src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwNone_GeomW4W4_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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            None
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset1;
  81     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  82     int              vdwioffset2;
  83     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  84     int              vdwioffset3;
  85     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  86     int              vdwjidx1A,vdwjidx1B;
  87     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  88     int              vdwjidx2A,vdwjidx2B;
  89     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  90     int              vdwjidx3A,vdwjidx3B;
  91     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  92     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  93     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  94     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  95     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  96     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  97     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  98     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  99     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 100     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     __m128d          dummy_mask,cutoff_mask;
 104     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 105     __m128d          one     = _mm_set1_pd(1.0);
 106     __m128d          two     = _mm_set1_pd(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_pd(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120
 121     /* Setup water-specific parameters */
 122     inr              = nlist->iinr[0];
 123     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 124     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 125     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 126
 127     jq1              = _mm_set1_pd(charge[inr+1]);
 128     jq2              = _mm_set1_pd(charge[inr+2]);
 129     jq3              = _mm_set1_pd(charge[inr+3]);
 130     qq11             = _mm_mul_pd(iq1,jq1);
 131     qq12             = _mm_mul_pd(iq1,jq2);
 132     qq13             = _mm_mul_pd(iq1,jq3);
 133     qq21             = _mm_mul_pd(iq2,jq1);
 134     qq22             = _mm_mul_pd(iq2,jq2);
 135     qq23             = _mm_mul_pd(iq2,jq3);
 136     qq31             = _mm_mul_pd(iq3,jq1);
 137     qq32             = _mm_mul_pd(iq3,jq2);
 138     qq33             = _mm_mul_pd(iq3,jq3);
 139
 140     /* Avoid stupid compiler warnings */
 141     jnrA = jnrB = 0;
 142     j_coord_offsetA = 0;
 143     j_coord_offsetB = 0;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 164                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 165
 166         fix1             = _mm_setzero_pd();
 167         fiy1             = _mm_setzero_pd();
 168         fiz1             = _mm_setzero_pd();
 169         fix2             = _mm_setzero_pd();
 170         fiy2             = _mm_setzero_pd();
 171         fiz2             = _mm_setzero_pd();
 172         fix3             = _mm_setzero_pd();
 173         fiy3             = _mm_setzero_pd();
 174         fiz3             = _mm_setzero_pd();
 175
 176         /* Reset potential sums */
 177         velecsum         = _mm_setzero_pd();
 178
 179         /* Start inner kernel loop */
 180         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 181         {
 182
 183             /* Get j neighbor index, and coordinate index */
 184             jnrA             = jjnr[jidx];
 185             jnrB             = jjnr[jidx+1];
 186             j_coord_offsetA  = DIM*jnrA;
 187             j_coord_offsetB  = DIM*jnrB;
 188
 189             /* load j atom coordinates */
 190             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 191                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 192
 193             /* Calculate displacement vector */
 194             dx11             = _mm_sub_pd(ix1,jx1);
 195             dy11             = _mm_sub_pd(iy1,jy1);
 196             dz11             = _mm_sub_pd(iz1,jz1);
 197             dx12             = _mm_sub_pd(ix1,jx2);
 198             dy12             = _mm_sub_pd(iy1,jy2);
 199             dz12             = _mm_sub_pd(iz1,jz2);
 200             dx13             = _mm_sub_pd(ix1,jx3);
 201             dy13             = _mm_sub_pd(iy1,jy3);
 202             dz13             = _mm_sub_pd(iz1,jz3);
 203             dx21             = _mm_sub_pd(ix2,jx1);
 204             dy21             = _mm_sub_pd(iy2,jy1);
 205             dz21             = _mm_sub_pd(iz2,jz1);
 206             dx22             = _mm_sub_pd(ix2,jx2);
 207             dy22             = _mm_sub_pd(iy2,jy2);
 208             dz22             = _mm_sub_pd(iz2,jz2);
 209             dx23             = _mm_sub_pd(ix2,jx3);
 210             dy23             = _mm_sub_pd(iy2,jy3);
 211             dz23             = _mm_sub_pd(iz2,jz3);
 212             dx31             = _mm_sub_pd(ix3,jx1);
 213             dy31             = _mm_sub_pd(iy3,jy1);
 214             dz31             = _mm_sub_pd(iz3,jz1);
 215             dx32             = _mm_sub_pd(ix3,jx2);
 216             dy32             = _mm_sub_pd(iy3,jy2);
 217             dz32             = _mm_sub_pd(iz3,jz2);
 218             dx33             = _mm_sub_pd(ix3,jx3);
 219             dy33             = _mm_sub_pd(iy3,jy3);
 220             dz33             = _mm_sub_pd(iz3,jz3);
 221
 222             /* Calculate squared distance and things based on it */
 223             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 224             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 225             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 226             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 227             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 228             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 229             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 230             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 231             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 232
 233             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 234             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 235             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 236             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 237             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 238             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 239             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 240             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 241             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 242
 243             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 244             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 245             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 246             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 247             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 248             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 249             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 250             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 251             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 252
 253             fjx1             = _mm_setzero_pd();
 254             fjy1             = _mm_setzero_pd();
 255             fjz1             = _mm_setzero_pd();
 256             fjx2             = _mm_setzero_pd();
 257             fjy2             = _mm_setzero_pd();
 258             fjz2             = _mm_setzero_pd();
 259             fjx3             = _mm_setzero_pd();
 260             fjy3             = _mm_setzero_pd();
 261             fjz3             = _mm_setzero_pd();
 262
 263             /**************************
 264              * CALCULATE INTERACTIONS *
 265              **************************/
 266
 267             /* COULOMB ELECTROSTATICS */
 268             velec            = _mm_mul_pd(qq11,rinv11);
 269             felec            = _mm_mul_pd(velec,rinvsq11);
 270
 271             /* Update potential sum for this i atom from the interaction with this j atom. */
 272             velecsum         = _mm_add_pd(velecsum,velec);
 273
 274             fscal            = felec;
 275
 276             /* Update vectorial force */
 277             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 278             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 279             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 280
 281             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 282             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 283             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 284
 285             /**************************
 286              * CALCULATE INTERACTIONS *
 287              **************************/
 288
 289             /* COULOMB ELECTROSTATICS */
 290             velec            = _mm_mul_pd(qq12,rinv12);
 291             felec            = _mm_mul_pd(velec,rinvsq12);
 292
 293             /* Update potential sum for this i atom from the interaction with this j atom. */
 294             velecsum         = _mm_add_pd(velecsum,velec);
 295
 296             fscal            = felec;
 297
 298             /* Update vectorial force */
 299             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 300             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 301             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 302
 303             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 304             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 305             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 306
 307             /**************************
 308              * CALCULATE INTERACTIONS *
 309              **************************/
 310
 311             /* COULOMB ELECTROSTATICS */
 312             velec            = _mm_mul_pd(qq13,rinv13);
 313             felec            = _mm_mul_pd(velec,rinvsq13);
 314
 315             /* Update potential sum for this i atom from the interaction with this j atom. */
 316             velecsum         = _mm_add_pd(velecsum,velec);
 317
 318             fscal            = felec;
 319
 320             /* Update vectorial force */
 321             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 322             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 323             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 324
 325             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 326             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 327             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             /* COULOMB ELECTROSTATICS */
 334             velec            = _mm_mul_pd(qq21,rinv21);
 335             felec            = _mm_mul_pd(velec,rinvsq21);
 336
 337             /* Update potential sum for this i atom from the interaction with this j atom. */
 338             velecsum         = _mm_add_pd(velecsum,velec);
 339
 340             fscal            = felec;
 341
 342             /* Update vectorial force */
 343             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 344             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 345             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 346
 347             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 348             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 349             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             /* COULOMB ELECTROSTATICS */
 356             velec            = _mm_mul_pd(qq22,rinv22);
 357             felec            = _mm_mul_pd(velec,rinvsq22);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velecsum         = _mm_add_pd(velecsum,velec);
 361
 362             fscal            = felec;
 363
 364             /* Update vectorial force */
 365             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 366             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 367             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 368
 369             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 370             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 371             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 372
 373             /**************************
 374              * CALCULATE INTERACTIONS *
 375              **************************/
 376
 377             /* COULOMB ELECTROSTATICS */
 378             velec            = _mm_mul_pd(qq23,rinv23);
 379             felec            = _mm_mul_pd(velec,rinvsq23);
 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             /* Update vectorial force */
 387             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 388             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 389             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 390
 391             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 392             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 393             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             /* COULOMB ELECTROSTATICS */
 400             velec            = _mm_mul_pd(qq31,rinv31);
 401             felec            = _mm_mul_pd(velec,rinvsq31);
 402
 403             /* Update potential sum for this i atom from the interaction with this j atom. */
 404             velecsum         = _mm_add_pd(velecsum,velec);
 405
 406             fscal            = felec;
 407
 408             /* Update vectorial force */
 409             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 410             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 411             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 412
 413             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 414             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 415             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 416
 417             /**************************
 418              * CALCULATE INTERACTIONS *
 419              **************************/
 420
 421             /* COULOMB ELECTROSTATICS */
 422             velec            = _mm_mul_pd(qq32,rinv32);
 423             felec            = _mm_mul_pd(velec,rinvsq32);
 424
 425             /* Update potential sum for this i atom from the interaction with this j atom. */
 426             velecsum         = _mm_add_pd(velecsum,velec);
 427
 428             fscal            = felec;
 429
 430             /* Update vectorial force */
 431             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 432             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 433             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 434
 435             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 436             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 437             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 438
 439             /**************************
 440              * CALCULATE INTERACTIONS *
 441              **************************/
 442
 443             /* COULOMB ELECTROSTATICS */
 444             velec            = _mm_mul_pd(qq33,rinv33);
 445             felec            = _mm_mul_pd(velec,rinvsq33);
 446
 447             /* Update potential sum for this i atom from the interaction with this j atom. */
 448             velecsum         = _mm_add_pd(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             /* Update vectorial force */
 453             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 454             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 455             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 456
 457             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 458             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 459             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 460
 461             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);
 462
 463             /* Inner loop uses 279 flops */
 464         }
 465
 466         if(jidx<j_index_end)
 467         {
 468
 469             jnrA             = jjnr[jidx];
 470             j_coord_offsetA  = DIM*jnrA;
 471
 472             /* load j atom coordinates */
 473             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 474                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 475
 476             /* Calculate displacement vector */
 477             dx11             = _mm_sub_pd(ix1,jx1);
 478             dy11             = _mm_sub_pd(iy1,jy1);
 479             dz11             = _mm_sub_pd(iz1,jz1);
 480             dx12             = _mm_sub_pd(ix1,jx2);
 481             dy12             = _mm_sub_pd(iy1,jy2);
 482             dz12             = _mm_sub_pd(iz1,jz2);
 483             dx13             = _mm_sub_pd(ix1,jx3);
 484             dy13             = _mm_sub_pd(iy1,jy3);
 485             dz13             = _mm_sub_pd(iz1,jz3);
 486             dx21             = _mm_sub_pd(ix2,jx1);
 487             dy21             = _mm_sub_pd(iy2,jy1);
 488             dz21             = _mm_sub_pd(iz2,jz1);
 489             dx22             = _mm_sub_pd(ix2,jx2);
 490             dy22             = _mm_sub_pd(iy2,jy2);
 491             dz22             = _mm_sub_pd(iz2,jz2);
 492             dx23             = _mm_sub_pd(ix2,jx3);
 493             dy23             = _mm_sub_pd(iy2,jy3);
 494             dz23             = _mm_sub_pd(iz2,jz3);
 495             dx31             = _mm_sub_pd(ix3,jx1);
 496             dy31             = _mm_sub_pd(iy3,jy1);
 497             dz31             = _mm_sub_pd(iz3,jz1);
 498             dx32             = _mm_sub_pd(ix3,jx2);
 499             dy32             = _mm_sub_pd(iy3,jy2);
 500             dz32             = _mm_sub_pd(iz3,jz2);
 501             dx33             = _mm_sub_pd(ix3,jx3);
 502             dy33             = _mm_sub_pd(iy3,jy3);
 503             dz33             = _mm_sub_pd(iz3,jz3);
 504
 505             /* Calculate squared distance and things based on it */
 506             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 507             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 508             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 509             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 510             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 511             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 512             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 513             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 514             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 515
 516             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 517             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 518             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 519             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 520             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 521             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 522             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 523             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 524             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 525
 526             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 527             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 528             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 529             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 530             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 531             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 532             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 533             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 534             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 535
 536             fjx1             = _mm_setzero_pd();
 537             fjy1             = _mm_setzero_pd();
 538             fjz1             = _mm_setzero_pd();
 539             fjx2             = _mm_setzero_pd();
 540             fjy2             = _mm_setzero_pd();
 541             fjz2             = _mm_setzero_pd();
 542             fjx3             = _mm_setzero_pd();
 543             fjy3             = _mm_setzero_pd();
 544             fjz3             = _mm_setzero_pd();
 545
 546             /**************************
 547              * CALCULATE INTERACTIONS *
 548              **************************/
 549
 550             /* COULOMB ELECTROSTATICS */
 551             velec            = _mm_mul_pd(qq11,rinv11);
 552             felec            = _mm_mul_pd(velec,rinvsq11);
 553
 554             /* Update potential sum for this i atom from the interaction with this j atom. */
 555             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 556             velecsum         = _mm_add_pd(velecsum,velec);
 557
 558             fscal            = felec;
 559
 560             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 561
 562             /* Update vectorial force */
 563             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 564             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 565             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 566
 567             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 568             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 569             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 570
 571             /**************************
 572              * CALCULATE INTERACTIONS *
 573              **************************/
 574
 575             /* COULOMB ELECTROSTATICS */
 576             velec            = _mm_mul_pd(qq12,rinv12);
 577             felec            = _mm_mul_pd(velec,rinvsq12);
 578
 579             /* Update potential sum for this i atom from the interaction with this j atom. */
 580             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 581             velecsum         = _mm_add_pd(velecsum,velec);
 582
 583             fscal            = felec;
 584
 585             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 586
 587             /* Update vectorial force */
 588             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 589             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 590             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 591
 592             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 593             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 594             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 595
 596             /**************************
 597              * CALCULATE INTERACTIONS *
 598              **************************/
 599
 600             /* COULOMB ELECTROSTATICS */
 601             velec            = _mm_mul_pd(qq13,rinv13);
 602             felec            = _mm_mul_pd(velec,rinvsq13);
 603
 604             /* Update potential sum for this i atom from the interaction with this j atom. */
 605             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 606             velecsum         = _mm_add_pd(velecsum,velec);
 607
 608             fscal            = felec;
 609
 610             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 611
 612             /* Update vectorial force */
 613             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 614             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 615             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 616
 617             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 618             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 619             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 620
 621             /**************************
 622              * CALCULATE INTERACTIONS *
 623              **************************/
 624
 625             /* COULOMB ELECTROSTATICS */
 626             velec            = _mm_mul_pd(qq21,rinv21);
 627             felec            = _mm_mul_pd(velec,rinvsq21);
 628
 629             /* Update potential sum for this i atom from the interaction with this j atom. */
 630             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 631             velecsum         = _mm_add_pd(velecsum,velec);
 632
 633             fscal            = felec;
 634
 635             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 636
 637             /* Update vectorial force */
 638             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 639             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 640             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 641
 642             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 643             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 644             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 645
 646             /**************************
 647              * CALCULATE INTERACTIONS *
 648              **************************/
 649
 650             /* COULOMB ELECTROSTATICS */
 651             velec            = _mm_mul_pd(qq22,rinv22);
 652             felec            = _mm_mul_pd(velec,rinvsq22);
 653
 654             /* Update potential sum for this i atom from the interaction with this j atom. */
 655             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 656             velecsum         = _mm_add_pd(velecsum,velec);
 657
 658             fscal            = felec;
 659
 660             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 661
 662             /* Update vectorial force */
 663             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 664             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 665             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 666
 667             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 668             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 669             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             /* COULOMB ELECTROSTATICS */
 676             velec            = _mm_mul_pd(qq23,rinv23);
 677             felec            = _mm_mul_pd(velec,rinvsq23);
 678
 679             /* Update potential sum for this i atom from the interaction with this j atom. */
 680             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 681             velecsum         = _mm_add_pd(velecsum,velec);
 682
 683             fscal            = felec;
 684
 685             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 686
 687             /* Update vectorial force */
 688             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 689             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 690             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 691
 692             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 693             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 694             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 695
 696             /**************************
 697              * CALCULATE INTERACTIONS *
 698              **************************/
 699
 700             /* COULOMB ELECTROSTATICS */
 701             velec            = _mm_mul_pd(qq31,rinv31);
 702             felec            = _mm_mul_pd(velec,rinvsq31);
 703
 704             /* Update potential sum for this i atom from the interaction with this j atom. */
 705             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 706             velecsum         = _mm_add_pd(velecsum,velec);
 707
 708             fscal            = felec;
 709
 710             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 711
 712             /* Update vectorial force */
 713             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 714             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 715             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 716
 717             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 718             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 719             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             /* COULOMB ELECTROSTATICS */
 726             velec            = _mm_mul_pd(qq32,rinv32);
 727             felec            = _mm_mul_pd(velec,rinvsq32);
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 731             velecsum         = _mm_add_pd(velecsum,velec);
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 736
 737             /* Update vectorial force */
 738             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 739             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 740             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 741
 742             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 743             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 744             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 745
 746             /**************************
 747              * CALCULATE INTERACTIONS *
 748              **************************/
 749
 750             /* COULOMB ELECTROSTATICS */
 751             velec            = _mm_mul_pd(qq33,rinv33);
 752             felec            = _mm_mul_pd(velec,rinvsq33);
 753
 754             /* Update potential sum for this i atom from the interaction with this j atom. */
 755             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 756             velecsum         = _mm_add_pd(velecsum,velec);
 757
 758             fscal            = felec;
 759
 760             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 761
 762             /* Update vectorial force */
 763             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 764             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 765             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 766
 767             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 768             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 769             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 770
 771             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 772
 773             /* Inner loop uses 279 flops */
 774         }
 775
 776         /* End of innermost loop */
 777
 778         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 779                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 780
 781         ggid                        = gid[iidx];
 782         /* Update potential energies */
 783         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 784
 785         /* Increment number of inner iterations */
 786         inneriter                  += j_index_end - j_index_start;
 787
 788         /* Outer loop uses 19 flops */
 789     }
 790
 791     /* Increment number of outer iterations */
 792     outeriter        += nri;
 793
 794     /* Update outer/inner flops */
 795
 796     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*279);
 797 }
 798 /*
 799  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_double
 800  * Electrostatics interaction: Coulomb
 801  * VdW interaction:            None
 802  * Geometry:                   Water4-Water4
 803  * Calculate force/pot:        Force
 804  */
 805 void
 806 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_double
 807                     (t_nblist                    * gmx_restrict       nlist,
 808                      rvec                        * gmx_restrict          xx,
 809                      rvec                        * gmx_restrict          ff,
 810                      t_forcerec                  * gmx_restrict          fr,
 811                      t_mdatoms                   * gmx_restrict     mdatoms,
 812                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 813                      t_nrnb                      * gmx_restrict        nrnb)
 814 {
 815     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 816      * just 0 for non-waters.
 817      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 818      * jnr indices corresponding to data put in the four positions in the SIMD register.
 819      */
 820     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 821     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 822     int              jnrA,jnrB;
 823     int              j_coord_offsetA,j_coord_offsetB;
 824     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 825     real             rcutoff_scalar;
 826     real             *shiftvec,*fshift,*x,*f;
 827     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 828     int              vdwioffset1;
 829     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 830     int              vdwioffset2;
 831     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 832     int              vdwioffset3;
 833     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 834     int              vdwjidx1A,vdwjidx1B;
 835     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 836     int              vdwjidx2A,vdwjidx2B;
 837     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 838     int              vdwjidx3A,vdwjidx3B;
 839     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 840     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 841     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 842     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 843     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 844     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 845     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 846     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 847     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 848     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 849     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 850     real             *charge;
 851     __m128d          dummy_mask,cutoff_mask;
 852     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 853     __m128d          one     = _mm_set1_pd(1.0);
 854     __m128d          two     = _mm_set1_pd(2.0);
 855     x                = xx[0];
 856     f                = ff[0];
 857
 858     nri              = nlist->nri;
 859     iinr             = nlist->iinr;
 860     jindex           = nlist->jindex;
 861     jjnr             = nlist->jjnr;
 862     shiftidx         = nlist->shift;
 863     gid              = nlist->gid;
 864     shiftvec         = fr->shift_vec[0];
 865     fshift           = fr->fshift[0];
 866     facel            = _mm_set1_pd(fr->epsfac);
 867     charge           = mdatoms->chargeA;
 868
 869     /* Setup water-specific parameters */
 870     inr              = nlist->iinr[0];
 871     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 872     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 873     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 874
 875     jq1              = _mm_set1_pd(charge[inr+1]);
 876     jq2              = _mm_set1_pd(charge[inr+2]);
 877     jq3              = _mm_set1_pd(charge[inr+3]);
 878     qq11             = _mm_mul_pd(iq1,jq1);
 879     qq12             = _mm_mul_pd(iq1,jq2);
 880     qq13             = _mm_mul_pd(iq1,jq3);
 881     qq21             = _mm_mul_pd(iq2,jq1);
 882     qq22             = _mm_mul_pd(iq2,jq2);
 883     qq23             = _mm_mul_pd(iq2,jq3);
 884     qq31             = _mm_mul_pd(iq3,jq1);
 885     qq32             = _mm_mul_pd(iq3,jq2);
 886     qq33             = _mm_mul_pd(iq3,jq3);
 887
 888     /* Avoid stupid compiler warnings */
 889     jnrA = jnrB = 0;
 890     j_coord_offsetA = 0;
 891     j_coord_offsetB = 0;
 892
 893     outeriter        = 0;
 894     inneriter        = 0;
 895
 896     /* Start outer loop over neighborlists */
 897     for(iidx=0; iidx<nri; iidx++)
 898     {
 899         /* Load shift vector for this list */
 900         i_shift_offset   = DIM*shiftidx[iidx];
 901
 902         /* Load limits for loop over neighbors */
 903         j_index_start    = jindex[iidx];
 904         j_index_end      = jindex[iidx+1];
 905
 906         /* Get outer coordinate index */
 907         inr              = iinr[iidx];
 908         i_coord_offset   = DIM*inr;
 909
 910         /* Load i particle coords and add shift vector */
 911         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 912                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 913
 914         fix1             = _mm_setzero_pd();
 915         fiy1             = _mm_setzero_pd();
 916         fiz1             = _mm_setzero_pd();
 917         fix2             = _mm_setzero_pd();
 918         fiy2             = _mm_setzero_pd();
 919         fiz2             = _mm_setzero_pd();
 920         fix3             = _mm_setzero_pd();
 921         fiy3             = _mm_setzero_pd();
 922         fiz3             = _mm_setzero_pd();
 923
 924         /* Start inner kernel loop */
 925         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 926         {
 927
 928             /* Get j neighbor index, and coordinate index */
 929             jnrA             = jjnr[jidx];
 930             jnrB             = jjnr[jidx+1];
 931             j_coord_offsetA  = DIM*jnrA;
 932             j_coord_offsetB  = DIM*jnrB;
 933
 934             /* load j atom coordinates */
 935             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 936                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 937
 938             /* Calculate displacement vector */
 939             dx11             = _mm_sub_pd(ix1,jx1);
 940             dy11             = _mm_sub_pd(iy1,jy1);
 941             dz11             = _mm_sub_pd(iz1,jz1);
 942             dx12             = _mm_sub_pd(ix1,jx2);
 943             dy12             = _mm_sub_pd(iy1,jy2);
 944             dz12             = _mm_sub_pd(iz1,jz2);
 945             dx13             = _mm_sub_pd(ix1,jx3);
 946             dy13             = _mm_sub_pd(iy1,jy3);
 947             dz13             = _mm_sub_pd(iz1,jz3);
 948             dx21             = _mm_sub_pd(ix2,jx1);
 949             dy21             = _mm_sub_pd(iy2,jy1);
 950             dz21             = _mm_sub_pd(iz2,jz1);
 951             dx22             = _mm_sub_pd(ix2,jx2);
 952             dy22             = _mm_sub_pd(iy2,jy2);
 953             dz22             = _mm_sub_pd(iz2,jz2);
 954             dx23             = _mm_sub_pd(ix2,jx3);
 955             dy23             = _mm_sub_pd(iy2,jy3);
 956             dz23             = _mm_sub_pd(iz2,jz3);
 957             dx31             = _mm_sub_pd(ix3,jx1);
 958             dy31             = _mm_sub_pd(iy3,jy1);
 959             dz31             = _mm_sub_pd(iz3,jz1);
 960             dx32             = _mm_sub_pd(ix3,jx2);
 961             dy32             = _mm_sub_pd(iy3,jy2);
 962             dz32             = _mm_sub_pd(iz3,jz2);
 963             dx33             = _mm_sub_pd(ix3,jx3);
 964             dy33             = _mm_sub_pd(iy3,jy3);
 965             dz33             = _mm_sub_pd(iz3,jz3);
 966
 967             /* Calculate squared distance and things based on it */
 968             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 969             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 970             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 971             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 972             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 973             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 974             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 975             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 976             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 977
 978             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 979             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 980             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 981             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 982             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 983             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 984             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 985             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 986             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 987
 988             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 989             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 990             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 991             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 992             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 993             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 994             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 995             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 996             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 997
 998             fjx1             = _mm_setzero_pd();
 999             fjy1             = _mm_setzero_pd();
1000             fjz1             = _mm_setzero_pd();
1001             fjx2             = _mm_setzero_pd();
1002             fjy2             = _mm_setzero_pd();
1003             fjz2             = _mm_setzero_pd();
1004             fjx3             = _mm_setzero_pd();
1005             fjy3             = _mm_setzero_pd();
1006             fjz3             = _mm_setzero_pd();
1007
1008             /**************************
1009              * CALCULATE INTERACTIONS *
1010              **************************/
1011
1012             /* COULOMB ELECTROSTATICS */
1013             velec            = _mm_mul_pd(qq11,rinv11);
1014             felec            = _mm_mul_pd(velec,rinvsq11);
1015
1016             fscal            = felec;
1017
1018             /* Update vectorial force */
1019             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1020             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1021             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1022
1023             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1024             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1025             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1026
1027             /**************************
1028              * CALCULATE INTERACTIONS *
1029              **************************/
1030
1031             /* COULOMB ELECTROSTATICS */
1032             velec            = _mm_mul_pd(qq12,rinv12);
1033             felec            = _mm_mul_pd(velec,rinvsq12);
1034
1035             fscal            = felec;
1036
1037             /* Update vectorial force */
1038             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1039             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1040             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1041
1042             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1043             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1044             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1045
1046             /**************************
1047              * CALCULATE INTERACTIONS *
1048              **************************/
1049
1050             /* COULOMB ELECTROSTATICS */
1051             velec            = _mm_mul_pd(qq13,rinv13);
1052             felec            = _mm_mul_pd(velec,rinvsq13);
1053
1054             fscal            = felec;
1055
1056             /* Update vectorial force */
1057             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1058             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1059             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1060
1061             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1062             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1063             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1064
1065             /**************************
1066              * CALCULATE INTERACTIONS *
1067              **************************/
1068
1069             /* COULOMB ELECTROSTATICS */
1070             velec            = _mm_mul_pd(qq21,rinv21);
1071             felec            = _mm_mul_pd(velec,rinvsq21);
1072
1073             fscal            = felec;
1074
1075             /* Update vectorial force */
1076             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1077             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1078             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1079
1080             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1081             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1082             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1083
1084             /**************************
1085              * CALCULATE INTERACTIONS *
1086              **************************/
1087
1088             /* COULOMB ELECTROSTATICS */
1089             velec            = _mm_mul_pd(qq22,rinv22);
1090             felec            = _mm_mul_pd(velec,rinvsq22);
1091
1092             fscal            = felec;
1093
1094             /* Update vectorial force */
1095             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1096             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1097             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1098
1099             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1100             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1101             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1102
1103             /**************************
1104              * CALCULATE INTERACTIONS *
1105              **************************/
1106
1107             /* COULOMB ELECTROSTATICS */
1108             velec            = _mm_mul_pd(qq23,rinv23);
1109             felec            = _mm_mul_pd(velec,rinvsq23);
1110
1111             fscal            = felec;
1112
1113             /* Update vectorial force */
1114             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1115             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1116             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1117
1118             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1119             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1120             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1121
1122             /**************************
1123              * CALCULATE INTERACTIONS *
1124              **************************/
1125
1126             /* COULOMB ELECTROSTATICS */
1127             velec            = _mm_mul_pd(qq31,rinv31);
1128             felec            = _mm_mul_pd(velec,rinvsq31);
1129
1130             fscal            = felec;
1131
1132             /* Update vectorial force */
1133             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1134             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1135             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1136
1137             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1138             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1139             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1140
1141             /**************************
1142              * CALCULATE INTERACTIONS *
1143              **************************/
1144
1145             /* COULOMB ELECTROSTATICS */
1146             velec            = _mm_mul_pd(qq32,rinv32);
1147             felec            = _mm_mul_pd(velec,rinvsq32);
1148
1149             fscal            = felec;
1150
1151             /* Update vectorial force */
1152             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1153             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1154             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1155
1156             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1157             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1158             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1159
1160             /**************************
1161              * CALCULATE INTERACTIONS *
1162              **************************/
1163
1164             /* COULOMB ELECTROSTATICS */
1165             velec            = _mm_mul_pd(qq33,rinv33);
1166             felec            = _mm_mul_pd(velec,rinvsq33);
1167
1168             fscal            = felec;
1169
1170             /* Update vectorial force */
1171             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1172             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1173             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1174
1175             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1176             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1177             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1178
1179             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);
1180
1181             /* Inner loop uses 270 flops */
1182         }
1183
1184         if(jidx<j_index_end)
1185         {
1186
1187             jnrA             = jjnr[jidx];
1188             j_coord_offsetA  = DIM*jnrA;
1189
1190             /* load j atom coordinates */
1191             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1192                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1193
1194             /* Calculate displacement vector */
1195             dx11             = _mm_sub_pd(ix1,jx1);
1196             dy11             = _mm_sub_pd(iy1,jy1);
1197             dz11             = _mm_sub_pd(iz1,jz1);
1198             dx12             = _mm_sub_pd(ix1,jx2);
1199             dy12             = _mm_sub_pd(iy1,jy2);
1200             dz12             = _mm_sub_pd(iz1,jz2);
1201             dx13             = _mm_sub_pd(ix1,jx3);
1202             dy13             = _mm_sub_pd(iy1,jy3);
1203             dz13             = _mm_sub_pd(iz1,jz3);
1204             dx21             = _mm_sub_pd(ix2,jx1);
1205             dy21             = _mm_sub_pd(iy2,jy1);
1206             dz21             = _mm_sub_pd(iz2,jz1);
1207             dx22             = _mm_sub_pd(ix2,jx2);
1208             dy22             = _mm_sub_pd(iy2,jy2);
1209             dz22             = _mm_sub_pd(iz2,jz2);
1210             dx23             = _mm_sub_pd(ix2,jx3);
1211             dy23             = _mm_sub_pd(iy2,jy3);
1212             dz23             = _mm_sub_pd(iz2,jz3);
1213             dx31             = _mm_sub_pd(ix3,jx1);
1214             dy31             = _mm_sub_pd(iy3,jy1);
1215             dz31             = _mm_sub_pd(iz3,jz1);
1216             dx32             = _mm_sub_pd(ix3,jx2);
1217             dy32             = _mm_sub_pd(iy3,jy2);
1218             dz32             = _mm_sub_pd(iz3,jz2);
1219             dx33             = _mm_sub_pd(ix3,jx3);
1220             dy33             = _mm_sub_pd(iy3,jy3);
1221             dz33             = _mm_sub_pd(iz3,jz3);
1222
1223             /* Calculate squared distance and things based on it */
1224             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1225             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1226             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1227             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1228             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1229             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1230             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1231             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1232             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1233
1234             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1235             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1236             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1237             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1238             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1239             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1240             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1241             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1242             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1243
1244             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1245             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1246             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1247             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1248             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1249             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1250             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1251             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1252             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1253
1254             fjx1             = _mm_setzero_pd();
1255             fjy1             = _mm_setzero_pd();
1256             fjz1             = _mm_setzero_pd();
1257             fjx2             = _mm_setzero_pd();
1258             fjy2             = _mm_setzero_pd();
1259             fjz2             = _mm_setzero_pd();
1260             fjx3             = _mm_setzero_pd();
1261             fjy3             = _mm_setzero_pd();
1262             fjz3             = _mm_setzero_pd();
1263
1264             /**************************
1265              * CALCULATE INTERACTIONS *
1266              **************************/
1267
1268             /* COULOMB ELECTROSTATICS */
1269             velec            = _mm_mul_pd(qq11,rinv11);
1270             felec            = _mm_mul_pd(velec,rinvsq11);
1271
1272             fscal            = felec;
1273
1274             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1275
1276             /* Update vectorial force */
1277             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1278             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1279             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1280
1281             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1282             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1283             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1284
1285             /**************************
1286              * CALCULATE INTERACTIONS *
1287              **************************/
1288
1289             /* COULOMB ELECTROSTATICS */
1290             velec            = _mm_mul_pd(qq12,rinv12);
1291             felec            = _mm_mul_pd(velec,rinvsq12);
1292
1293             fscal            = felec;
1294
1295             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1296
1297             /* Update vectorial force */
1298             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1299             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1300             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1301
1302             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1303             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1304             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* COULOMB ELECTROSTATICS */
1311             velec            = _mm_mul_pd(qq13,rinv13);
1312             felec            = _mm_mul_pd(velec,rinvsq13);
1313
1314             fscal            = felec;
1315
1316             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1317
1318             /* Update vectorial force */
1319             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1320             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1321             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1322
1323             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1324             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1325             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1326
1327             /**************************
1328              * CALCULATE INTERACTIONS *
1329              **************************/
1330
1331             /* COULOMB ELECTROSTATICS */
1332             velec            = _mm_mul_pd(qq21,rinv21);
1333             felec            = _mm_mul_pd(velec,rinvsq21);
1334
1335             fscal            = felec;
1336
1337             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1338
1339             /* Update vectorial force */
1340             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1341             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1342             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1343
1344             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1345             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1346             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* COULOMB ELECTROSTATICS */
1353             velec            = _mm_mul_pd(qq22,rinv22);
1354             felec            = _mm_mul_pd(velec,rinvsq22);
1355
1356             fscal            = felec;
1357
1358             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1359
1360             /* Update vectorial force */
1361             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1362             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1363             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1364
1365             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1366             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1367             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1368
1369             /**************************
1370              * CALCULATE INTERACTIONS *
1371              **************************/
1372
1373             /* COULOMB ELECTROSTATICS */
1374             velec            = _mm_mul_pd(qq23,rinv23);
1375             felec            = _mm_mul_pd(velec,rinvsq23);
1376
1377             fscal            = felec;
1378
1379             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1380
1381             /* Update vectorial force */
1382             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1383             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1384             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1385
1386             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1387             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1388             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1389
1390             /**************************
1391              * CALCULATE INTERACTIONS *
1392              **************************/
1393
1394             /* COULOMB ELECTROSTATICS */
1395             velec            = _mm_mul_pd(qq31,rinv31);
1396             felec            = _mm_mul_pd(velec,rinvsq31);
1397
1398             fscal            = felec;
1399
1400             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1401
1402             /* Update vectorial force */
1403             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1404             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1405             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1406
1407             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1408             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1409             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1410
1411             /**************************
1412              * CALCULATE INTERACTIONS *
1413              **************************/
1414
1415             /* COULOMB ELECTROSTATICS */
1416             velec            = _mm_mul_pd(qq32,rinv32);
1417             felec            = _mm_mul_pd(velec,rinvsq32);
1418
1419             fscal            = felec;
1420
1421             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1422
1423             /* Update vectorial force */
1424             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1425             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1426             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1427
1428             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1429             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1430             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1431
1432             /**************************
1433              * CALCULATE INTERACTIONS *
1434              **************************/
1435
1436             /* COULOMB ELECTROSTATICS */
1437             velec            = _mm_mul_pd(qq33,rinv33);
1438             felec            = _mm_mul_pd(velec,rinvsq33);
1439
1440             fscal            = felec;
1441
1442             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1443
1444             /* Update vectorial force */
1445             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1446             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1447             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1448
1449             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1450             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1451             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1452
1453             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1454
1455             /* Inner loop uses 270 flops */
1456         }
1457
1458         /* End of innermost loop */
1459
1460         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1461                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1462
1463         /* Increment number of inner iterations */
1464         inneriter                  += j_index_end - j_index_start;
1465
1466         /* Outer loop uses 18 flops */
1467     }
1468
1469     /* Increment number of outer iterations */
1470     outeriter        += nri;
1471
1472     /* Update outer/inner flops */
1473
1474     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1475 }