src/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecCoul_VdwNone_GeomW4W4_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse4_1_double
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse4_1_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset1;
  67     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  68     int              vdwioffset2;
  69     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  70     int              vdwioffset3;
  71     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  72     int              vdwjidx1A,vdwjidx1B;
  73     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  74     int              vdwjidx2A,vdwjidx2B;
  75     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  76     int              vdwjidx3A,vdwjidx3B;
  77     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  78     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  79     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  80     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  81     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  82     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  83     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  84     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  85     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  86     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128d          dummy_mask,cutoff_mask;
  90     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  91     __m128d          one     = _mm_set1_pd(1.0);
  92     __m128d          two     = _mm_set1_pd(2.0);
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = _mm_set1_pd(fr->epsfac);
 105     charge           = mdatoms->chargeA;
 106
 107     /* Setup water-specific parameters */
 108     inr              = nlist->iinr[0];
 109     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 110     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 111     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 112
 113     jq1              = _mm_set1_pd(charge[inr+1]);
 114     jq2              = _mm_set1_pd(charge[inr+2]);
 115     jq3              = _mm_set1_pd(charge[inr+3]);
 116     qq11             = _mm_mul_pd(iq1,jq1);
 117     qq12             = _mm_mul_pd(iq1,jq2);
 118     qq13             = _mm_mul_pd(iq1,jq3);
 119     qq21             = _mm_mul_pd(iq2,jq1);
 120     qq22             = _mm_mul_pd(iq2,jq2);
 121     qq23             = _mm_mul_pd(iq2,jq3);
 122     qq31             = _mm_mul_pd(iq3,jq1);
 123     qq32             = _mm_mul_pd(iq3,jq2);
 124     qq33             = _mm_mul_pd(iq3,jq3);
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     /* Start outer loop over neighborlists */
 135     for(iidx=0; iidx<nri; iidx++)
 136     {
 137         /* Load shift vector for this list */
 138         i_shift_offset   = DIM*shiftidx[iidx];
 139
 140         /* Load limits for loop over neighbors */
 141         j_index_start    = jindex[iidx];
 142         j_index_end      = jindex[iidx+1];
 143
 144         /* Get outer coordinate index */
 145         inr              = iinr[iidx];
 146         i_coord_offset   = DIM*inr;
 147
 148         /* Load i particle coords and add shift vector */
 149         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 150                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 151
 152         fix1             = _mm_setzero_pd();
 153         fiy1             = _mm_setzero_pd();
 154         fiz1             = _mm_setzero_pd();
 155         fix2             = _mm_setzero_pd();
 156         fiy2             = _mm_setzero_pd();
 157         fiz2             = _mm_setzero_pd();
 158         fix3             = _mm_setzero_pd();
 159         fiy3             = _mm_setzero_pd();
 160         fiz3             = _mm_setzero_pd();
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm_setzero_pd();
 164
 165         /* Start inner kernel loop */
 166         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 167         {
 168
 169             /* Get j neighbor index, and coordinate index */
 170             jnrA             = jjnr[jidx];
 171             jnrB             = jjnr[jidx+1];
 172             j_coord_offsetA  = DIM*jnrA;
 173             j_coord_offsetB  = DIM*jnrB;
 174
 175             /* load j atom coordinates */
 176             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 177                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 178
 179             /* Calculate displacement vector */
 180             dx11             = _mm_sub_pd(ix1,jx1);
 181             dy11             = _mm_sub_pd(iy1,jy1);
 182             dz11             = _mm_sub_pd(iz1,jz1);
 183             dx12             = _mm_sub_pd(ix1,jx2);
 184             dy12             = _mm_sub_pd(iy1,jy2);
 185             dz12             = _mm_sub_pd(iz1,jz2);
 186             dx13             = _mm_sub_pd(ix1,jx3);
 187             dy13             = _mm_sub_pd(iy1,jy3);
 188             dz13             = _mm_sub_pd(iz1,jz3);
 189             dx21             = _mm_sub_pd(ix2,jx1);
 190             dy21             = _mm_sub_pd(iy2,jy1);
 191             dz21             = _mm_sub_pd(iz2,jz1);
 192             dx22             = _mm_sub_pd(ix2,jx2);
 193             dy22             = _mm_sub_pd(iy2,jy2);
 194             dz22             = _mm_sub_pd(iz2,jz2);
 195             dx23             = _mm_sub_pd(ix2,jx3);
 196             dy23             = _mm_sub_pd(iy2,jy3);
 197             dz23             = _mm_sub_pd(iz2,jz3);
 198             dx31             = _mm_sub_pd(ix3,jx1);
 199             dy31             = _mm_sub_pd(iy3,jy1);
 200             dz31             = _mm_sub_pd(iz3,jz1);
 201             dx32             = _mm_sub_pd(ix3,jx2);
 202             dy32             = _mm_sub_pd(iy3,jy2);
 203             dz32             = _mm_sub_pd(iz3,jz2);
 204             dx33             = _mm_sub_pd(ix3,jx3);
 205             dy33             = _mm_sub_pd(iy3,jy3);
 206             dz33             = _mm_sub_pd(iz3,jz3);
 207
 208             /* Calculate squared distance and things based on it */
 209             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 210             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 211             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 212             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 213             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 214             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 215             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 216             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 217             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 218
 219             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 220             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 221             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 222             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 223             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 224             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 225             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 226             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 227             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 228
 229             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 230             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 231             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 232             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 233             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 234             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 235             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 236             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 237             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 238
 239             fjx1             = _mm_setzero_pd();
 240             fjy1             = _mm_setzero_pd();
 241             fjz1             = _mm_setzero_pd();
 242             fjx2             = _mm_setzero_pd();
 243             fjy2             = _mm_setzero_pd();
 244             fjz2             = _mm_setzero_pd();
 245             fjx3             = _mm_setzero_pd();
 246             fjy3             = _mm_setzero_pd();
 247             fjz3             = _mm_setzero_pd();
 248
 249             /**************************
 250              * CALCULATE INTERACTIONS *
 251              **************************/
 252
 253             /* COULOMB ELECTROSTATICS */
 254             velec            = _mm_mul_pd(qq11,rinv11);
 255             felec            = _mm_mul_pd(velec,rinvsq11);
 256
 257             /* Update potential sum for this i atom from the interaction with this j atom. */
 258             velecsum         = _mm_add_pd(velecsum,velec);
 259
 260             fscal            = felec;
 261
 262             /* Calculate temporary vectorial force */
 263             tx               = _mm_mul_pd(fscal,dx11);
 264             ty               = _mm_mul_pd(fscal,dy11);
 265             tz               = _mm_mul_pd(fscal,dz11);
 266
 267             /* Update vectorial force */
 268             fix1             = _mm_add_pd(fix1,tx);
 269             fiy1             = _mm_add_pd(fiy1,ty);
 270             fiz1             = _mm_add_pd(fiz1,tz);
 271
 272             fjx1             = _mm_add_pd(fjx1,tx);
 273             fjy1             = _mm_add_pd(fjy1,ty);
 274             fjz1             = _mm_add_pd(fjz1,tz);
 275
 276             /**************************
 277              * CALCULATE INTERACTIONS *
 278              **************************/
 279
 280             /* COULOMB ELECTROSTATICS */
 281             velec            = _mm_mul_pd(qq12,rinv12);
 282             felec            = _mm_mul_pd(velec,rinvsq12);
 283
 284             /* Update potential sum for this i atom from the interaction with this j atom. */
 285             velecsum         = _mm_add_pd(velecsum,velec);
 286
 287             fscal            = felec;
 288
 289             /* Calculate temporary vectorial force */
 290             tx               = _mm_mul_pd(fscal,dx12);
 291             ty               = _mm_mul_pd(fscal,dy12);
 292             tz               = _mm_mul_pd(fscal,dz12);
 293
 294             /* Update vectorial force */
 295             fix1             = _mm_add_pd(fix1,tx);
 296             fiy1             = _mm_add_pd(fiy1,ty);
 297             fiz1             = _mm_add_pd(fiz1,tz);
 298
 299             fjx2             = _mm_add_pd(fjx2,tx);
 300             fjy2             = _mm_add_pd(fjy2,ty);
 301             fjz2             = _mm_add_pd(fjz2,tz);
 302
 303             /**************************
 304              * CALCULATE INTERACTIONS *
 305              **************************/
 306
 307             /* COULOMB ELECTROSTATICS */
 308             velec            = _mm_mul_pd(qq13,rinv13);
 309             felec            = _mm_mul_pd(velec,rinvsq13);
 310
 311             /* Update potential sum for this i atom from the interaction with this j atom. */
 312             velecsum         = _mm_add_pd(velecsum,velec);
 313
 314             fscal            = felec;
 315
 316             /* Calculate temporary vectorial force */
 317             tx               = _mm_mul_pd(fscal,dx13);
 318             ty               = _mm_mul_pd(fscal,dy13);
 319             tz               = _mm_mul_pd(fscal,dz13);
 320
 321             /* Update vectorial force */
 322             fix1             = _mm_add_pd(fix1,tx);
 323             fiy1             = _mm_add_pd(fiy1,ty);
 324             fiz1             = _mm_add_pd(fiz1,tz);
 325
 326             fjx3             = _mm_add_pd(fjx3,tx);
 327             fjy3             = _mm_add_pd(fjy3,ty);
 328             fjz3             = _mm_add_pd(fjz3,tz);
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             /* COULOMB ELECTROSTATICS */
 335             velec            = _mm_mul_pd(qq21,rinv21);
 336             felec            = _mm_mul_pd(velec,rinvsq21);
 337
 338             /* Update potential sum for this i atom from the interaction with this j atom. */
 339             velecsum         = _mm_add_pd(velecsum,velec);
 340
 341             fscal            = felec;
 342
 343             /* Calculate temporary vectorial force */
 344             tx               = _mm_mul_pd(fscal,dx21);
 345             ty               = _mm_mul_pd(fscal,dy21);
 346             tz               = _mm_mul_pd(fscal,dz21);
 347
 348             /* Update vectorial force */
 349             fix2             = _mm_add_pd(fix2,tx);
 350             fiy2             = _mm_add_pd(fiy2,ty);
 351             fiz2             = _mm_add_pd(fiz2,tz);
 352
 353             fjx1             = _mm_add_pd(fjx1,tx);
 354             fjy1             = _mm_add_pd(fjy1,ty);
 355             fjz1             = _mm_add_pd(fjz1,tz);
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             /* COULOMB ELECTROSTATICS */
 362             velec            = _mm_mul_pd(qq22,rinv22);
 363             felec            = _mm_mul_pd(velec,rinvsq22);
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             velecsum         = _mm_add_pd(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370             /* Calculate temporary vectorial force */
 371             tx               = _mm_mul_pd(fscal,dx22);
 372             ty               = _mm_mul_pd(fscal,dy22);
 373             tz               = _mm_mul_pd(fscal,dz22);
 374
 375             /* Update vectorial force */
 376             fix2             = _mm_add_pd(fix2,tx);
 377             fiy2             = _mm_add_pd(fiy2,ty);
 378             fiz2             = _mm_add_pd(fiz2,tz);
 379
 380             fjx2             = _mm_add_pd(fjx2,tx);
 381             fjy2             = _mm_add_pd(fjy2,ty);
 382             fjz2             = _mm_add_pd(fjz2,tz);
 383
 384             /**************************
 385              * CALCULATE INTERACTIONS *
 386              **************************/
 387
 388             /* COULOMB ELECTROSTATICS */
 389             velec            = _mm_mul_pd(qq23,rinv23);
 390             felec            = _mm_mul_pd(velec,rinvsq23);
 391
 392             /* Update potential sum for this i atom from the interaction with this j atom. */
 393             velecsum         = _mm_add_pd(velecsum,velec);
 394
 395             fscal            = felec;
 396
 397             /* Calculate temporary vectorial force */
 398             tx               = _mm_mul_pd(fscal,dx23);
 399             ty               = _mm_mul_pd(fscal,dy23);
 400             tz               = _mm_mul_pd(fscal,dz23);
 401
 402             /* Update vectorial force */
 403             fix2             = _mm_add_pd(fix2,tx);
 404             fiy2             = _mm_add_pd(fiy2,ty);
 405             fiz2             = _mm_add_pd(fiz2,tz);
 406
 407             fjx3             = _mm_add_pd(fjx3,tx);
 408             fjy3             = _mm_add_pd(fjy3,ty);
 409             fjz3             = _mm_add_pd(fjz3,tz);
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             /* COULOMB ELECTROSTATICS */
 416             velec            = _mm_mul_pd(qq31,rinv31);
 417             felec            = _mm_mul_pd(velec,rinvsq31);
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velecsum         = _mm_add_pd(velecsum,velec);
 421
 422             fscal            = felec;
 423
 424             /* Calculate temporary vectorial force */
 425             tx               = _mm_mul_pd(fscal,dx31);
 426             ty               = _mm_mul_pd(fscal,dy31);
 427             tz               = _mm_mul_pd(fscal,dz31);
 428
 429             /* Update vectorial force */
 430             fix3             = _mm_add_pd(fix3,tx);
 431             fiy3             = _mm_add_pd(fiy3,ty);
 432             fiz3             = _mm_add_pd(fiz3,tz);
 433
 434             fjx1             = _mm_add_pd(fjx1,tx);
 435             fjy1             = _mm_add_pd(fjy1,ty);
 436             fjz1             = _mm_add_pd(fjz1,tz);
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             /* COULOMB ELECTROSTATICS */
 443             velec            = _mm_mul_pd(qq32,rinv32);
 444             felec            = _mm_mul_pd(velec,rinvsq32);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velecsum         = _mm_add_pd(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             /* Calculate temporary vectorial force */
 452             tx               = _mm_mul_pd(fscal,dx32);
 453             ty               = _mm_mul_pd(fscal,dy32);
 454             tz               = _mm_mul_pd(fscal,dz32);
 455
 456             /* Update vectorial force */
 457             fix3             = _mm_add_pd(fix3,tx);
 458             fiy3             = _mm_add_pd(fiy3,ty);
 459             fiz3             = _mm_add_pd(fiz3,tz);
 460
 461             fjx2             = _mm_add_pd(fjx2,tx);
 462             fjy2             = _mm_add_pd(fjy2,ty);
 463             fjz2             = _mm_add_pd(fjz2,tz);
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             /* COULOMB ELECTROSTATICS */
 470             velec            = _mm_mul_pd(qq33,rinv33);
 471             felec            = _mm_mul_pd(velec,rinvsq33);
 472
 473             /* Update potential sum for this i atom from the interaction with this j atom. */
 474             velecsum         = _mm_add_pd(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478             /* Calculate temporary vectorial force */
 479             tx               = _mm_mul_pd(fscal,dx33);
 480             ty               = _mm_mul_pd(fscal,dy33);
 481             tz               = _mm_mul_pd(fscal,dz33);
 482
 483             /* Update vectorial force */
 484             fix3             = _mm_add_pd(fix3,tx);
 485             fiy3             = _mm_add_pd(fiy3,ty);
 486             fiz3             = _mm_add_pd(fiz3,tz);
 487
 488             fjx3             = _mm_add_pd(fjx3,tx);
 489             fjy3             = _mm_add_pd(fjy3,ty);
 490             fjz3             = _mm_add_pd(fjz3,tz);
 491
 492             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);
 493
 494             /* Inner loop uses 252 flops */
 495         }
 496
 497         if(jidx<j_index_end)
 498         {
 499
 500             jnrA             = jjnr[jidx];
 501             j_coord_offsetA  = DIM*jnrA;
 502
 503             /* load j atom coordinates */
 504             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 505                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 506
 507             /* Calculate displacement vector */
 508             dx11             = _mm_sub_pd(ix1,jx1);
 509             dy11             = _mm_sub_pd(iy1,jy1);
 510             dz11             = _mm_sub_pd(iz1,jz1);
 511             dx12             = _mm_sub_pd(ix1,jx2);
 512             dy12             = _mm_sub_pd(iy1,jy2);
 513             dz12             = _mm_sub_pd(iz1,jz2);
 514             dx13             = _mm_sub_pd(ix1,jx3);
 515             dy13             = _mm_sub_pd(iy1,jy3);
 516             dz13             = _mm_sub_pd(iz1,jz3);
 517             dx21             = _mm_sub_pd(ix2,jx1);
 518             dy21             = _mm_sub_pd(iy2,jy1);
 519             dz21             = _mm_sub_pd(iz2,jz1);
 520             dx22             = _mm_sub_pd(ix2,jx2);
 521             dy22             = _mm_sub_pd(iy2,jy2);
 522             dz22             = _mm_sub_pd(iz2,jz2);
 523             dx23             = _mm_sub_pd(ix2,jx3);
 524             dy23             = _mm_sub_pd(iy2,jy3);
 525             dz23             = _mm_sub_pd(iz2,jz3);
 526             dx31             = _mm_sub_pd(ix3,jx1);
 527             dy31             = _mm_sub_pd(iy3,jy1);
 528             dz31             = _mm_sub_pd(iz3,jz1);
 529             dx32             = _mm_sub_pd(ix3,jx2);
 530             dy32             = _mm_sub_pd(iy3,jy2);
 531             dz32             = _mm_sub_pd(iz3,jz2);
 532             dx33             = _mm_sub_pd(ix3,jx3);
 533             dy33             = _mm_sub_pd(iy3,jy3);
 534             dz33             = _mm_sub_pd(iz3,jz3);
 535
 536             /* Calculate squared distance and things based on it */
 537             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 538             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 539             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 540             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 541             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 542             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 543             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 544             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 545             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 546
 547             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 548             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 549             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 550             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 551             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 552             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 553             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 554             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 555             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 556
 557             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 558             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 559             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 560             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 561             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 562             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 563             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 564             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 565             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 566
 567             fjx1             = _mm_setzero_pd();
 568             fjy1             = _mm_setzero_pd();
 569             fjz1             = _mm_setzero_pd();
 570             fjx2             = _mm_setzero_pd();
 571             fjy2             = _mm_setzero_pd();
 572             fjz2             = _mm_setzero_pd();
 573             fjx3             = _mm_setzero_pd();
 574             fjy3             = _mm_setzero_pd();
 575             fjz3             = _mm_setzero_pd();
 576
 577             /**************************
 578              * CALCULATE INTERACTIONS *
 579              **************************/
 580
 581             /* COULOMB ELECTROSTATICS */
 582             velec            = _mm_mul_pd(qq11,rinv11);
 583             felec            = _mm_mul_pd(velec,rinvsq11);
 584
 585             /* Update potential sum for this i atom from the interaction with this j atom. */
 586             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 587             velecsum         = _mm_add_pd(velecsum,velec);
 588
 589             fscal            = felec;
 590
 591             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 592
 593             /* Calculate temporary vectorial force */
 594             tx               = _mm_mul_pd(fscal,dx11);
 595             ty               = _mm_mul_pd(fscal,dy11);
 596             tz               = _mm_mul_pd(fscal,dz11);
 597
 598             /* Update vectorial force */
 599             fix1             = _mm_add_pd(fix1,tx);
 600             fiy1             = _mm_add_pd(fiy1,ty);
 601             fiz1             = _mm_add_pd(fiz1,tz);
 602
 603             fjx1             = _mm_add_pd(fjx1,tx);
 604             fjy1             = _mm_add_pd(fjy1,ty);
 605             fjz1             = _mm_add_pd(fjz1,tz);
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             /* COULOMB ELECTROSTATICS */
 612             velec            = _mm_mul_pd(qq12,rinv12);
 613             felec            = _mm_mul_pd(velec,rinvsq12);
 614
 615             /* Update potential sum for this i atom from the interaction with this j atom. */
 616             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 617             velecsum         = _mm_add_pd(velecsum,velec);
 618
 619             fscal            = felec;
 620
 621             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 622
 623             /* Calculate temporary vectorial force */
 624             tx               = _mm_mul_pd(fscal,dx12);
 625             ty               = _mm_mul_pd(fscal,dy12);
 626             tz               = _mm_mul_pd(fscal,dz12);
 627
 628             /* Update vectorial force */
 629             fix1             = _mm_add_pd(fix1,tx);
 630             fiy1             = _mm_add_pd(fiy1,ty);
 631             fiz1             = _mm_add_pd(fiz1,tz);
 632
 633             fjx2             = _mm_add_pd(fjx2,tx);
 634             fjy2             = _mm_add_pd(fjy2,ty);
 635             fjz2             = _mm_add_pd(fjz2,tz);
 636
 637             /**************************
 638              * CALCULATE INTERACTIONS *
 639              **************************/
 640
 641             /* COULOMB ELECTROSTATICS */
 642             velec            = _mm_mul_pd(qq13,rinv13);
 643             felec            = _mm_mul_pd(velec,rinvsq13);
 644
 645             /* Update potential sum for this i atom from the interaction with this j atom. */
 646             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 647             velecsum         = _mm_add_pd(velecsum,velec);
 648
 649             fscal            = felec;
 650
 651             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 652
 653             /* Calculate temporary vectorial force */
 654             tx               = _mm_mul_pd(fscal,dx13);
 655             ty               = _mm_mul_pd(fscal,dy13);
 656             tz               = _mm_mul_pd(fscal,dz13);
 657
 658             /* Update vectorial force */
 659             fix1             = _mm_add_pd(fix1,tx);
 660             fiy1             = _mm_add_pd(fiy1,ty);
 661             fiz1             = _mm_add_pd(fiz1,tz);
 662
 663             fjx3             = _mm_add_pd(fjx3,tx);
 664             fjy3             = _mm_add_pd(fjy3,ty);
 665             fjz3             = _mm_add_pd(fjz3,tz);
 666
 667             /**************************
 668              * CALCULATE INTERACTIONS *
 669              **************************/
 670
 671             /* COULOMB ELECTROSTATICS */
 672             velec            = _mm_mul_pd(qq21,rinv21);
 673             felec            = _mm_mul_pd(velec,rinvsq21);
 674
 675             /* Update potential sum for this i atom from the interaction with this j atom. */
 676             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 677             velecsum         = _mm_add_pd(velecsum,velec);
 678
 679             fscal            = felec;
 680
 681             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 682
 683             /* Calculate temporary vectorial force */
 684             tx               = _mm_mul_pd(fscal,dx21);
 685             ty               = _mm_mul_pd(fscal,dy21);
 686             tz               = _mm_mul_pd(fscal,dz21);
 687
 688             /* Update vectorial force */
 689             fix2             = _mm_add_pd(fix2,tx);
 690             fiy2             = _mm_add_pd(fiy2,ty);
 691             fiz2             = _mm_add_pd(fiz2,tz);
 692
 693             fjx1             = _mm_add_pd(fjx1,tx);
 694             fjy1             = _mm_add_pd(fjy1,ty);
 695             fjz1             = _mm_add_pd(fjz1,tz);
 696
 697             /**************************
 698              * CALCULATE INTERACTIONS *
 699              **************************/
 700
 701             /* COULOMB ELECTROSTATICS */
 702             velec            = _mm_mul_pd(qq22,rinv22);
 703             felec            = _mm_mul_pd(velec,rinvsq22);
 704
 705             /* Update potential sum for this i atom from the interaction with this j atom. */
 706             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 707             velecsum         = _mm_add_pd(velecsum,velec);
 708
 709             fscal            = felec;
 710
 711             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 712
 713             /* Calculate temporary vectorial force */
 714             tx               = _mm_mul_pd(fscal,dx22);
 715             ty               = _mm_mul_pd(fscal,dy22);
 716             tz               = _mm_mul_pd(fscal,dz22);
 717
 718             /* Update vectorial force */
 719             fix2             = _mm_add_pd(fix2,tx);
 720             fiy2             = _mm_add_pd(fiy2,ty);
 721             fiz2             = _mm_add_pd(fiz2,tz);
 722
 723             fjx2             = _mm_add_pd(fjx2,tx);
 724             fjy2             = _mm_add_pd(fjy2,ty);
 725             fjz2             = _mm_add_pd(fjz2,tz);
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             /* COULOMB ELECTROSTATICS */
 732             velec            = _mm_mul_pd(qq23,rinv23);
 733             felec            = _mm_mul_pd(velec,rinvsq23);
 734
 735             /* Update potential sum for this i atom from the interaction with this j atom. */
 736             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 737             velecsum         = _mm_add_pd(velecsum,velec);
 738
 739             fscal            = felec;
 740
 741             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 742
 743             /* Calculate temporary vectorial force */
 744             tx               = _mm_mul_pd(fscal,dx23);
 745             ty               = _mm_mul_pd(fscal,dy23);
 746             tz               = _mm_mul_pd(fscal,dz23);
 747
 748             /* Update vectorial force */
 749             fix2             = _mm_add_pd(fix2,tx);
 750             fiy2             = _mm_add_pd(fiy2,ty);
 751             fiz2             = _mm_add_pd(fiz2,tz);
 752
 753             fjx3             = _mm_add_pd(fjx3,tx);
 754             fjy3             = _mm_add_pd(fjy3,ty);
 755             fjz3             = _mm_add_pd(fjz3,tz);
 756
 757             /**************************
 758              * CALCULATE INTERACTIONS *
 759              **************************/
 760
 761             /* COULOMB ELECTROSTATICS */
 762             velec            = _mm_mul_pd(qq31,rinv31);
 763             felec            = _mm_mul_pd(velec,rinvsq31);
 764
 765             /* Update potential sum for this i atom from the interaction with this j atom. */
 766             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 767             velecsum         = _mm_add_pd(velecsum,velec);
 768
 769             fscal            = felec;
 770
 771             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 772
 773             /* Calculate temporary vectorial force */
 774             tx               = _mm_mul_pd(fscal,dx31);
 775             ty               = _mm_mul_pd(fscal,dy31);
 776             tz               = _mm_mul_pd(fscal,dz31);
 777
 778             /* Update vectorial force */
 779             fix3             = _mm_add_pd(fix3,tx);
 780             fiy3             = _mm_add_pd(fiy3,ty);
 781             fiz3             = _mm_add_pd(fiz3,tz);
 782
 783             fjx1             = _mm_add_pd(fjx1,tx);
 784             fjy1             = _mm_add_pd(fjy1,ty);
 785             fjz1             = _mm_add_pd(fjz1,tz);
 786
 787             /**************************
 788              * CALCULATE INTERACTIONS *
 789              **************************/
 790
 791             /* COULOMB ELECTROSTATICS */
 792             velec            = _mm_mul_pd(qq32,rinv32);
 793             felec            = _mm_mul_pd(velec,rinvsq32);
 794
 795             /* Update potential sum for this i atom from the interaction with this j atom. */
 796             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 797             velecsum         = _mm_add_pd(velecsum,velec);
 798
 799             fscal            = felec;
 800
 801             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 802
 803             /* Calculate temporary vectorial force */
 804             tx               = _mm_mul_pd(fscal,dx32);
 805             ty               = _mm_mul_pd(fscal,dy32);
 806             tz               = _mm_mul_pd(fscal,dz32);
 807
 808             /* Update vectorial force */
 809             fix3             = _mm_add_pd(fix3,tx);
 810             fiy3             = _mm_add_pd(fiy3,ty);
 811             fiz3             = _mm_add_pd(fiz3,tz);
 812
 813             fjx2             = _mm_add_pd(fjx2,tx);
 814             fjy2             = _mm_add_pd(fjy2,ty);
 815             fjz2             = _mm_add_pd(fjz2,tz);
 816
 817             /**************************
 818              * CALCULATE INTERACTIONS *
 819              **************************/
 820
 821             /* COULOMB ELECTROSTATICS */
 822             velec            = _mm_mul_pd(qq33,rinv33);
 823             felec            = _mm_mul_pd(velec,rinvsq33);
 824
 825             /* Update potential sum for this i atom from the interaction with this j atom. */
 826             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 827             velecsum         = _mm_add_pd(velecsum,velec);
 828
 829             fscal            = felec;
 830
 831             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 832
 833             /* Calculate temporary vectorial force */
 834             tx               = _mm_mul_pd(fscal,dx33);
 835             ty               = _mm_mul_pd(fscal,dy33);
 836             tz               = _mm_mul_pd(fscal,dz33);
 837
 838             /* Update vectorial force */
 839             fix3             = _mm_add_pd(fix3,tx);
 840             fiy3             = _mm_add_pd(fiy3,ty);
 841             fiz3             = _mm_add_pd(fiz3,tz);
 842
 843             fjx3             = _mm_add_pd(fjx3,tx);
 844             fjy3             = _mm_add_pd(fjy3,ty);
 845             fjz3             = _mm_add_pd(fjz3,tz);
 846
 847             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 848
 849             /* Inner loop uses 252 flops */
 850         }
 851
 852         /* End of innermost loop */
 853
 854         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 855                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 856
 857         ggid                        = gid[iidx];
 858         /* Update potential energies */
 859         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 860
 861         /* Increment number of inner iterations */
 862         inneriter                  += j_index_end - j_index_start;
 863
 864         /* Outer loop uses 19 flops */
 865     }
 866
 867     /* Increment number of outer iterations */
 868     outeriter        += nri;
 869
 870     /* Update outer/inner flops */
 871
 872     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*252);
 873 }
 874 /*
 875  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse4_1_double
 876  * Electrostatics interaction: Coulomb
 877  * VdW interaction:            None
 878  * Geometry:                   Water4-Water4
 879  * Calculate force/pot:        Force
 880  */
 881 void
 882 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse4_1_double
 883                     (t_nblist * gmx_restrict                nlist,
 884                      rvec * gmx_restrict                    xx,
 885                      rvec * gmx_restrict                    ff,
 886                      t_forcerec * gmx_restrict              fr,
 887                      t_mdatoms * gmx_restrict               mdatoms,
 888                      nb_kernel_data_t * gmx_restrict        kernel_data,
 889                      t_nrnb * gmx_restrict                  nrnb)
 890 {
 891     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 892      * just 0 for non-waters.
 893      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 894      * jnr indices corresponding to data put in the four positions in the SIMD register.
 895      */
 896     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 897     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 898     int              jnrA,jnrB;
 899     int              j_coord_offsetA,j_coord_offsetB;
 900     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 901     real             rcutoff_scalar;
 902     real             *shiftvec,*fshift,*x,*f;
 903     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 904     int              vdwioffset1;
 905     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 906     int              vdwioffset2;
 907     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 908     int              vdwioffset3;
 909     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 910     int              vdwjidx1A,vdwjidx1B;
 911     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 912     int              vdwjidx2A,vdwjidx2B;
 913     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 914     int              vdwjidx3A,vdwjidx3B;
 915     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 916     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 917     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 918     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 919     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 920     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 921     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 922     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 923     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 924     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 925     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 926     real             *charge;
 927     __m128d          dummy_mask,cutoff_mask;
 928     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 929     __m128d          one     = _mm_set1_pd(1.0);
 930     __m128d          two     = _mm_set1_pd(2.0);
 931     x                = xx[0];
 932     f                = ff[0];
 933
 934     nri              = nlist->nri;
 935     iinr             = nlist->iinr;
 936     jindex           = nlist->jindex;
 937     jjnr             = nlist->jjnr;
 938     shiftidx         = nlist->shift;
 939     gid              = nlist->gid;
 940     shiftvec         = fr->shift_vec[0];
 941     fshift           = fr->fshift[0];
 942     facel            = _mm_set1_pd(fr->epsfac);
 943     charge           = mdatoms->chargeA;
 944
 945     /* Setup water-specific parameters */
 946     inr              = nlist->iinr[0];
 947     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 948     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 949     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 950
 951     jq1              = _mm_set1_pd(charge[inr+1]);
 952     jq2              = _mm_set1_pd(charge[inr+2]);
 953     jq3              = _mm_set1_pd(charge[inr+3]);
 954     qq11             = _mm_mul_pd(iq1,jq1);
 955     qq12             = _mm_mul_pd(iq1,jq2);
 956     qq13             = _mm_mul_pd(iq1,jq3);
 957     qq21             = _mm_mul_pd(iq2,jq1);
 958     qq22             = _mm_mul_pd(iq2,jq2);
 959     qq23             = _mm_mul_pd(iq2,jq3);
 960     qq31             = _mm_mul_pd(iq3,jq1);
 961     qq32             = _mm_mul_pd(iq3,jq2);
 962     qq33             = _mm_mul_pd(iq3,jq3);
 963
 964     /* Avoid stupid compiler warnings */
 965     jnrA = jnrB = 0;
 966     j_coord_offsetA = 0;
 967     j_coord_offsetB = 0;
 968
 969     outeriter        = 0;
 970     inneriter        = 0;
 971
 972     /* Start outer loop over neighborlists */
 973     for(iidx=0; iidx<nri; iidx++)
 974     {
 975         /* Load shift vector for this list */
 976         i_shift_offset   = DIM*shiftidx[iidx];
 977
 978         /* Load limits for loop over neighbors */
 979         j_index_start    = jindex[iidx];
 980         j_index_end      = jindex[iidx+1];
 981
 982         /* Get outer coordinate index */
 983         inr              = iinr[iidx];
 984         i_coord_offset   = DIM*inr;
 985
 986         /* Load i particle coords and add shift vector */
 987         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 988                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 989
 990         fix1             = _mm_setzero_pd();
 991         fiy1             = _mm_setzero_pd();
 992         fiz1             = _mm_setzero_pd();
 993         fix2             = _mm_setzero_pd();
 994         fiy2             = _mm_setzero_pd();
 995         fiz2             = _mm_setzero_pd();
 996         fix3             = _mm_setzero_pd();
 997         fiy3             = _mm_setzero_pd();
 998         fiz3             = _mm_setzero_pd();
 999
1000         /* Start inner kernel loop */
1001         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1002         {
1003
1004             /* Get j neighbor index, and coordinate index */
1005             jnrA             = jjnr[jidx];
1006             jnrB             = jjnr[jidx+1];
1007             j_coord_offsetA  = DIM*jnrA;
1008             j_coord_offsetB  = DIM*jnrB;
1009
1010             /* load j atom coordinates */
1011             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1012                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1013
1014             /* Calculate displacement vector */
1015             dx11             = _mm_sub_pd(ix1,jx1);
1016             dy11             = _mm_sub_pd(iy1,jy1);
1017             dz11             = _mm_sub_pd(iz1,jz1);
1018             dx12             = _mm_sub_pd(ix1,jx2);
1019             dy12             = _mm_sub_pd(iy1,jy2);
1020             dz12             = _mm_sub_pd(iz1,jz2);
1021             dx13             = _mm_sub_pd(ix1,jx3);
1022             dy13             = _mm_sub_pd(iy1,jy3);
1023             dz13             = _mm_sub_pd(iz1,jz3);
1024             dx21             = _mm_sub_pd(ix2,jx1);
1025             dy21             = _mm_sub_pd(iy2,jy1);
1026             dz21             = _mm_sub_pd(iz2,jz1);
1027             dx22             = _mm_sub_pd(ix2,jx2);
1028             dy22             = _mm_sub_pd(iy2,jy2);
1029             dz22             = _mm_sub_pd(iz2,jz2);
1030             dx23             = _mm_sub_pd(ix2,jx3);
1031             dy23             = _mm_sub_pd(iy2,jy3);
1032             dz23             = _mm_sub_pd(iz2,jz3);
1033             dx31             = _mm_sub_pd(ix3,jx1);
1034             dy31             = _mm_sub_pd(iy3,jy1);
1035             dz31             = _mm_sub_pd(iz3,jz1);
1036             dx32             = _mm_sub_pd(ix3,jx2);
1037             dy32             = _mm_sub_pd(iy3,jy2);
1038             dz32             = _mm_sub_pd(iz3,jz2);
1039             dx33             = _mm_sub_pd(ix3,jx3);
1040             dy33             = _mm_sub_pd(iy3,jy3);
1041             dz33             = _mm_sub_pd(iz3,jz3);
1042
1043             /* Calculate squared distance and things based on it */
1044             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1045             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1046             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1047             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1048             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1049             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1050             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1051             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1052             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1053
1054             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1055             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1056             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1057             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1058             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1059             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1060             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1061             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1062             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1063
1064             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1065             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1066             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1067             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1068             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1069             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1070             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1071             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1072             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1073
1074             fjx1             = _mm_setzero_pd();
1075             fjy1             = _mm_setzero_pd();
1076             fjz1             = _mm_setzero_pd();
1077             fjx2             = _mm_setzero_pd();
1078             fjy2             = _mm_setzero_pd();
1079             fjz2             = _mm_setzero_pd();
1080             fjx3             = _mm_setzero_pd();
1081             fjy3             = _mm_setzero_pd();
1082             fjz3             = _mm_setzero_pd();
1083
1084             /**************************
1085              * CALCULATE INTERACTIONS *
1086              **************************/
1087
1088             /* COULOMB ELECTROSTATICS */
1089             velec            = _mm_mul_pd(qq11,rinv11);
1090             felec            = _mm_mul_pd(velec,rinvsq11);
1091
1092             fscal            = felec;
1093
1094             /* Calculate temporary vectorial force */
1095             tx               = _mm_mul_pd(fscal,dx11);
1096             ty               = _mm_mul_pd(fscal,dy11);
1097             tz               = _mm_mul_pd(fscal,dz11);
1098
1099             /* Update vectorial force */
1100             fix1             = _mm_add_pd(fix1,tx);
1101             fiy1             = _mm_add_pd(fiy1,ty);
1102             fiz1             = _mm_add_pd(fiz1,tz);
1103
1104             fjx1             = _mm_add_pd(fjx1,tx);
1105             fjy1             = _mm_add_pd(fjy1,ty);
1106             fjz1             = _mm_add_pd(fjz1,tz);
1107
1108             /**************************
1109              * CALCULATE INTERACTIONS *
1110              **************************/
1111
1112             /* COULOMB ELECTROSTATICS */
1113             velec            = _mm_mul_pd(qq12,rinv12);
1114             felec            = _mm_mul_pd(velec,rinvsq12);
1115
1116             fscal            = felec;
1117
1118             /* Calculate temporary vectorial force */
1119             tx               = _mm_mul_pd(fscal,dx12);
1120             ty               = _mm_mul_pd(fscal,dy12);
1121             tz               = _mm_mul_pd(fscal,dz12);
1122
1123             /* Update vectorial force */
1124             fix1             = _mm_add_pd(fix1,tx);
1125             fiy1             = _mm_add_pd(fiy1,ty);
1126             fiz1             = _mm_add_pd(fiz1,tz);
1127
1128             fjx2             = _mm_add_pd(fjx2,tx);
1129             fjy2             = _mm_add_pd(fjy2,ty);
1130             fjz2             = _mm_add_pd(fjz2,tz);
1131
1132             /**************************
1133              * CALCULATE INTERACTIONS *
1134              **************************/
1135
1136             /* COULOMB ELECTROSTATICS */
1137             velec            = _mm_mul_pd(qq13,rinv13);
1138             felec            = _mm_mul_pd(velec,rinvsq13);
1139
1140             fscal            = felec;
1141
1142             /* Calculate temporary vectorial force */
1143             tx               = _mm_mul_pd(fscal,dx13);
1144             ty               = _mm_mul_pd(fscal,dy13);
1145             tz               = _mm_mul_pd(fscal,dz13);
1146
1147             /* Update vectorial force */
1148             fix1             = _mm_add_pd(fix1,tx);
1149             fiy1             = _mm_add_pd(fiy1,ty);
1150             fiz1             = _mm_add_pd(fiz1,tz);
1151
1152             fjx3             = _mm_add_pd(fjx3,tx);
1153             fjy3             = _mm_add_pd(fjy3,ty);
1154             fjz3             = _mm_add_pd(fjz3,tz);
1155
1156             /**************************
1157              * CALCULATE INTERACTIONS *
1158              **************************/
1159
1160             /* COULOMB ELECTROSTATICS */
1161             velec            = _mm_mul_pd(qq21,rinv21);
1162             felec            = _mm_mul_pd(velec,rinvsq21);
1163
1164             fscal            = felec;
1165
1166             /* Calculate temporary vectorial force */
1167             tx               = _mm_mul_pd(fscal,dx21);
1168             ty               = _mm_mul_pd(fscal,dy21);
1169             tz               = _mm_mul_pd(fscal,dz21);
1170
1171             /* Update vectorial force */
1172             fix2             = _mm_add_pd(fix2,tx);
1173             fiy2             = _mm_add_pd(fiy2,ty);
1174             fiz2             = _mm_add_pd(fiz2,tz);
1175
1176             fjx1             = _mm_add_pd(fjx1,tx);
1177             fjy1             = _mm_add_pd(fjy1,ty);
1178             fjz1             = _mm_add_pd(fjz1,tz);
1179
1180             /**************************
1181              * CALCULATE INTERACTIONS *
1182              **************************/
1183
1184             /* COULOMB ELECTROSTATICS */
1185             velec            = _mm_mul_pd(qq22,rinv22);
1186             felec            = _mm_mul_pd(velec,rinvsq22);
1187
1188             fscal            = felec;
1189
1190             /* Calculate temporary vectorial force */
1191             tx               = _mm_mul_pd(fscal,dx22);
1192             ty               = _mm_mul_pd(fscal,dy22);
1193             tz               = _mm_mul_pd(fscal,dz22);
1194
1195             /* Update vectorial force */
1196             fix2             = _mm_add_pd(fix2,tx);
1197             fiy2             = _mm_add_pd(fiy2,ty);
1198             fiz2             = _mm_add_pd(fiz2,tz);
1199
1200             fjx2             = _mm_add_pd(fjx2,tx);
1201             fjy2             = _mm_add_pd(fjy2,ty);
1202             fjz2             = _mm_add_pd(fjz2,tz);
1203
1204             /**************************
1205              * CALCULATE INTERACTIONS *
1206              **************************/
1207
1208             /* COULOMB ELECTROSTATICS */
1209             velec            = _mm_mul_pd(qq23,rinv23);
1210             felec            = _mm_mul_pd(velec,rinvsq23);
1211
1212             fscal            = felec;
1213
1214             /* Calculate temporary vectorial force */
1215             tx               = _mm_mul_pd(fscal,dx23);
1216             ty               = _mm_mul_pd(fscal,dy23);
1217             tz               = _mm_mul_pd(fscal,dz23);
1218
1219             /* Update vectorial force */
1220             fix2             = _mm_add_pd(fix2,tx);
1221             fiy2             = _mm_add_pd(fiy2,ty);
1222             fiz2             = _mm_add_pd(fiz2,tz);
1223
1224             fjx3             = _mm_add_pd(fjx3,tx);
1225             fjy3             = _mm_add_pd(fjy3,ty);
1226             fjz3             = _mm_add_pd(fjz3,tz);
1227
1228             /**************************
1229              * CALCULATE INTERACTIONS *
1230              **************************/
1231
1232             /* COULOMB ELECTROSTATICS */
1233             velec            = _mm_mul_pd(qq31,rinv31);
1234             felec            = _mm_mul_pd(velec,rinvsq31);
1235
1236             fscal            = felec;
1237
1238             /* Calculate temporary vectorial force */
1239             tx               = _mm_mul_pd(fscal,dx31);
1240             ty               = _mm_mul_pd(fscal,dy31);
1241             tz               = _mm_mul_pd(fscal,dz31);
1242
1243             /* Update vectorial force */
1244             fix3             = _mm_add_pd(fix3,tx);
1245             fiy3             = _mm_add_pd(fiy3,ty);
1246             fiz3             = _mm_add_pd(fiz3,tz);
1247
1248             fjx1             = _mm_add_pd(fjx1,tx);
1249             fjy1             = _mm_add_pd(fjy1,ty);
1250             fjz1             = _mm_add_pd(fjz1,tz);
1251
1252             /**************************
1253              * CALCULATE INTERACTIONS *
1254              **************************/
1255
1256             /* COULOMB ELECTROSTATICS */
1257             velec            = _mm_mul_pd(qq32,rinv32);
1258             felec            = _mm_mul_pd(velec,rinvsq32);
1259
1260             fscal            = felec;
1261
1262             /* Calculate temporary vectorial force */
1263             tx               = _mm_mul_pd(fscal,dx32);
1264             ty               = _mm_mul_pd(fscal,dy32);
1265             tz               = _mm_mul_pd(fscal,dz32);
1266
1267             /* Update vectorial force */
1268             fix3             = _mm_add_pd(fix3,tx);
1269             fiy3             = _mm_add_pd(fiy3,ty);
1270             fiz3             = _mm_add_pd(fiz3,tz);
1271
1272             fjx2             = _mm_add_pd(fjx2,tx);
1273             fjy2             = _mm_add_pd(fjy2,ty);
1274             fjz2             = _mm_add_pd(fjz2,tz);
1275
1276             /**************************
1277              * CALCULATE INTERACTIONS *
1278              **************************/
1279
1280             /* COULOMB ELECTROSTATICS */
1281             velec            = _mm_mul_pd(qq33,rinv33);
1282             felec            = _mm_mul_pd(velec,rinvsq33);
1283
1284             fscal            = felec;
1285
1286             /* Calculate temporary vectorial force */
1287             tx               = _mm_mul_pd(fscal,dx33);
1288             ty               = _mm_mul_pd(fscal,dy33);
1289             tz               = _mm_mul_pd(fscal,dz33);
1290
1291             /* Update vectorial force */
1292             fix3             = _mm_add_pd(fix3,tx);
1293             fiy3             = _mm_add_pd(fiy3,ty);
1294             fiz3             = _mm_add_pd(fiz3,tz);
1295
1296             fjx3             = _mm_add_pd(fjx3,tx);
1297             fjy3             = _mm_add_pd(fjy3,ty);
1298             fjz3             = _mm_add_pd(fjz3,tz);
1299
1300             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);
1301
1302             /* Inner loop uses 243 flops */
1303         }
1304
1305         if(jidx<j_index_end)
1306         {
1307
1308             jnrA             = jjnr[jidx];
1309             j_coord_offsetA  = DIM*jnrA;
1310
1311             /* load j atom coordinates */
1312             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1313                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1314
1315             /* Calculate displacement vector */
1316             dx11             = _mm_sub_pd(ix1,jx1);
1317             dy11             = _mm_sub_pd(iy1,jy1);
1318             dz11             = _mm_sub_pd(iz1,jz1);
1319             dx12             = _mm_sub_pd(ix1,jx2);
1320             dy12             = _mm_sub_pd(iy1,jy2);
1321             dz12             = _mm_sub_pd(iz1,jz2);
1322             dx13             = _mm_sub_pd(ix1,jx3);
1323             dy13             = _mm_sub_pd(iy1,jy3);
1324             dz13             = _mm_sub_pd(iz1,jz3);
1325             dx21             = _mm_sub_pd(ix2,jx1);
1326             dy21             = _mm_sub_pd(iy2,jy1);
1327             dz21             = _mm_sub_pd(iz2,jz1);
1328             dx22             = _mm_sub_pd(ix2,jx2);
1329             dy22             = _mm_sub_pd(iy2,jy2);
1330             dz22             = _mm_sub_pd(iz2,jz2);
1331             dx23             = _mm_sub_pd(ix2,jx3);
1332             dy23             = _mm_sub_pd(iy2,jy3);
1333             dz23             = _mm_sub_pd(iz2,jz3);
1334             dx31             = _mm_sub_pd(ix3,jx1);
1335             dy31             = _mm_sub_pd(iy3,jy1);
1336             dz31             = _mm_sub_pd(iz3,jz1);
1337             dx32             = _mm_sub_pd(ix3,jx2);
1338             dy32             = _mm_sub_pd(iy3,jy2);
1339             dz32             = _mm_sub_pd(iz3,jz2);
1340             dx33             = _mm_sub_pd(ix3,jx3);
1341             dy33             = _mm_sub_pd(iy3,jy3);
1342             dz33             = _mm_sub_pd(iz3,jz3);
1343
1344             /* Calculate squared distance and things based on it */
1345             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1346             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1347             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1348             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1349             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1350             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1351             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1352             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1353             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1354
1355             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1356             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1357             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1358             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1359             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1360             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1361             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1362             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1363             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1364
1365             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1366             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1367             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1368             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1369             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1370             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1371             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1372             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1373             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1374
1375             fjx1             = _mm_setzero_pd();
1376             fjy1             = _mm_setzero_pd();
1377             fjz1             = _mm_setzero_pd();
1378             fjx2             = _mm_setzero_pd();
1379             fjy2             = _mm_setzero_pd();
1380             fjz2             = _mm_setzero_pd();
1381             fjx3             = _mm_setzero_pd();
1382             fjy3             = _mm_setzero_pd();
1383             fjz3             = _mm_setzero_pd();
1384
1385             /**************************
1386              * CALCULATE INTERACTIONS *
1387              **************************/
1388
1389             /* COULOMB ELECTROSTATICS */
1390             velec            = _mm_mul_pd(qq11,rinv11);
1391             felec            = _mm_mul_pd(velec,rinvsq11);
1392
1393             fscal            = felec;
1394
1395             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1396
1397             /* Calculate temporary vectorial force */
1398             tx               = _mm_mul_pd(fscal,dx11);
1399             ty               = _mm_mul_pd(fscal,dy11);
1400             tz               = _mm_mul_pd(fscal,dz11);
1401
1402             /* Update vectorial force */
1403             fix1             = _mm_add_pd(fix1,tx);
1404             fiy1             = _mm_add_pd(fiy1,ty);
1405             fiz1             = _mm_add_pd(fiz1,tz);
1406
1407             fjx1             = _mm_add_pd(fjx1,tx);
1408             fjy1             = _mm_add_pd(fjy1,ty);
1409             fjz1             = _mm_add_pd(fjz1,tz);
1410
1411             /**************************
1412              * CALCULATE INTERACTIONS *
1413              **************************/
1414
1415             /* COULOMB ELECTROSTATICS */
1416             velec            = _mm_mul_pd(qq12,rinv12);
1417             felec            = _mm_mul_pd(velec,rinvsq12);
1418
1419             fscal            = felec;
1420
1421             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1422
1423             /* Calculate temporary vectorial force */
1424             tx               = _mm_mul_pd(fscal,dx12);
1425             ty               = _mm_mul_pd(fscal,dy12);
1426             tz               = _mm_mul_pd(fscal,dz12);
1427
1428             /* Update vectorial force */
1429             fix1             = _mm_add_pd(fix1,tx);
1430             fiy1             = _mm_add_pd(fiy1,ty);
1431             fiz1             = _mm_add_pd(fiz1,tz);
1432
1433             fjx2             = _mm_add_pd(fjx2,tx);
1434             fjy2             = _mm_add_pd(fjy2,ty);
1435             fjz2             = _mm_add_pd(fjz2,tz);
1436
1437             /**************************
1438              * CALCULATE INTERACTIONS *
1439              **************************/
1440
1441             /* COULOMB ELECTROSTATICS */
1442             velec            = _mm_mul_pd(qq13,rinv13);
1443             felec            = _mm_mul_pd(velec,rinvsq13);
1444
1445             fscal            = felec;
1446
1447             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1448
1449             /* Calculate temporary vectorial force */
1450             tx               = _mm_mul_pd(fscal,dx13);
1451             ty               = _mm_mul_pd(fscal,dy13);
1452             tz               = _mm_mul_pd(fscal,dz13);
1453
1454             /* Update vectorial force */
1455             fix1             = _mm_add_pd(fix1,tx);
1456             fiy1             = _mm_add_pd(fiy1,ty);
1457             fiz1             = _mm_add_pd(fiz1,tz);
1458
1459             fjx3             = _mm_add_pd(fjx3,tx);
1460             fjy3             = _mm_add_pd(fjy3,ty);
1461             fjz3             = _mm_add_pd(fjz3,tz);
1462
1463             /**************************
1464              * CALCULATE INTERACTIONS *
1465              **************************/
1466
1467             /* COULOMB ELECTROSTATICS */
1468             velec            = _mm_mul_pd(qq21,rinv21);
1469             felec            = _mm_mul_pd(velec,rinvsq21);
1470
1471             fscal            = felec;
1472
1473             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1474
1475             /* Calculate temporary vectorial force */
1476             tx               = _mm_mul_pd(fscal,dx21);
1477             ty               = _mm_mul_pd(fscal,dy21);
1478             tz               = _mm_mul_pd(fscal,dz21);
1479
1480             /* Update vectorial force */
1481             fix2             = _mm_add_pd(fix2,tx);
1482             fiy2             = _mm_add_pd(fiy2,ty);
1483             fiz2             = _mm_add_pd(fiz2,tz);
1484
1485             fjx1             = _mm_add_pd(fjx1,tx);
1486             fjy1             = _mm_add_pd(fjy1,ty);
1487             fjz1             = _mm_add_pd(fjz1,tz);
1488
1489             /**************************
1490              * CALCULATE INTERACTIONS *
1491              **************************/
1492
1493             /* COULOMB ELECTROSTATICS */
1494             velec            = _mm_mul_pd(qq22,rinv22);
1495             felec            = _mm_mul_pd(velec,rinvsq22);
1496
1497             fscal            = felec;
1498
1499             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1500
1501             /* Calculate temporary vectorial force */
1502             tx               = _mm_mul_pd(fscal,dx22);
1503             ty               = _mm_mul_pd(fscal,dy22);
1504             tz               = _mm_mul_pd(fscal,dz22);
1505
1506             /* Update vectorial force */
1507             fix2             = _mm_add_pd(fix2,tx);
1508             fiy2             = _mm_add_pd(fiy2,ty);
1509             fiz2             = _mm_add_pd(fiz2,tz);
1510
1511             fjx2             = _mm_add_pd(fjx2,tx);
1512             fjy2             = _mm_add_pd(fjy2,ty);
1513             fjz2             = _mm_add_pd(fjz2,tz);
1514
1515             /**************************
1516              * CALCULATE INTERACTIONS *
1517              **************************/
1518
1519             /* COULOMB ELECTROSTATICS */
1520             velec            = _mm_mul_pd(qq23,rinv23);
1521             felec            = _mm_mul_pd(velec,rinvsq23);
1522
1523             fscal            = felec;
1524
1525             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1526
1527             /* Calculate temporary vectorial force */
1528             tx               = _mm_mul_pd(fscal,dx23);
1529             ty               = _mm_mul_pd(fscal,dy23);
1530             tz               = _mm_mul_pd(fscal,dz23);
1531
1532             /* Update vectorial force */
1533             fix2             = _mm_add_pd(fix2,tx);
1534             fiy2             = _mm_add_pd(fiy2,ty);
1535             fiz2             = _mm_add_pd(fiz2,tz);
1536
1537             fjx3             = _mm_add_pd(fjx3,tx);
1538             fjy3             = _mm_add_pd(fjy3,ty);
1539             fjz3             = _mm_add_pd(fjz3,tz);
1540
1541             /**************************
1542              * CALCULATE INTERACTIONS *
1543              **************************/
1544
1545             /* COULOMB ELECTROSTATICS */
1546             velec            = _mm_mul_pd(qq31,rinv31);
1547             felec            = _mm_mul_pd(velec,rinvsq31);
1548
1549             fscal            = felec;
1550
1551             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1552
1553             /* Calculate temporary vectorial force */
1554             tx               = _mm_mul_pd(fscal,dx31);
1555             ty               = _mm_mul_pd(fscal,dy31);
1556             tz               = _mm_mul_pd(fscal,dz31);
1557
1558             /* Update vectorial force */
1559             fix3             = _mm_add_pd(fix3,tx);
1560             fiy3             = _mm_add_pd(fiy3,ty);
1561             fiz3             = _mm_add_pd(fiz3,tz);
1562
1563             fjx1             = _mm_add_pd(fjx1,tx);
1564             fjy1             = _mm_add_pd(fjy1,ty);
1565             fjz1             = _mm_add_pd(fjz1,tz);
1566
1567             /**************************
1568              * CALCULATE INTERACTIONS *
1569              **************************/
1570
1571             /* COULOMB ELECTROSTATICS */
1572             velec            = _mm_mul_pd(qq32,rinv32);
1573             felec            = _mm_mul_pd(velec,rinvsq32);
1574
1575             fscal            = felec;
1576
1577             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1578
1579             /* Calculate temporary vectorial force */
1580             tx               = _mm_mul_pd(fscal,dx32);
1581             ty               = _mm_mul_pd(fscal,dy32);
1582             tz               = _mm_mul_pd(fscal,dz32);
1583
1584             /* Update vectorial force */
1585             fix3             = _mm_add_pd(fix3,tx);
1586             fiy3             = _mm_add_pd(fiy3,ty);
1587             fiz3             = _mm_add_pd(fiz3,tz);
1588
1589             fjx2             = _mm_add_pd(fjx2,tx);
1590             fjy2             = _mm_add_pd(fjy2,ty);
1591             fjz2             = _mm_add_pd(fjz2,tz);
1592
1593             /**************************
1594              * CALCULATE INTERACTIONS *
1595              **************************/
1596
1597             /* COULOMB ELECTROSTATICS */
1598             velec            = _mm_mul_pd(qq33,rinv33);
1599             felec            = _mm_mul_pd(velec,rinvsq33);
1600
1601             fscal            = felec;
1602
1603             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1604
1605             /* Calculate temporary vectorial force */
1606             tx               = _mm_mul_pd(fscal,dx33);
1607             ty               = _mm_mul_pd(fscal,dy33);
1608             tz               = _mm_mul_pd(fscal,dz33);
1609
1610             /* Update vectorial force */
1611             fix3             = _mm_add_pd(fix3,tx);
1612             fiy3             = _mm_add_pd(fiy3,ty);
1613             fiz3             = _mm_add_pd(fiz3,tz);
1614
1615             fjx3             = _mm_add_pd(fjx3,tx);
1616             fjy3             = _mm_add_pd(fjy3,ty);
1617             fjz3             = _mm_add_pd(fjz3,tz);
1618
1619             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1620
1621             /* Inner loop uses 243 flops */
1622         }
1623
1624         /* End of innermost loop */
1625
1626         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1627                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1628
1629         /* Increment number of inner iterations */
1630         inneriter                  += j_index_end - j_index_start;
1631
1632         /* Outer loop uses 18 flops */
1633     }
1634
1635     /* Increment number of outer iterations */
1636     outeriter        += nri;
1637
1638     /* Update outer/inner flops */
1639
1640     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1641 }