src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwNone_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_double
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            None
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_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              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 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     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 126     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 127     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 128
 129     jq0              = _mm_set1_pd(charge[inr+0]);
 130     jq1              = _mm_set1_pd(charge[inr+1]);
 131     jq2              = _mm_set1_pd(charge[inr+2]);
 132     qq00             = _mm_mul_pd(iq0,jq0);
 133     qq01             = _mm_mul_pd(iq0,jq1);
 134     qq02             = _mm_mul_pd(iq0,jq2);
 135     qq10             = _mm_mul_pd(iq1,jq0);
 136     qq11             = _mm_mul_pd(iq1,jq1);
 137     qq12             = _mm_mul_pd(iq1,jq2);
 138     qq20             = _mm_mul_pd(iq2,jq0);
 139     qq21             = _mm_mul_pd(iq2,jq1);
 140     qq22             = _mm_mul_pd(iq2,jq2);
 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,
 166                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 167
 168         fix0             = _mm_setzero_pd();
 169         fiy0             = _mm_setzero_pd();
 170         fiz0             = _mm_setzero_pd();
 171         fix1             = _mm_setzero_pd();
 172         fiy1             = _mm_setzero_pd();
 173         fiz1             = _mm_setzero_pd();
 174         fix2             = _mm_setzero_pd();
 175         fiy2             = _mm_setzero_pd();
 176         fiz2             = _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,x+j_coord_offsetB,
 193                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 194
 195             /* Calculate displacement vector */
 196             dx00             = _mm_sub_pd(ix0,jx0);
 197             dy00             = _mm_sub_pd(iy0,jy0);
 198             dz00             = _mm_sub_pd(iz0,jz0);
 199             dx01             = _mm_sub_pd(ix0,jx1);
 200             dy01             = _mm_sub_pd(iy0,jy1);
 201             dz01             = _mm_sub_pd(iz0,jz1);
 202             dx02             = _mm_sub_pd(ix0,jx2);
 203             dy02             = _mm_sub_pd(iy0,jy2);
 204             dz02             = _mm_sub_pd(iz0,jz2);
 205             dx10             = _mm_sub_pd(ix1,jx0);
 206             dy10             = _mm_sub_pd(iy1,jy0);
 207             dz10             = _mm_sub_pd(iz1,jz0);
 208             dx11             = _mm_sub_pd(ix1,jx1);
 209             dy11             = _mm_sub_pd(iy1,jy1);
 210             dz11             = _mm_sub_pd(iz1,jz1);
 211             dx12             = _mm_sub_pd(ix1,jx2);
 212             dy12             = _mm_sub_pd(iy1,jy2);
 213             dz12             = _mm_sub_pd(iz1,jz2);
 214             dx20             = _mm_sub_pd(ix2,jx0);
 215             dy20             = _mm_sub_pd(iy2,jy0);
 216             dz20             = _mm_sub_pd(iz2,jz0);
 217             dx21             = _mm_sub_pd(ix2,jx1);
 218             dy21             = _mm_sub_pd(iy2,jy1);
 219             dz21             = _mm_sub_pd(iz2,jz1);
 220             dx22             = _mm_sub_pd(ix2,jx2);
 221             dy22             = _mm_sub_pd(iy2,jy2);
 222             dz22             = _mm_sub_pd(iz2,jz2);
 223
 224             /* Calculate squared distance and things based on it */
 225             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 226             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 227             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 228             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 229             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 230             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 231             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 232             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 233             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 234
 235             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 236             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 237             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 238             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 239             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 240             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 241             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 242             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 243             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 244
 245             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 246             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 247             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 248             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 249             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 250             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 251             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 252             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 253             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 254
 255             fjx0             = _mm_setzero_pd();
 256             fjy0             = _mm_setzero_pd();
 257             fjz0             = _mm_setzero_pd();
 258             fjx1             = _mm_setzero_pd();
 259             fjy1             = _mm_setzero_pd();
 260             fjz1             = _mm_setzero_pd();
 261             fjx2             = _mm_setzero_pd();
 262             fjy2             = _mm_setzero_pd();
 263             fjz2             = _mm_setzero_pd();
 264
 265             /**************************
 266              * CALCULATE INTERACTIONS *
 267              **************************/
 268
 269             /* COULOMB ELECTROSTATICS */
 270             velec            = _mm_mul_pd(qq00,rinv00);
 271             felec            = _mm_mul_pd(velec,rinvsq00);
 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             /* Update vectorial force */
 279             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 280             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 281             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 282
 283             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 284             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 285             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 286
 287             /**************************
 288              * CALCULATE INTERACTIONS *
 289              **************************/
 290
 291             /* COULOMB ELECTROSTATICS */
 292             velec            = _mm_mul_pd(qq01,rinv01);
 293             felec            = _mm_mul_pd(velec,rinvsq01);
 294
 295             /* Update potential sum for this i atom from the interaction with this j atom. */
 296             velecsum         = _mm_add_pd(velecsum,velec);
 297
 298             fscal            = felec;
 299
 300             /* Update vectorial force */
 301             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 302             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 303             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 304
 305             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 306             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 307             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             /* COULOMB ELECTROSTATICS */
 314             velec            = _mm_mul_pd(qq02,rinv02);
 315             felec            = _mm_mul_pd(velec,rinvsq02);
 316
 317             /* Update potential sum for this i atom from the interaction with this j atom. */
 318             velecsum         = _mm_add_pd(velecsum,velec);
 319
 320             fscal            = felec;
 321
 322             /* Update vectorial force */
 323             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 324             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 325             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 326
 327             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 328             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 329             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 330
 331             /**************************
 332              * CALCULATE INTERACTIONS *
 333              **************************/
 334
 335             /* COULOMB ELECTROSTATICS */
 336             velec            = _mm_mul_pd(qq10,rinv10);
 337             felec            = _mm_mul_pd(velec,rinvsq10);
 338
 339             /* Update potential sum for this i atom from the interaction with this j atom. */
 340             velecsum         = _mm_add_pd(velecsum,velec);
 341
 342             fscal            = felec;
 343
 344             /* Update vectorial force */
 345             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 346             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 347             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 348
 349             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 350             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 351             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             /* COULOMB ELECTROSTATICS */
 358             velec            = _mm_mul_pd(qq11,rinv11);
 359             felec            = _mm_mul_pd(velec,rinvsq11);
 360
 361             /* Update potential sum for this i atom from the interaction with this j atom. */
 362             velecsum         = _mm_add_pd(velecsum,velec);
 363
 364             fscal            = felec;
 365
 366             /* Update vectorial force */
 367             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 368             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 369             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 370
 371             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 372             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 373             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 374
 375             /**************************
 376              * CALCULATE INTERACTIONS *
 377              **************************/
 378
 379             /* COULOMB ELECTROSTATICS */
 380             velec            = _mm_mul_pd(qq12,rinv12);
 381             felec            = _mm_mul_pd(velec,rinvsq12);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velecsum         = _mm_add_pd(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             /* Update vectorial force */
 389             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 390             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 391             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 392
 393             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 394             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 395             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 396
 397             /**************************
 398              * CALCULATE INTERACTIONS *
 399              **************************/
 400
 401             /* COULOMB ELECTROSTATICS */
 402             velec            = _mm_mul_pd(qq20,rinv20);
 403             felec            = _mm_mul_pd(velec,rinvsq20);
 404
 405             /* Update potential sum for this i atom from the interaction with this j atom. */
 406             velecsum         = _mm_add_pd(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             /* Update vectorial force */
 411             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 412             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 413             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 414
 415             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 416             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 417             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             /* COULOMB ELECTROSTATICS */
 424             velec            = _mm_mul_pd(qq21,rinv21);
 425             felec            = _mm_mul_pd(velec,rinvsq21);
 426
 427             /* Update potential sum for this i atom from the interaction with this j atom. */
 428             velecsum         = _mm_add_pd(velecsum,velec);
 429
 430             fscal            = felec;
 431
 432             /* Update vectorial force */
 433             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 434             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 435             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 436
 437             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 438             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 439             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 440
 441             /**************************
 442              * CALCULATE INTERACTIONS *
 443              **************************/
 444
 445             /* COULOMB ELECTROSTATICS */
 446             velec            = _mm_mul_pd(qq22,rinv22);
 447             felec            = _mm_mul_pd(velec,rinvsq22);
 448
 449             /* Update potential sum for this i atom from the interaction with this j atom. */
 450             velecsum         = _mm_add_pd(velecsum,velec);
 451
 452             fscal            = felec;
 453
 454             /* Update vectorial force */
 455             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 456             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 457             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 458
 459             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 460             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 461             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 462
 463             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 464
 465             /* Inner loop uses 279 flops */
 466         }
 467
 468         if(jidx<j_index_end)
 469         {
 470
 471             jnrA             = jjnr[jidx];
 472             j_coord_offsetA  = DIM*jnrA;
 473
 474             /* load j atom coordinates */
 475             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 476                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 477
 478             /* Calculate displacement vector */
 479             dx00             = _mm_sub_pd(ix0,jx0);
 480             dy00             = _mm_sub_pd(iy0,jy0);
 481             dz00             = _mm_sub_pd(iz0,jz0);
 482             dx01             = _mm_sub_pd(ix0,jx1);
 483             dy01             = _mm_sub_pd(iy0,jy1);
 484             dz01             = _mm_sub_pd(iz0,jz1);
 485             dx02             = _mm_sub_pd(ix0,jx2);
 486             dy02             = _mm_sub_pd(iy0,jy2);
 487             dz02             = _mm_sub_pd(iz0,jz2);
 488             dx10             = _mm_sub_pd(ix1,jx0);
 489             dy10             = _mm_sub_pd(iy1,jy0);
 490             dz10             = _mm_sub_pd(iz1,jz0);
 491             dx11             = _mm_sub_pd(ix1,jx1);
 492             dy11             = _mm_sub_pd(iy1,jy1);
 493             dz11             = _mm_sub_pd(iz1,jz1);
 494             dx12             = _mm_sub_pd(ix1,jx2);
 495             dy12             = _mm_sub_pd(iy1,jy2);
 496             dz12             = _mm_sub_pd(iz1,jz2);
 497             dx20             = _mm_sub_pd(ix2,jx0);
 498             dy20             = _mm_sub_pd(iy2,jy0);
 499             dz20             = _mm_sub_pd(iz2,jz0);
 500             dx21             = _mm_sub_pd(ix2,jx1);
 501             dy21             = _mm_sub_pd(iy2,jy1);
 502             dz21             = _mm_sub_pd(iz2,jz1);
 503             dx22             = _mm_sub_pd(ix2,jx2);
 504             dy22             = _mm_sub_pd(iy2,jy2);
 505             dz22             = _mm_sub_pd(iz2,jz2);
 506
 507             /* Calculate squared distance and things based on it */
 508             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 509             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 510             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 511             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 512             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 513             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 514             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 515             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 516             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 517
 518             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 519             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 520             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 521             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 522             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 523             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 524             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 525             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 526             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 527
 528             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 529             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 530             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 531             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 532             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 533             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 534             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 535             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 536             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 537
 538             fjx0             = _mm_setzero_pd();
 539             fjy0             = _mm_setzero_pd();
 540             fjz0             = _mm_setzero_pd();
 541             fjx1             = _mm_setzero_pd();
 542             fjy1             = _mm_setzero_pd();
 543             fjz1             = _mm_setzero_pd();
 544             fjx2             = _mm_setzero_pd();
 545             fjy2             = _mm_setzero_pd();
 546             fjz2             = _mm_setzero_pd();
 547
 548             /**************************
 549              * CALCULATE INTERACTIONS *
 550              **************************/
 551
 552             /* COULOMB ELECTROSTATICS */
 553             velec            = _mm_mul_pd(qq00,rinv00);
 554             felec            = _mm_mul_pd(velec,rinvsq00);
 555
 556             /* Update potential sum for this i atom from the interaction with this j atom. */
 557             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 558             velecsum         = _mm_add_pd(velecsum,velec);
 559
 560             fscal            = felec;
 561
 562             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 563
 564             /* Update vectorial force */
 565             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 566             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 567             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 568
 569             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 570             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 571             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 572
 573             /**************************
 574              * CALCULATE INTERACTIONS *
 575              **************************/
 576
 577             /* COULOMB ELECTROSTATICS */
 578             velec            = _mm_mul_pd(qq01,rinv01);
 579             felec            = _mm_mul_pd(velec,rinvsq01);
 580
 581             /* Update potential sum for this i atom from the interaction with this j atom. */
 582             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 583             velecsum         = _mm_add_pd(velecsum,velec);
 584
 585             fscal            = felec;
 586
 587             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 588
 589             /* Update vectorial force */
 590             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 591             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 592             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 593
 594             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 595             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 596             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 597
 598             /**************************
 599              * CALCULATE INTERACTIONS *
 600              **************************/
 601
 602             /* COULOMB ELECTROSTATICS */
 603             velec            = _mm_mul_pd(qq02,rinv02);
 604             felec            = _mm_mul_pd(velec,rinvsq02);
 605
 606             /* Update potential sum for this i atom from the interaction with this j atom. */
 607             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 608             velecsum         = _mm_add_pd(velecsum,velec);
 609
 610             fscal            = felec;
 611
 612             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 613
 614             /* Update vectorial force */
 615             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 616             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 617             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 618
 619             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 620             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 621             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 622
 623             /**************************
 624              * CALCULATE INTERACTIONS *
 625              **************************/
 626
 627             /* COULOMB ELECTROSTATICS */
 628             velec            = _mm_mul_pd(qq10,rinv10);
 629             felec            = _mm_mul_pd(velec,rinvsq10);
 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             /* Update vectorial force */
 640             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 641             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 642             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 643
 644             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 645             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 646             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 647
 648             /**************************
 649              * CALCULATE INTERACTIONS *
 650              **************************/
 651
 652             /* COULOMB ELECTROSTATICS */
 653             velec            = _mm_mul_pd(qq11,rinv11);
 654             felec            = _mm_mul_pd(velec,rinvsq11);
 655
 656             /* Update potential sum for this i atom from the interaction with this j atom. */
 657             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 658             velecsum         = _mm_add_pd(velecsum,velec);
 659
 660             fscal            = felec;
 661
 662             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 663
 664             /* Update vectorial force */
 665             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 666             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 667             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 668
 669             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 670             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 671             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 672
 673             /**************************
 674              * CALCULATE INTERACTIONS *
 675              **************************/
 676
 677             /* COULOMB ELECTROSTATICS */
 678             velec            = _mm_mul_pd(qq12,rinv12);
 679             felec            = _mm_mul_pd(velec,rinvsq12);
 680
 681             /* Update potential sum for this i atom from the interaction with this j atom. */
 682             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 683             velecsum         = _mm_add_pd(velecsum,velec);
 684
 685             fscal            = felec;
 686
 687             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 688
 689             /* Update vectorial force */
 690             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 691             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 692             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 693
 694             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 695             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 696             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 697
 698             /**************************
 699              * CALCULATE INTERACTIONS *
 700              **************************/
 701
 702             /* COULOMB ELECTROSTATICS */
 703             velec            = _mm_mul_pd(qq20,rinv20);
 704             felec            = _mm_mul_pd(velec,rinvsq20);
 705
 706             /* Update potential sum for this i atom from the interaction with this j atom. */
 707             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 708             velecsum         = _mm_add_pd(velecsum,velec);
 709
 710             fscal            = felec;
 711
 712             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 713
 714             /* Update vectorial force */
 715             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 716             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 717             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 718
 719             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 720             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 721             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 722
 723             /**************************
 724              * CALCULATE INTERACTIONS *
 725              **************************/
 726
 727             /* COULOMB ELECTROSTATICS */
 728             velec            = _mm_mul_pd(qq21,rinv21);
 729             felec            = _mm_mul_pd(velec,rinvsq21);
 730
 731             /* Update potential sum for this i atom from the interaction with this j atom. */
 732             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 733             velecsum         = _mm_add_pd(velecsum,velec);
 734
 735             fscal            = felec;
 736
 737             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 738
 739             /* Update vectorial force */
 740             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 741             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 742             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 743
 744             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 745             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 746             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             /* COULOMB ELECTROSTATICS */
 753             velec            = _mm_mul_pd(qq22,rinv22);
 754             felec            = _mm_mul_pd(velec,rinvsq22);
 755
 756             /* Update potential sum for this i atom from the interaction with this j atom. */
 757             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 758             velecsum         = _mm_add_pd(velecsum,velec);
 759
 760             fscal            = felec;
 761
 762             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 763
 764             /* Update vectorial force */
 765             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 766             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 767             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 768
 769             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 770             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 771             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 772
 773             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 774
 775             /* Inner loop uses 279 flops */
 776         }
 777
 778         /* End of innermost loop */
 779
 780         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 781                                               f+i_coord_offset,fshift+i_shift_offset);
 782
 783         ggid                        = gid[iidx];
 784         /* Update potential energies */
 785         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 786
 787         /* Increment number of inner iterations */
 788         inneriter                  += j_index_end - j_index_start;
 789
 790         /* Outer loop uses 19 flops */
 791     }
 792
 793     /* Increment number of outer iterations */
 794     outeriter        += nri;
 795
 796     /* Update outer/inner flops */
 797
 798     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*279);
 799 }
 800 /*
 801  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_double
 802  * Electrostatics interaction: Coulomb
 803  * VdW interaction:            None
 804  * Geometry:                   Water3-Water3
 805  * Calculate force/pot:        Force
 806  */
 807 void
 808 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_double
 809                     (t_nblist                    * gmx_restrict       nlist,
 810                      rvec                        * gmx_restrict          xx,
 811                      rvec                        * gmx_restrict          ff,
 812                      t_forcerec                  * gmx_restrict          fr,
 813                      t_mdatoms                   * gmx_restrict     mdatoms,
 814                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 815                      t_nrnb                      * gmx_restrict        nrnb)
 816 {
 817     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 818      * just 0 for non-waters.
 819      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 820      * jnr indices corresponding to data put in the four positions in the SIMD register.
 821      */
 822     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 823     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 824     int              jnrA,jnrB;
 825     int              j_coord_offsetA,j_coord_offsetB;
 826     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 827     real             rcutoff_scalar;
 828     real             *shiftvec,*fshift,*x,*f;
 829     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 830     int              vdwioffset0;
 831     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 832     int              vdwioffset1;
 833     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 834     int              vdwioffset2;
 835     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 836     int              vdwjidx0A,vdwjidx0B;
 837     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 838     int              vdwjidx1A,vdwjidx1B;
 839     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 840     int              vdwjidx2A,vdwjidx2B;
 841     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 842     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 843     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 844     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 845     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 846     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 847     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 848     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 849     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 850     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 851     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 852     real             *charge;
 853     __m128d          dummy_mask,cutoff_mask;
 854     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 855     __m128d          one     = _mm_set1_pd(1.0);
 856     __m128d          two     = _mm_set1_pd(2.0);
 857     x                = xx[0];
 858     f                = ff[0];
 859
 860     nri              = nlist->nri;
 861     iinr             = nlist->iinr;
 862     jindex           = nlist->jindex;
 863     jjnr             = nlist->jjnr;
 864     shiftidx         = nlist->shift;
 865     gid              = nlist->gid;
 866     shiftvec         = fr->shift_vec[0];
 867     fshift           = fr->fshift[0];
 868     facel            = _mm_set1_pd(fr->epsfac);
 869     charge           = mdatoms->chargeA;
 870
 871     /* Setup water-specific parameters */
 872     inr              = nlist->iinr[0];
 873     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 874     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 875     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 876
 877     jq0              = _mm_set1_pd(charge[inr+0]);
 878     jq1              = _mm_set1_pd(charge[inr+1]);
 879     jq2              = _mm_set1_pd(charge[inr+2]);
 880     qq00             = _mm_mul_pd(iq0,jq0);
 881     qq01             = _mm_mul_pd(iq0,jq1);
 882     qq02             = _mm_mul_pd(iq0,jq2);
 883     qq10             = _mm_mul_pd(iq1,jq0);
 884     qq11             = _mm_mul_pd(iq1,jq1);
 885     qq12             = _mm_mul_pd(iq1,jq2);
 886     qq20             = _mm_mul_pd(iq2,jq0);
 887     qq21             = _mm_mul_pd(iq2,jq1);
 888     qq22             = _mm_mul_pd(iq2,jq2);
 889
 890     /* Avoid stupid compiler warnings */
 891     jnrA = jnrB = 0;
 892     j_coord_offsetA = 0;
 893     j_coord_offsetB = 0;
 894
 895     outeriter        = 0;
 896     inneriter        = 0;
 897
 898     /* Start outer loop over neighborlists */
 899     for(iidx=0; iidx<nri; iidx++)
 900     {
 901         /* Load shift vector for this list */
 902         i_shift_offset   = DIM*shiftidx[iidx];
 903
 904         /* Load limits for loop over neighbors */
 905         j_index_start    = jindex[iidx];
 906         j_index_end      = jindex[iidx+1];
 907
 908         /* Get outer coordinate index */
 909         inr              = iinr[iidx];
 910         i_coord_offset   = DIM*inr;
 911
 912         /* Load i particle coords and add shift vector */
 913         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 914                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 915
 916         fix0             = _mm_setzero_pd();
 917         fiy0             = _mm_setzero_pd();
 918         fiz0             = _mm_setzero_pd();
 919         fix1             = _mm_setzero_pd();
 920         fiy1             = _mm_setzero_pd();
 921         fiz1             = _mm_setzero_pd();
 922         fix2             = _mm_setzero_pd();
 923         fiy2             = _mm_setzero_pd();
 924         fiz2             = _mm_setzero_pd();
 925
 926         /* Start inner kernel loop */
 927         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 928         {
 929
 930             /* Get j neighbor index, and coordinate index */
 931             jnrA             = jjnr[jidx];
 932             jnrB             = jjnr[jidx+1];
 933             j_coord_offsetA  = DIM*jnrA;
 934             j_coord_offsetB  = DIM*jnrB;
 935
 936             /* load j atom coordinates */
 937             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 938                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 939
 940             /* Calculate displacement vector */
 941             dx00             = _mm_sub_pd(ix0,jx0);
 942             dy00             = _mm_sub_pd(iy0,jy0);
 943             dz00             = _mm_sub_pd(iz0,jz0);
 944             dx01             = _mm_sub_pd(ix0,jx1);
 945             dy01             = _mm_sub_pd(iy0,jy1);
 946             dz01             = _mm_sub_pd(iz0,jz1);
 947             dx02             = _mm_sub_pd(ix0,jx2);
 948             dy02             = _mm_sub_pd(iy0,jy2);
 949             dz02             = _mm_sub_pd(iz0,jz2);
 950             dx10             = _mm_sub_pd(ix1,jx0);
 951             dy10             = _mm_sub_pd(iy1,jy0);
 952             dz10             = _mm_sub_pd(iz1,jz0);
 953             dx11             = _mm_sub_pd(ix1,jx1);
 954             dy11             = _mm_sub_pd(iy1,jy1);
 955             dz11             = _mm_sub_pd(iz1,jz1);
 956             dx12             = _mm_sub_pd(ix1,jx2);
 957             dy12             = _mm_sub_pd(iy1,jy2);
 958             dz12             = _mm_sub_pd(iz1,jz2);
 959             dx20             = _mm_sub_pd(ix2,jx0);
 960             dy20             = _mm_sub_pd(iy2,jy0);
 961             dz20             = _mm_sub_pd(iz2,jz0);
 962             dx21             = _mm_sub_pd(ix2,jx1);
 963             dy21             = _mm_sub_pd(iy2,jy1);
 964             dz21             = _mm_sub_pd(iz2,jz1);
 965             dx22             = _mm_sub_pd(ix2,jx2);
 966             dy22             = _mm_sub_pd(iy2,jy2);
 967             dz22             = _mm_sub_pd(iz2,jz2);
 968
 969             /* Calculate squared distance and things based on it */
 970             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 971             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 972             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 973             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 974             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 975             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 976             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 977             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 978             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 979
 980             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 981             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 982             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 983             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 984             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 985             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 986             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 987             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 988             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 989
 990             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 991             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 992             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 993             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 994             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 995             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 996             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 997             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 998             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 999
1000             fjx0             = _mm_setzero_pd();
1001             fjy0             = _mm_setzero_pd();
1002             fjz0             = _mm_setzero_pd();
1003             fjx1             = _mm_setzero_pd();
1004             fjy1             = _mm_setzero_pd();
1005             fjz1             = _mm_setzero_pd();
1006             fjx2             = _mm_setzero_pd();
1007             fjy2             = _mm_setzero_pd();
1008             fjz2             = _mm_setzero_pd();
1009
1010             /**************************
1011              * CALCULATE INTERACTIONS *
1012              **************************/
1013
1014             /* COULOMB ELECTROSTATICS */
1015             velec            = _mm_mul_pd(qq00,rinv00);
1016             felec            = _mm_mul_pd(velec,rinvsq00);
1017
1018             fscal            = felec;
1019
1020             /* Update vectorial force */
1021             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1022             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1023             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1024
1025             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1026             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1027             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1028
1029             /**************************
1030              * CALCULATE INTERACTIONS *
1031              **************************/
1032
1033             /* COULOMB ELECTROSTATICS */
1034             velec            = _mm_mul_pd(qq01,rinv01);
1035             felec            = _mm_mul_pd(velec,rinvsq01);
1036
1037             fscal            = felec;
1038
1039             /* Update vectorial force */
1040             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1041             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1042             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1043
1044             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1045             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1046             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1047
1048             /**************************
1049              * CALCULATE INTERACTIONS *
1050              **************************/
1051
1052             /* COULOMB ELECTROSTATICS */
1053             velec            = _mm_mul_pd(qq02,rinv02);
1054             felec            = _mm_mul_pd(velec,rinvsq02);
1055
1056             fscal            = felec;
1057
1058             /* Update vectorial force */
1059             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1060             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1061             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1062
1063             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1064             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1065             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1066
1067             /**************************
1068              * CALCULATE INTERACTIONS *
1069              **************************/
1070
1071             /* COULOMB ELECTROSTATICS */
1072             velec            = _mm_mul_pd(qq10,rinv10);
1073             felec            = _mm_mul_pd(velec,rinvsq10);
1074
1075             fscal            = felec;
1076
1077             /* Update vectorial force */
1078             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1079             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1080             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1081
1082             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1083             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1084             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1085
1086             /**************************
1087              * CALCULATE INTERACTIONS *
1088              **************************/
1089
1090             /* COULOMB ELECTROSTATICS */
1091             velec            = _mm_mul_pd(qq11,rinv11);
1092             felec            = _mm_mul_pd(velec,rinvsq11);
1093
1094             fscal            = felec;
1095
1096             /* Update vectorial force */
1097             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1098             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1099             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1100
1101             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1102             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1103             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1104
1105             /**************************
1106              * CALCULATE INTERACTIONS *
1107              **************************/
1108
1109             /* COULOMB ELECTROSTATICS */
1110             velec            = _mm_mul_pd(qq12,rinv12);
1111             felec            = _mm_mul_pd(velec,rinvsq12);
1112
1113             fscal            = felec;
1114
1115             /* Update vectorial force */
1116             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1117             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1118             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1119
1120             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1121             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1122             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1123
1124             /**************************
1125              * CALCULATE INTERACTIONS *
1126              **************************/
1127
1128             /* COULOMB ELECTROSTATICS */
1129             velec            = _mm_mul_pd(qq20,rinv20);
1130             felec            = _mm_mul_pd(velec,rinvsq20);
1131
1132             fscal            = felec;
1133
1134             /* Update vectorial force */
1135             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1136             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1137             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1138
1139             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1140             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1141             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             /* COULOMB ELECTROSTATICS */
1148             velec            = _mm_mul_pd(qq21,rinv21);
1149             felec            = _mm_mul_pd(velec,rinvsq21);
1150
1151             fscal            = felec;
1152
1153             /* Update vectorial force */
1154             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1155             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1156             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1157
1158             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1159             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1160             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1161
1162             /**************************
1163              * CALCULATE INTERACTIONS *
1164              **************************/
1165
1166             /* COULOMB ELECTROSTATICS */
1167             velec            = _mm_mul_pd(qq22,rinv22);
1168             felec            = _mm_mul_pd(velec,rinvsq22);
1169
1170             fscal            = felec;
1171
1172             /* Update vectorial force */
1173             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1174             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1175             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1176
1177             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1178             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1179             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1180
1181             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1182
1183             /* Inner loop uses 270 flops */
1184         }
1185
1186         if(jidx<j_index_end)
1187         {
1188
1189             jnrA             = jjnr[jidx];
1190             j_coord_offsetA  = DIM*jnrA;
1191
1192             /* load j atom coordinates */
1193             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1194                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1195
1196             /* Calculate displacement vector */
1197             dx00             = _mm_sub_pd(ix0,jx0);
1198             dy00             = _mm_sub_pd(iy0,jy0);
1199             dz00             = _mm_sub_pd(iz0,jz0);
1200             dx01             = _mm_sub_pd(ix0,jx1);
1201             dy01             = _mm_sub_pd(iy0,jy1);
1202             dz01             = _mm_sub_pd(iz0,jz1);
1203             dx02             = _mm_sub_pd(ix0,jx2);
1204             dy02             = _mm_sub_pd(iy0,jy2);
1205             dz02             = _mm_sub_pd(iz0,jz2);
1206             dx10             = _mm_sub_pd(ix1,jx0);
1207             dy10             = _mm_sub_pd(iy1,jy0);
1208             dz10             = _mm_sub_pd(iz1,jz0);
1209             dx11             = _mm_sub_pd(ix1,jx1);
1210             dy11             = _mm_sub_pd(iy1,jy1);
1211             dz11             = _mm_sub_pd(iz1,jz1);
1212             dx12             = _mm_sub_pd(ix1,jx2);
1213             dy12             = _mm_sub_pd(iy1,jy2);
1214             dz12             = _mm_sub_pd(iz1,jz2);
1215             dx20             = _mm_sub_pd(ix2,jx0);
1216             dy20             = _mm_sub_pd(iy2,jy0);
1217             dz20             = _mm_sub_pd(iz2,jz0);
1218             dx21             = _mm_sub_pd(ix2,jx1);
1219             dy21             = _mm_sub_pd(iy2,jy1);
1220             dz21             = _mm_sub_pd(iz2,jz1);
1221             dx22             = _mm_sub_pd(ix2,jx2);
1222             dy22             = _mm_sub_pd(iy2,jy2);
1223             dz22             = _mm_sub_pd(iz2,jz2);
1224
1225             /* Calculate squared distance and things based on it */
1226             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1227             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1228             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1229             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1230             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1231             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1232             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1233             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1234             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1235
1236             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1237             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1238             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1239             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1240             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1241             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1242             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1243             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1244             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1245
1246             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1247             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1248             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1249             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1250             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1251             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1252             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1253             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1254             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1255
1256             fjx0             = _mm_setzero_pd();
1257             fjy0             = _mm_setzero_pd();
1258             fjz0             = _mm_setzero_pd();
1259             fjx1             = _mm_setzero_pd();
1260             fjy1             = _mm_setzero_pd();
1261             fjz1             = _mm_setzero_pd();
1262             fjx2             = _mm_setzero_pd();
1263             fjy2             = _mm_setzero_pd();
1264             fjz2             = _mm_setzero_pd();
1265
1266             /**************************
1267              * CALCULATE INTERACTIONS *
1268              **************************/
1269
1270             /* COULOMB ELECTROSTATICS */
1271             velec            = _mm_mul_pd(qq00,rinv00);
1272             felec            = _mm_mul_pd(velec,rinvsq00);
1273
1274             fscal            = felec;
1275
1276             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1277
1278             /* Update vectorial force */
1279             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1280             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1281             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1282
1283             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1284             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1285             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             /* COULOMB ELECTROSTATICS */
1292             velec            = _mm_mul_pd(qq01,rinv01);
1293             felec            = _mm_mul_pd(velec,rinvsq01);
1294
1295             fscal            = felec;
1296
1297             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1298
1299             /* Update vectorial force */
1300             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1301             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1302             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1303
1304             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1305             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1306             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1307
1308             /**************************
1309              * CALCULATE INTERACTIONS *
1310              **************************/
1311
1312             /* COULOMB ELECTROSTATICS */
1313             velec            = _mm_mul_pd(qq02,rinv02);
1314             felec            = _mm_mul_pd(velec,rinvsq02);
1315
1316             fscal            = felec;
1317
1318             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1319
1320             /* Update vectorial force */
1321             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1322             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1323             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1324
1325             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1326             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1327             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1328
1329             /**************************
1330              * CALCULATE INTERACTIONS *
1331              **************************/
1332
1333             /* COULOMB ELECTROSTATICS */
1334             velec            = _mm_mul_pd(qq10,rinv10);
1335             felec            = _mm_mul_pd(velec,rinvsq10);
1336
1337             fscal            = felec;
1338
1339             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1340
1341             /* Update vectorial force */
1342             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1343             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1344             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1345
1346             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1347             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1348             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1349
1350             /**************************
1351              * CALCULATE INTERACTIONS *
1352              **************************/
1353
1354             /* COULOMB ELECTROSTATICS */
1355             velec            = _mm_mul_pd(qq11,rinv11);
1356             felec            = _mm_mul_pd(velec,rinvsq11);
1357
1358             fscal            = felec;
1359
1360             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1361
1362             /* Update vectorial force */
1363             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1364             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1365             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1366
1367             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1368             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1369             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1370
1371             /**************************
1372              * CALCULATE INTERACTIONS *
1373              **************************/
1374
1375             /* COULOMB ELECTROSTATICS */
1376             velec            = _mm_mul_pd(qq12,rinv12);
1377             felec            = _mm_mul_pd(velec,rinvsq12);
1378
1379             fscal            = felec;
1380
1381             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1382
1383             /* Update vectorial force */
1384             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1385             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1386             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1387
1388             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1389             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1390             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1391
1392             /**************************
1393              * CALCULATE INTERACTIONS *
1394              **************************/
1395
1396             /* COULOMB ELECTROSTATICS */
1397             velec            = _mm_mul_pd(qq20,rinv20);
1398             felec            = _mm_mul_pd(velec,rinvsq20);
1399
1400             fscal            = felec;
1401
1402             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1403
1404             /* Update vectorial force */
1405             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1406             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1407             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1408
1409             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1410             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1411             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1412
1413             /**************************
1414              * CALCULATE INTERACTIONS *
1415              **************************/
1416
1417             /* COULOMB ELECTROSTATICS */
1418             velec            = _mm_mul_pd(qq21,rinv21);
1419             felec            = _mm_mul_pd(velec,rinvsq21);
1420
1421             fscal            = felec;
1422
1423             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1424
1425             /* Update vectorial force */
1426             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1427             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1428             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1429
1430             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1431             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1432             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1433
1434             /**************************
1435              * CALCULATE INTERACTIONS *
1436              **************************/
1437
1438             /* COULOMB ELECTROSTATICS */
1439             velec            = _mm_mul_pd(qq22,rinv22);
1440             felec            = _mm_mul_pd(velec,rinvsq22);
1441
1442             fscal            = felec;
1443
1444             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1445
1446             /* Update vectorial force */
1447             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1448             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1449             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1450
1451             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1452             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1453             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1454
1455             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1456
1457             /* Inner loop uses 270 flops */
1458         }
1459
1460         /* End of innermost loop */
1461
1462         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1463                                               f+i_coord_offset,fshift+i_shift_offset);
1464
1465         /* Increment number of inner iterations */
1466         inneriter                  += j_index_end - j_index_start;
1467
1468         /* Outer loop uses 18 flops */
1469     }
1470
1471     /* Increment number of outer iterations */
1472     outeriter        += nri;
1473
1474     /* Update outer/inner flops */
1475
1476     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
1477 }