ty);
1310             fiz3             = _mm_add_pd(fiz3,tz);
1311
1312             fjx3             = _mm_add_pd(fjx3,tx);
1313             fjy3             = _mm_add_pd(fjy3,ty);
1314             fjz3             = _mm_add_pd(fjz3,tz);
1315
1316             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1317
1318             /* Inner loop uses 243 flops */
1319         }
1320
1321         if(jidx<j_index_end)
1322         {
1323
1324             jnrA             = jjnr[jidx];
1325             j_coord_offsetA  = DIM*jnrA;
1326
1327             /* load j atom coordinates */
1328             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1329                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1330
1331             /* Calculate displacement vector */
1332             dx11             = _mm_sub_pd(ix1,jx1);
1333             dy11             = _mm_sub_pd(iy1,jy1);
1334             dz11             = _mm_sub_pd(iz1,jz1);
1335             dx12             = _mm_sub_pd(ix1,jx2);
1336             dy12             = _mm_sub_pd(iy1,jy2);
1337             dz12             = _mm_sub_pd(iz1,jz2);
1338             dx13             = _mm_sub_pd(ix1,jx3);
1339             dy13             = _mm_sub_pd(iy1,jy3);
1340             dz13             = _mm_sub_pd(iz1,jz3);
1341             dx21             = _mm_sub_pd(ix2,jx1);
1342             dy21             = _mm_sub_pd(iy2,jy1);
1343             dz21             = _mm_sub_pd(iz2,jz1);
1344             dx22             = _mm_sub_pd(ix2,jx2);
1345             dy22             = _mm_sub_pd(iy2,jy2);
1346             dz22             = _mm_sub_pd(iz2,jz2);
1347             dx23             = _mm_sub_pd(ix2,jx3);
1348             dy23             = _mm_sub_pd(iy2,jy3);
1349             dz23             = _mm_sub_pd(iz2,jz3);
1350             dx31             = _mm_sub_pd(ix3,jx1);
1351             dy31             = _mm_sub_pd(iy3,jy1);
1352             dz31             = _mm_sub_pd(iz3,jz1);
1353             dx32             = _mm_sub_pd(ix3,jx2);
1354             dy32             = _mm_sub_pd(iy3,jy2);
1355             dz32             = _mm_sub_pd(iz3,jz2);
1356             dx33             = _mm_sub_pd(ix3,jx3);
1357             dy33             = _mm_sub_pd(iy3,jy3);
1358             dz33             = _mm_sub_pd(iz3,jz3);
1359
1360             /* Calculate squared distance and things based on it */
1361             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1362             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1363             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1364             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1365             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1366             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1367             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1368             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1369             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1370
1371             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1372             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1373             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1374             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1375             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1376             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1377             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1378             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1379             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1380
1381             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1382             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1383             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1384             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1385             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1386             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1387             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1388             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1389             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1390
1391             fjx1             = _mm_setzero_pd();
1392             fjy1             = _mm_setzero_pd();
1393             fjz1             = _mm_setzero_pd();
1394             fjx2             = _mm_setzero_pd();
1395             fjy2             = _mm_setzero_pd();
1396             fjz2             = _mm_setzero_pd();
1397             fjx3             = _mm_setzero_pd();
1398             fjy3             = _mm_setzero_pd();
1399             fjz3             = _mm_setzero_pd();
1400
1401             /**************************
1402              * CALCULATE INTERACTIONS *
1403              **************************/
1404
1405             /* COULOMB ELECTROSTATICS */
1406             velec            = _mm_mul_pd(qq11,rinv11);
1407             felec            = _mm_mul_pd(velec,rinvsq11);
1408
1409             fscal            = felec;
1410
1411             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1412
1413             /* Calculate temporary vectorial force */
1414             tx               = _mm_mul_pd(fscal,dx11);
1415             ty               = _mm_mul_pd(fscal,dy11);
1416             tz               = _mm_mul_pd(fscal,dz11);
1417
1418             /* Update vectorial force */
1419             fix1             = _mm_add_pd(fix1,tx);
1420             fiy1             = _mm_add_pd(fiy1,ty);
1421             fiz1             = _mm_add_pd(fiz1,tz);
1422
1423             fjx1             = _mm_add_pd(fjx1,tx);
1424             fjy1             = _mm_add_pd(fjy1,ty);
1425             fjz1             = _mm_add_pd(fjz1,tz);
1426
1427             /**************************
1428              * CALCULATE INTERACTIONS *
1429              **************************/
1430
1431             /* COULOMB ELECTROSTATICS */
1432             velec            = _mm_mul_pd(qq12,rinv12);
1433             felec            = _mm_mul_pd(velec,rinvsq12);
1434
1435             fscal            = felec;
1436
1437             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1438
1439             /* Calculate temporary vectorial force */
1440             tx               = _mm_mul_pd(fscal,dx12);
1441             ty               = _mm_mul_pd(fscal,dy12);
1442             tz               = _mm_mul_pd(fscal,dz12);
1443
1444             /* Update vectorial force */
1445             fix1             = _mm_add_pd(fix1,tx);
1446             fiy1             = _mm_add_pd(fiy1,ty);
1447             fiz1             = _mm_add_pd(fiz1,tz);
1448
1449             fjx2             = _mm_add_pd(fjx2,tx);
1450             fjy2             = _mm_add_pd(fjy2,ty);
1451             fjz2             = _mm_add_pd(fjz2,tz);
1452
1453             /**************************
1454              * CALCULATE INTERACTIONS *
1455              **************************/
1456
1457             /* COULOMB ELECTROSTATICS */
1458             velec            = _mm_mul_pd(qq13,rinv13);
1459             felec            = _mm_mul_pd(velec,rinvsq13);
1460
1461             fscal            = felec;
1462
1463             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1464
1465             /* Calculate temporary vectorial force */
1466             tx               = _mm_mul_pd(fscal,dx13);
1467             ty               = _mm_mul_pd(fscal,dy13);
1468             tz               = _mm_mul_pd(fscal,dz13);
1469
1470             /* Update vectorial force */
1471             fix1             = _mm_add_pd(fix1,tx);
1472             fiy1             = _mm_add_pd(fiy1,ty);
1473             fiz1             = _mm_add_pd(fiz1,tz);
1474
1475             fjx3             = _mm_add_pd(fjx3,tx);
1476             fjy3             = _mm_add_pd(fjy3,ty);
1477             fjz3             = _mm_add_pd(fjz3,tz);
1478
1479             /**************************
1480              * CALCULATE INTERACTIONS *
1481              **************************/
1482
1483             /* COULOMB ELECTROSTATICS */
1484             velec            = _mm_mul_pd(qq21,rinv21);
1485             felec            = _mm_mul_pd(velec,rinvsq21);
1486
1487             fscal            = felec;
1488
1489             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1490
1491             /* Calculate temporary vectorial force */
1492             tx               = _mm_mul_pd(fscal,dx21);
1493             ty               = _mm_mul_pd(fscal,dy21);
1494             tz               = _mm_mul_pd(fscal,dz21);
1495
1496             /* Update vectorial force */
1497             fix2             = _mm_add_pd(fix2,tx);
1498             fiy2             = _mm_add_pd(fiy2,ty);
1499             fiz2             = _mm_add_pd(fiz2,tz);
1500
1501             fjx1             = _mm_add_pd(fjx1,tx);
1502             fjy1             = _mm_add_pd(fjy1,ty);
1503             fjz1             = _mm_add_pd(fjz1,tz);
1504
1505             /**************************
1506              * CALCULATE INTERACTIONS *
1507              **************************/
1508
1509             /* COULOMB ELECTROSTATICS */
1510             velec            = _mm_mul_pd(qq22,rinv22);
1511             felec            = _mm_mul_pd(velec,rinvsq22);
1512
1513             fscal            = felec;
1514
1515             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1516
1517             /* Calculate temporary vectorial force */
1518             tx               = _mm_mul_pd(fscal,dx22);
1519             ty               = _mm_mul_pd(fscal,dy22);
1520             tz               = _mm_mul_pd(fscal,dz22);
1521
1522             /* Update vectorial force */
1523             fix2             = _mm_add_pd(fix2,tx);
1524             fiy2             = _mm_add_pd(fiy2,ty);
1525             fiz2             = _mm_add_pd(fiz2,tz);
1526
1527             fjx2             = _mm_add_pd(fjx2,tx);
1528             fjy2             = _mm_add_pd(fjy2,ty);
1529             fjz2             = _mm_add_pd(fjz2,tz);
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             /* COULOMB ELECTROSTATICS */
1536             velec            = _mm_mul_pd(qq23,rinv23);
1537             felec            = _mm_mul_pd(velec,rinvsq23);
1538
1539             fscal            = felec;
1540
1541             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1542
1543             /* Calculate temporary vectorial force */
1544             tx               = _mm_mul_pd(fscal,dx23);
1545             ty               = _mm_mul_pd(fscal,dy23);
1546             tz               = _mm_mul_pd(fscal,dz23);
1547
1548             /* Update vectorial force */
1549             fix2             = _mm_add_pd(fix2,tx);
1550             fiy2             = _mm_add_pd(fiy2,ty);
1551             fiz2             = _mm_add_pd(fiz2,tz);
1552
1553             fjx3             = _mm_add_pd(fjx3,tx);
1554             fjy3             = _mm_add_pd(fjy3,ty);
1555             fjz3             = _mm_add_pd(fjz3,tz);
1556
1557             /**************************
1558              * CALCULATE INTERACTIONS *
1559              **************************/
1560
1561             /* COULOMB ELECTROSTATICS */
1562             velec            = _mm_mul_pd(qq31,rinv31);
1563             felec            = _mm_mul_pd(velec,rinvsq31);
1564
1565             fscal            = felec;
1566
1567             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1568
1569             /* Calculate temporary vectorial force */
1570             tx               = _mm_mul_pd(fscal,dx31);
1571             ty               = _mm_mul_pd(fscal,dy31);
1572             tz               = _mm_mul_pd(fscal,dz31);
1573
1574             /* Update vectorial force */
1575             fix3             = _mm_add_pd(fix3,tx);
1576             fiy3             = _mm_add_pd(fiy3,ty);
1577             fiz3             = _mm_add_pd(fiz3,tz);
1578
1579             fjx1             = _mm_add_pd(fjx1,tx);
1580             fjy1             = _mm_add_pd(fjy1,ty);
1581             fjz1             = _mm_add_pd(fjz1,tz);
1582
1583             /**************************
1584              * CALCULATE INTERACTIONS *
1585              **************************/
1586
1587             /* COULOMB ELECTROSTATICS */
1588             velec            = _mm_mul_pd(qq32,rinv32);
1589             felec            = _mm_mul_pd(velec,rinvsq32);
1590
1591             fscal            = felec;
1592
1593             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1594
1595             /* Calculate temporary vectorial force */
1596             tx               = _mm_mul_pd(fscal,dx32);
1597             ty               = _mm_mul_pd(fscal,dy32);
1598             tz               = _mm_mul_pd(fscal,dz32);
1599
1600             /* Update vectorial force */
1601             fix3             = _mm_add_pd(fix3,tx);
1602             fiy3             = _mm_add_pd(fiy3,ty);
1603             fiz3             = _mm_add_pd(fiz3,tz);
1604
1605             fjx2             = _mm_add_pd(fjx2,tx);
1606             fjy2             = _mm_add_pd(fjy2,ty);
1607             fjz2             = _mm_add_pd(fjz2,tz);
1608
1609             /**************************
1610              * CALCULATE INTERACTIONS *
1611              **************************/
1612
1613             /* COULOMB ELECTROSTATICS */
1614             velec            = _mm_mul_pd(qq33,rinv33);
1615             felec            = _mm_mul_pd(velec,rinvsq33);
1616
1617             fscal            = felec;
1618
1619             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1620
1621             /* Calculate temporary vectorial force */
1622             tx               = _mm_mul_pd(fscal,dx33);
1623             ty               = _mm_mul_pd(fscal,dy33);
1624             tz               = _mm_mul_pd(fscal,dz33);
1625
1626             /* Update vectorial force */
1627             fix3             = _mm_add_pd(fix3,tx);
1628             fiy3             = _mm_add_pd(fiy3,ty);
1629             fiz3             = _mm_add_pd(fiz3,tz);
1630
1631             fjx3             = _mm_add_pd(fjx3,tx);
1632             fjy3             = _mm_add_pd(fjy3,ty);
1633             fjz3             = _mm_add_pd(fjz3,tz);
1634
1635             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1636
1637             /* Inner loop uses 243 flops */
1638         }
1639
1640         /* End of innermost loop */
1641
1642         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1643                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1644
1645         /* Increment number of inner iterations */
1646         inneriter                  += j_index_end - j_index_start;
1647
1648         /* Outer loop uses 18 flops */
1649     }
1650
1651     /* Increment number of outer iterations */
1652     outeriter        += nri;
1653
1654     /* Update outer/inner flops */
1655
1656     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1657 }