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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "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_GeomW3W3_VF_avx_128_fma_double
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            None
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwNone_GeomW3W3_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              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwjidx0A,vdwjidx0B;
  87     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  94     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  95     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 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     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 124     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 125     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 126
 127     jq0              = _mm_set1_pd(charge[inr+0]);
 128     jq1              = _mm_set1_pd(charge[inr+1]);
 129     jq2              = _mm_set1_pd(charge[inr+2]);
 130     qq00             = _mm_mul_pd(iq0,jq0);
 131     qq01             = _mm_mul_pd(iq0,jq1);
 132     qq02             = _mm_mul_pd(iq0,jq2);
 133     qq10             = _mm_mul_pd(iq1,jq0);
 134     qq11             = _mm_mul_pd(iq1,jq1);
 135     qq12             = _mm_mul_pd(iq1,jq2);
 136     qq20             = _mm_mul_pd(iq2,jq0);
 137     qq21             = _mm_mul_pd(iq2,jq1);
 138     qq22             = _mm_mul_pd(iq2,jq2);
 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,
 164                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 165
 166         fix0             = _mm_setzero_pd();
 167         fiy0             = _mm_setzero_pd();
 168         fiz0             = _mm_setzero_pd();
 169         fix1             = _mm_setzero_pd();
 170         fiy1             = _mm_setzero_pd();
 171         fiz1             = _mm_setzero_pd();
 172         fix2             = _mm_setzero_pd();
 173         fiy2             = _mm_setzero_pd();
 174         fiz2             = _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,x+j_coord_offsetB,
 191                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 192
 193             /* Calculate displacement vector */
 194             dx00             = _mm_sub_pd(ix0,jx0);
 195             dy00             = _mm_sub_pd(iy0,jy0);
 196             dz00             = _mm_sub_pd(iz0,jz0);
 197             dx01             = _mm_sub_pd(ix0,jx1);
 198             dy01             = _mm_sub_pd(iy0,jy1);
 199             dz01             = _mm_sub_pd(iz0,jz1);
 200             dx02             = _mm_sub_pd(ix0,jx2);
 201             dy02             = _mm_sub_pd(iy0,jy2);
 202             dz02             = _mm_sub_pd(iz0,jz2);
 203             dx10             = _mm_sub_pd(ix1,jx0);
 204             dy10             = _mm_sub_pd(iy1,jy0);
 205             dz10             = _mm_sub_pd(iz1,jz0);
 206             dx11             = _mm_sub_pd(ix1,jx1);
 207             dy11             = _mm_sub_pd(iy1,jy1);
 208             dz11             = _mm_sub_pd(iz1,jz1);
 209             dx12             = _mm_sub_pd(ix1,jx2);
 210             dy12             = _mm_sub_pd(iy1,jy2);
 211             dz12             = _mm_sub_pd(iz1,jz2);
 212             dx20             = _mm_sub_pd(ix2,jx0);
 213             dy20             = _mm_sub_pd(iy2,jy0);
 214             dz20             = _mm_sub_pd(iz2,jz0);
 215             dx21             = _mm_sub_pd(ix2,jx1);
 216             dy21             = _mm_sub_pd(iy2,jy1);
 217             dz21             = _mm_sub_pd(iz2,jz1);
 218             dx22             = _mm_sub_pd(ix2,jx2);
 219             dy22             = _mm_sub_pd(iy2,jy2);
 220             dz22             = _mm_sub_pd(iz2,jz2);
 221
 222             /* Calculate squared distance and things based on it */
 223             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 224             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 225             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 226             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 227             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 228             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 229             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 230             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 231             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 232
 233             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 234             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 235             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 236             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 237             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 238             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 239             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 240             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 241             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 242
 243             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 244             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 245             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 246             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 247             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 248             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 249             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 250             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 251             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 252
 253             fjx0             = _mm_setzero_pd();
 254             fjy0             = _mm_setzero_pd();
 255             fjz0             = _mm_setzero_pd();
 256             fjx1             = _mm_setzero_pd();
 257             fjy1             = _mm_setzero_pd();
 258             fjz1             = _mm_setzero_pd();
 259             fjx2             = _mm_setzero_pd();
 260             fjy2             = _mm_setzero_pd();
 261             fjz2             = _mm_setzero_pd();
 262
 263             /**************************
 264              * CALCULATE INTERACTIONS *
 265              **************************/
 266
 267             /* COULOMB ELECTROSTATICS */
 268             velec            = _mm_mul_pd(qq00,rinv00);
 269             felec            = _mm_mul_pd(velec,rinvsq00);
 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             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 278             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 279             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 280
 281             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 282             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 283             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 284
 285             /**************************
 286              * CALCULATE INTERACTIONS *
 287              **************************/
 288
 289             /* COULOMB ELECTROSTATICS */
 290             velec            = _mm_mul_pd(qq01,rinv01);
 291             felec            = _mm_mul_pd(velec,rinvsq01);
 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             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 300             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 301             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 302
 303             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 304             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 305             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 306
 307             /**************************
 308              * CALCULATE INTERACTIONS *
 309              **************************/
 310
 311             /* COULOMB ELECTROSTATICS */
 312             velec            = _mm_mul_pd(qq02,rinv02);
 313             felec            = _mm_mul_pd(velec,rinvsq02);
 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             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 322             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 323             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 324
 325             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 326             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 327             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             /* COULOMB ELECTROSTATICS */
 334             velec            = _mm_mul_pd(qq10,rinv10);
 335             felec            = _mm_mul_pd(velec,rinvsq10);
 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             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 344             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 345             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 346
 347             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 348             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 349             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             /* COULOMB ELECTROSTATICS */
 356             velec            = _mm_mul_pd(qq11,rinv11);
 357             felec            = _mm_mul_pd(velec,rinvsq11);
 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             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 366             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 367             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 368
 369             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 370             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 371             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 372
 373             /**************************
 374              * CALCULATE INTERACTIONS *
 375              **************************/
 376
 377             /* COULOMB ELECTROSTATICS */
 378             velec            = _mm_mul_pd(qq12,rinv12);
 379             felec            = _mm_mul_pd(velec,rinvsq12);
 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             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 388             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 389             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 390
 391             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 392             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 393             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             /* COULOMB ELECTROSTATICS */
 400             velec            = _mm_mul_pd(qq20,rinv20);
 401             felec            = _mm_mul_pd(velec,rinvsq20);
 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             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 410             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 411             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 412
 413             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 414             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 415             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 416
 417             /**************************
 418              * CALCULATE INTERACTIONS *
 419              **************************/
 420
 421             /* COULOMB ELECTROSTATICS */
 422             velec            = _mm_mul_pd(qq21,rinv21);
 423             felec            = _mm_mul_pd(velec,rinvsq21);
 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             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 432             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 433             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 434
 435             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 436             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 437             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 438
 439             /**************************
 440              * CALCULATE INTERACTIONS *
 441              **************************/
 442
 443             /* COULOMB ELECTROSTATICS */
 444             velec            = _mm_mul_pd(qq22,rinv22);
 445             felec            = _mm_mul_pd(velec,rinvsq22);
 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             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 454             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 455             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 456
 457             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 458             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 459             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 460
 461             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 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,
 474                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 475
 476             /* Calculate displacement vector */
 477             dx00             = _mm_sub_pd(ix0,jx0);
 478             dy00             = _mm_sub_pd(iy0,jy0);
 479             dz00             = _mm_sub_pd(iz0,jz0);
 480             dx01             = _mm_sub_pd(ix0,jx1);
 481             dy01             = _mm_sub_pd(iy0,jy1);
 482             dz01             = _mm_sub_pd(iz0,jz1);
 483             dx02             = _mm_sub_pd(ix0,jx2);
 484             dy02             = _mm_sub_pd(iy0,jy2);
 485             dz02             = _mm_sub_pd(iz0,jz2);
 486             dx10             = _mm_sub_pd(ix1,jx0);
 487             dy10             = _mm_sub_pd(iy1,jy0);
 488             dz10             = _mm_sub_pd(iz1,jz0);
 489             dx11             = _mm_sub_pd(ix1,jx1);
 490             dy11             = _mm_sub_pd(iy1,jy1);
 491             dz11             = _mm_sub_pd(iz1,jz1);
 492             dx12             = _mm_sub_pd(ix1,jx2);
 493             dy12             = _mm_sub_pd(iy1,jy2);
 494             dz12             = _mm_sub_pd(iz1,jz2);
 495             dx20             = _mm_sub_pd(ix2,jx0);
 496             dy20             = _mm_sub_pd(iy2,jy0);
 497             dz20             = _mm_sub_pd(iz2,jz0);
 498             dx21             = _mm_sub_pd(ix2,jx1);
 499             dy21             = _mm_sub_pd(iy2,jy1);
 500             dz21             = _mm_sub_pd(iz2,jz1);
 501             dx22             = _mm_sub_pd(ix2,jx2);
 502             dy22             = _mm_sub_pd(iy2,jy2);
 503             dz22             = _mm_sub_pd(iz2,jz2);
 504
 505             /* Calculate squared distance and things based on it */
 506             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 507             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 508             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 509             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 510             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 511             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 512             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 513             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 514             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 515
 516             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 517             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 518             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 519             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 520             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 521             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 522             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 523             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 524             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 525
 526             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 527             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 528             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 529             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 530             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 531             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 532             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 533             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 534             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 535
 536             fjx0             = _mm_setzero_pd();
 537             fjy0             = _mm_setzero_pd();
 538             fjz0             = _mm_setzero_pd();
 539             fjx1             = _mm_setzero_pd();
 540             fjy1             = _mm_setzero_pd();
 541             fjz1             = _mm_setzero_pd();
 542             fjx2             = _mm_setzero_pd();
 543             fjy2             = _mm_setzero_pd();
 544             fjz2             = _mm_setzero_pd();
 545
 546             /**************************
 547              * CALCULATE INTERACTIONS *
 548              **************************/
 549
 550             /* COULOMB ELECTROSTATICS */
 551             velec            = _mm_mul_pd(qq00,rinv00);
 552             felec            = _mm_mul_pd(velec,rinvsq00);
 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             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 564             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 565             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 566
 567             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 568             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 569             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 570
 571             /**************************
 572              * CALCULATE INTERACTIONS *
 573              **************************/
 574
 575             /* COULOMB ELECTROSTATICS */
 576             velec            = _mm_mul_pd(qq01,rinv01);
 577             felec            = _mm_mul_pd(velec,rinvsq01);
 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             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 589             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 590             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 591
 592             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 593             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 594             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 595
 596             /**************************
 597              * CALCULATE INTERACTIONS *
 598              **************************/
 599
 600             /* COULOMB ELECTROSTATICS */
 601             velec            = _mm_mul_pd(qq02,rinv02);
 602             felec            = _mm_mul_pd(velec,rinvsq02);
 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             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 614             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 615             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 616
 617             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 618             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 619             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 620
 621             /**************************
 622              * CALCULATE INTERACTIONS *
 623              **************************/
 624
 625             /* COULOMB ELECTROSTATICS */
 626             velec            = _mm_mul_pd(qq10,rinv10);
 627             felec            = _mm_mul_pd(velec,rinvsq10);
 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             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 639             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 640             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 641
 642             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 643             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 644             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 645
 646             /**************************
 647              * CALCULATE INTERACTIONS *
 648              **************************/
 649
 650             /* COULOMB ELECTROSTATICS */
 651             velec            = _mm_mul_pd(qq11,rinv11);
 652             felec            = _mm_mul_pd(velec,rinvsq11);
 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             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 664             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 665             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 666
 667             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 668             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 669             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             /* COULOMB ELECTROSTATICS */
 676             velec            = _mm_mul_pd(qq12,rinv12);
 677             felec            = _mm_mul_pd(velec,rinvsq12);
 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             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 689             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 690             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 691
 692             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 693             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 694             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 695
 696             /**************************
 697              * CALCULATE INTERACTIONS *
 698              **************************/
 699
 700             /* COULOMB ELECTROSTATICS */
 701             velec            = _mm_mul_pd(qq20,rinv20);
 702             felec            = _mm_mul_pd(velec,rinvsq20);
 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             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 714             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 715             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 716
 717             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 718             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 719             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             /* COULOMB ELECTROSTATICS */
 726             velec            = _mm_mul_pd(qq21,rinv21);
 727             felec            = _mm_mul_pd(velec,rinvsq21);
 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             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 739             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 740             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 741
 742             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 743             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 744             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 745
 746             /**************************
 747              * CALCULATE INTERACTIONS *
 748              **************************/
 749
 750             /* COULOMB ELECTROSTATICS */
 751             velec            = _mm_mul_pd(qq22,rinv22);
 752             felec            = _mm_mul_pd(velec,rinvsq22);
 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             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 764             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 765             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 766
 767             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 768             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 769             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 770
 771             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 772
 773             /* Inner loop uses 279 flops */
 774         }
 775
 776         /* End of innermost loop */
 777
 778         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 779                                               f+i_coord_offset,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_W3W3_VF,outeriter*19 + inneriter*279);
 797 }
 798 /*
 799  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_double
 800  * Electrostatics interaction: Coulomb
 801  * VdW interaction:            None
 802  * Geometry:                   Water3-Water3
 803  * Calculate force/pot:        Force
 804  */
 805 void
 806 nb_kernel_ElecCoul_VdwNone_GeomW3W3_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              vdwioffset0;
 829     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 830     int              vdwioffset1;
 831     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 832     int              vdwioffset2;
 833     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 834     int              vdwjidx0A,vdwjidx0B;
 835     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 836     int              vdwjidx1A,vdwjidx1B;
 837     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 838     int              vdwjidx2A,vdwjidx2B;
 839     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 840     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 841     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 842     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 843     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 844     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 845     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 846     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 847     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 848     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 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     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 872     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 873     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 874
 875     jq0              = _mm_set1_pd(charge[inr+0]);
 876     jq1              = _mm_set1_pd(charge[inr+1]);
 877     jq2              = _mm_set1_pd(charge[inr+2]);
 878     qq00             = _mm_mul_pd(iq0,jq0);
 879     qq01             = _mm_mul_pd(iq0,jq1);
 880     qq02             = _mm_mul_pd(iq0,jq2);
 881     qq10             = _mm_mul_pd(iq1,jq0);
 882     qq11             = _mm_mul_pd(iq1,jq1);
 883     qq12             = _mm_mul_pd(iq1,jq2);
 884     qq20             = _mm_mul_pd(iq2,jq0);
 885     qq21             = _mm_mul_pd(iq2,jq1);
 886     qq22             = _mm_mul_pd(iq2,jq2);
 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,
 912                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 913
 914         fix0             = _mm_setzero_pd();
 915         fiy0             = _mm_setzero_pd();
 916         fiz0             = _mm_setzero_pd();
 917         fix1             = _mm_setzero_pd();
 918         fiy1             = _mm_setzero_pd();
 919         fiz1             = _mm_setzero_pd();
 920         fix2             = _mm_setzero_pd();
 921         fiy2             = _mm_setzero_pd();
 922         fiz2             = _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,x+j_coord_offsetB,
 936                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 937
 938             /* Calculate displacement vector */
 939             dx00             = _mm_sub_pd(ix0,jx0);
 940             dy00             = _mm_sub_pd(iy0,jy0);
 941             dz00             = _mm_sub_pd(iz0,jz0);
 942             dx01             = _mm_sub_pd(ix0,jx1);
 943             dy01             = _mm_sub_pd(iy0,jy1);
 944             dz01             = _mm_sub_pd(iz0,jz1);
 945             dx02             = _mm_sub_pd(ix0,jx2);
 946             dy02             = _mm_sub_pd(iy0,jy2);
 947             dz02             = _mm_sub_pd(iz0,jz2);
 948             dx10             = _mm_sub_pd(ix1,jx0);
 949             dy10             = _mm_sub_pd(iy1,jy0);
 950             dz10             = _mm_sub_pd(iz1,jz0);
 951             dx11             = _mm_sub_pd(ix1,jx1);
 952             dy11             = _mm_sub_pd(iy1,jy1);
 953             dz11             = _mm_sub_pd(iz1,jz1);
 954             dx12             = _mm_sub_pd(ix1,jx2);
 955             dy12             = _mm_sub_pd(iy1,jy2);
 956             dz12             = _mm_sub_pd(iz1,jz2);
 957             dx20             = _mm_sub_pd(ix2,jx0);
 958             dy20             = _mm_sub_pd(iy2,jy0);
 959             dz20             = _mm_sub_pd(iz2,jz0);
 960             dx21             = _mm_sub_pd(ix2,jx1);
 961             dy21             = _mm_sub_pd(iy2,jy1);
 962             dz21             = _mm_sub_pd(iz2,jz1);
 963             dx22             = _mm_sub_pd(ix2,jx2);
 964             dy22             = _mm_sub_pd(iy2,jy2);
 965             dz22             = _mm_sub_pd(iz2,jz2);
 966
 967             /* Calculate squared distance and things based on it */
 968             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 969             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 970             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 971             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 972             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 973             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 974             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 975             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 976             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 977
 978             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 979             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 980             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 981             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 982             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 983             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 984             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 985             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 986             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 987
 988             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 989             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 990             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 991             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 992             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 993             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 994             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 995             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 996             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 997
 998             fjx0             = _mm_setzero_pd();
 999             fjy0             = _mm_setzero_pd();
1000             fjz0             = _mm_setzero_pd();
1001             fjx1             = _mm_setzero_pd();
1002             fjy1             = _mm_setzero_pd();
1003             fjz1             = _mm_setzero_pd();
1004             fjx2             = _mm_setzero_pd();
1005             fjy2             = _mm_setzero_pd();
1006             fjz2             = _mm_setzero_pd();
1007
1008             /**************************
1009              * CALCULATE INTERACTIONS *
1010              **************************/
1011
1012             /* COULOMB ELECTROSTATICS */
1013             velec            = _mm_mul_pd(qq00,rinv00);
1014             felec            = _mm_mul_pd(velec,rinvsq00);
1015
1016             fscal            = felec;
1017
1018             /* Update vectorial force */
1019             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1020             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1021             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1022
1023             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1024             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1025             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1026
1027             /**************************
1028              * CALCULATE INTERACTIONS *
1029              **************************/
1030
1031             /* COULOMB ELECTROSTATICS */
1032             velec            = _mm_mul_pd(qq01,rinv01);
1033             felec            = _mm_mul_pd(velec,rinvsq01);
1034
1035             fscal            = felec;
1036
1037             /* Update vectorial force */
1038             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1039             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1040             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1041
1042             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1043             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1044             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1045
1046             /**************************
1047              * CALCULATE INTERACTIONS *
1048              **************************/
1049
1050             /* COULOMB ELECTROSTATICS */
1051             velec            = _mm_mul_pd(qq02,rinv02);
1052             felec            = _mm_mul_pd(velec,rinvsq02);
1053
1054             fscal            = felec;
1055
1056             /* Update vectorial force */
1057             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1058             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1059             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1060
1061             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1062             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1063             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1064
1065             /**************************
1066              * CALCULATE INTERACTIONS *
1067              **************************/
1068
1069             /* COULOMB ELECTROSTATICS */
1070             velec            = _mm_mul_pd(qq10,rinv10);
1071             felec            = _mm_mul_pd(velec,rinvsq10);
1072
1073             fscal            = felec;
1074
1075             /* Update vectorial force */
1076             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1077             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1078             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1079
1080             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1081             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1082             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1083
1084             /**************************
1085              * CALCULATE INTERACTIONS *
1086              **************************/
1087
1088             /* COULOMB ELECTROSTATICS */
1089             velec            = _mm_mul_pd(qq11,rinv11);
1090             felec            = _mm_mul_pd(velec,rinvsq11);
1091
1092             fscal            = felec;
1093
1094             /* Update vectorial force */
1095             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1096             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1097             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1098
1099             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1100             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1101             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1102
1103             /**************************
1104              * CALCULATE INTERACTIONS *
1105              **************************/
1106
1107             /* COULOMB ELECTROSTATICS */
1108             velec            = _mm_mul_pd(qq12,rinv12);
1109             felec            = _mm_mul_pd(velec,rinvsq12);
1110
1111             fscal            = felec;
1112
1113             /* Update vectorial force */
1114             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1115             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1116             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1117
1118             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1119             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1120             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1121
1122             /**************************
1123              * CALCULATE INTERACTIONS *
1124              **************************/
1125
1126             /* COULOMB ELECTROSTATICS */
1127             velec            = _mm_mul_pd(qq20,rinv20);
1128             felec            = _mm_mul_pd(velec,rinvsq20);
1129
1130             fscal            = felec;
1131
1132             /* Update vectorial force */
1133             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1134             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1135             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1136
1137             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1138             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1139             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1140
1141             /**************************
1142              * CALCULATE INTERACTIONS *
1143              **************************/
1144
1145             /* COULOMB ELECTROSTATICS */
1146             velec            = _mm_mul_pd(qq21,rinv21);
1147             felec            = _mm_mul_pd(velec,rinvsq21);
1148
1149             fscal            = felec;
1150
1151             /* Update vectorial force */
1152             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1153             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1154             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1155
1156             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1157             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1158             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1159
1160             /**************************
1161              * CALCULATE INTERACTIONS *
1162              **************************/
1163
1164             /* COULOMB ELECTROSTATICS */
1165             velec            = _mm_mul_pd(qq22,rinv22);
1166             felec            = _mm_mul_pd(velec,rinvsq22);
1167
1168             fscal            = felec;
1169
1170             /* Update vectorial force */
1171             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1172             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1173             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1174
1175             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1176             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1177             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1178
1179             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
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,
1192                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1193
1194             /* Calculate displacement vector */
1195             dx00             = _mm_sub_pd(ix0,jx0);
1196             dy00             = _mm_sub_pd(iy0,jy0);
1197             dz00             = _mm_sub_pd(iz0,jz0);
1198             dx01             = _mm_sub_pd(ix0,jx1);
1199             dy01             = _mm_sub_pd(iy0,jy1);
1200             dz01             = _mm_sub_pd(iz0,jz1);
1201             dx02             = _mm_sub_pd(ix0,jx2);
1202             dy02             = _mm_sub_pd(iy0,jy2);
1203             dz02             = _mm_sub_pd(iz0,jz2);
1204             dx10             = _mm_sub_pd(ix1,jx0);
1205             dy10             = _mm_sub_pd(iy1,jy0);
1206             dz10             = _mm_sub_pd(iz1,jz0);
1207             dx11             = _mm_sub_pd(ix1,jx1);
1208             dy11             = _mm_sub_pd(iy1,jy1);
1209             dz11             = _mm_sub_pd(iz1,jz1);
1210             dx12             = _mm_sub_pd(ix1,jx2);
1211             dy12             = _mm_sub_pd(iy1,jy2);
1212             dz12             = _mm_sub_pd(iz1,jz2);
1213             dx20             = _mm_sub_pd(ix2,jx0);
1214             dy20             = _mm_sub_pd(iy2,jy0);
1215             dz20             = _mm_sub_pd(iz2,jz0);
1216             dx21             = _mm_sub_pd(ix2,jx1);
1217             dy21             = _mm_sub_pd(iy2,jy1);
1218             dz21             = _mm_sub_pd(iz2,jz1);
1219             dx22             = _mm_sub_pd(ix2,jx2);
1220             dy22             = _mm_sub_pd(iy2,jy2);
1221             dz22             = _mm_sub_pd(iz2,jz2);
1222
1223             /* Calculate squared distance and things based on it */
1224             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1225             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1226             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1227             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1228             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1229             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1230             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1231             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1232             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1233
1234             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1235             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1236             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1237             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1238             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1239             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1240             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1241             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1242             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1243
1244             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1245             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1246             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1247             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1248             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1249             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1250             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1251             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1252             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1253
1254             fjx0             = _mm_setzero_pd();
1255             fjy0             = _mm_setzero_pd();
1256             fjz0             = _mm_setzero_pd();
1257             fjx1             = _mm_setzero_pd();
1258             fjy1             = _mm_setzero_pd();
1259             fjz1             = _mm_setzero_pd();
1260             fjx2             = _mm_setzero_pd();
1261             fjy2             = _mm_setzero_pd();
1262             fjz2             = _mm_setzero_pd();
1263
1264             /**************************
1265              * CALCULATE INTERACTIONS *
1266              **************************/
1267
1268             /* COULOMB ELECTROSTATICS */
1269             velec            = _mm_mul_pd(qq00,rinv00);
1270             felec            = _mm_mul_pd(velec,rinvsq00);
1271
1272             fscal            = felec;
1273
1274             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1275
1276             /* Update vectorial force */
1277             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1278             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1279             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1280
1281             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1282             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1283             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1284
1285             /**************************
1286              * CALCULATE INTERACTIONS *
1287              **************************/
1288
1289             /* COULOMB ELECTROSTATICS */
1290             velec            = _mm_mul_pd(qq01,rinv01);
1291             felec            = _mm_mul_pd(velec,rinvsq01);
1292
1293             fscal            = felec;
1294
1295             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1296
1297             /* Update vectorial force */
1298             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1299             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1300             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1301
1302             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1303             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1304             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* COULOMB ELECTROSTATICS */
1311             velec            = _mm_mul_pd(qq02,rinv02);
1312             felec            = _mm_mul_pd(velec,rinvsq02);
1313
1314             fscal            = felec;
1315
1316             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1317
1318             /* Update vectorial force */
1319             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1320             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1321             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1322
1323             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1324             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1325             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1326
1327             /**************************
1328              * CALCULATE INTERACTIONS *
1329              **************************/
1330
1331             /* COULOMB ELECTROSTATICS */
1332             velec            = _mm_mul_pd(qq10,rinv10);
1333             felec            = _mm_mul_pd(velec,rinvsq10);
1334
1335             fscal            = felec;
1336
1337             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1338
1339             /* Update vectorial force */
1340             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1341             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1342             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1343
1344             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1345             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1346             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* COULOMB ELECTROSTATICS */
1353             velec            = _mm_mul_pd(qq11,rinv11);
1354             felec            = _mm_mul_pd(velec,rinvsq11);
1355
1356             fscal            = felec;
1357
1358             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1359
1360             /* Update vectorial force */
1361             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1362             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1363             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1364
1365             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1366             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1367             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1368
1369             /**************************
1370              * CALCULATE INTERACTIONS *
1371              **************************/
1372
1373             /* COULOMB ELECTROSTATICS */
1374             velec            = _mm_mul_pd(qq12,rinv12);
1375             felec            = _mm_mul_pd(velec,rinvsq12);
1376
1377             fscal            = felec;
1378
1379             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1380
1381             /* Update vectorial force */
1382             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1383             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1384             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1385
1386             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1387             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1388             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1389
1390             /**************************
1391              * CALCULATE INTERACTIONS *
1392              **************************/
1393
1394             /* COULOMB ELECTROSTATICS */
1395             velec            = _mm_mul_pd(qq20,rinv20);
1396             felec            = _mm_mul_pd(velec,rinvsq20);
1397
1398             fscal            = felec;
1399
1400             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1401
1402             /* Update vectorial force */
1403             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1404             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1405             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1406
1407             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1408             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1409             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1410
1411             /**************************
1412              * CALCULATE INTERACTIONS *
1413              **************************/
1414
1415             /* COULOMB ELECTROSTATICS */
1416             velec            = _mm_mul_pd(qq21,rinv21);
1417             felec            = _mm_mul_pd(velec,rinvsq21);
1418
1419             fscal            = felec;
1420
1421             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1422
1423             /* Update vectorial force */
1424             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1425             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1426             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1427
1428             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1429             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1430             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1431
1432             /**************************
1433              * CALCULATE INTERACTIONS *
1434              **************************/
1435
1436             /* COULOMB ELECTROSTATICS */
1437             velec            = _mm_mul_pd(qq22,rinv22);
1438             felec            = _mm_mul_pd(velec,rinvsq22);
1439
1440             fscal            = felec;
1441
1442             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1443
1444             /* Update vectorial force */
1445             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1446             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1447             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1448
1449             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1450             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1451             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1452
1453             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1454
1455             /* Inner loop uses 270 flops */
1456         }
1457
1458         /* End of innermost loop */
1459
1460         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1461                                               f+i_coord_offset,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_W3W3_F,outeriter*18 + inneriter*270);
1475 }