src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecCoul_VdwNone_GeomW4W4_sse4_1_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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 sse4_1_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse4_1_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse4_1_double
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            None
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse4_1_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset1;
  80     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  81     int              vdwioffset2;
  82     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  83     int              vdwioffset3;
  84     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  85     int              vdwjidx1A,vdwjidx1B;
  86     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  87     int              vdwjidx2A,vdwjidx2B;
  88     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  89     int              vdwjidx3A,vdwjidx3B;
  90     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  91     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  92     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  93     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  94     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  95     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  96     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  97     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  98     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  99     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 100     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     __m128d          dummy_mask,cutoff_mask;
 103     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 104     __m128d          one     = _mm_set1_pd(1.0);
 105     __m128d          two     = _mm_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_pd(fr->ic->epsfac);
 118     charge           = mdatoms->chargeA;
 119
 120     /* Setup water-specific parameters */
 121     inr              = nlist->iinr[0];
 122     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 123     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 124     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 125
 126     jq1              = _mm_set1_pd(charge[inr+1]);
 127     jq2              = _mm_set1_pd(charge[inr+2]);
 128     jq3              = _mm_set1_pd(charge[inr+3]);
 129     qq11             = _mm_mul_pd(iq1,jq1);
 130     qq12             = _mm_mul_pd(iq1,jq2);
 131     qq13             = _mm_mul_pd(iq1,jq3);
 132     qq21             = _mm_mul_pd(iq2,jq1);
 133     qq22             = _mm_mul_pd(iq2,jq2);
 134     qq23             = _mm_mul_pd(iq2,jq3);
 135     qq31             = _mm_mul_pd(iq3,jq1);
 136     qq32             = _mm_mul_pd(iq3,jq2);
 137     qq33             = _mm_mul_pd(iq3,jq3);
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 163                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 164
 165         fix1             = _mm_setzero_pd();
 166         fiy1             = _mm_setzero_pd();
 167         fiz1             = _mm_setzero_pd();
 168         fix2             = _mm_setzero_pd();
 169         fiy2             = _mm_setzero_pd();
 170         fiz2             = _mm_setzero_pd();
 171         fix3             = _mm_setzero_pd();
 172         fiy3             = _mm_setzero_pd();
 173         fiz3             = _mm_setzero_pd();
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_pd();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187
 188             /* load j atom coordinates */
 189             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 190                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 191
 192             /* Calculate displacement vector */
 193             dx11             = _mm_sub_pd(ix1,jx1);
 194             dy11             = _mm_sub_pd(iy1,jy1);
 195             dz11             = _mm_sub_pd(iz1,jz1);
 196             dx12             = _mm_sub_pd(ix1,jx2);
 197             dy12             = _mm_sub_pd(iy1,jy2);
 198             dz12             = _mm_sub_pd(iz1,jz2);
 199             dx13             = _mm_sub_pd(ix1,jx3);
 200             dy13             = _mm_sub_pd(iy1,jy3);
 201             dz13             = _mm_sub_pd(iz1,jz3);
 202             dx21             = _mm_sub_pd(ix2,jx1);
 203             dy21             = _mm_sub_pd(iy2,jy1);
 204             dz21             = _mm_sub_pd(iz2,jz1);
 205             dx22             = _mm_sub_pd(ix2,jx2);
 206             dy22             = _mm_sub_pd(iy2,jy2);
 207             dz22             = _mm_sub_pd(iz2,jz2);
 208             dx23             = _mm_sub_pd(ix2,jx3);
 209             dy23             = _mm_sub_pd(iy2,jy3);
 210             dz23             = _mm_sub_pd(iz2,jz3);
 211             dx31             = _mm_sub_pd(ix3,jx1);
 212             dy31             = _mm_sub_pd(iy3,jy1);
 213             dz31             = _mm_sub_pd(iz3,jz1);
 214             dx32             = _mm_sub_pd(ix3,jx2);
 215             dy32             = _mm_sub_pd(iy3,jy2);
 216             dz32             = _mm_sub_pd(iz3,jz2);
 217             dx33             = _mm_sub_pd(ix3,jx3);
 218             dy33             = _mm_sub_pd(iy3,jy3);
 219             dz33             = _mm_sub_pd(iz3,jz3);
 220
 221             /* Calculate squared distance and things based on it */
 222             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 223             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 224             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 225             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 226             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 227             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 228             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 229             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 230             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 231
 232             rinv11           = sse41_invsqrt_d(rsq11);
 233             rinv12           = sse41_invsqrt_d(rsq12);
 234             rinv13           = sse41_invsqrt_d(rsq13);
 235             rinv21           = sse41_invsqrt_d(rsq21);
 236             rinv22           = sse41_invsqrt_d(rsq22);
 237             rinv23           = sse41_invsqrt_d(rsq23);
 238             rinv31           = sse41_invsqrt_d(rsq31);
 239             rinv32           = sse41_invsqrt_d(rsq32);
 240             rinv33           = sse41_invsqrt_d(rsq33);
 241
 242             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 243             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 244             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 245             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 246             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 247             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 248             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 249             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 250             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 251
 252             fjx1             = _mm_setzero_pd();
 253             fjy1             = _mm_setzero_pd();
 254             fjz1             = _mm_setzero_pd();
 255             fjx2             = _mm_setzero_pd();
 256             fjy2             = _mm_setzero_pd();
 257             fjz2             = _mm_setzero_pd();
 258             fjx3             = _mm_setzero_pd();
 259             fjy3             = _mm_setzero_pd();
 260             fjz3             = _mm_setzero_pd();
 261
 262             /**************************
 263              * CALCULATE INTERACTIONS *
 264              **************************/
 265
 266             /* COULOMB ELECTROSTATICS */
 267             velec            = _mm_mul_pd(qq11,rinv11);
 268             felec            = _mm_mul_pd(velec,rinvsq11);
 269
 270             /* Update potential sum for this i atom from the interaction with this j atom. */
 271             velecsum         = _mm_add_pd(velecsum,velec);
 272
 273             fscal            = felec;
 274
 275             /* Calculate temporary vectorial force */
 276             tx               = _mm_mul_pd(fscal,dx11);
 277             ty               = _mm_mul_pd(fscal,dy11);
 278             tz               = _mm_mul_pd(fscal,dz11);
 279
 280             /* Update vectorial force */
 281             fix1             = _mm_add_pd(fix1,tx);
 282             fiy1             = _mm_add_pd(fiy1,ty);
 283             fiz1             = _mm_add_pd(fiz1,tz);
 284
 285             fjx1             = _mm_add_pd(fjx1,tx);
 286             fjy1             = _mm_add_pd(fjy1,ty);
 287             fjz1             = _mm_add_pd(fjz1,tz);
 288
 289             /**************************
 290              * CALCULATE INTERACTIONS *
 291              **************************/
 292
 293             /* COULOMB ELECTROSTATICS */
 294             velec            = _mm_mul_pd(qq12,rinv12);
 295             felec            = _mm_mul_pd(velec,rinvsq12);
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _mm_add_pd(velecsum,velec);
 299
 300             fscal            = felec;
 301
 302             /* Calculate temporary vectorial force */
 303             tx               = _mm_mul_pd(fscal,dx12);
 304             ty               = _mm_mul_pd(fscal,dy12);
 305             tz               = _mm_mul_pd(fscal,dz12);
 306
 307             /* Update vectorial force */
 308             fix1             = _mm_add_pd(fix1,tx);
 309             fiy1             = _mm_add_pd(fiy1,ty);
 310             fiz1             = _mm_add_pd(fiz1,tz);
 311
 312             fjx2             = _mm_add_pd(fjx2,tx);
 313             fjy2             = _mm_add_pd(fjy2,ty);
 314             fjz2             = _mm_add_pd(fjz2,tz);
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             /* COULOMB ELECTROSTATICS */
 321             velec            = _mm_mul_pd(qq13,rinv13);
 322             felec            = _mm_mul_pd(velec,rinvsq13);
 323
 324             /* Update potential sum for this i atom from the interaction with this j atom. */
 325             velecsum         = _mm_add_pd(velecsum,velec);
 326
 327             fscal            = felec;
 328
 329             /* Calculate temporary vectorial force */
 330             tx               = _mm_mul_pd(fscal,dx13);
 331             ty               = _mm_mul_pd(fscal,dy13);
 332             tz               = _mm_mul_pd(fscal,dz13);
 333
 334             /* Update vectorial force */
 335             fix1             = _mm_add_pd(fix1,tx);
 336             fiy1             = _mm_add_pd(fiy1,ty);
 337             fiz1             = _mm_add_pd(fiz1,tz);
 338
 339             fjx3             = _mm_add_pd(fjx3,tx);
 340             fjy3             = _mm_add_pd(fjy3,ty);
 341             fjz3             = _mm_add_pd(fjz3,tz);
 342
 343             /**************************
 344              * CALCULATE INTERACTIONS *
 345              **************************/
 346
 347             /* COULOMB ELECTROSTATICS */
 348             velec            = _mm_mul_pd(qq21,rinv21);
 349             felec            = _mm_mul_pd(velec,rinvsq21);
 350
 351             /* Update potential sum for this i atom from the interaction with this j atom. */
 352             velecsum         = _mm_add_pd(velecsum,velec);
 353
 354             fscal            = felec;
 355
 356             /* Calculate temporary vectorial force */
 357             tx               = _mm_mul_pd(fscal,dx21);
 358             ty               = _mm_mul_pd(fscal,dy21);
 359             tz               = _mm_mul_pd(fscal,dz21);
 360
 361             /* Update vectorial force */
 362             fix2             = _mm_add_pd(fix2,tx);
 363             fiy2             = _mm_add_pd(fiy2,ty);
 364             fiz2             = _mm_add_pd(fiz2,tz);
 365
 366             fjx1             = _mm_add_pd(fjx1,tx);
 367             fjy1             = _mm_add_pd(fjy1,ty);
 368             fjz1             = _mm_add_pd(fjz1,tz);
 369
 370             /**************************
 371              * CALCULATE INTERACTIONS *
 372              **************************/
 373
 374             /* COULOMB ELECTROSTATICS */
 375             velec            = _mm_mul_pd(qq22,rinv22);
 376             felec            = _mm_mul_pd(velec,rinvsq22);
 377
 378             /* Update potential sum for this i atom from the interaction with this j atom. */
 379             velecsum         = _mm_add_pd(velecsum,velec);
 380
 381             fscal            = felec;
 382
 383             /* Calculate temporary vectorial force */
 384             tx               = _mm_mul_pd(fscal,dx22);
 385             ty               = _mm_mul_pd(fscal,dy22);
 386             tz               = _mm_mul_pd(fscal,dz22);
 387
 388             /* Update vectorial force */
 389             fix2             = _mm_add_pd(fix2,tx);
 390             fiy2             = _mm_add_pd(fiy2,ty);
 391             fiz2             = _mm_add_pd(fiz2,tz);
 392
 393             fjx2             = _mm_add_pd(fjx2,tx);
 394             fjy2             = _mm_add_pd(fjy2,ty);
 395             fjz2             = _mm_add_pd(fjz2,tz);
 396
 397             /**************************
 398              * CALCULATE INTERACTIONS *
 399              **************************/
 400
 401             /* COULOMB ELECTROSTATICS */
 402             velec            = _mm_mul_pd(qq23,rinv23);
 403             felec            = _mm_mul_pd(velec,rinvsq23);
 404
 405             /* Update potential sum for this i atom from the interaction with this j atom. */
 406             velecsum         = _mm_add_pd(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             /* Calculate temporary vectorial force */
 411             tx               = _mm_mul_pd(fscal,dx23);
 412             ty               = _mm_mul_pd(fscal,dy23);
 413             tz               = _mm_mul_pd(fscal,dz23);
 414
 415             /* Update vectorial force */
 416             fix2             = _mm_add_pd(fix2,tx);
 417             fiy2             = _mm_add_pd(fiy2,ty);
 418             fiz2             = _mm_add_pd(fiz2,tz);
 419
 420             fjx3             = _mm_add_pd(fjx3,tx);
 421             fjy3             = _mm_add_pd(fjy3,ty);
 422             fjz3             = _mm_add_pd(fjz3,tz);
 423
 424             /**************************
 425              * CALCULATE INTERACTIONS *
 426              **************************/
 427
 428             /* COULOMB ELECTROSTATICS */
 429             velec            = _mm_mul_pd(qq31,rinv31);
 430             felec            = _mm_mul_pd(velec,rinvsq31);
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velecsum         = _mm_add_pd(velecsum,velec);
 434
 435             fscal            = felec;
 436
 437             /* Calculate temporary vectorial force */
 438             tx               = _mm_mul_pd(fscal,dx31);
 439             ty               = _mm_mul_pd(fscal,dy31);
 440             tz               = _mm_mul_pd(fscal,dz31);
 441
 442             /* Update vectorial force */
 443             fix3             = _mm_add_pd(fix3,tx);
 444             fiy3             = _mm_add_pd(fiy3,ty);
 445             fiz3             = _mm_add_pd(fiz3,tz);
 446
 447             fjx1             = _mm_add_pd(fjx1,tx);
 448             fjy1             = _mm_add_pd(fjy1,ty);
 449             fjz1             = _mm_add_pd(fjz1,tz);
 450
 451             /**************************
 452              * CALCULATE INTERACTIONS *
 453              **************************/
 454
 455             /* COULOMB ELECTROSTATICS */
 456             velec            = _mm_mul_pd(qq32,rinv32);
 457             felec            = _mm_mul_pd(velec,rinvsq32);
 458
 459             /* Update potential sum for this i atom from the interaction with this j atom. */
 460             velecsum         = _mm_add_pd(velecsum,velec);
 461
 462             fscal            = felec;
 463
 464             /* Calculate temporary vectorial force */
 465             tx               = _mm_mul_pd(fscal,dx32);
 466             ty               = _mm_mul_pd(fscal,dy32);
 467             tz               = _mm_mul_pd(fscal,dz32);
 468
 469             /* Update vectorial force */
 470             fix3             = _mm_add_pd(fix3,tx);
 471             fiy3             = _mm_add_pd(fiy3,ty);
 472             fiz3             = _mm_add_pd(fiz3,tz);
 473
 474             fjx2             = _mm_add_pd(fjx2,tx);
 475             fjy2             = _mm_add_pd(fjy2,ty);
 476             fjz2             = _mm_add_pd(fjz2,tz);
 477
 478             /**************************
 479              * CALCULATE INTERACTIONS *
 480              **************************/
 481
 482             /* COULOMB ELECTROSTATICS */
 483             velec            = _mm_mul_pd(qq33,rinv33);
 484             felec            = _mm_mul_pd(velec,rinvsq33);
 485
 486             /* Update potential sum for this i atom from the interaction with this j atom. */
 487             velecsum         = _mm_add_pd(velecsum,velec);
 488
 489             fscal            = felec;
 490
 491             /* Calculate temporary vectorial force */
 492             tx               = _mm_mul_pd(fscal,dx33);
 493             ty               = _mm_mul_pd(fscal,dy33);
 494             tz               = _mm_mul_pd(fscal,dz33);
 495
 496             /* Update vectorial force */
 497             fix3             = _mm_add_pd(fix3,tx);
 498             fiy3             = _mm_add_pd(fiy3,ty);
 499             fiz3             = _mm_add_pd(fiz3,tz);
 500
 501             fjx3             = _mm_add_pd(fjx3,tx);
 502             fjy3             = _mm_add_pd(fjy3,ty);
 503             fjz3             = _mm_add_pd(fjz3,tz);
 504
 505             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 506
 507             /* Inner loop uses 252 flops */
 508         }
 509
 510         if(jidx<j_index_end)
 511         {
 512
 513             jnrA             = jjnr[jidx];
 514             j_coord_offsetA  = DIM*jnrA;
 515
 516             /* load j atom coordinates */
 517             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 518                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 519
 520             /* Calculate displacement vector */
 521             dx11             = _mm_sub_pd(ix1,jx1);
 522             dy11             = _mm_sub_pd(iy1,jy1);
 523             dz11             = _mm_sub_pd(iz1,jz1);
 524             dx12             = _mm_sub_pd(ix1,jx2);
 525             dy12             = _mm_sub_pd(iy1,jy2);
 526             dz12             = _mm_sub_pd(iz1,jz2);
 527             dx13             = _mm_sub_pd(ix1,jx3);
 528             dy13             = _mm_sub_pd(iy1,jy3);
 529             dz13             = _mm_sub_pd(iz1,jz3);
 530             dx21             = _mm_sub_pd(ix2,jx1);
 531             dy21             = _mm_sub_pd(iy2,jy1);
 532             dz21             = _mm_sub_pd(iz2,jz1);
 533             dx22             = _mm_sub_pd(ix2,jx2);
 534             dy22             = _mm_sub_pd(iy2,jy2);
 535             dz22             = _mm_sub_pd(iz2,jz2);
 536             dx23             = _mm_sub_pd(ix2,jx3);
 537             dy23             = _mm_sub_pd(iy2,jy3);
 538             dz23             = _mm_sub_pd(iz2,jz3);
 539             dx31             = _mm_sub_pd(ix3,jx1);
 540             dy31             = _mm_sub_pd(iy3,jy1);
 541             dz31             = _mm_sub_pd(iz3,jz1);
 542             dx32             = _mm_sub_pd(ix3,jx2);
 543             dy32             = _mm_sub_pd(iy3,jy2);
 544             dz32             = _mm_sub_pd(iz3,jz2);
 545             dx33             = _mm_sub_pd(ix3,jx3);
 546             dy33             = _mm_sub_pd(iy3,jy3);
 547             dz33             = _mm_sub_pd(iz3,jz3);
 548
 549             /* Calculate squared distance and things based on it */
 550             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 551             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 552             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 553             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 554             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 555             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 556             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 557             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 558             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 559
 560             rinv11           = sse41_invsqrt_d(rsq11);
 561             rinv12           = sse41_invsqrt_d(rsq12);
 562             rinv13           = sse41_invsqrt_d(rsq13);
 563             rinv21           = sse41_invsqrt_d(rsq21);
 564             rinv22           = sse41_invsqrt_d(rsq22);
 565             rinv23           = sse41_invsqrt_d(rsq23);
 566             rinv31           = sse41_invsqrt_d(rsq31);
 567             rinv32           = sse41_invsqrt_d(rsq32);
 568             rinv33           = sse41_invsqrt_d(rsq33);
 569
 570             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 571             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 572             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 573             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 574             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 575             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 576             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 577             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 578             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 579
 580             fjx1             = _mm_setzero_pd();
 581             fjy1             = _mm_setzero_pd();
 582             fjz1             = _mm_setzero_pd();
 583             fjx2             = _mm_setzero_pd();
 584             fjy2             = _mm_setzero_pd();
 585             fjz2             = _mm_setzero_pd();
 586             fjx3             = _mm_setzero_pd();
 587             fjy3             = _mm_setzero_pd();
 588             fjz3             = _mm_setzero_pd();
 589
 590             /**************************
 591              * CALCULATE INTERACTIONS *
 592              **************************/
 593
 594             /* COULOMB ELECTROSTATICS */
 595             velec            = _mm_mul_pd(qq11,rinv11);
 596             felec            = _mm_mul_pd(velec,rinvsq11);
 597
 598             /* Update potential sum for this i atom from the interaction with this j atom. */
 599             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 600             velecsum         = _mm_add_pd(velecsum,velec);
 601
 602             fscal            = felec;
 603
 604             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 605
 606             /* Calculate temporary vectorial force */
 607             tx               = _mm_mul_pd(fscal,dx11);
 608             ty               = _mm_mul_pd(fscal,dy11);
 609             tz               = _mm_mul_pd(fscal,dz11);
 610
 611             /* Update vectorial force */
 612             fix1             = _mm_add_pd(fix1,tx);
 613             fiy1             = _mm_add_pd(fiy1,ty);
 614             fiz1             = _mm_add_pd(fiz1,tz);
 615
 616             fjx1             = _mm_add_pd(fjx1,tx);
 617             fjy1             = _mm_add_pd(fjy1,ty);
 618             fjz1             = _mm_add_pd(fjz1,tz);
 619
 620             /**************************
 621              * CALCULATE INTERACTIONS *
 622              **************************/
 623
 624             /* COULOMB ELECTROSTATICS */
 625             velec            = _mm_mul_pd(qq12,rinv12);
 626             felec            = _mm_mul_pd(velec,rinvsq12);
 627
 628             /* Update potential sum for this i atom from the interaction with this j atom. */
 629             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 630             velecsum         = _mm_add_pd(velecsum,velec);
 631
 632             fscal            = felec;
 633
 634             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 635
 636             /* Calculate temporary vectorial force */
 637             tx               = _mm_mul_pd(fscal,dx12);
 638             ty               = _mm_mul_pd(fscal,dy12);
 639             tz               = _mm_mul_pd(fscal,dz12);
 640
 641             /* Update vectorial force */
 642             fix1             = _mm_add_pd(fix1,tx);
 643             fiy1             = _mm_add_pd(fiy1,ty);
 644             fiz1             = _mm_add_pd(fiz1,tz);
 645
 646             fjx2             = _mm_add_pd(fjx2,tx);
 647             fjy2             = _mm_add_pd(fjy2,ty);
 648             fjz2             = _mm_add_pd(fjz2,tz);
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             /* COULOMB ELECTROSTATICS */
 655             velec            = _mm_mul_pd(qq13,rinv13);
 656             felec            = _mm_mul_pd(velec,rinvsq13);
 657
 658             /* Update potential sum for this i atom from the interaction with this j atom. */
 659             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 660             velecsum         = _mm_add_pd(velecsum,velec);
 661
 662             fscal            = felec;
 663
 664             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 665
 666             /* Calculate temporary vectorial force */
 667             tx               = _mm_mul_pd(fscal,dx13);
 668             ty               = _mm_mul_pd(fscal,dy13);
 669             tz               = _mm_mul_pd(fscal,dz13);
 670
 671             /* Update vectorial force */
 672             fix1             = _mm_add_pd(fix1,tx);
 673             fiy1             = _mm_add_pd(fiy1,ty);
 674             fiz1             = _mm_add_pd(fiz1,tz);
 675
 676             fjx3             = _mm_add_pd(fjx3,tx);
 677             fjy3             = _mm_add_pd(fjy3,ty);
 678             fjz3             = _mm_add_pd(fjz3,tz);
 679
 680             /**************************
 681              * CALCULATE INTERACTIONS *
 682              **************************/
 683
 684             /* COULOMB ELECTROSTATICS */
 685             velec            = _mm_mul_pd(qq21,rinv21);
 686             felec            = _mm_mul_pd(velec,rinvsq21);
 687
 688             /* Update potential sum for this i atom from the interaction with this j atom. */
 689             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 690             velecsum         = _mm_add_pd(velecsum,velec);
 691
 692             fscal            = felec;
 693
 694             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 695
 696             /* Calculate temporary vectorial force */
 697             tx               = _mm_mul_pd(fscal,dx21);
 698             ty               = _mm_mul_pd(fscal,dy21);
 699             tz               = _mm_mul_pd(fscal,dz21);
 700
 701             /* Update vectorial force */
 702             fix2             = _mm_add_pd(fix2,tx);
 703             fiy2             = _mm_add_pd(fiy2,ty);
 704             fiz2             = _mm_add_pd(fiz2,tz);
 705
 706             fjx1             = _mm_add_pd(fjx1,tx);
 707             fjy1             = _mm_add_pd(fjy1,ty);
 708             fjz1             = _mm_add_pd(fjz1,tz);
 709
 710             /**************************
 711              * CALCULATE INTERACTIONS *
 712              **************************/
 713
 714             /* COULOMB ELECTROSTATICS */
 715             velec            = _mm_mul_pd(qq22,rinv22);
 716             felec            = _mm_mul_pd(velec,rinvsq22);
 717
 718             /* Update potential sum for this i atom from the interaction with this j atom. */
 719             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 720             velecsum         = _mm_add_pd(velecsum,velec);
 721
 722             fscal            = felec;
 723
 724             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 725
 726             /* Calculate temporary vectorial force */
 727             tx               = _mm_mul_pd(fscal,dx22);
 728             ty               = _mm_mul_pd(fscal,dy22);
 729             tz               = _mm_mul_pd(fscal,dz22);
 730
 731             /* Update vectorial force */
 732             fix2             = _mm_add_pd(fix2,tx);
 733             fiy2             = _mm_add_pd(fiy2,ty);
 734             fiz2             = _mm_add_pd(fiz2,tz);
 735
 736             fjx2             = _mm_add_pd(fjx2,tx);
 737             fjy2             = _mm_add_pd(fjy2,ty);
 738             fjz2             = _mm_add_pd(fjz2,tz);
 739
 740             /**************************
 741              * CALCULATE INTERACTIONS *
 742              **************************/
 743
 744             /* COULOMB ELECTROSTATICS */
 745             velec            = _mm_mul_pd(qq23,rinv23);
 746             felec            = _mm_mul_pd(velec,rinvsq23);
 747
 748             /* Update potential sum for this i atom from the interaction with this j atom. */
 749             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 750             velecsum         = _mm_add_pd(velecsum,velec);
 751
 752             fscal            = felec;
 753
 754             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 755
 756             /* Calculate temporary vectorial force */
 757             tx               = _mm_mul_pd(fscal,dx23);
 758             ty               = _mm_mul_pd(fscal,dy23);
 759             tz               = _mm_mul_pd(fscal,dz23);
 760
 761             /* Update vectorial force */
 762             fix2             = _mm_add_pd(fix2,tx);
 763             fiy2             = _mm_add_pd(fiy2,ty);
 764             fiz2             = _mm_add_pd(fiz2,tz);
 765
 766             fjx3             = _mm_add_pd(fjx3,tx);
 767             fjy3             = _mm_add_pd(fjy3,ty);
 768             fjz3             = _mm_add_pd(fjz3,tz);
 769
 770             /**************************
 771              * CALCULATE INTERACTIONS *
 772              **************************/
 773
 774             /* COULOMB ELECTROSTATICS */
 775             velec            = _mm_mul_pd(qq31,rinv31);
 776             felec            = _mm_mul_pd(velec,rinvsq31);
 777
 778             /* Update potential sum for this i atom from the interaction with this j atom. */
 779             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 780             velecsum         = _mm_add_pd(velecsum,velec);
 781
 782             fscal            = felec;
 783
 784             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 785
 786             /* Calculate temporary vectorial force */
 787             tx               = _mm_mul_pd(fscal,dx31);
 788             ty               = _mm_mul_pd(fscal,dy31);
 789             tz               = _mm_mul_pd(fscal,dz31);
 790
 791             /* Update vectorial force */
 792             fix3             = _mm_add_pd(fix3,tx);
 793             fiy3             = _mm_add_pd(fiy3,ty);
 794             fiz3             = _mm_add_pd(fiz3,tz);
 795
 796             fjx1             = _mm_add_pd(fjx1,tx);
 797             fjy1             = _mm_add_pd(fjy1,ty);
 798             fjz1             = _mm_add_pd(fjz1,tz);
 799
 800             /**************************
 801              * CALCULATE INTERACTIONS *
 802              **************************/
 803
 804             /* COULOMB ELECTROSTATICS */
 805             velec            = _mm_mul_pd(qq32,rinv32);
 806             felec            = _mm_mul_pd(velec,rinvsq32);
 807
 808             /* Update potential sum for this i atom from the interaction with this j atom. */
 809             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 810             velecsum         = _mm_add_pd(velecsum,velec);
 811
 812             fscal            = felec;
 813
 814             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 815
 816             /* Calculate temporary vectorial force */
 817             tx               = _mm_mul_pd(fscal,dx32);
 818             ty               = _mm_mul_pd(fscal,dy32);
 819             tz               = _mm_mul_pd(fscal,dz32);
 820
 821             /* Update vectorial force */
 822             fix3             = _mm_add_pd(fix3,tx);
 823             fiy3             = _mm_add_pd(fiy3,ty);
 824             fiz3             = _mm_add_pd(fiz3,tz);
 825
 826             fjx2             = _mm_add_pd(fjx2,tx);
 827             fjy2             = _mm_add_pd(fjy2,ty);
 828             fjz2             = _mm_add_pd(fjz2,tz);
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             /* COULOMB ELECTROSTATICS */
 835             velec            = _mm_mul_pd(qq33,rinv33);
 836             felec            = _mm_mul_pd(velec,rinvsq33);
 837
 838             /* Update potential sum for this i atom from the interaction with this j atom. */
 839             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 840             velecsum         = _mm_add_pd(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 845
 846             /* Calculate temporary vectorial force */
 847             tx               = _mm_mul_pd(fscal,dx33);
 848             ty               = _mm_mul_pd(fscal,dy33);
 849             tz               = _mm_mul_pd(fscal,dz33);
 850
 851             /* Update vectorial force */
 852             fix3             = _mm_add_pd(fix3,tx);
 853             fiy3             = _mm_add_pd(fiy3,ty);
 854             fiz3             = _mm_add_pd(fiz3,tz);
 855
 856             fjx3             = _mm_add_pd(fjx3,tx);
 857             fjy3             = _mm_add_pd(fjy3,ty);
 858             fjz3             = _mm_add_pd(fjz3,tz);
 859
 860             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 861
 862             /* Inner loop uses 252 flops */
 863         }
 864
 865         /* End of innermost loop */
 866
 867         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 868                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 869
 870         ggid                        = gid[iidx];
 871         /* Update potential energies */
 872         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 873
 874         /* Increment number of inner iterations */
 875         inneriter                  += j_index_end - j_index_start;
 876
 877         /* Outer loop uses 19 flops */
 878     }
 879
 880     /* Increment number of outer iterations */
 881     outeriter        += nri;
 882
 883     /* Update outer/inner flops */
 884
 885     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*252);
 886 }
 887 /*
 888  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse4_1_double
 889  * Electrostatics interaction: Coulomb
 890  * VdW interaction:            None
 891  * Geometry:                   Water4-Water4
 892  * Calculate force/pot:        Force
 893  */
 894 void
 895 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse4_1_double
 896                     (t_nblist                    * gmx_restrict       nlist,
 897                      rvec                        * gmx_restrict          xx,
 898                      rvec                        * gmx_restrict          ff,
 899                      struct t_forcerec           * gmx_restrict          fr,
 900                      t_mdatoms                   * gmx_restrict     mdatoms,
 901                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 902                      t_nrnb                      * gmx_restrict        nrnb)
 903 {
 904     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 905      * just 0 for non-waters.
 906      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 907      * jnr indices corresponding to data put in the four positions in the SIMD register.
 908      */
 909     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 910     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 911     int              jnrA,jnrB;
 912     int              j_coord_offsetA,j_coord_offsetB;
 913     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 914     real             rcutoff_scalar;
 915     real             *shiftvec,*fshift,*x,*f;
 916     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 917     int              vdwioffset1;
 918     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 919     int              vdwioffset2;
 920     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 921     int              vdwioffset3;
 922     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 923     int              vdwjidx1A,vdwjidx1B;
 924     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 925     int              vdwjidx2A,vdwjidx2B;
 926     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 927     int              vdwjidx3A,vdwjidx3B;
 928     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 929     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 930     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 931     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 932     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 933     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 934     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 935     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 936     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 937     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 938     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 939     real             *charge;
 940     __m128d          dummy_mask,cutoff_mask;
 941     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 942     __m128d          one     = _mm_set1_pd(1.0);
 943     __m128d          two     = _mm_set1_pd(2.0);
 944     x                = xx[0];
 945     f                = ff[0];
 946
 947     nri              = nlist->nri;
 948     iinr             = nlist->iinr;
 949     jindex           = nlist->jindex;
 950     jjnr             = nlist->jjnr;
 951     shiftidx         = nlist->shift;
 952     gid              = nlist->gid;
 953     shiftvec         = fr->shift_vec[0];
 954     fshift           = fr->fshift[0];
 955     facel            = _mm_set1_pd(fr->ic->epsfac);
 956     charge           = mdatoms->chargeA;
 957
 958     /* Setup water-specific parameters */
 959     inr              = nlist->iinr[0];
 960     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 961     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 962     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 963
 964     jq1              = _mm_set1_pd(charge[inr+1]);
 965     jq2              = _mm_set1_pd(charge[inr+2]);
 966     jq3              = _mm_set1_pd(charge[inr+3]);
 967     qq11             = _mm_mul_pd(iq1,jq1);
 968     qq12             = _mm_mul_pd(iq1,jq2);
 969     qq13             = _mm_mul_pd(iq1,jq3);
 970     qq21             = _mm_mul_pd(iq2,jq1);
 971     qq22             = _mm_mul_pd(iq2,jq2);
 972     qq23             = _mm_mul_pd(iq2,jq3);
 973     qq31             = _mm_mul_pd(iq3,jq1);
 974     qq32             = _mm_mul_pd(iq3,jq2);
 975     qq33             = _mm_mul_pd(iq3,jq3);
 976
 977     /* Avoid stupid compiler warnings */
 978     jnrA = jnrB = 0;
 979     j_coord_offsetA = 0;
 980     j_coord_offsetB = 0;
 981
 982     outeriter        = 0;
 983     inneriter        = 0;
 984
 985     /* Start outer loop over neighborlists */
 986     for(iidx=0; iidx<nri; iidx++)
 987     {
 988         /* Load shift vector for this list */
 989         i_shift_offset   = DIM*shiftidx[iidx];
 990
 991         /* Load limits for loop over neighbors */
 992         j_index_start    = jindex[iidx];
 993         j_index_end      = jindex[iidx+1];
 994
 995         /* Get outer coordinate index */
 996         inr              = iinr[iidx];
 997         i_coord_offset   = DIM*inr;
 998
 999         /* Load i particle coords and add shift vector */
1000         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1001                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1002
1003         fix1             = _mm_setzero_pd();
1004         fiy1             = _mm_setzero_pd();
1005         fiz1             = _mm_setzero_pd();
1006         fix2             = _mm_setzero_pd();
1007         fiy2             = _mm_setzero_pd();
1008         fiz2             = _mm_setzero_pd();
1009         fix3             = _mm_setzero_pd();
1010         fiy3             = _mm_setzero_pd();
1011         fiz3             = _mm_setzero_pd();
1012
1013         /* Start inner kernel loop */
1014         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1015         {
1016
1017             /* Get j neighbor index, and coordinate index */
1018             jnrA             = jjnr[jidx];
1019             jnrB             = jjnr[jidx+1];
1020             j_coord_offsetA  = DIM*jnrA;
1021             j_coord_offsetB  = DIM*jnrB;
1022
1023             /* load j atom coordinates */
1024             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1025                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1026
1027             /* Calculate displacement vector */
1028             dx11             = _mm_sub_pd(ix1,jx1);
1029             dy11             = _mm_sub_pd(iy1,jy1);
1030             dz11             = _mm_sub_pd(iz1,jz1);
1031             dx12             = _mm_sub_pd(ix1,jx2);
1032             dy12             = _mm_sub_pd(iy1,jy2);
1033             dz12             = _mm_sub_pd(iz1,jz2);
1034             dx13             = _mm_sub_pd(ix1,jx3);
1035             dy13             = _mm_sub_pd(iy1,jy3);
1036             dz13             = _mm_sub_pd(iz1,jz3);
1037             dx21             = _mm_sub_pd(ix2,jx1);
1038             dy21             = _mm_sub_pd(iy2,jy1);
1039             dz21             = _mm_sub_pd(iz2,jz1);
1040             dx22             = _mm_sub_pd(ix2,jx2);
1041             dy22             = _mm_sub_pd(iy2,jy2);
1042             dz22             = _mm_sub_pd(iz2,jz2);
1043             dx23             = _mm_sub_pd(ix2,jx3);
1044             dy23             = _mm_sub_pd(iy2,jy3);
1045             dz23             = _mm_sub_pd(iz2,jz3);
1046             dx31             = _mm_sub_pd(ix3,jx1);
1047             dy31             = _mm_sub_pd(iy3,jy1);
1048             dz31             = _mm_sub_pd(iz3,jz1);
1049             dx32             = _mm_sub_pd(ix3,jx2);
1050             dy32             = _mm_sub_pd(iy3,jy2);
1051             dz32             = _mm_sub_pd(iz3,jz2);
1052             dx33             = _mm_sub_pd(ix3,jx3);
1053             dy33             = _mm_sub_pd(iy3,jy3);
1054             dz33             = _mm_sub_pd(iz3,jz3);
1055
1056             /* Calculate squared distance and things based on it */
1057             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1058             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1059             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1060             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1061             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1062             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1063             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1064             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1065             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1066
1067             rinv11           = sse41_invsqrt_d(rsq11);
1068             rinv12           = sse41_invsqrt_d(rsq12);
1069             rinv13           = sse41_invsqrt_d(rsq13);
1070             rinv21           = sse41_invsqrt_d(rsq21);
1071             rinv22           = sse41_invsqrt_d(rsq22);
1072             rinv23           = sse41_invsqrt_d(rsq23);
1073             rinv31           = sse41_invsqrt_d(rsq31);
1074             rinv32           = sse41_invsqrt_d(rsq32);
1075             rinv33           = sse41_invsqrt_d(rsq33);
1076
1077             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1078             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1079             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1080             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1081             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1082             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1083             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1084             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1085             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1086
1087             fjx1             = _mm_setzero_pd();
1088             fjy1             = _mm_setzero_pd();
1089             fjz1             = _mm_setzero_pd();
1090             fjx2             = _mm_setzero_pd();
1091             fjy2             = _mm_setzero_pd();
1092             fjz2             = _mm_setzero_pd();
1093             fjx3             = _mm_setzero_pd();
1094             fjy3             = _mm_setzero_pd();
1095             fjz3             = _mm_setzero_pd();
1096
1097             /**************************
1098              * CALCULATE INTERACTIONS *
1099              **************************/
1100
1101             /* COULOMB ELECTROSTATICS */
1102             velec            = _mm_mul_pd(qq11,rinv11);
1103             felec            = _mm_mul_pd(velec,rinvsq11);
1104
1105             fscal            = felec;
1106
1107             /* Calculate temporary vectorial force */
1108             tx               = _mm_mul_pd(fscal,dx11);
1109             ty               = _mm_mul_pd(fscal,dy11);
1110             tz               = _mm_mul_pd(fscal,dz11);
1111
1112             /* Update vectorial force */
1113             fix1             = _mm_add_pd(fix1,tx);
1114             fiy1             = _mm_add_pd(fiy1,ty);
1115             fiz1             = _mm_add_pd(fiz1,tz);
1116
1117             fjx1             = _mm_add_pd(fjx1,tx);
1118             fjy1             = _mm_add_pd(fjy1,ty);
1119             fjz1             = _mm_add_pd(fjz1,tz);
1120
1121             /**************************
1122              * CALCULATE INTERACTIONS *
1123              **************************/
1124
1125             /* COULOMB ELECTROSTATICS */
1126             velec            = _mm_mul_pd(qq12,rinv12);
1127             felec            = _mm_mul_pd(velec,rinvsq12);
1128
1129             fscal            = felec;
1130
1131             /* Calculate temporary vectorial force */
1132             tx               = _mm_mul_pd(fscal,dx12);
1133             ty               = _mm_mul_pd(fscal,dy12);
1134             tz               = _mm_mul_pd(fscal,dz12);
1135
1136             /* Update vectorial force */
1137             fix1             = _mm_add_pd(fix1,tx);
1138             fiy1             = _mm_add_pd(fiy1,ty);
1139             fiz1             = _mm_add_pd(fiz1,tz);
1140
1141             fjx2             = _mm_add_pd(fjx2,tx);
1142             fjy2             = _mm_add_pd(fjy2,ty);
1143             fjz2             = _mm_add_pd(fjz2,tz);
1144
1145             /**************************
1146              * CALCULATE INTERACTIONS *
1147              **************************/
1148
1149             /* COULOMB ELECTROSTATICS */
1150             velec            = _mm_mul_pd(qq13,rinv13);
1151             felec            = _mm_mul_pd(velec,rinvsq13);
1152
1153             fscal            = felec;
1154
1155             /* Calculate temporary vectorial force */
1156             tx               = _mm_mul_pd(fscal,dx13);
1157             ty               = _mm_mul_pd(fscal,dy13);
1158             tz               = _mm_mul_pd(fscal,dz13);
1159
1160             /* Update vectorial force */
1161             fix1             = _mm_add_pd(fix1,tx);
1162             fiy1             = _mm_add_pd(fiy1,ty);
1163             fiz1             = _mm_add_pd(fiz1,tz);
1164
1165             fjx3             = _mm_add_pd(fjx3,tx);
1166             fjy3             = _mm_add_pd(fjy3,ty);
1167             fjz3             = _mm_add_pd(fjz3,tz);
1168
1169             /**************************
1170              * CALCULATE INTERACTIONS *
1171              **************************/
1172
1173             /* COULOMB ELECTROSTATICS */
1174             velec            = _mm_mul_pd(qq21,rinv21);
1175             felec            = _mm_mul_pd(velec,rinvsq21);
1176
1177             fscal            = felec;
1178
1179             /* Calculate temporary vectorial force */
1180             tx               = _mm_mul_pd(fscal,dx21);
1181             ty               = _mm_mul_pd(fscal,dy21);
1182             tz               = _mm_mul_pd(fscal,dz21);
1183
1184             /* Update vectorial force */
1185             fix2             = _mm_add_pd(fix2,tx);
1186             fiy2             = _mm_add_pd(fiy2,ty);
1187             fiz2             = _mm_add_pd(fiz2,tz);
1188
1189             fjx1             = _mm_add_pd(fjx1,tx);
1190             fjy1             = _mm_add_pd(fjy1,ty);
1191             fjz1             = _mm_add_pd(fjz1,tz);
1192
1193             /**************************
1194              * CALCULATE INTERACTIONS *
1195              **************************/
1196
1197             /* COULOMB ELECTROSTATICS */
1198             velec            = _mm_mul_pd(qq22,rinv22);
1199             felec            = _mm_mul_pd(velec,rinvsq22);
1200
1201             fscal            = felec;
1202
1203             /* Calculate temporary vectorial force */
1204             tx               = _mm_mul_pd(fscal,dx22);
1205             ty               = _mm_mul_pd(fscal,dy22);
1206             tz               = _mm_mul_pd(fscal,dz22);
1207
1208             /* Update vectorial force */
1209             fix2             = _mm_add_pd(fix2,tx);
1210             fiy2             = _mm_add_pd(fiy2,ty);
1211             fiz2             = _mm_add_pd(fiz2,tz);
1212
1213             fjx2             = _mm_add_pd(fjx2,tx);
1214             fjy2             = _mm_add_pd(fjy2,ty);
1215             fjz2             = _mm_add_pd(fjz2,tz);
1216
1217             /**************************
1218              * CALCULATE INTERACTIONS *
1219              **************************/
1220
1221             /* COULOMB ELECTROSTATICS */
1222             velec            = _mm_mul_pd(qq23,rinv23);
1223             felec            = _mm_mul_pd(velec,rinvsq23);
1224
1225             fscal            = felec;
1226
1227             /* Calculate temporary vectorial force */
1228             tx               = _mm_mul_pd(fscal,dx23);
1229             ty               = _mm_mul_pd(fscal,dy23);
1230             tz               = _mm_mul_pd(fscal,dz23);
1231
1232             /* Update vectorial force */
1233             fix2             = _mm_add_pd(fix2,tx);
1234             fiy2             = _mm_add_pd(fiy2,ty);
1235             fiz2             = _mm_add_pd(fiz2,tz);
1236
1237             fjx3             = _mm_add_pd(fjx3,tx);
1238             fjy3             = _mm_add_pd(fjy3,ty);
1239             fjz3             = _mm_add_pd(fjz3,tz);
1240
1241             /**************************
1242              * CALCULATE INTERACTIONS *
1243              **************************/
1244
1245             /* COULOMB ELECTROSTATICS */
1246             velec            = _mm_mul_pd(qq31,rinv31);
1247             felec            = _mm_mul_pd(velec,rinvsq31);
1248
1249             fscal            = felec;
1250
1251             /* Calculate temporary vectorial force */
1252             tx               = _mm_mul_pd(fscal,dx31);
1253             ty               = _mm_mul_pd(fscal,dy31);
1254             tz               = _mm_mul_pd(fscal,dz31);
1255
1256             /* Update vectorial force */
1257             fix3             = _mm_add_pd(fix3,tx);
1258             fiy3             = _mm_add_pd(fiy3,ty);
1259             fiz3             = _mm_add_pd(fiz3,tz);
1260
1261             fjx1             = _mm_add_pd(fjx1,tx);
1262             fjy1             = _mm_add_pd(fjy1,ty);
1263             fjz1             = _mm_add_pd(fjz1,tz);
1264
1265             /**************************
1266              * CALCULATE INTERACTIONS *
1267              **************************/
1268
1269             /* COULOMB ELECTROSTATICS */
1270             velec            = _mm_mul_pd(qq32,rinv32);
1271             felec            = _mm_mul_pd(velec,rinvsq32);
1272
1273             fscal            = felec;
1274
1275             /* Calculate temporary vectorial force */
1276             tx               = _mm_mul_pd(fscal,dx32);
1277             ty               = _mm_mul_pd(fscal,dy32);
1278             tz               = _mm_mul_pd(fscal,dz32);
1279
1280             /* Update vectorial force */
1281             fix3             = _mm_add_pd(fix3,tx);
1282             fiy3             = _mm_add_pd(fiy3,ty);
1283             fiz3             = _mm_add_pd(fiz3,tz);
1284
1285             fjx2             = _mm_add_pd(fjx2,tx);
1286             fjy2             = _mm_add_pd(fjy2,ty);
1287             fjz2             = _mm_add_pd(fjz2,tz);
1288
1289             /**************************
1290              * CALCULATE INTERACTIONS *
1291              **************************/
1292
1293             /* COULOMB ELECTROSTATICS */
1294             velec            = _mm_mul_pd(qq33,rinv33);
1295             felec            = _mm_mul_pd(velec,rinvsq33);
1296
1297             fscal            = felec;
1298
1299             /* Calculate temporary vectorial force */
1300             tx               = _mm_mul_pd(fscal,dx33);
1301             ty               = _mm_mul_pd(fscal,dy33);
1302             tz               = _mm_mul_pd(fscal,dz33);
1303
1304             /* Update vectorial force */
1305             fix3             = _mm_add_pd(fix3,tx);
1306             fiy3             = _mm_add_pd(fiy3,ty);
1307             fiz3             = _mm_add_pd(fiz3,tz);
1308
1309             fjx3             = _mm_add_pd(fjx3,tx);
1310             fjy3             = _mm_add_pd(fjy3,ty);
1311             fjz3             = _mm_add_pd(fjz3,tz);
1312
1313             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1314
1315             /* Inner loop uses 243 flops */
1316         }
1317
1318         if(jidx<j_index_end)
1319         {
1320
1321             jnrA             = jjnr[jidx];
1322             j_coord_offsetA  = DIM*jnrA;
1323
1324             /* load j atom coordinates */
1325             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1326                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1327
1328             /* Calculate displacement vector */
1329             dx11             = _mm_sub_pd(ix1,jx1);
1330             dy11             = _mm_sub_pd(iy1,jy1);
1331             dz11             = _mm_sub_pd(iz1,jz1);
1332             dx12             = _mm_sub_pd(ix1,jx2);
1333             dy12             = _mm_sub_pd(iy1,jy2);
1334             dz12             = _mm_sub_pd(iz1,jz2);
1335             dx13             = _mm_sub_pd(ix1,jx3);
1336             dy13             = _mm_sub_pd(iy1,jy3);
1337             dz13             = _mm_sub_pd(iz1,jz3);
1338             dx21             = _mm_sub_pd(ix2,jx1);
1339             dy21             = _mm_sub_pd(iy2,jy1);
1340             dz21             = _mm_sub_pd(iz2,jz1);
1341             dx22             = _mm_sub_pd(ix2,jx2);
1342             dy22             = _mm_sub_pd(iy2,jy2);
1343             dz22             = _mm_sub_pd(iz2,jz2);
1344             dx23             = _mm_sub_pd(ix2,jx3);
1345             dy23             = _mm_sub_pd(iy2,jy3);
1346             dz23             = _mm_sub_pd(iz2,jz3);
1347             dx31             = _mm_sub_pd(ix3,jx1);
1348             dy31             = _mm_sub_pd(iy3,jy1);
1349             dz31             = _mm_sub_pd(iz3,jz1);
1350             dx32             = _mm_sub_pd(ix3,jx2);
1351             dy32             = _mm_sub_pd(iy3,jy2);
1352             dz32             = _mm_sub_pd(iz3,jz2);
1353             dx33             = _mm_sub_pd(ix3,jx3);
1354             dy33             = _mm_sub_pd(iy3,jy3);
1355             dz33             = _mm_sub_pd(iz3,jz3);
1356
1357             /* Calculate squared distance and things based on it */
1358             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1359             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1360             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1361             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1362             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1363             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1364             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1365             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1366             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1367
1368             rinv11           = sse41_invsqrt_d(rsq11);
1369             rinv12           = sse41_invsqrt_d(rsq12);
1370             rinv13           = sse41_invsqrt_d(rsq13);
1371             rinv21           = sse41_invsqrt_d(rsq21);
1372             rinv22           = sse41_invsqrt_d(rsq22);
1373             rinv23           = sse41_invsqrt_d(rsq23);
1374             rinv31           = sse41_invsqrt_d(rsq31);
1375             rinv32           = sse41_invsqrt_d(rsq32);
1376             rinv33           = sse41_invsqrt_d(rsq33);
1377
1378             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1379             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1380             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1381             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1382             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1383             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1384             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1385             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1386             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1387
1388             fjx1             = _mm_setzero_pd();
1389             fjy1             = _mm_setzero_pd();
1390             fjz1             = _mm_setzero_pd();
1391             fjx2             = _mm_setzero_pd();
1392             fjy2             = _mm_setzero_pd();
1393             fjz2             = _mm_setzero_pd();
1394             fjx3             = _mm_setzero_pd();
1395             fjy3             = _mm_setzero_pd();
1396             fjz3             = _mm_setzero_pd();
1397
1398             /**************************
1399              * CALCULATE INTERACTIONS *
1400              **************************/
1401
1402             /* COULOMB ELECTROSTATICS */
1403             velec            = _mm_mul_pd(qq11,rinv11);
1404             felec            = _mm_mul_pd(velec,rinvsq11);
1405
1406             fscal            = felec;
1407
1408             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1409
1410             /* Calculate temporary vectorial force */
1411             tx               = _mm_mul_pd(fscal,dx11);
1412             ty               = _mm_mul_pd(fscal,dy11);
1413             tz               = _mm_mul_pd(fscal,dz11);
1414
1415             /* Update vectorial force */
1416             fix1             = _mm_add_pd(fix1,tx);
1417             fiy1             = _mm_add_pd(fiy1,ty);
1418             fiz1             = _mm_add_pd(fiz1,tz);
1419
1420             fjx1             = _mm_add_pd(fjx1,tx);
1421             fjy1             = _mm_add_pd(fjy1,ty);
1422             fjz1             = _mm_add_pd(fjz1,tz);
1423
1424             /**************************
1425              * CALCULATE INTERACTIONS *
1426              **************************/
1427
1428             /* COULOMB ELECTROSTATICS */
1429             velec            = _mm_mul_pd(qq12,rinv12);
1430             felec            = _mm_mul_pd(velec,rinvsq12);
1431
1432             fscal            = felec;
1433
1434             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1435
1436             /* Calculate temporary vectorial force */
1437             tx               = _mm_mul_pd(fscal,dx12);
1438             ty               = _mm_mul_pd(fscal,dy12);
1439             tz               = _mm_mul_pd(fscal,dz12);
1440
1441             /* Update vectorial force */
1442             fix1             = _mm_add_pd(fix1,tx);
1443             fiy1             = _mm_add_pd(fiy1,ty);
1444             fiz1             = _mm_add_pd(fiz1,tz);
1445
1446             fjx2             = _mm_add_pd(fjx2,tx);
1447             fjy2             = _mm_add_pd(fjy2,ty);
1448             fjz2             = _mm_add_pd(fjz2,tz);
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             /* COULOMB ELECTROSTATICS */
1455             velec            = _mm_mul_pd(qq13,rinv13);
1456             felec            = _mm_mul_pd(velec,rinvsq13);
1457
1458             fscal            = felec;
1459
1460             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1461
1462             /* Calculate temporary vectorial force */
1463             tx               = _mm_mul_pd(fscal,dx13);
1464             ty               = _mm_mul_pd(fscal,dy13);
1465             tz               = _mm_mul_pd(fscal,dz13);
1466
1467             /* Update vectorial force */
1468             fix1             = _mm_add_pd(fix1,tx);
1469             fiy1             = _mm_add_pd(fiy1,ty);
1470             fiz1             = _mm_add_pd(fiz1,tz);
1471
1472             fjx3             = _mm_add_pd(fjx3,tx);
1473             fjy3             = _mm_add_pd(fjy3,ty);
1474             fjz3             = _mm_add_pd(fjz3,tz);
1475
1476             /**************************
1477              * CALCULATE INTERACTIONS *
1478              **************************/
1479
1480             /* COULOMB ELECTROSTATICS */
1481             velec            = _mm_mul_pd(qq21,rinv21);
1482             felec            = _mm_mul_pd(velec,rinvsq21);
1483
1484             fscal            = felec;
1485
1486             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1487
1488             /* Calculate temporary vectorial force */
1489             tx               = _mm_mul_pd(fscal,dx21);
1490             ty               = _mm_mul_pd(fscal,dy21);
1491             tz               = _mm_mul_pd(fscal,dz21);
1492
1493             /* Update vectorial force */
1494             fix2             = _mm_add_pd(fix2,tx);
1495             fiy2             = _mm_add_pd(fiy2,ty);
1496             fiz2             = _mm_add_pd(fiz2,tz);
1497
1498             fjx1             = _mm_add_pd(fjx1,tx);
1499             fjy1             = _mm_add_pd(fjy1,ty);
1500             fjz1             = _mm_add_pd(fjz1,tz);
1501
1502             /**************************
1503              * CALCULATE INTERACTIONS *
1504              **************************/
1505
1506             /* COULOMB ELECTROSTATICS */
1507             velec            = _mm_mul_pd(qq22,rinv22);
1508             felec            = _mm_mul_pd(velec,rinvsq22);
1509
1510             fscal            = felec;
1511
1512             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1513
1514             /* Calculate temporary vectorial force */
1515             tx               = _mm_mul_pd(fscal,dx22);
1516             ty               = _mm_mul_pd(fscal,dy22);
1517             tz               = _mm_mul_pd(fscal,dz22);
1518
1519             /* Update vectorial force */
1520             fix2             = _mm_add_pd(fix2,tx);
1521             fiy2             = _mm_add_pd(fiy2,ty);
1522             fiz2             = _mm_add_pd(fiz2,tz);
1523
1524             fjx2             = _mm_add_pd(fjx2,tx);
1525             fjy2             = _mm_add_pd(fjy2,ty);
1526             fjz2             = _mm_add_pd(fjz2,tz);
1527
1528             /**************************
1529              * CALCULATE INTERACTIONS *
1530              **************************/
1531
1532             /* COULOMB ELECTROSTATICS */
1533             velec            = _mm_mul_pd(qq23,rinv23);
1534             felec            = _mm_mul_pd(velec,rinvsq23);
1535
1536             fscal            = felec;
1537
1538             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1539
1540             /* Calculate temporary vectorial force */
1541             tx               = _mm_mul_pd(fscal,dx23);
1542             ty               = _mm_mul_pd(fscal,dy23);
1543             tz               = _mm_mul_pd(fscal,dz23);
1544
1545             /* Update vectorial force */
1546             fix2             = _mm_add_pd(fix2,tx);
1547             fiy2             = _mm_add_pd(fiy2,ty);
1548             fiz2             = _mm_add_pd(fiz2,tz);
1549
1550             fjx3             = _mm_add_pd(fjx3,tx);
1551             fjy3             = _mm_add_pd(fjy3,ty);
1552             fjz3             = _mm_add_pd(fjz3,tz);
1553
1554             /**************************
1555              * CALCULATE INTERACTIONS *
1556              **************************/
1557
1558             /* COULOMB ELECTROSTATICS */
1559             velec            = _mm_mul_pd(qq31,rinv31);
1560             felec            = _mm_mul_pd(velec,rinvsq31);
1561
1562             fscal            = felec;
1563
1564             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1565
1566             /* Calculate temporary vectorial force */
1567             tx               = _mm_mul_pd(fscal,dx31);
1568             ty               = _mm_mul_pd(fscal,dy31);
1569             tz               = _mm_mul_pd(fscal,dz31);
1570
1571             /* Update vectorial force */
1572             fix3             = _mm_add_pd(fix3,tx);
1573             fiy3             = _mm_add_pd(fiy3,ty);
1574             fiz3             = _mm_add_pd(fiz3,tz);
1575
1576             fjx1             = _mm_add_pd(fjx1,tx);
1577             fjy1             = _mm_add_pd(fjy1,ty);
1578             fjz1             = _mm_add_pd(fjz1,tz);
1579
1580             /**************************
1581              * CALCULATE INTERACTIONS *
1582              **************************/
1583
1584             /* COULOMB ELECTROSTATICS */
1585             velec            = _mm_mul_pd(qq32,rinv32);
1586             felec            = _mm_mul_pd(velec,rinvsq32);
1587
1588             fscal            = felec;
1589
1590             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1591
1592             /* Calculate temporary vectorial force */
1593             tx               = _mm_mul_pd(fscal,dx32);
1594             ty               = _mm_mul_pd(fscal,dy32);
1595             tz               = _mm_mul_pd(fscal,dz32);
1596
1597             /* Update vectorial force */
1598             fix3             = _mm_add_pd(fix3,tx);
1599             fiy3             = _mm_add_pd(fiy3,ty);
1600             fiz3             = _mm_add_pd(fiz3,tz);
1601
1602             fjx2             = _mm_add_pd(fjx2,tx);
1603             fjy2             = _mm_add_pd(fjy2,ty);
1604             fjz2             = _mm_add_pd(fjz2,tz);
1605
1606             /**************************
1607              * CALCULATE INTERACTIONS *
1608              **************************/
1609
1610             /* COULOMB ELECTROSTATICS */
1611             velec            = _mm_mul_pd(qq33,rinv33);
1612             felec            = _mm_mul_pd(velec,rinvsq33);
1613
1614             fscal            = felec;
1615
1616             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1617
1618             /* Calculate temporary vectorial force */
1619             tx               = _mm_mul_pd(fscal,dx33);
1620             ty               = _mm_mul_pd(fscal,dy33);
1621             tz               = _mm_mul_pd(fscal,dz33);
1622
1623             /* Update vectorial force */
1624             fix3             = _mm_add_pd(fix3,tx);
1625             fiy3             = _mm_add_pd(fiy3,ty);
1626             fiz3             = _mm_add_pd(fiz3,tz);
1627
1628             fjx3             = _mm_add_pd(fjx3,tx);
1629             fjy3             = _mm_add_pd(fjy3,ty);
1630             fjz3             = _mm_add_pd(fjz3,tz);
1631
1632             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1633
1634             /* Inner loop uses 243 flops */
1635         }
1636
1637         /* End of innermost loop */
1638
1639         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1640                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1641
1642         /* Increment number of inner iterations */
1643         inneriter                  += j_index_end - j_index_start;
1644
1645         /* Outer loop uses 18 flops */
1646     }
1647
1648     /* Increment number of outer iterations */
1649     outeriter        += nri;
1650
1651     /* Update outer/inner flops */
1652
1653     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1654 }