src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecCoul_VdwNone_GeomW3W3_sse2_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 sse2_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_sse2_double.h"
  48 #include "kernelutil_x86_sse2_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_sse2_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_sse2_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             /* Calculate temporary vectorial force */
 277             tx               = _mm_mul_pd(fscal,dx00);
 278             ty               = _mm_mul_pd(fscal,dy00);
 279             tz               = _mm_mul_pd(fscal,dz00);
 280
 281             /* Update vectorial force */
 282             fix0             = _mm_add_pd(fix0,tx);
 283             fiy0             = _mm_add_pd(fiy0,ty);
 284             fiz0             = _mm_add_pd(fiz0,tz);
 285
 286             fjx0             = _mm_add_pd(fjx0,tx);
 287             fjy0             = _mm_add_pd(fjy0,ty);
 288             fjz0             = _mm_add_pd(fjz0,tz);
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             /* COULOMB ELECTROSTATICS */
 295             velec            = _mm_mul_pd(qq01,rinv01);
 296             felec            = _mm_mul_pd(velec,rinvsq01);
 297
 298             /* Update potential sum for this i atom from the interaction with this j atom. */
 299             velecsum         = _mm_add_pd(velecsum,velec);
 300
 301             fscal            = felec;
 302
 303             /* Calculate temporary vectorial force */
 304             tx               = _mm_mul_pd(fscal,dx01);
 305             ty               = _mm_mul_pd(fscal,dy01);
 306             tz               = _mm_mul_pd(fscal,dz01);
 307
 308             /* Update vectorial force */
 309             fix0             = _mm_add_pd(fix0,tx);
 310             fiy0             = _mm_add_pd(fiy0,ty);
 311             fiz0             = _mm_add_pd(fiz0,tz);
 312
 313             fjx1             = _mm_add_pd(fjx1,tx);
 314             fjy1             = _mm_add_pd(fjy1,ty);
 315             fjz1             = _mm_add_pd(fjz1,tz);
 316
 317             /**************************
 318              * CALCULATE INTERACTIONS *
 319              **************************/
 320
 321             /* COULOMB ELECTROSTATICS */
 322             velec            = _mm_mul_pd(qq02,rinv02);
 323             felec            = _mm_mul_pd(velec,rinvsq02);
 324
 325             /* Update potential sum for this i atom from the interaction with this j atom. */
 326             velecsum         = _mm_add_pd(velecsum,velec);
 327
 328             fscal            = felec;
 329
 330             /* Calculate temporary vectorial force */
 331             tx               = _mm_mul_pd(fscal,dx02);
 332             ty               = _mm_mul_pd(fscal,dy02);
 333             tz               = _mm_mul_pd(fscal,dz02);
 334
 335             /* Update vectorial force */
 336             fix0             = _mm_add_pd(fix0,tx);
 337             fiy0             = _mm_add_pd(fiy0,ty);
 338             fiz0             = _mm_add_pd(fiz0,tz);
 339
 340             fjx2             = _mm_add_pd(fjx2,tx);
 341             fjy2             = _mm_add_pd(fjy2,ty);
 342             fjz2             = _mm_add_pd(fjz2,tz);
 343
 344             /**************************
 345              * CALCULATE INTERACTIONS *
 346              **************************/
 347
 348             /* COULOMB ELECTROSTATICS */
 349             velec            = _mm_mul_pd(qq10,rinv10);
 350             felec            = _mm_mul_pd(velec,rinvsq10);
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             velecsum         = _mm_add_pd(velecsum,velec);
 354
 355             fscal            = felec;
 356
 357             /* Calculate temporary vectorial force */
 358             tx               = _mm_mul_pd(fscal,dx10);
 359             ty               = _mm_mul_pd(fscal,dy10);
 360             tz               = _mm_mul_pd(fscal,dz10);
 361
 362             /* Update vectorial force */
 363             fix1             = _mm_add_pd(fix1,tx);
 364             fiy1             = _mm_add_pd(fiy1,ty);
 365             fiz1             = _mm_add_pd(fiz1,tz);
 366
 367             fjx0             = _mm_add_pd(fjx0,tx);
 368             fjy0             = _mm_add_pd(fjy0,ty);
 369             fjz0             = _mm_add_pd(fjz0,tz);
 370
 371             /**************************
 372              * CALCULATE INTERACTIONS *
 373              **************************/
 374
 375             /* COULOMB ELECTROSTATICS */
 376             velec            = _mm_mul_pd(qq11,rinv11);
 377             felec            = _mm_mul_pd(velec,rinvsq11);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velecsum         = _mm_add_pd(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             /* Calculate temporary vectorial force */
 385             tx               = _mm_mul_pd(fscal,dx11);
 386             ty               = _mm_mul_pd(fscal,dy11);
 387             tz               = _mm_mul_pd(fscal,dz11);
 388
 389             /* Update vectorial force */
 390             fix1             = _mm_add_pd(fix1,tx);
 391             fiy1             = _mm_add_pd(fiy1,ty);
 392             fiz1             = _mm_add_pd(fiz1,tz);
 393
 394             fjx1             = _mm_add_pd(fjx1,tx);
 395             fjy1             = _mm_add_pd(fjy1,ty);
 396             fjz1             = _mm_add_pd(fjz1,tz);
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             /* COULOMB ELECTROSTATICS */
 403             velec            = _mm_mul_pd(qq12,rinv12);
 404             felec            = _mm_mul_pd(velec,rinvsq12);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velecsum         = _mm_add_pd(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411             /* Calculate temporary vectorial force */
 412             tx               = _mm_mul_pd(fscal,dx12);
 413             ty               = _mm_mul_pd(fscal,dy12);
 414             tz               = _mm_mul_pd(fscal,dz12);
 415
 416             /* Update vectorial force */
 417             fix1             = _mm_add_pd(fix1,tx);
 418             fiy1             = _mm_add_pd(fiy1,ty);
 419             fiz1             = _mm_add_pd(fiz1,tz);
 420
 421             fjx2             = _mm_add_pd(fjx2,tx);
 422             fjy2             = _mm_add_pd(fjy2,ty);
 423             fjz2             = _mm_add_pd(fjz2,tz);
 424
 425             /**************************
 426              * CALCULATE INTERACTIONS *
 427              **************************/
 428
 429             /* COULOMB ELECTROSTATICS */
 430             velec            = _mm_mul_pd(qq20,rinv20);
 431             felec            = _mm_mul_pd(velec,rinvsq20);
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velecsum         = _mm_add_pd(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             /* Calculate temporary vectorial force */
 439             tx               = _mm_mul_pd(fscal,dx20);
 440             ty               = _mm_mul_pd(fscal,dy20);
 441             tz               = _mm_mul_pd(fscal,dz20);
 442
 443             /* Update vectorial force */
 444             fix2             = _mm_add_pd(fix2,tx);
 445             fiy2             = _mm_add_pd(fiy2,ty);
 446             fiz2             = _mm_add_pd(fiz2,tz);
 447
 448             fjx0             = _mm_add_pd(fjx0,tx);
 449             fjy0             = _mm_add_pd(fjy0,ty);
 450             fjz0             = _mm_add_pd(fjz0,tz);
 451
 452             /**************************
 453              * CALCULATE INTERACTIONS *
 454              **************************/
 455
 456             /* COULOMB ELECTROSTATICS */
 457             velec            = _mm_mul_pd(qq21,rinv21);
 458             felec            = _mm_mul_pd(velec,rinvsq21);
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velecsum         = _mm_add_pd(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             /* Calculate temporary vectorial force */
 466             tx               = _mm_mul_pd(fscal,dx21);
 467             ty               = _mm_mul_pd(fscal,dy21);
 468             tz               = _mm_mul_pd(fscal,dz21);
 469
 470             /* Update vectorial force */
 471             fix2             = _mm_add_pd(fix2,tx);
 472             fiy2             = _mm_add_pd(fiy2,ty);
 473             fiz2             = _mm_add_pd(fiz2,tz);
 474
 475             fjx1             = _mm_add_pd(fjx1,tx);
 476             fjy1             = _mm_add_pd(fjy1,ty);
 477             fjz1             = _mm_add_pd(fjz1,tz);
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             /* COULOMB ELECTROSTATICS */
 484             velec            = _mm_mul_pd(qq22,rinv22);
 485             felec            = _mm_mul_pd(velec,rinvsq22);
 486
 487             /* Update potential sum for this i atom from the interaction with this j atom. */
 488             velecsum         = _mm_add_pd(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492             /* Calculate temporary vectorial force */
 493             tx               = _mm_mul_pd(fscal,dx22);
 494             ty               = _mm_mul_pd(fscal,dy22);
 495             tz               = _mm_mul_pd(fscal,dz22);
 496
 497             /* Update vectorial force */
 498             fix2             = _mm_add_pd(fix2,tx);
 499             fiy2             = _mm_add_pd(fiy2,ty);
 500             fiz2             = _mm_add_pd(fiz2,tz);
 501
 502             fjx2             = _mm_add_pd(fjx2,tx);
 503             fjy2             = _mm_add_pd(fjy2,ty);
 504             fjz2             = _mm_add_pd(fjz2,tz);
 505
 506             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 507
 508             /* Inner loop uses 252 flops */
 509         }
 510
 511         if(jidx<j_index_end)
 512         {
 513
 514             jnrA             = jjnr[jidx];
 515             j_coord_offsetA  = DIM*jnrA;
 516
 517             /* load j atom coordinates */
 518             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 519                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 520
 521             /* Calculate displacement vector */
 522             dx00             = _mm_sub_pd(ix0,jx0);
 523             dy00             = _mm_sub_pd(iy0,jy0);
 524             dz00             = _mm_sub_pd(iz0,jz0);
 525             dx01             = _mm_sub_pd(ix0,jx1);
 526             dy01             = _mm_sub_pd(iy0,jy1);
 527             dz01             = _mm_sub_pd(iz0,jz1);
 528             dx02             = _mm_sub_pd(ix0,jx2);
 529             dy02             = _mm_sub_pd(iy0,jy2);
 530             dz02             = _mm_sub_pd(iz0,jz2);
 531             dx10             = _mm_sub_pd(ix1,jx0);
 532             dy10             = _mm_sub_pd(iy1,jy0);
 533             dz10             = _mm_sub_pd(iz1,jz0);
 534             dx11             = _mm_sub_pd(ix1,jx1);
 535             dy11             = _mm_sub_pd(iy1,jy1);
 536             dz11             = _mm_sub_pd(iz1,jz1);
 537             dx12             = _mm_sub_pd(ix1,jx2);
 538             dy12             = _mm_sub_pd(iy1,jy2);
 539             dz12             = _mm_sub_pd(iz1,jz2);
 540             dx20             = _mm_sub_pd(ix2,jx0);
 541             dy20             = _mm_sub_pd(iy2,jy0);
 542             dz20             = _mm_sub_pd(iz2,jz0);
 543             dx21             = _mm_sub_pd(ix2,jx1);
 544             dy21             = _mm_sub_pd(iy2,jy1);
 545             dz21             = _mm_sub_pd(iz2,jz1);
 546             dx22             = _mm_sub_pd(ix2,jx2);
 547             dy22             = _mm_sub_pd(iy2,jy2);
 548             dz22             = _mm_sub_pd(iz2,jz2);
 549
 550             /* Calculate squared distance and things based on it */
 551             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 552             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 553             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 554             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 555             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 556             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 557             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 558             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 559             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 560
 561             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 562             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 563             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 564             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 565             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 566             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 567             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 568             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 569             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 570
 571             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 572             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 573             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 574             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 575             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 576             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 577             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 578             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 579             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 580
 581             fjx0             = _mm_setzero_pd();
 582             fjy0             = _mm_setzero_pd();
 583             fjz0             = _mm_setzero_pd();
 584             fjx1             = _mm_setzero_pd();
 585             fjy1             = _mm_setzero_pd();
 586             fjz1             = _mm_setzero_pd();
 587             fjx2             = _mm_setzero_pd();
 588             fjy2             = _mm_setzero_pd();
 589             fjz2             = _mm_setzero_pd();
 590
 591             /**************************
 592              * CALCULATE INTERACTIONS *
 593              **************************/
 594
 595             /* COULOMB ELECTROSTATICS */
 596             velec            = _mm_mul_pd(qq00,rinv00);
 597             felec            = _mm_mul_pd(velec,rinvsq00);
 598
 599             /* Update potential sum for this i atom from the interaction with this j atom. */
 600             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 601             velecsum         = _mm_add_pd(velecsum,velec);
 602
 603             fscal            = felec;
 604
 605             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 606
 607             /* Calculate temporary vectorial force */
 608             tx               = _mm_mul_pd(fscal,dx00);
 609             ty               = _mm_mul_pd(fscal,dy00);
 610             tz               = _mm_mul_pd(fscal,dz00);
 611
 612             /* Update vectorial force */
 613             fix0             = _mm_add_pd(fix0,tx);
 614             fiy0             = _mm_add_pd(fiy0,ty);
 615             fiz0             = _mm_add_pd(fiz0,tz);
 616
 617             fjx0             = _mm_add_pd(fjx0,tx);
 618             fjy0             = _mm_add_pd(fjy0,ty);
 619             fjz0             = _mm_add_pd(fjz0,tz);
 620
 621             /**************************
 622              * CALCULATE INTERACTIONS *
 623              **************************/
 624
 625             /* COULOMB ELECTROSTATICS */
 626             velec            = _mm_mul_pd(qq01,rinv01);
 627             felec            = _mm_mul_pd(velec,rinvsq01);
 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             /* Calculate temporary vectorial force */
 638             tx               = _mm_mul_pd(fscal,dx01);
 639             ty               = _mm_mul_pd(fscal,dy01);
 640             tz               = _mm_mul_pd(fscal,dz01);
 641
 642             /* Update vectorial force */
 643             fix0             = _mm_add_pd(fix0,tx);
 644             fiy0             = _mm_add_pd(fiy0,ty);
 645             fiz0             = _mm_add_pd(fiz0,tz);
 646
 647             fjx1             = _mm_add_pd(fjx1,tx);
 648             fjy1             = _mm_add_pd(fjy1,ty);
 649             fjz1             = _mm_add_pd(fjz1,tz);
 650
 651             /**************************
 652              * CALCULATE INTERACTIONS *
 653              **************************/
 654
 655             /* COULOMB ELECTROSTATICS */
 656             velec            = _mm_mul_pd(qq02,rinv02);
 657             felec            = _mm_mul_pd(velec,rinvsq02);
 658
 659             /* Update potential sum for this i atom from the interaction with this j atom. */
 660             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 661             velecsum         = _mm_add_pd(velecsum,velec);
 662
 663             fscal            = felec;
 664
 665             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 666
 667             /* Calculate temporary vectorial force */
 668             tx               = _mm_mul_pd(fscal,dx02);
 669             ty               = _mm_mul_pd(fscal,dy02);
 670             tz               = _mm_mul_pd(fscal,dz02);
 671
 672             /* Update vectorial force */
 673             fix0             = _mm_add_pd(fix0,tx);
 674             fiy0             = _mm_add_pd(fiy0,ty);
 675             fiz0             = _mm_add_pd(fiz0,tz);
 676
 677             fjx2             = _mm_add_pd(fjx2,tx);
 678             fjy2             = _mm_add_pd(fjy2,ty);
 679             fjz2             = _mm_add_pd(fjz2,tz);
 680
 681             /**************************
 682              * CALCULATE INTERACTIONS *
 683              **************************/
 684
 685             /* COULOMB ELECTROSTATICS */
 686             velec            = _mm_mul_pd(qq10,rinv10);
 687             felec            = _mm_mul_pd(velec,rinvsq10);
 688
 689             /* Update potential sum for this i atom from the interaction with this j atom. */
 690             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 691             velecsum         = _mm_add_pd(velecsum,velec);
 692
 693             fscal            = felec;
 694
 695             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 696
 697             /* Calculate temporary vectorial force */
 698             tx               = _mm_mul_pd(fscal,dx10);
 699             ty               = _mm_mul_pd(fscal,dy10);
 700             tz               = _mm_mul_pd(fscal,dz10);
 701
 702             /* Update vectorial force */
 703             fix1             = _mm_add_pd(fix1,tx);
 704             fiy1             = _mm_add_pd(fiy1,ty);
 705             fiz1             = _mm_add_pd(fiz1,tz);
 706
 707             fjx0             = _mm_add_pd(fjx0,tx);
 708             fjy0             = _mm_add_pd(fjy0,ty);
 709             fjz0             = _mm_add_pd(fjz0,tz);
 710
 711             /**************************
 712              * CALCULATE INTERACTIONS *
 713              **************************/
 714
 715             /* COULOMB ELECTROSTATICS */
 716             velec            = _mm_mul_pd(qq11,rinv11);
 717             felec            = _mm_mul_pd(velec,rinvsq11);
 718
 719             /* Update potential sum for this i atom from the interaction with this j atom. */
 720             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 721             velecsum         = _mm_add_pd(velecsum,velec);
 722
 723             fscal            = felec;
 724
 725             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 726
 727             /* Calculate temporary vectorial force */
 728             tx               = _mm_mul_pd(fscal,dx11);
 729             ty               = _mm_mul_pd(fscal,dy11);
 730             tz               = _mm_mul_pd(fscal,dz11);
 731
 732             /* Update vectorial force */
 733             fix1             = _mm_add_pd(fix1,tx);
 734             fiy1             = _mm_add_pd(fiy1,ty);
 735             fiz1             = _mm_add_pd(fiz1,tz);
 736
 737             fjx1             = _mm_add_pd(fjx1,tx);
 738             fjy1             = _mm_add_pd(fjy1,ty);
 739             fjz1             = _mm_add_pd(fjz1,tz);
 740
 741             /**************************
 742              * CALCULATE INTERACTIONS *
 743              **************************/
 744
 745             /* COULOMB ELECTROSTATICS */
 746             velec            = _mm_mul_pd(qq12,rinv12);
 747             felec            = _mm_mul_pd(velec,rinvsq12);
 748
 749             /* Update potential sum for this i atom from the interaction with this j atom. */
 750             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 751             velecsum         = _mm_add_pd(velecsum,velec);
 752
 753             fscal            = felec;
 754
 755             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 756
 757             /* Calculate temporary vectorial force */
 758             tx               = _mm_mul_pd(fscal,dx12);
 759             ty               = _mm_mul_pd(fscal,dy12);
 760             tz               = _mm_mul_pd(fscal,dz12);
 761
 762             /* Update vectorial force */
 763             fix1             = _mm_add_pd(fix1,tx);
 764             fiy1             = _mm_add_pd(fiy1,ty);
 765             fiz1             = _mm_add_pd(fiz1,tz);
 766
 767             fjx2             = _mm_add_pd(fjx2,tx);
 768             fjy2             = _mm_add_pd(fjy2,ty);
 769             fjz2             = _mm_add_pd(fjz2,tz);
 770
 771             /**************************
 772              * CALCULATE INTERACTIONS *
 773              **************************/
 774
 775             /* COULOMB ELECTROSTATICS */
 776             velec            = _mm_mul_pd(qq20,rinv20);
 777             felec            = _mm_mul_pd(velec,rinvsq20);
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 781             velecsum         = _mm_add_pd(velecsum,velec);
 782
 783             fscal            = felec;
 784
 785             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 786
 787             /* Calculate temporary vectorial force */
 788             tx               = _mm_mul_pd(fscal,dx20);
 789             ty               = _mm_mul_pd(fscal,dy20);
 790             tz               = _mm_mul_pd(fscal,dz20);
 791
 792             /* Update vectorial force */
 793             fix2             = _mm_add_pd(fix2,tx);
 794             fiy2             = _mm_add_pd(fiy2,ty);
 795             fiz2             = _mm_add_pd(fiz2,tz);
 796
 797             fjx0             = _mm_add_pd(fjx0,tx);
 798             fjy0             = _mm_add_pd(fjy0,ty);
 799             fjz0             = _mm_add_pd(fjz0,tz);
 800
 801             /**************************
 802              * CALCULATE INTERACTIONS *
 803              **************************/
 804
 805             /* COULOMB ELECTROSTATICS */
 806             velec            = _mm_mul_pd(qq21,rinv21);
 807             felec            = _mm_mul_pd(velec,rinvsq21);
 808
 809             /* Update potential sum for this i atom from the interaction with this j atom. */
 810             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 811             velecsum         = _mm_add_pd(velecsum,velec);
 812
 813             fscal            = felec;
 814
 815             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 816
 817             /* Calculate temporary vectorial force */
 818             tx               = _mm_mul_pd(fscal,dx21);
 819             ty               = _mm_mul_pd(fscal,dy21);
 820             tz               = _mm_mul_pd(fscal,dz21);
 821
 822             /* Update vectorial force */
 823             fix2             = _mm_add_pd(fix2,tx);
 824             fiy2             = _mm_add_pd(fiy2,ty);
 825             fiz2             = _mm_add_pd(fiz2,tz);
 826
 827             fjx1             = _mm_add_pd(fjx1,tx);
 828             fjy1             = _mm_add_pd(fjy1,ty);
 829             fjz1             = _mm_add_pd(fjz1,tz);
 830
 831             /**************************
 832              * CALCULATE INTERACTIONS *
 833              **************************/
 834
 835             /* COULOMB ELECTROSTATICS */
 836             velec            = _mm_mul_pd(qq22,rinv22);
 837             felec            = _mm_mul_pd(velec,rinvsq22);
 838
 839             /* Update potential sum for this i atom from the interaction with this j atom. */
 840             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 841             velecsum         = _mm_add_pd(velecsum,velec);
 842
 843             fscal            = felec;
 844
 845             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 846
 847             /* Calculate temporary vectorial force */
 848             tx               = _mm_mul_pd(fscal,dx22);
 849             ty               = _mm_mul_pd(fscal,dy22);
 850             tz               = _mm_mul_pd(fscal,dz22);
 851
 852             /* Update vectorial force */
 853             fix2             = _mm_add_pd(fix2,tx);
 854             fiy2             = _mm_add_pd(fiy2,ty);
 855             fiz2             = _mm_add_pd(fiz2,tz);
 856
 857             fjx2             = _mm_add_pd(fjx2,tx);
 858             fjy2             = _mm_add_pd(fjy2,ty);
 859             fjz2             = _mm_add_pd(fjz2,tz);
 860
 861             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 862
 863             /* Inner loop uses 252 flops */
 864         }
 865
 866         /* End of innermost loop */
 867
 868         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 869                                               f+i_coord_offset,fshift+i_shift_offset);
 870
 871         ggid                        = gid[iidx];
 872         /* Update potential energies */
 873         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 874
 875         /* Increment number of inner iterations */
 876         inneriter                  += j_index_end - j_index_start;
 877
 878         /* Outer loop uses 19 flops */
 879     }
 880
 881     /* Increment number of outer iterations */
 882     outeriter        += nri;
 883
 884     /* Update outer/inner flops */
 885
 886     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*252);
 887 }
 888 /*
 889  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_sse2_double
 890  * Electrostatics interaction: Coulomb
 891  * VdW interaction:            None
 892  * Geometry:                   Water3-Water3
 893  * Calculate force/pot:        Force
 894  */
 895 void
 896 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_sse2_double
 897                     (t_nblist                    * gmx_restrict       nlist,
 898                      rvec                        * gmx_restrict          xx,
 899                      rvec                        * gmx_restrict          ff,
 900                      t_forcerec                  * gmx_restrict          fr,
 901                      t_mdatoms                   * gmx_restrict     mdatoms,
 902                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 903                      t_nrnb                      * gmx_restrict        nrnb)
 904 {
 905     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 906      * just 0 for non-waters.
 907      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 908      * jnr indices corresponding to data put in the four positions in the SIMD register.
 909      */
 910     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 911     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 912     int              jnrA,jnrB;
 913     int              j_coord_offsetA,j_coord_offsetB;
 914     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 915     real             rcutoff_scalar;
 916     real             *shiftvec,*fshift,*x,*f;
 917     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 918     int              vdwioffset0;
 919     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 920     int              vdwioffset1;
 921     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 922     int              vdwioffset2;
 923     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 924     int              vdwjidx0A,vdwjidx0B;
 925     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 926     int              vdwjidx1A,vdwjidx1B;
 927     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 928     int              vdwjidx2A,vdwjidx2B;
 929     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 930     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 931     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 932     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 933     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 934     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 935     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 936     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 937     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 938     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 939     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 940     real             *charge;
 941     __m128d          dummy_mask,cutoff_mask;
 942     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 943     __m128d          one     = _mm_set1_pd(1.0);
 944     __m128d          two     = _mm_set1_pd(2.0);
 945     x                = xx[0];
 946     f                = ff[0];
 947
 948     nri              = nlist->nri;
 949     iinr             = nlist->iinr;
 950     jindex           = nlist->jindex;
 951     jjnr             = nlist->jjnr;
 952     shiftidx         = nlist->shift;
 953     gid              = nlist->gid;
 954     shiftvec         = fr->shift_vec[0];
 955     fshift           = fr->fshift[0];
 956     facel            = _mm_set1_pd(fr->epsfac);
 957     charge           = mdatoms->chargeA;
 958
 959     /* Setup water-specific parameters */
 960     inr              = nlist->iinr[0];
 961     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 962     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 963     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 964
 965     jq0              = _mm_set1_pd(charge[inr+0]);
 966     jq1              = _mm_set1_pd(charge[inr+1]);
 967     jq2              = _mm_set1_pd(charge[inr+2]);
 968     qq00             = _mm_mul_pd(iq0,jq0);
 969     qq01             = _mm_mul_pd(iq0,jq1);
 970     qq02             = _mm_mul_pd(iq0,jq2);
 971     qq10             = _mm_mul_pd(iq1,jq0);
 972     qq11             = _mm_mul_pd(iq1,jq1);
 973     qq12             = _mm_mul_pd(iq1,jq2);
 974     qq20             = _mm_mul_pd(iq2,jq0);
 975     qq21             = _mm_mul_pd(iq2,jq1);
 976     qq22             = _mm_mul_pd(iq2,jq2);
 977
 978     /* Avoid stupid compiler warnings */
 979     jnrA = jnrB = 0;
 980     j_coord_offsetA = 0;
 981     j_coord_offsetB = 0;
 982
 983     outeriter        = 0;
 984     inneriter        = 0;
 985
 986     /* Start outer loop over neighborlists */
 987     for(iidx=0; iidx<nri; iidx++)
 988     {
 989         /* Load shift vector for this list */
 990         i_shift_offset   = DIM*shiftidx[iidx];
 991
 992         /* Load limits for loop over neighbors */
 993         j_index_start    = jindex[iidx];
 994         j_index_end      = jindex[iidx+1];
 995
 996         /* Get outer coordinate index */
 997         inr              = iinr[iidx];
 998         i_coord_offset   = DIM*inr;
 999
1000         /* Load i particle coords and add shift vector */
1001         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1002                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1003
1004         fix0             = _mm_setzero_pd();
1005         fiy0             = _mm_setzero_pd();
1006         fiz0             = _mm_setzero_pd();
1007         fix1             = _mm_setzero_pd();
1008         fiy1             = _mm_setzero_pd();
1009         fiz1             = _mm_setzero_pd();
1010         fix2             = _mm_setzero_pd();
1011         fiy2             = _mm_setzero_pd();
1012         fiz2             = _mm_setzero_pd();
1013
1014         /* Start inner kernel loop */
1015         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1016         {
1017
1018             /* Get j neighbor index, and coordinate index */
1019             jnrA             = jjnr[jidx];
1020             jnrB             = jjnr[jidx+1];
1021             j_coord_offsetA  = DIM*jnrA;
1022             j_coord_offsetB  = DIM*jnrB;
1023
1024             /* load j atom coordinates */
1025             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1026                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1027
1028             /* Calculate displacement vector */
1029             dx00             = _mm_sub_pd(ix0,jx0);
1030             dy00             = _mm_sub_pd(iy0,jy0);
1031             dz00             = _mm_sub_pd(iz0,jz0);
1032             dx01             = _mm_sub_pd(ix0,jx1);
1033             dy01             = _mm_sub_pd(iy0,jy1);
1034             dz01             = _mm_sub_pd(iz0,jz1);
1035             dx02             = _mm_sub_pd(ix0,jx2);
1036             dy02             = _mm_sub_pd(iy0,jy2);
1037             dz02             = _mm_sub_pd(iz0,jz2);
1038             dx10             = _mm_sub_pd(ix1,jx0);
1039             dy10             = _mm_sub_pd(iy1,jy0);
1040             dz10             = _mm_sub_pd(iz1,jz0);
1041             dx11             = _mm_sub_pd(ix1,jx1);
1042             dy11             = _mm_sub_pd(iy1,jy1);
1043             dz11             = _mm_sub_pd(iz1,jz1);
1044             dx12             = _mm_sub_pd(ix1,jx2);
1045             dy12             = _mm_sub_pd(iy1,jy2);
1046             dz12             = _mm_sub_pd(iz1,jz2);
1047             dx20             = _mm_sub_pd(ix2,jx0);
1048             dy20             = _mm_sub_pd(iy2,jy0);
1049             dz20             = _mm_sub_pd(iz2,jz0);
1050             dx21             = _mm_sub_pd(ix2,jx1);
1051             dy21             = _mm_sub_pd(iy2,jy1);
1052             dz21             = _mm_sub_pd(iz2,jz1);
1053             dx22             = _mm_sub_pd(ix2,jx2);
1054             dy22             = _mm_sub_pd(iy2,jy2);
1055             dz22             = _mm_sub_pd(iz2,jz2);
1056
1057             /* Calculate squared distance and things based on it */
1058             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1059             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1060             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1061             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1062             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1063             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1064             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1065             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1066             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1067
1068             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1069             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1070             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1071             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1072             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1073             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1074             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1075             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1076             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1077
1078             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1079             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1080             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1081             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1082             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1083             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1084             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1085             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1086             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1087
1088             fjx0             = _mm_setzero_pd();
1089             fjy0             = _mm_setzero_pd();
1090             fjz0             = _mm_setzero_pd();
1091             fjx1             = _mm_setzero_pd();
1092             fjy1             = _mm_setzero_pd();
1093             fjz1             = _mm_setzero_pd();
1094             fjx2             = _mm_setzero_pd();
1095             fjy2             = _mm_setzero_pd();
1096             fjz2             = _mm_setzero_pd();
1097
1098             /**************************
1099              * CALCULATE INTERACTIONS *
1100              **************************/
1101
1102             /* COULOMB ELECTROSTATICS */
1103             velec            = _mm_mul_pd(qq00,rinv00);
1104             felec            = _mm_mul_pd(velec,rinvsq00);
1105
1106             fscal            = felec;
1107
1108             /* Calculate temporary vectorial force */
1109             tx               = _mm_mul_pd(fscal,dx00);
1110             ty               = _mm_mul_pd(fscal,dy00);
1111             tz               = _mm_mul_pd(fscal,dz00);
1112
1113             /* Update vectorial force */
1114             fix0             = _mm_add_pd(fix0,tx);
1115             fiy0             = _mm_add_pd(fiy0,ty);
1116             fiz0             = _mm_add_pd(fiz0,tz);
1117
1118             fjx0             = _mm_add_pd(fjx0,tx);
1119             fjy0             = _mm_add_pd(fjy0,ty);
1120             fjz0             = _mm_add_pd(fjz0,tz);
1121
1122             /**************************
1123              * CALCULATE INTERACTIONS *
1124              **************************/
1125
1126             /* COULOMB ELECTROSTATICS */
1127             velec            = _mm_mul_pd(qq01,rinv01);
1128             felec            = _mm_mul_pd(velec,rinvsq01);
1129
1130             fscal            = felec;
1131
1132             /* Calculate temporary vectorial force */
1133             tx               = _mm_mul_pd(fscal,dx01);
1134             ty               = _mm_mul_pd(fscal,dy01);
1135             tz               = _mm_mul_pd(fscal,dz01);
1136
1137             /* Update vectorial force */
1138             fix0             = _mm_add_pd(fix0,tx);
1139             fiy0             = _mm_add_pd(fiy0,ty);
1140             fiz0             = _mm_add_pd(fiz0,tz);
1141
1142             fjx1             = _mm_add_pd(fjx1,tx);
1143             fjy1             = _mm_add_pd(fjy1,ty);
1144             fjz1             = _mm_add_pd(fjz1,tz);
1145
1146             /**************************
1147              * CALCULATE INTERACTIONS *
1148              **************************/
1149
1150             /* COULOMB ELECTROSTATICS */
1151             velec            = _mm_mul_pd(qq02,rinv02);
1152             felec            = _mm_mul_pd(velec,rinvsq02);
1153
1154             fscal            = felec;
1155
1156             /* Calculate temporary vectorial force */
1157             tx               = _mm_mul_pd(fscal,dx02);
1158             ty               = _mm_mul_pd(fscal,dy02);
1159             tz               = _mm_mul_pd(fscal,dz02);
1160
1161             /* Update vectorial force */
1162             fix0             = _mm_add_pd(fix0,tx);
1163             fiy0             = _mm_add_pd(fiy0,ty);
1164             fiz0             = _mm_add_pd(fiz0,tz);
1165
1166             fjx2             = _mm_add_pd(fjx2,tx);
1167             fjy2             = _mm_add_pd(fjy2,ty);
1168             fjz2             = _mm_add_pd(fjz2,tz);
1169
1170             /**************************
1171              * CALCULATE INTERACTIONS *
1172              **************************/
1173
1174             /* COULOMB ELECTROSTATICS */
1175             velec            = _mm_mul_pd(qq10,rinv10);
1176             felec            = _mm_mul_pd(velec,rinvsq10);
1177
1178             fscal            = felec;
1179
1180             /* Calculate temporary vectorial force */
1181             tx               = _mm_mul_pd(fscal,dx10);
1182             ty               = _mm_mul_pd(fscal,dy10);
1183             tz               = _mm_mul_pd(fscal,dz10);
1184
1185             /* Update vectorial force */
1186             fix1             = _mm_add_pd(fix1,tx);
1187             fiy1             = _mm_add_pd(fiy1,ty);
1188             fiz1             = _mm_add_pd(fiz1,tz);
1189
1190             fjx0             = _mm_add_pd(fjx0,tx);
1191             fjy0             = _mm_add_pd(fjy0,ty);
1192             fjz0             = _mm_add_pd(fjz0,tz);
1193
1194             /**************************
1195              * CALCULATE INTERACTIONS *
1196              **************************/
1197
1198             /* COULOMB ELECTROSTATICS */
1199             velec            = _mm_mul_pd(qq11,rinv11);
1200             felec            = _mm_mul_pd(velec,rinvsq11);
1201
1202             fscal            = felec;
1203
1204             /* Calculate temporary vectorial force */
1205             tx               = _mm_mul_pd(fscal,dx11);
1206             ty               = _mm_mul_pd(fscal,dy11);
1207             tz               = _mm_mul_pd(fscal,dz11);
1208
1209             /* Update vectorial force */
1210             fix1             = _mm_add_pd(fix1,tx);
1211             fiy1             = _mm_add_pd(fiy1,ty);
1212             fiz1             = _mm_add_pd(fiz1,tz);
1213
1214             fjx1             = _mm_add_pd(fjx1,tx);
1215             fjy1             = _mm_add_pd(fjy1,ty);
1216             fjz1             = _mm_add_pd(fjz1,tz);
1217
1218             /**************************
1219              * CALCULATE INTERACTIONS *
1220              **************************/
1221
1222             /* COULOMB ELECTROSTATICS */
1223             velec            = _mm_mul_pd(qq12,rinv12);
1224             felec            = _mm_mul_pd(velec,rinvsq12);
1225
1226             fscal            = felec;
1227
1228             /* Calculate temporary vectorial force */
1229             tx               = _mm_mul_pd(fscal,dx12);
1230             ty               = _mm_mul_pd(fscal,dy12);
1231             tz               = _mm_mul_pd(fscal,dz12);
1232
1233             /* Update vectorial force */
1234             fix1             = _mm_add_pd(fix1,tx);
1235             fiy1             = _mm_add_pd(fiy1,ty);
1236             fiz1             = _mm_add_pd(fiz1,tz);
1237
1238             fjx2             = _mm_add_pd(fjx2,tx);
1239             fjy2             = _mm_add_pd(fjy2,ty);
1240             fjz2             = _mm_add_pd(fjz2,tz);
1241
1242             /**************************
1243              * CALCULATE INTERACTIONS *
1244              **************************/
1245
1246             /* COULOMB ELECTROSTATICS */
1247             velec            = _mm_mul_pd(qq20,rinv20);
1248             felec            = _mm_mul_pd(velec,rinvsq20);
1249
1250             fscal            = felec;
1251
1252             /* Calculate temporary vectorial force */
1253             tx               = _mm_mul_pd(fscal,dx20);
1254             ty               = _mm_mul_pd(fscal,dy20);
1255             tz               = _mm_mul_pd(fscal,dz20);
1256
1257             /* Update vectorial force */
1258             fix2             = _mm_add_pd(fix2,tx);
1259             fiy2             = _mm_add_pd(fiy2,ty);
1260             fiz2             = _mm_add_pd(fiz2,tz);
1261
1262             fjx0             = _mm_add_pd(fjx0,tx);
1263             fjy0             = _mm_add_pd(fjy0,ty);
1264             fjz0             = _mm_add_pd(fjz0,tz);
1265
1266             /**************************
1267              * CALCULATE INTERACTIONS *
1268              **************************/
1269
1270             /* COULOMB ELECTROSTATICS */
1271             velec            = _mm_mul_pd(qq21,rinv21);
1272             felec            = _mm_mul_pd(velec,rinvsq21);
1273
1274             fscal            = felec;
1275
1276             /* Calculate temporary vectorial force */
1277             tx               = _mm_mul_pd(fscal,dx21);
1278             ty               = _mm_mul_pd(fscal,dy21);
1279             tz               = _mm_mul_pd(fscal,dz21);
1280
1281             /* Update vectorial force */
1282             fix2             = _mm_add_pd(fix2,tx);
1283             fiy2             = _mm_add_pd(fiy2,ty);
1284             fiz2             = _mm_add_pd(fiz2,tz);
1285
1286             fjx1             = _mm_add_pd(fjx1,tx);
1287             fjy1             = _mm_add_pd(fjy1,ty);
1288             fjz1             = _mm_add_pd(fjz1,tz);
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             /* COULOMB ELECTROSTATICS */
1295             velec            = _mm_mul_pd(qq22,rinv22);
1296             felec            = _mm_mul_pd(velec,rinvsq22);
1297
1298             fscal            = felec;
1299
1300             /* Calculate temporary vectorial force */
1301             tx               = _mm_mul_pd(fscal,dx22);
1302             ty               = _mm_mul_pd(fscal,dy22);
1303             tz               = _mm_mul_pd(fscal,dz22);
1304
1305             /* Update vectorial force */
1306             fix2             = _mm_add_pd(fix2,tx);
1307             fiy2             = _mm_add_pd(fiy2,ty);
1308             fiz2             = _mm_add_pd(fiz2,tz);
1309
1310             fjx2             = _mm_add_pd(fjx2,tx);
1311             fjy2             = _mm_add_pd(fjy2,ty);
1312             fjz2             = _mm_add_pd(fjz2,tz);
1313
1314             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1315
1316             /* Inner loop uses 243 flops */
1317         }
1318
1319         if(jidx<j_index_end)
1320         {
1321
1322             jnrA             = jjnr[jidx];
1323             j_coord_offsetA  = DIM*jnrA;
1324
1325             /* load j atom coordinates */
1326             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1327                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1328
1329             /* Calculate displacement vector */
1330             dx00             = _mm_sub_pd(ix0,jx0);
1331             dy00             = _mm_sub_pd(iy0,jy0);
1332             dz00             = _mm_sub_pd(iz0,jz0);
1333             dx01             = _mm_sub_pd(ix0,jx1);
1334             dy01             = _mm_sub_pd(iy0,jy1);
1335             dz01             = _mm_sub_pd(iz0,jz1);
1336             dx02             = _mm_sub_pd(ix0,jx2);
1337             dy02             = _mm_sub_pd(iy0,jy2);
1338             dz02             = _mm_sub_pd(iz0,jz2);
1339             dx10             = _mm_sub_pd(ix1,jx0);
1340             dy10             = _mm_sub_pd(iy1,jy0);
1341             dz10             = _mm_sub_pd(iz1,jz0);
1342             dx11             = _mm_sub_pd(ix1,jx1);
1343             dy11             = _mm_sub_pd(iy1,jy1);
1344             dz11             = _mm_sub_pd(iz1,jz1);
1345             dx12             = _mm_sub_pd(ix1,jx2);
1346             dy12             = _mm_sub_pd(iy1,jy2);
1347             dz12             = _mm_sub_pd(iz1,jz2);
1348             dx20             = _mm_sub_pd(ix2,jx0);
1349             dy20             = _mm_sub_pd(iy2,jy0);
1350             dz20             = _mm_sub_pd(iz2,jz0);
1351             dx21             = _mm_sub_pd(ix2,jx1);
1352             dy21             = _mm_sub_pd(iy2,jy1);
1353             dz21             = _mm_sub_pd(iz2,jz1);
1354             dx22             = _mm_sub_pd(ix2,jx2);
1355             dy22             = _mm_sub_pd(iy2,jy2);
1356             dz22             = _mm_sub_pd(iz2,jz2);
1357
1358             /* Calculate squared distance and things based on it */
1359             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1360             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1361             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1362             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1363             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1364             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1365             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1366             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1367             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1368
1369             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1370             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1371             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1372             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1373             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1374             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1375             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1376             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1377             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1378
1379             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1380             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1381             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1382             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1383             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1384             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1385             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1386             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1387             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1388
1389             fjx0             = _mm_setzero_pd();
1390             fjy0             = _mm_setzero_pd();
1391             fjz0             = _mm_setzero_pd();
1392             fjx1             = _mm_setzero_pd();
1393             fjy1             = _mm_setzero_pd();
1394             fjz1             = _mm_setzero_pd();
1395             fjx2             = _mm_setzero_pd();
1396             fjy2             = _mm_setzero_pd();
1397             fjz2             = _mm_setzero_pd();
1398
1399             /**************************
1400              * CALCULATE INTERACTIONS *
1401              **************************/
1402
1403             /* COULOMB ELECTROSTATICS */
1404             velec            = _mm_mul_pd(qq00,rinv00);
1405             felec            = _mm_mul_pd(velec,rinvsq00);
1406
1407             fscal            = felec;
1408
1409             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1410
1411             /* Calculate temporary vectorial force */
1412             tx               = _mm_mul_pd(fscal,dx00);
1413             ty               = _mm_mul_pd(fscal,dy00);
1414             tz               = _mm_mul_pd(fscal,dz00);
1415
1416             /* Update vectorial force */
1417             fix0             = _mm_add_pd(fix0,tx);
1418             fiy0             = _mm_add_pd(fiy0,ty);
1419             fiz0             = _mm_add_pd(fiz0,tz);
1420
1421             fjx0             = _mm_add_pd(fjx0,tx);
1422             fjy0             = _mm_add_pd(fjy0,ty);
1423             fjz0             = _mm_add_pd(fjz0,tz);
1424
1425             /**************************
1426              * CALCULATE INTERACTIONS *
1427              **************************/
1428
1429             /* COULOMB ELECTROSTATICS */
1430             velec            = _mm_mul_pd(qq01,rinv01);
1431             felec            = _mm_mul_pd(velec,rinvsq01);
1432
1433             fscal            = felec;
1434
1435             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1436
1437             /* Calculate temporary vectorial force */
1438             tx               = _mm_mul_pd(fscal,dx01);
1439             ty               = _mm_mul_pd(fscal,dy01);
1440             tz               = _mm_mul_pd(fscal,dz01);
1441
1442             /* Update vectorial force */
1443             fix0             = _mm_add_pd(fix0,tx);
1444             fiy0             = _mm_add_pd(fiy0,ty);
1445             fiz0             = _mm_add_pd(fiz0,tz);
1446
1447             fjx1             = _mm_add_pd(fjx1,tx);
1448             fjy1             = _mm_add_pd(fjy1,ty);
1449             fjz1             = _mm_add_pd(fjz1,tz);
1450
1451             /**************************
1452              * CALCULATE INTERACTIONS *
1453              **************************/
1454
1455             /* COULOMB ELECTROSTATICS */
1456             velec            = _mm_mul_pd(qq02,rinv02);
1457             felec            = _mm_mul_pd(velec,rinvsq02);
1458
1459             fscal            = felec;
1460
1461             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1462
1463             /* Calculate temporary vectorial force */
1464             tx               = _mm_mul_pd(fscal,dx02);
1465             ty               = _mm_mul_pd(fscal,dy02);
1466             tz               = _mm_mul_pd(fscal,dz02);
1467
1468             /* Update vectorial force */
1469             fix0             = _mm_add_pd(fix0,tx);
1470             fiy0             = _mm_add_pd(fiy0,ty);
1471             fiz0             = _mm_add_pd(fiz0,tz);
1472
1473             fjx2             = _mm_add_pd(fjx2,tx);
1474             fjy2             = _mm_add_pd(fjy2,ty);
1475             fjz2             = _mm_add_pd(fjz2,tz);
1476
1477             /**************************
1478              * CALCULATE INTERACTIONS *
1479              **************************/
1480
1481             /* COULOMB ELECTROSTATICS */
1482             velec            = _mm_mul_pd(qq10,rinv10);
1483             felec            = _mm_mul_pd(velec,rinvsq10);
1484
1485             fscal            = felec;
1486
1487             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1488
1489             /* Calculate temporary vectorial force */
1490             tx               = _mm_mul_pd(fscal,dx10);
1491             ty               = _mm_mul_pd(fscal,dy10);
1492             tz               = _mm_mul_pd(fscal,dz10);
1493
1494             /* Update vectorial force */
1495             fix1             = _mm_add_pd(fix1,tx);
1496             fiy1             = _mm_add_pd(fiy1,ty);
1497             fiz1             = _mm_add_pd(fiz1,tz);
1498
1499             fjx0             = _mm_add_pd(fjx0,tx);
1500             fjy0             = _mm_add_pd(fjy0,ty);
1501             fjz0             = _mm_add_pd(fjz0,tz);
1502
1503             /**************************
1504              * CALCULATE INTERACTIONS *
1505              **************************/
1506
1507             /* COULOMB ELECTROSTATICS */
1508             velec            = _mm_mul_pd(qq11,rinv11);
1509             felec            = _mm_mul_pd(velec,rinvsq11);
1510
1511             fscal            = felec;
1512
1513             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1514
1515             /* Calculate temporary vectorial force */
1516             tx               = _mm_mul_pd(fscal,dx11);
1517             ty               = _mm_mul_pd(fscal,dy11);
1518             tz               = _mm_mul_pd(fscal,dz11);
1519
1520             /* Update vectorial force */
1521             fix1             = _mm_add_pd(fix1,tx);
1522             fiy1             = _mm_add_pd(fiy1,ty);
1523             fiz1             = _mm_add_pd(fiz1,tz);
1524
1525             fjx1             = _mm_add_pd(fjx1,tx);
1526             fjy1             = _mm_add_pd(fjy1,ty);
1527             fjz1             = _mm_add_pd(fjz1,tz);
1528
1529             /**************************
1530              * CALCULATE INTERACTIONS *
1531              **************************/
1532
1533             /* COULOMB ELECTROSTATICS */
1534             velec            = _mm_mul_pd(qq12,rinv12);
1535             felec            = _mm_mul_pd(velec,rinvsq12);
1536
1537             fscal            = felec;
1538
1539             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1540
1541             /* Calculate temporary vectorial force */
1542             tx               = _mm_mul_pd(fscal,dx12);
1543             ty               = _mm_mul_pd(fscal,dy12);
1544             tz               = _mm_mul_pd(fscal,dz12);
1545
1546             /* Update vectorial force */
1547             fix1             = _mm_add_pd(fix1,tx);
1548             fiy1             = _mm_add_pd(fiy1,ty);
1549             fiz1             = _mm_add_pd(fiz1,tz);
1550
1551             fjx2             = _mm_add_pd(fjx2,tx);
1552             fjy2             = _mm_add_pd(fjy2,ty);
1553             fjz2             = _mm_add_pd(fjz2,tz);
1554
1555             /**************************
1556              * CALCULATE INTERACTIONS *
1557              **************************/
1558
1559             /* COULOMB ELECTROSTATICS */
1560             velec            = _mm_mul_pd(qq20,rinv20);
1561             felec            = _mm_mul_pd(velec,rinvsq20);
1562
1563             fscal            = felec;
1564
1565             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1566
1567             /* Calculate temporary vectorial force */
1568             tx               = _mm_mul_pd(fscal,dx20);
1569             ty               = _mm_mul_pd(fscal,dy20);
1570             tz               = _mm_mul_pd(fscal,dz20);
1571
1572             /* Update vectorial force */
1573             fix2             = _mm_add_pd(fix2,tx);
1574             fiy2             = _mm_add_pd(fiy2,ty);
1575             fiz2             = _mm_add_pd(fiz2,tz);
1576
1577             fjx0             = _mm_add_pd(fjx0,tx);
1578             fjy0             = _mm_add_pd(fjy0,ty);
1579             fjz0             = _mm_add_pd(fjz0,tz);
1580
1581             /**************************
1582              * CALCULATE INTERACTIONS *
1583              **************************/
1584
1585             /* COULOMB ELECTROSTATICS */
1586             velec            = _mm_mul_pd(qq21,rinv21);
1587             felec            = _mm_mul_pd(velec,rinvsq21);
1588
1589             fscal            = felec;
1590
1591             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1592
1593             /* Calculate temporary vectorial force */
1594             tx               = _mm_mul_pd(fscal,dx21);
1595             ty               = _mm_mul_pd(fscal,dy21);
1596             tz               = _mm_mul_pd(fscal,dz21);
1597
1598             /* Update vectorial force */
1599             fix2             = _mm_add_pd(fix2,tx);
1600             fiy2             = _mm_add_pd(fiy2,ty);
1601             fiz2             = _mm_add_pd(fiz2,tz);
1602
1603             fjx1             = _mm_add_pd(fjx1,tx);
1604             fjy1             = _mm_add_pd(fjy1,ty);
1605             fjz1             = _mm_add_pd(fjz1,tz);
1606
1607             /**************************
1608              * CALCULATE INTERACTIONS *
1609              **************************/
1610
1611             /* COULOMB ELECTROSTATICS */
1612             velec            = _mm_mul_pd(qq22,rinv22);
1613             felec            = _mm_mul_pd(velec,rinvsq22);
1614
1615             fscal            = felec;
1616
1617             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1618
1619             /* Calculate temporary vectorial force */
1620             tx               = _mm_mul_pd(fscal,dx22);
1621             ty               = _mm_mul_pd(fscal,dy22);
1622             tz               = _mm_mul_pd(fscal,dz22);
1623
1624             /* Update vectorial force */
1625             fix2             = _mm_add_pd(fix2,tx);
1626             fiy2             = _mm_add_pd(fiy2,ty);
1627             fiz2             = _mm_add_pd(fiz2,tz);
1628
1629             fjx2             = _mm_add_pd(fjx2,tx);
1630             fjy2             = _mm_add_pd(fjy2,ty);
1631             fjz2             = _mm_add_pd(fjz2,tz);
1632
1633             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1634
1635             /* Inner loop uses 243 flops */
1636         }
1637
1638         /* End of innermost loop */
1639
1640         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1641                                               f+i_coord_offset,fshift+i_shift_offset);
1642
1643         /* Increment number of inner iterations */
1644         inneriter                  += j_index_end - j_index_start;
1645
1646         /* Outer loop uses 18 flops */
1647     }
1648
1649     /* Increment number of outer iterations */
1650     outeriter        += nri;
1651
1652     /* Update outer/inner flops */
1653
1654     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*243);
1655 }