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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_double
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            None
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwjidx0A,vdwjidx0B;
  86     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     int              vdwjidx1A,vdwjidx1B;
  88     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     __m128d          dummy_mask,cutoff_mask;
 103     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 104     __m128d          one     = _mm_set1_pd(1.0);
 105     __m128d          two     = _mm_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_pd(fr->ic->epsfac);
 118     charge           = mdatoms->chargeA;
 119
 120     /* Setup water-specific parameters */
 121     inr              = nlist->iinr[0];
 122     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+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
 126     jq0              = _mm_set1_pd(charge[inr+0]);
 127     jq1              = _mm_set1_pd(charge[inr+1]);
 128     jq2              = _mm_set1_pd(charge[inr+2]);
 129     qq00             = _mm_mul_pd(iq0,jq0);
 130     qq01             = _mm_mul_pd(iq0,jq1);
 131     qq02             = _mm_mul_pd(iq0,jq2);
 132     qq10             = _mm_mul_pd(iq1,jq0);
 133     qq11             = _mm_mul_pd(iq1,jq1);
 134     qq12             = _mm_mul_pd(iq1,jq2);
 135     qq20             = _mm_mul_pd(iq2,jq0);
 136     qq21             = _mm_mul_pd(iq2,jq1);
 137     qq22             = _mm_mul_pd(iq2,jq2);
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 163                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 164
 165         fix0             = _mm_setzero_pd();
 166         fiy0             = _mm_setzero_pd();
 167         fiz0             = _mm_setzero_pd();
 168         fix1             = _mm_setzero_pd();
 169         fiy1             = _mm_setzero_pd();
 170         fiz1             = _mm_setzero_pd();
 171         fix2             = _mm_setzero_pd();
 172         fiy2             = _mm_setzero_pd();
 173         fiz2             = _mm_setzero_pd();
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_pd();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187
 188             /* load j atom coordinates */
 189             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 190                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm_sub_pd(ix0,jx0);
 194             dy00             = _mm_sub_pd(iy0,jy0);
 195             dz00             = _mm_sub_pd(iz0,jz0);
 196             dx01             = _mm_sub_pd(ix0,jx1);
 197             dy01             = _mm_sub_pd(iy0,jy1);
 198             dz01             = _mm_sub_pd(iz0,jz1);
 199             dx02             = _mm_sub_pd(ix0,jx2);
 200             dy02             = _mm_sub_pd(iy0,jy2);
 201             dz02             = _mm_sub_pd(iz0,jz2);
 202             dx10             = _mm_sub_pd(ix1,jx0);
 203             dy10             = _mm_sub_pd(iy1,jy0);
 204             dz10             = _mm_sub_pd(iz1,jz0);
 205             dx11             = _mm_sub_pd(ix1,jx1);
 206             dy11             = _mm_sub_pd(iy1,jy1);
 207             dz11             = _mm_sub_pd(iz1,jz1);
 208             dx12             = _mm_sub_pd(ix1,jx2);
 209             dy12             = _mm_sub_pd(iy1,jy2);
 210             dz12             = _mm_sub_pd(iz1,jz2);
 211             dx20             = _mm_sub_pd(ix2,jx0);
 212             dy20             = _mm_sub_pd(iy2,jy0);
 213             dz20             = _mm_sub_pd(iz2,jz0);
 214             dx21             = _mm_sub_pd(ix2,jx1);
 215             dy21             = _mm_sub_pd(iy2,jy1);
 216             dz21             = _mm_sub_pd(iz2,jz1);
 217             dx22             = _mm_sub_pd(ix2,jx2);
 218             dy22             = _mm_sub_pd(iy2,jy2);
 219             dz22             = _mm_sub_pd(iz2,jz2);
 220
 221             /* Calculate squared distance and things based on it */
 222             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 223             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 224             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 225             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 226             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 227             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 228             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 229             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 230             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 231
 232             rinv00           = avx128fma_invsqrt_d(rsq00);
 233             rinv01           = avx128fma_invsqrt_d(rsq01);
 234             rinv02           = avx128fma_invsqrt_d(rsq02);
 235             rinv10           = avx128fma_invsqrt_d(rsq10);
 236             rinv11           = avx128fma_invsqrt_d(rsq11);
 237             rinv12           = avx128fma_invsqrt_d(rsq12);
 238             rinv20           = avx128fma_invsqrt_d(rsq20);
 239             rinv21           = avx128fma_invsqrt_d(rsq21);
 240             rinv22           = avx128fma_invsqrt_d(rsq22);
 241
 242             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 243             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 244             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 245             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 246             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 247             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 248             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 249             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 250             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 251
 252             fjx0             = _mm_setzero_pd();
 253             fjy0             = _mm_setzero_pd();
 254             fjz0             = _mm_setzero_pd();
 255             fjx1             = _mm_setzero_pd();
 256             fjy1             = _mm_setzero_pd();
 257             fjz1             = _mm_setzero_pd();
 258             fjx2             = _mm_setzero_pd();
 259             fjy2             = _mm_setzero_pd();
 260             fjz2             = _mm_setzero_pd();
 261
 262             /**************************
 263              * CALCULATE INTERACTIONS *
 264              **************************/
 265
 266             /* COULOMB ELECTROSTATICS */
 267             velec            = _mm_mul_pd(qq00,rinv00);
 268             felec            = _mm_mul_pd(velec,rinvsq00);
 269
 270             /* Update potential sum for this i atom from the interaction with this j atom. */
 271             velecsum         = _mm_add_pd(velecsum,velec);
 272
 273             fscal            = felec;
 274
 275             /* Update vectorial force */
 276             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 277             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 278             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 279
 280             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 281             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 282             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 283
 284             /**************************
 285              * CALCULATE INTERACTIONS *
 286              **************************/
 287
 288             /* COULOMB ELECTROSTATICS */
 289             velec            = _mm_mul_pd(qq01,rinv01);
 290             felec            = _mm_mul_pd(velec,rinvsq01);
 291
 292             /* Update potential sum for this i atom from the interaction with this j atom. */
 293             velecsum         = _mm_add_pd(velecsum,velec);
 294
 295             fscal            = felec;
 296
 297             /* Update vectorial force */
 298             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 299             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 300             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 301
 302             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 303             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 304             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             /* COULOMB ELECTROSTATICS */
 311             velec            = _mm_mul_pd(qq02,rinv02);
 312             felec            = _mm_mul_pd(velec,rinvsq02);
 313
 314             /* Update potential sum for this i atom from the interaction with this j atom. */
 315             velecsum         = _mm_add_pd(velecsum,velec);
 316
 317             fscal            = felec;
 318
 319             /* Update vectorial force */
 320             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 321             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 322             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 323
 324             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 325             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 326             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             /* COULOMB ELECTROSTATICS */
 333             velec            = _mm_mul_pd(qq10,rinv10);
 334             felec            = _mm_mul_pd(velec,rinvsq10);
 335
 336             /* Update potential sum for this i atom from the interaction with this j atom. */
 337             velecsum         = _mm_add_pd(velecsum,velec);
 338
 339             fscal            = felec;
 340
 341             /* Update vectorial force */
 342             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 343             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 344             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 345
 346             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 347             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 348             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 349
 350             /**************************
 351              * CALCULATE INTERACTIONS *
 352              **************************/
 353
 354             /* COULOMB ELECTROSTATICS */
 355             velec            = _mm_mul_pd(qq11,rinv11);
 356             felec            = _mm_mul_pd(velec,rinvsq11);
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velecsum         = _mm_add_pd(velecsum,velec);
 360
 361             fscal            = felec;
 362
 363             /* Update vectorial force */
 364             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 365             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 366             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 367
 368             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 369             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 370             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             /* COULOMB ELECTROSTATICS */
 377             velec            = _mm_mul_pd(qq12,rinv12);
 378             felec            = _mm_mul_pd(velec,rinvsq12);
 379
 380             /* Update potential sum for this i atom from the interaction with this j atom. */
 381             velecsum         = _mm_add_pd(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             /* Update vectorial force */
 386             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 387             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 388             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 389
 390             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 391             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 392             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             /* COULOMB ELECTROSTATICS */
 399             velec            = _mm_mul_pd(qq20,rinv20);
 400             felec            = _mm_mul_pd(velec,rinvsq20);
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             velecsum         = _mm_add_pd(velecsum,velec);
 404
 405             fscal            = felec;
 406
 407             /* Update vectorial force */
 408             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 409             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 410             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 411
 412             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 413             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 414             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             /* COULOMB ELECTROSTATICS */
 421             velec            = _mm_mul_pd(qq21,rinv21);
 422             felec            = _mm_mul_pd(velec,rinvsq21);
 423
 424             /* Update potential sum for this i atom from the interaction with this j atom. */
 425             velecsum         = _mm_add_pd(velecsum,velec);
 426
 427             fscal            = felec;
 428
 429             /* Update vectorial force */
 430             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 431             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 432             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 433
 434             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 435             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 436             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             /* COULOMB ELECTROSTATICS */
 443             velec            = _mm_mul_pd(qq22,rinv22);
 444             felec            = _mm_mul_pd(velec,rinvsq22);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velecsum         = _mm_add_pd(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             /* Update vectorial force */
 452             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 453             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 454             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 455
 456             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 457             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 458             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 459
 460             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 461
 462             /* Inner loop uses 279 flops */
 463         }
 464
 465         if(jidx<j_index_end)
 466         {
 467
 468             jnrA             = jjnr[jidx];
 469             j_coord_offsetA  = DIM*jnrA;
 470
 471             /* load j atom coordinates */
 472             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 473                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 474
 475             /* Calculate displacement vector */
 476             dx00             = _mm_sub_pd(ix0,jx0);
 477             dy00             = _mm_sub_pd(iy0,jy0);
 478             dz00             = _mm_sub_pd(iz0,jz0);
 479             dx01             = _mm_sub_pd(ix0,jx1);
 480             dy01             = _mm_sub_pd(iy0,jy1);
 481             dz01             = _mm_sub_pd(iz0,jz1);
 482             dx02             = _mm_sub_pd(ix0,jx2);
 483             dy02             = _mm_sub_pd(iy0,jy2);
 484             dz02             = _mm_sub_pd(iz0,jz2);
 485             dx10             = _mm_sub_pd(ix1,jx0);
 486             dy10             = _mm_sub_pd(iy1,jy0);
 487             dz10             = _mm_sub_pd(iz1,jz0);
 488             dx11             = _mm_sub_pd(ix1,jx1);
 489             dy11             = _mm_sub_pd(iy1,jy1);
 490             dz11             = _mm_sub_pd(iz1,jz1);
 491             dx12             = _mm_sub_pd(ix1,jx2);
 492             dy12             = _mm_sub_pd(iy1,jy2);
 493             dz12             = _mm_sub_pd(iz1,jz2);
 494             dx20             = _mm_sub_pd(ix2,jx0);
 495             dy20             = _mm_sub_pd(iy2,jy0);
 496             dz20             = _mm_sub_pd(iz2,jz0);
 497             dx21             = _mm_sub_pd(ix2,jx1);
 498             dy21             = _mm_sub_pd(iy2,jy1);
 499             dz21             = _mm_sub_pd(iz2,jz1);
 500             dx22             = _mm_sub_pd(ix2,jx2);
 501             dy22             = _mm_sub_pd(iy2,jy2);
 502             dz22             = _mm_sub_pd(iz2,jz2);
 503
 504             /* Calculate squared distance and things based on it */
 505             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 506             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 507             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 508             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 509             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 510             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 511             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 512             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 513             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 514
 515             rinv00           = avx128fma_invsqrt_d(rsq00);
 516             rinv01           = avx128fma_invsqrt_d(rsq01);
 517             rinv02           = avx128fma_invsqrt_d(rsq02);
 518             rinv10           = avx128fma_invsqrt_d(rsq10);
 519             rinv11           = avx128fma_invsqrt_d(rsq11);
 520             rinv12           = avx128fma_invsqrt_d(rsq12);
 521             rinv20           = avx128fma_invsqrt_d(rsq20);
 522             rinv21           = avx128fma_invsqrt_d(rsq21);
 523             rinv22           = avx128fma_invsqrt_d(rsq22);
 524
 525             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 526             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 527             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 528             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 529             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 530             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 531             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 532             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 533             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 534
 535             fjx0             = _mm_setzero_pd();
 536             fjy0             = _mm_setzero_pd();
 537             fjz0             = _mm_setzero_pd();
 538             fjx1             = _mm_setzero_pd();
 539             fjy1             = _mm_setzero_pd();
 540             fjz1             = _mm_setzero_pd();
 541             fjx2             = _mm_setzero_pd();
 542             fjy2             = _mm_setzero_pd();
 543             fjz2             = _mm_setzero_pd();
 544
 545             /**************************
 546              * CALCULATE INTERACTIONS *
 547              **************************/
 548
 549             /* COULOMB ELECTROSTATICS */
 550             velec            = _mm_mul_pd(qq00,rinv00);
 551             felec            = _mm_mul_pd(velec,rinvsq00);
 552
 553             /* Update potential sum for this i atom from the interaction with this j atom. */
 554             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 555             velecsum         = _mm_add_pd(velecsum,velec);
 556
 557             fscal            = felec;
 558
 559             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 560
 561             /* Update vectorial force */
 562             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 563             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 564             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 565
 566             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 567             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 568             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             /* COULOMB ELECTROSTATICS */
 575             velec            = _mm_mul_pd(qq01,rinv01);
 576             felec            = _mm_mul_pd(velec,rinvsq01);
 577
 578             /* Update potential sum for this i atom from the interaction with this j atom. */
 579             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 580             velecsum         = _mm_add_pd(velecsum,velec);
 581
 582             fscal            = felec;
 583
 584             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 585
 586             /* Update vectorial force */
 587             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 588             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 589             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 590
 591             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 592             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 593             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             /* COULOMB ELECTROSTATICS */
 600             velec            = _mm_mul_pd(qq02,rinv02);
 601             felec            = _mm_mul_pd(velec,rinvsq02);
 602
 603             /* Update potential sum for this i atom from the interaction with this j atom. */
 604             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 605             velecsum         = _mm_add_pd(velecsum,velec);
 606
 607             fscal            = felec;
 608
 609             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 610
 611             /* Update vectorial force */
 612             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 613             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 614             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 615
 616             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 617             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 618             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 619
 620             /**************************
 621              * CALCULATE INTERACTIONS *
 622              **************************/
 623
 624             /* COULOMB ELECTROSTATICS */
 625             velec            = _mm_mul_pd(qq10,rinv10);
 626             felec            = _mm_mul_pd(velec,rinvsq10);
 627
 628             /* Update potential sum for this i atom from the interaction with this j atom. */
 629             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 630             velecsum         = _mm_add_pd(velecsum,velec);
 631
 632             fscal            = felec;
 633
 634             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 635
 636             /* Update vectorial force */
 637             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 638             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 639             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 640
 641             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 642             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 643             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 644
 645             /**************************
 646              * CALCULATE INTERACTIONS *
 647              **************************/
 648
 649             /* COULOMB ELECTROSTATICS */
 650             velec            = _mm_mul_pd(qq11,rinv11);
 651             felec            = _mm_mul_pd(velec,rinvsq11);
 652
 653             /* Update potential sum for this i atom from the interaction with this j atom. */
 654             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 655             velecsum         = _mm_add_pd(velecsum,velec);
 656
 657             fscal            = felec;
 658
 659             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 660
 661             /* Update vectorial force */
 662             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 663             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 664             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 665
 666             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 667             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 668             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 669
 670             /**************************
 671              * CALCULATE INTERACTIONS *
 672              **************************/
 673
 674             /* COULOMB ELECTROSTATICS */
 675             velec            = _mm_mul_pd(qq12,rinv12);
 676             felec            = _mm_mul_pd(velec,rinvsq12);
 677
 678             /* Update potential sum for this i atom from the interaction with this j atom. */
 679             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 680             velecsum         = _mm_add_pd(velecsum,velec);
 681
 682             fscal            = felec;
 683
 684             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 685
 686             /* Update vectorial force */
 687             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 688             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 689             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 690
 691             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 692             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 693             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 694
 695             /**************************
 696              * CALCULATE INTERACTIONS *
 697              **************************/
 698
 699             /* COULOMB ELECTROSTATICS */
 700             velec            = _mm_mul_pd(qq20,rinv20);
 701             felec            = _mm_mul_pd(velec,rinvsq20);
 702
 703             /* Update potential sum for this i atom from the interaction with this j atom. */
 704             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 705             velecsum         = _mm_add_pd(velecsum,velec);
 706
 707             fscal            = felec;
 708
 709             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 710
 711             /* Update vectorial force */
 712             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 713             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 714             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 715
 716             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 717             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 718             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 719
 720             /**************************
 721              * CALCULATE INTERACTIONS *
 722              **************************/
 723
 724             /* COULOMB ELECTROSTATICS */
 725             velec            = _mm_mul_pd(qq21,rinv21);
 726             felec            = _mm_mul_pd(velec,rinvsq21);
 727
 728             /* Update potential sum for this i atom from the interaction with this j atom. */
 729             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 730             velecsum         = _mm_add_pd(velecsum,velec);
 731
 732             fscal            = felec;
 733
 734             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 735
 736             /* Update vectorial force */
 737             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 738             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 739             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 740
 741             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 742             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 743             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 744
 745             /**************************
 746              * CALCULATE INTERACTIONS *
 747              **************************/
 748
 749             /* COULOMB ELECTROSTATICS */
 750             velec            = _mm_mul_pd(qq22,rinv22);
 751             felec            = _mm_mul_pd(velec,rinvsq22);
 752
 753             /* Update potential sum for this i atom from the interaction with this j atom. */
 754             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 755             velecsum         = _mm_add_pd(velecsum,velec);
 756
 757             fscal            = felec;
 758
 759             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 760
 761             /* Update vectorial force */
 762             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 763             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 764             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 765
 766             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 767             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 768             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 769
 770             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 771
 772             /* Inner loop uses 279 flops */
 773         }
 774
 775         /* End of innermost loop */
 776
 777         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 778                                               f+i_coord_offset,fshift+i_shift_offset);
 779
 780         ggid                        = gid[iidx];
 781         /* Update potential energies */
 782         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 783
 784         /* Increment number of inner iterations */
 785         inneriter                  += j_index_end - j_index_start;
 786
 787         /* Outer loop uses 19 flops */
 788     }
 789
 790     /* Increment number of outer iterations */
 791     outeriter        += nri;
 792
 793     /* Update outer/inner flops */
 794
 795     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*279);
 796 }
 797 /*
 798  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_double
 799  * Electrostatics interaction: Coulomb
 800  * VdW interaction:            None
 801  * Geometry:                   Water3-Water3
 802  * Calculate force/pot:        Force
 803  */
 804 void
 805 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_double
 806                     (t_nblist                    * gmx_restrict       nlist,
 807                      rvec                        * gmx_restrict          xx,
 808                      rvec                        * gmx_restrict          ff,
 809                      struct t_forcerec           * gmx_restrict          fr,
 810                      t_mdatoms                   * gmx_restrict     mdatoms,
 811                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 812                      t_nrnb                      * gmx_restrict        nrnb)
 813 {
 814     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 815      * just 0 for non-waters.
 816      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 817      * jnr indices corresponding to data put in the four positions in the SIMD register.
 818      */
 819     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 820     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 821     int              jnrA,jnrB;
 822     int              j_coord_offsetA,j_coord_offsetB;
 823     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 824     real             rcutoff_scalar;
 825     real             *shiftvec,*fshift,*x,*f;
 826     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 827     int              vdwioffset0;
 828     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 829     int              vdwioffset1;
 830     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 831     int              vdwioffset2;
 832     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 833     int              vdwjidx0A,vdwjidx0B;
 834     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 835     int              vdwjidx1A,vdwjidx1B;
 836     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 837     int              vdwjidx2A,vdwjidx2B;
 838     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 839     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 840     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 841     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 842     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 843     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 844     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 845     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 846     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 847     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 848     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 849     real             *charge;
 850     __m128d          dummy_mask,cutoff_mask;
 851     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 852     __m128d          one     = _mm_set1_pd(1.0);
 853     __m128d          two     = _mm_set1_pd(2.0);
 854     x                = xx[0];
 855     f                = ff[0];
 856
 857     nri              = nlist->nri;
 858     iinr             = nlist->iinr;
 859     jindex           = nlist->jindex;
 860     jjnr             = nlist->jjnr;
 861     shiftidx         = nlist->shift;
 862     gid              = nlist->gid;
 863     shiftvec         = fr->shift_vec[0];
 864     fshift           = fr->fshift[0];
 865     facel            = _mm_set1_pd(fr->ic->epsfac);
 866     charge           = mdatoms->chargeA;
 867
 868     /* Setup water-specific parameters */
 869     inr              = nlist->iinr[0];
 870     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+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
 874     jq0              = _mm_set1_pd(charge[inr+0]);
 875     jq1              = _mm_set1_pd(charge[inr+1]);
 876     jq2              = _mm_set1_pd(charge[inr+2]);
 877     qq00             = _mm_mul_pd(iq0,jq0);
 878     qq01             = _mm_mul_pd(iq0,jq1);
 879     qq02             = _mm_mul_pd(iq0,jq2);
 880     qq10             = _mm_mul_pd(iq1,jq0);
 881     qq11             = _mm_mul_pd(iq1,jq1);
 882     qq12             = _mm_mul_pd(iq1,jq2);
 883     qq20             = _mm_mul_pd(iq2,jq0);
 884     qq21             = _mm_mul_pd(iq2,jq1);
 885     qq22             = _mm_mul_pd(iq2,jq2);
 886
 887     /* Avoid stupid compiler warnings */
 888     jnrA = jnrB = 0;
 889     j_coord_offsetA = 0;
 890     j_coord_offsetB = 0;
 891
 892     outeriter        = 0;
 893     inneriter        = 0;
 894
 895     /* Start outer loop over neighborlists */
 896     for(iidx=0; iidx<nri; iidx++)
 897     {
 898         /* Load shift vector for this list */
 899         i_shift_offset   = DIM*shiftidx[iidx];
 900
 901         /* Load limits for loop over neighbors */
 902         j_index_start    = jindex[iidx];
 903         j_index_end      = jindex[iidx+1];
 904
 905         /* Get outer coordinate index */
 906         inr              = iinr[iidx];
 907         i_coord_offset   = DIM*inr;
 908
 909         /* Load i particle coords and add shift vector */
 910         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 911                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 912
 913         fix0             = _mm_setzero_pd();
 914         fiy0             = _mm_setzero_pd();
 915         fiz0             = _mm_setzero_pd();
 916         fix1             = _mm_setzero_pd();
 917         fiy1             = _mm_setzero_pd();
 918         fiz1             = _mm_setzero_pd();
 919         fix2             = _mm_setzero_pd();
 920         fiy2             = _mm_setzero_pd();
 921         fiz2             = _mm_setzero_pd();
 922
 923         /* Start inner kernel loop */
 924         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 925         {
 926
 927             /* Get j neighbor index, and coordinate index */
 928             jnrA             = jjnr[jidx];
 929             jnrB             = jjnr[jidx+1];
 930             j_coord_offsetA  = DIM*jnrA;
 931             j_coord_offsetB  = DIM*jnrB;
 932
 933             /* load j atom coordinates */
 934             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 935                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 936
 937             /* Calculate displacement vector */
 938             dx00             = _mm_sub_pd(ix0,jx0);
 939             dy00             = _mm_sub_pd(iy0,jy0);
 940             dz00             = _mm_sub_pd(iz0,jz0);
 941             dx01             = _mm_sub_pd(ix0,jx1);
 942             dy01             = _mm_sub_pd(iy0,jy1);
 943             dz01             = _mm_sub_pd(iz0,jz1);
 944             dx02             = _mm_sub_pd(ix0,jx2);
 945             dy02             = _mm_sub_pd(iy0,jy2);
 946             dz02             = _mm_sub_pd(iz0,jz2);
 947             dx10             = _mm_sub_pd(ix1,jx0);
 948             dy10             = _mm_sub_pd(iy1,jy0);
 949             dz10             = _mm_sub_pd(iz1,jz0);
 950             dx11             = _mm_sub_pd(ix1,jx1);
 951             dy11             = _mm_sub_pd(iy1,jy1);
 952             dz11             = _mm_sub_pd(iz1,jz1);
 953             dx12             = _mm_sub_pd(ix1,jx2);
 954             dy12             = _mm_sub_pd(iy1,jy2);
 955             dz12             = _mm_sub_pd(iz1,jz2);
 956             dx20             = _mm_sub_pd(ix2,jx0);
 957             dy20             = _mm_sub_pd(iy2,jy0);
 958             dz20             = _mm_sub_pd(iz2,jz0);
 959             dx21             = _mm_sub_pd(ix2,jx1);
 960             dy21             = _mm_sub_pd(iy2,jy1);
 961             dz21             = _mm_sub_pd(iz2,jz1);
 962             dx22             = _mm_sub_pd(ix2,jx2);
 963             dy22             = _mm_sub_pd(iy2,jy2);
 964             dz22             = _mm_sub_pd(iz2,jz2);
 965
 966             /* Calculate squared distance and things based on it */
 967             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 968             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 969             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 970             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 971             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 972             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 973             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 974             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 975             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 976
 977             rinv00           = avx128fma_invsqrt_d(rsq00);
 978             rinv01           = avx128fma_invsqrt_d(rsq01);
 979             rinv02           = avx128fma_invsqrt_d(rsq02);
 980             rinv10           = avx128fma_invsqrt_d(rsq10);
 981             rinv11           = avx128fma_invsqrt_d(rsq11);
 982             rinv12           = avx128fma_invsqrt_d(rsq12);
 983             rinv20           = avx128fma_invsqrt_d(rsq20);
 984             rinv21           = avx128fma_invsqrt_d(rsq21);
 985             rinv22           = avx128fma_invsqrt_d(rsq22);
 986
 987             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 988             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 989             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 990             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 991             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 992             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 993             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 994             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 995             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 996
 997             fjx0             = _mm_setzero_pd();
 998             fjy0             = _mm_setzero_pd();
 999             fjz0             = _mm_setzero_pd();
1000             fjx1             = _mm_setzero_pd();
1001             fjy1             = _mm_setzero_pd();
1002             fjz1             = _mm_setzero_pd();
1003             fjx2             = _mm_setzero_pd();
1004             fjy2             = _mm_setzero_pd();
1005             fjz2             = _mm_setzero_pd();
1006
1007             /**************************
1008              * CALCULATE INTERACTIONS *
1009              **************************/
1010
1011             /* COULOMB ELECTROSTATICS */
1012             velec            = _mm_mul_pd(qq00,rinv00);
1013             felec            = _mm_mul_pd(velec,rinvsq00);
1014
1015             fscal            = felec;
1016
1017             /* Update vectorial force */
1018             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1019             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1020             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1021
1022             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1023             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1024             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1025
1026             /**************************
1027              * CALCULATE INTERACTIONS *
1028              **************************/
1029
1030             /* COULOMB ELECTROSTATICS */
1031             velec            = _mm_mul_pd(qq01,rinv01);
1032             felec            = _mm_mul_pd(velec,rinvsq01);
1033
1034             fscal            = felec;
1035
1036             /* Update vectorial force */
1037             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1038             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1039             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1040
1041             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1042             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1043             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1044
1045             /**************************
1046              * CALCULATE INTERACTIONS *
1047              **************************/
1048
1049             /* COULOMB ELECTROSTATICS */
1050             velec            = _mm_mul_pd(qq02,rinv02);
1051             felec            = _mm_mul_pd(velec,rinvsq02);
1052
1053             fscal            = felec;
1054
1055             /* Update vectorial force */
1056             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1057             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1058             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1059
1060             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1061             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1062             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1063
1064             /**************************
1065              * CALCULATE INTERACTIONS *
1066              **************************/
1067
1068             /* COULOMB ELECTROSTATICS */
1069             velec            = _mm_mul_pd(qq10,rinv10);
1070             felec            = _mm_mul_pd(velec,rinvsq10);
1071
1072             fscal            = felec;
1073
1074             /* Update vectorial force */
1075             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1076             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1077             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1078
1079             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1080             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1081             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1082
1083             /**************************
1084              * CALCULATE INTERACTIONS *
1085              **************************/
1086
1087             /* COULOMB ELECTROSTATICS */
1088             velec            = _mm_mul_pd(qq11,rinv11);
1089             felec            = _mm_mul_pd(velec,rinvsq11);
1090
1091             fscal            = felec;
1092
1093             /* Update vectorial force */
1094             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1095             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1096             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1097
1098             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1099             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1100             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1101
1102             /**************************
1103              * CALCULATE INTERACTIONS *
1104              **************************/
1105
1106             /* COULOMB ELECTROSTATICS */
1107             velec            = _mm_mul_pd(qq12,rinv12);
1108             felec            = _mm_mul_pd(velec,rinvsq12);
1109
1110             fscal            = felec;
1111
1112             /* Update vectorial force */
1113             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1114             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1115             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1116
1117             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1118             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1119             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1120
1121             /**************************
1122              * CALCULATE INTERACTIONS *
1123              **************************/
1124
1125             /* COULOMB ELECTROSTATICS */
1126             velec            = _mm_mul_pd(qq20,rinv20);
1127             felec            = _mm_mul_pd(velec,rinvsq20);
1128
1129             fscal            = felec;
1130
1131             /* Update vectorial force */
1132             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1133             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1134             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1135
1136             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1137             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1138             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1139
1140             /**************************
1141              * CALCULATE INTERACTIONS *
1142              **************************/
1143
1144             /* COULOMB ELECTROSTATICS */
1145             velec            = _mm_mul_pd(qq21,rinv21);
1146             felec            = _mm_mul_pd(velec,rinvsq21);
1147
1148             fscal            = felec;
1149
1150             /* Update vectorial force */
1151             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1152             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1153             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1154
1155             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1156             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1157             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1158
1159             /**************************
1160              * CALCULATE INTERACTIONS *
1161              **************************/
1162
1163             /* COULOMB ELECTROSTATICS */
1164             velec            = _mm_mul_pd(qq22,rinv22);
1165             felec            = _mm_mul_pd(velec,rinvsq22);
1166
1167             fscal            = felec;
1168
1169             /* Update vectorial force */
1170             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1171             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1172             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1173
1174             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1175             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1176             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1177
1178             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1179
1180             /* Inner loop uses 270 flops */
1181         }
1182
1183         if(jidx<j_index_end)
1184         {
1185
1186             jnrA             = jjnr[jidx];
1187             j_coord_offsetA  = DIM*jnrA;
1188
1189             /* load j atom coordinates */
1190             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1191                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1192
1193             /* Calculate displacement vector */
1194             dx00             = _mm_sub_pd(ix0,jx0);
1195             dy00             = _mm_sub_pd(iy0,jy0);
1196             dz00             = _mm_sub_pd(iz0,jz0);
1197             dx01             = _mm_sub_pd(ix0,jx1);
1198             dy01             = _mm_sub_pd(iy0,jy1);
1199             dz01             = _mm_sub_pd(iz0,jz1);
1200             dx02             = _mm_sub_pd(ix0,jx2);
1201             dy02             = _mm_sub_pd(iy0,jy2);
1202             dz02             = _mm_sub_pd(iz0,jz2);
1203             dx10             = _mm_sub_pd(ix1,jx0);
1204             dy10             = _mm_sub_pd(iy1,jy0);
1205             dz10             = _mm_sub_pd(iz1,jz0);
1206             dx11             = _mm_sub_pd(ix1,jx1);
1207             dy11             = _mm_sub_pd(iy1,jy1);
1208             dz11             = _mm_sub_pd(iz1,jz1);
1209             dx12             = _mm_sub_pd(ix1,jx2);
1210             dy12             = _mm_sub_pd(iy1,jy2);
1211             dz12             = _mm_sub_pd(iz1,jz2);
1212             dx20             = _mm_sub_pd(ix2,jx0);
1213             dy20             = _mm_sub_pd(iy2,jy0);
1214             dz20             = _mm_sub_pd(iz2,jz0);
1215             dx21             = _mm_sub_pd(ix2,jx1);
1216             dy21             = _mm_sub_pd(iy2,jy1);
1217             dz21             = _mm_sub_pd(iz2,jz1);
1218             dx22             = _mm_sub_pd(ix2,jx2);
1219             dy22             = _mm_sub_pd(iy2,jy2);
1220             dz22             = _mm_sub_pd(iz2,jz2);
1221
1222             /* Calculate squared distance and things based on it */
1223             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1224             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1225             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1226             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1227             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1228             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1229             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1230             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1231             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1232
1233             rinv00           = avx128fma_invsqrt_d(rsq00);
1234             rinv01           = avx128fma_invsqrt_d(rsq01);
1235             rinv02           = avx128fma_invsqrt_d(rsq02);
1236             rinv10           = avx128fma_invsqrt_d(rsq10);
1237             rinv11           = avx128fma_invsqrt_d(rsq11);
1238             rinv12           = avx128fma_invsqrt_d(rsq12);
1239             rinv20           = avx128fma_invsqrt_d(rsq20);
1240             rinv21           = avx128fma_invsqrt_d(rsq21);
1241             rinv22           = avx128fma_invsqrt_d(rsq22);
1242
1243             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1244             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1245             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1246             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1247             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1248             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1249             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1250             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1251             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1252
1253             fjx0             = _mm_setzero_pd();
1254             fjy0             = _mm_setzero_pd();
1255             fjz0             = _mm_setzero_pd();
1256             fjx1             = _mm_setzero_pd();
1257             fjy1             = _mm_setzero_pd();
1258             fjz1             = _mm_setzero_pd();
1259             fjx2             = _mm_setzero_pd();
1260             fjy2             = _mm_setzero_pd();
1261             fjz2             = _mm_setzero_pd();
1262
1263             /**************************
1264              * CALCULATE INTERACTIONS *
1265              **************************/
1266
1267             /* COULOMB ELECTROSTATICS */
1268             velec            = _mm_mul_pd(qq00,rinv00);
1269             felec            = _mm_mul_pd(velec,rinvsq00);
1270
1271             fscal            = felec;
1272
1273             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1274
1275             /* Update vectorial force */
1276             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1277             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1278             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1279
1280             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1281             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1282             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1283
1284             /**************************
1285              * CALCULATE INTERACTIONS *
1286              **************************/
1287
1288             /* COULOMB ELECTROSTATICS */
1289             velec            = _mm_mul_pd(qq01,rinv01);
1290             felec            = _mm_mul_pd(velec,rinvsq01);
1291
1292             fscal            = felec;
1293
1294             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1295
1296             /* Update vectorial force */
1297             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1298             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1299             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1300
1301             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1302             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1303             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1304
1305             /**************************
1306              * CALCULATE INTERACTIONS *
1307              **************************/
1308
1309             /* COULOMB ELECTROSTATICS */
1310             velec            = _mm_mul_pd(qq02,rinv02);
1311             felec            = _mm_mul_pd(velec,rinvsq02);
1312
1313             fscal            = felec;
1314
1315             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1316
1317             /* Update vectorial force */
1318             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1319             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1320             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1321
1322             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1323             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1324             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1325
1326             /**************************
1327              * CALCULATE INTERACTIONS *
1328              **************************/
1329
1330             /* COULOMB ELECTROSTATICS */
1331             velec            = _mm_mul_pd(qq10,rinv10);
1332             felec            = _mm_mul_pd(velec,rinvsq10);
1333
1334             fscal            = felec;
1335
1336             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1337
1338             /* Update vectorial force */
1339             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1340             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1341             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1342
1343             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1344             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1345             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1346
1347             /**************************
1348              * CALCULATE INTERACTIONS *
1349              **************************/
1350
1351             /* COULOMB ELECTROSTATICS */
1352             velec            = _mm_mul_pd(qq11,rinv11);
1353             felec            = _mm_mul_pd(velec,rinvsq11);
1354
1355             fscal            = felec;
1356
1357             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1358
1359             /* Update vectorial force */
1360             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1361             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1362             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1363
1364             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1365             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1366             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1367
1368             /**************************
1369              * CALCULATE INTERACTIONS *
1370              **************************/
1371
1372             /* COULOMB ELECTROSTATICS */
1373             velec            = _mm_mul_pd(qq12,rinv12);
1374             felec            = _mm_mul_pd(velec,rinvsq12);
1375
1376             fscal            = felec;
1377
1378             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1379
1380             /* Update vectorial force */
1381             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1382             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1383             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1384
1385             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1386             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1387             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1388
1389             /**************************
1390              * CALCULATE INTERACTIONS *
1391              **************************/
1392
1393             /* COULOMB ELECTROSTATICS */
1394             velec            = _mm_mul_pd(qq20,rinv20);
1395             felec            = _mm_mul_pd(velec,rinvsq20);
1396
1397             fscal            = felec;
1398
1399             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1400
1401             /* Update vectorial force */
1402             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1403             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1404             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1405
1406             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1407             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1408             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1409
1410             /**************************
1411              * CALCULATE INTERACTIONS *
1412              **************************/
1413
1414             /* COULOMB ELECTROSTATICS */
1415             velec            = _mm_mul_pd(qq21,rinv21);
1416             felec            = _mm_mul_pd(velec,rinvsq21);
1417
1418             fscal            = felec;
1419
1420             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1421
1422             /* Update vectorial force */
1423             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1424             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1425             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1426
1427             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1428             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1429             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1430
1431             /**************************
1432              * CALCULATE INTERACTIONS *
1433              **************************/
1434
1435             /* COULOMB ELECTROSTATICS */
1436             velec            = _mm_mul_pd(qq22,rinv22);
1437             felec            = _mm_mul_pd(velec,rinvsq22);
1438
1439             fscal            = felec;
1440
1441             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1442
1443             /* Update vectorial force */
1444             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1445             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1446             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1447
1448             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1449             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1450             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1451
1452             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1453
1454             /* Inner loop uses 270 flops */
1455         }
1456
1457         /* End of innermost loop */
1458
1459         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1460                                               f+i_coord_offset,fshift+i_shift_offset);
1461
1462         /* Increment number of inner iterations */
1463         inneriter                  += j_index_end - j_index_start;
1464
1465         /* Outer loop uses 18 flops */
1466     }
1467
1468     /* Increment number of outer iterations */
1469     outeriter        += nri;
1470
1471     /* Update outer/inner flops */
1472
1473     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
1474 }