src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwNone_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_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_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_avx_128_fma_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_avx_128_fma_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             /* Update vectorial force */
 263             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 264             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 265             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 266
 267             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 268             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 269             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 270
 271             /**************************
 272              * CALCULATE INTERACTIONS *
 273              **************************/
 274
 275             /* COULOMB ELECTROSTATICS */
 276             velec            = _mm_mul_pd(qq12,rinv12);
 277             felec            = _mm_mul_pd(velec,rinvsq12);
 278
 279             /* Update potential sum for this i atom from the interaction with this j atom. */
 280             velecsum         = _mm_add_pd(velecsum,velec);
 281
 282             fscal            = felec;
 283
 284             /* Update vectorial force */
 285             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 286             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 287             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 288
 289             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 290             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 291             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 292
 293             /**************************
 294              * CALCULATE INTERACTIONS *
 295              **************************/
 296
 297             /* COULOMB ELECTROSTATICS */
 298             velec            = _mm_mul_pd(qq13,rinv13);
 299             felec            = _mm_mul_pd(velec,rinvsq13);
 300
 301             /* Update potential sum for this i atom from the interaction with this j atom. */
 302             velecsum         = _mm_add_pd(velecsum,velec);
 303
 304             fscal            = felec;
 305
 306             /* Update vectorial force */
 307             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 308             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 309             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 310
 311             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 312             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 313             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 314
 315             /**************************
 316              * CALCULATE INTERACTIONS *
 317              **************************/
 318
 319             /* COULOMB ELECTROSTATICS */
 320             velec            = _mm_mul_pd(qq21,rinv21);
 321             felec            = _mm_mul_pd(velec,rinvsq21);
 322
 323             /* Update potential sum for this i atom from the interaction with this j atom. */
 324             velecsum         = _mm_add_pd(velecsum,velec);
 325
 326             fscal            = felec;
 327
 328             /* Update vectorial force */
 329             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 330             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 331             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 332
 333             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 334             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 335             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             /* COULOMB ELECTROSTATICS */
 342             velec            = _mm_mul_pd(qq22,rinv22);
 343             felec            = _mm_mul_pd(velec,rinvsq22);
 344
 345             /* Update potential sum for this i atom from the interaction with this j atom. */
 346             velecsum         = _mm_add_pd(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             /* Update vectorial force */
 351             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 352             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 353             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 354
 355             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 356             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 357             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             /* COULOMB ELECTROSTATICS */
 364             velec            = _mm_mul_pd(qq23,rinv23);
 365             felec            = _mm_mul_pd(velec,rinvsq23);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velecsum         = _mm_add_pd(velecsum,velec);
 369
 370             fscal            = felec;
 371
 372             /* Update vectorial force */
 373             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 374             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 375             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 376
 377             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 378             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 379             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 380
 381             /**************************
 382              * CALCULATE INTERACTIONS *
 383              **************************/
 384
 385             /* COULOMB ELECTROSTATICS */
 386             velec            = _mm_mul_pd(qq31,rinv31);
 387             felec            = _mm_mul_pd(velec,rinvsq31);
 388
 389             /* Update potential sum for this i atom from the interaction with this j atom. */
 390             velecsum         = _mm_add_pd(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             /* Update vectorial force */
 395             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 396             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 397             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 398
 399             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 400             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 401             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 402
 403             /**************************
 404              * CALCULATE INTERACTIONS *
 405              **************************/
 406
 407             /* COULOMB ELECTROSTATICS */
 408             velec            = _mm_mul_pd(qq32,rinv32);
 409             felec            = _mm_mul_pd(velec,rinvsq32);
 410
 411             /* Update potential sum for this i atom from the interaction with this j atom. */
 412             velecsum         = _mm_add_pd(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             /* Update vectorial force */
 417             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 418             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 419             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 420
 421             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 422             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 423             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 424
 425             /**************************
 426              * CALCULATE INTERACTIONS *
 427              **************************/
 428
 429             /* COULOMB ELECTROSTATICS */
 430             velec            = _mm_mul_pd(qq33,rinv33);
 431             felec            = _mm_mul_pd(velec,rinvsq33);
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velecsum         = _mm_add_pd(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             /* Update vectorial force */
 439             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 440             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 441             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 442
 443             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 444             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 445             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 446
 447             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);
 448
 449             /* Inner loop uses 279 flops */
 450         }
 451
 452         if(jidx<j_index_end)
 453         {
 454
 455             jnrA             = jjnr[jidx];
 456             j_coord_offsetA  = DIM*jnrA;
 457
 458             /* load j atom coordinates */
 459             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 460                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 461
 462             /* Calculate displacement vector */
 463             dx11             = _mm_sub_pd(ix1,jx1);
 464             dy11             = _mm_sub_pd(iy1,jy1);
 465             dz11             = _mm_sub_pd(iz1,jz1);
 466             dx12             = _mm_sub_pd(ix1,jx2);
 467             dy12             = _mm_sub_pd(iy1,jy2);
 468             dz12             = _mm_sub_pd(iz1,jz2);
 469             dx13             = _mm_sub_pd(ix1,jx3);
 470             dy13             = _mm_sub_pd(iy1,jy3);
 471             dz13             = _mm_sub_pd(iz1,jz3);
 472             dx21             = _mm_sub_pd(ix2,jx1);
 473             dy21             = _mm_sub_pd(iy2,jy1);
 474             dz21             = _mm_sub_pd(iz2,jz1);
 475             dx22             = _mm_sub_pd(ix2,jx2);
 476             dy22             = _mm_sub_pd(iy2,jy2);
 477             dz22             = _mm_sub_pd(iz2,jz2);
 478             dx23             = _mm_sub_pd(ix2,jx3);
 479             dy23             = _mm_sub_pd(iy2,jy3);
 480             dz23             = _mm_sub_pd(iz2,jz3);
 481             dx31             = _mm_sub_pd(ix3,jx1);
 482             dy31             = _mm_sub_pd(iy3,jy1);
 483             dz31             = _mm_sub_pd(iz3,jz1);
 484             dx32             = _mm_sub_pd(ix3,jx2);
 485             dy32             = _mm_sub_pd(iy3,jy2);
 486             dz32             = _mm_sub_pd(iz3,jz2);
 487             dx33             = _mm_sub_pd(ix3,jx3);
 488             dy33             = _mm_sub_pd(iy3,jy3);
 489             dz33             = _mm_sub_pd(iz3,jz3);
 490
 491             /* Calculate squared distance and things based on it */
 492             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 493             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 494             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 495             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 496             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 497             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 498             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 499             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 500             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 501
 502             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 503             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 504             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 505             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 506             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 507             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 508             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 509             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 510             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 511
 512             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 513             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 514             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 515             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 516             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 517             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 518             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 519             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 520             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 521
 522             fjx1             = _mm_setzero_pd();
 523             fjy1             = _mm_setzero_pd();
 524             fjz1             = _mm_setzero_pd();
 525             fjx2             = _mm_setzero_pd();
 526             fjy2             = _mm_setzero_pd();
 527             fjz2             = _mm_setzero_pd();
 528             fjx3             = _mm_setzero_pd();
 529             fjy3             = _mm_setzero_pd();
 530             fjz3             = _mm_setzero_pd();
 531
 532             /**************************
 533              * CALCULATE INTERACTIONS *
 534              **************************/
 535
 536             /* COULOMB ELECTROSTATICS */
 537             velec            = _mm_mul_pd(qq11,rinv11);
 538             felec            = _mm_mul_pd(velec,rinvsq11);
 539
 540             /* Update potential sum for this i atom from the interaction with this j atom. */
 541             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 542             velecsum         = _mm_add_pd(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 547
 548             /* Update vectorial force */
 549             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 550             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 551             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 552
 553             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 554             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 555             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             /* COULOMB ELECTROSTATICS */
 562             velec            = _mm_mul_pd(qq12,rinv12);
 563             felec            = _mm_mul_pd(velec,rinvsq12);
 564
 565             /* Update potential sum for this i atom from the interaction with this j atom. */
 566             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 567             velecsum         = _mm_add_pd(velecsum,velec);
 568
 569             fscal            = felec;
 570
 571             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 572
 573             /* Update vectorial force */
 574             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 575             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 576             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 577
 578             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 579             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 580             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             /* COULOMB ELECTROSTATICS */
 587             velec            = _mm_mul_pd(qq13,rinv13);
 588             felec            = _mm_mul_pd(velec,rinvsq13);
 589
 590             /* Update potential sum for this i atom from the interaction with this j atom. */
 591             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 592             velecsum         = _mm_add_pd(velecsum,velec);
 593
 594             fscal            = felec;
 595
 596             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 597
 598             /* Update vectorial force */
 599             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 600             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 601             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 602
 603             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 604             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 605             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             /* COULOMB ELECTROSTATICS */
 612             velec            = _mm_mul_pd(qq21,rinv21);
 613             felec            = _mm_mul_pd(velec,rinvsq21);
 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             /* Update vectorial force */
 624             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 625             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 626             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 627
 628             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 629             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 630             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 631
 632             /**************************
 633              * CALCULATE INTERACTIONS *
 634              **************************/
 635
 636             /* COULOMB ELECTROSTATICS */
 637             velec            = _mm_mul_pd(qq22,rinv22);
 638             felec            = _mm_mul_pd(velec,rinvsq22);
 639
 640             /* Update potential sum for this i atom from the interaction with this j atom. */
 641             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 642             velecsum         = _mm_add_pd(velecsum,velec);
 643
 644             fscal            = felec;
 645
 646             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 647
 648             /* Update vectorial force */
 649             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 650             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 651             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 652
 653             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 654             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 655             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 656
 657             /**************************
 658              * CALCULATE INTERACTIONS *
 659              **************************/
 660
 661             /* COULOMB ELECTROSTATICS */
 662             velec            = _mm_mul_pd(qq23,rinv23);
 663             felec            = _mm_mul_pd(velec,rinvsq23);
 664
 665             /* Update potential sum for this i atom from the interaction with this j atom. */
 666             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 667             velecsum         = _mm_add_pd(velecsum,velec);
 668
 669             fscal            = felec;
 670
 671             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 672
 673             /* Update vectorial force */
 674             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 675             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 676             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 677
 678             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 679             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 680             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             /* COULOMB ELECTROSTATICS */
 687             velec            = _mm_mul_pd(qq31,rinv31);
 688             felec            = _mm_mul_pd(velec,rinvsq31);
 689
 690             /* Update potential sum for this i atom from the interaction with this j atom. */
 691             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 692             velecsum         = _mm_add_pd(velecsum,velec);
 693
 694             fscal            = felec;
 695
 696             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 697
 698             /* Update vectorial force */
 699             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 700             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 701             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 702
 703             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 704             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 705             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 706
 707             /**************************
 708              * CALCULATE INTERACTIONS *
 709              **************************/
 710
 711             /* COULOMB ELECTROSTATICS */
 712             velec            = _mm_mul_pd(qq32,rinv32);
 713             felec            = _mm_mul_pd(velec,rinvsq32);
 714
 715             /* Update potential sum for this i atom from the interaction with this j atom. */
 716             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 717             velecsum         = _mm_add_pd(velecsum,velec);
 718
 719             fscal            = felec;
 720
 721             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 722
 723             /* Update vectorial force */
 724             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 725             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 726             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 727
 728             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 729             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 730             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 731
 732             /**************************
 733              * CALCULATE INTERACTIONS *
 734              **************************/
 735
 736             /* COULOMB ELECTROSTATICS */
 737             velec            = _mm_mul_pd(qq33,rinv33);
 738             felec            = _mm_mul_pd(velec,rinvsq33);
 739
 740             /* Update potential sum for this i atom from the interaction with this j atom. */
 741             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 742             velecsum         = _mm_add_pd(velecsum,velec);
 743
 744             fscal            = felec;
 745
 746             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 747
 748             /* Update vectorial force */
 749             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 750             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 751             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 752
 753             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 754             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 755             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 756
 757             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 758
 759             /* Inner loop uses 279 flops */
 760         }
 761
 762         /* End of innermost loop */
 763
 764         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 765                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 766
 767         ggid                        = gid[iidx];
 768         /* Update potential energies */
 769         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 770
 771         /* Increment number of inner iterations */
 772         inneriter                  += j_index_end - j_index_start;
 773
 774         /* Outer loop uses 19 flops */
 775     }
 776
 777     /* Increment number of outer iterations */
 778     outeriter        += nri;
 779
 780     /* Update outer/inner flops */
 781
 782     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*279);
 783 }
 784 /*
 785  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_double
 786  * Electrostatics interaction: Coulomb
 787  * VdW interaction:            None
 788  * Geometry:                   Water4-Water4
 789  * Calculate force/pot:        Force
 790  */
 791 void
 792 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_double
 793                     (t_nblist * gmx_restrict                nlist,
 794                      rvec * gmx_restrict                    xx,
 795                      rvec * gmx_restrict                    ff,
 796                      t_forcerec * gmx_restrict              fr,
 797                      t_mdatoms * gmx_restrict               mdatoms,
 798                      nb_kernel_data_t * gmx_restrict        kernel_data,
 799                      t_nrnb * gmx_restrict                  nrnb)
 800 {
 801     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 802      * just 0 for non-waters.
 803      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 804      * jnr indices corresponding to data put in the four positions in the SIMD register.
 805      */
 806     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 807     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 808     int              jnrA,jnrB;
 809     int              j_coord_offsetA,j_coord_offsetB;
 810     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 811     real             rcutoff_scalar;
 812     real             *shiftvec,*fshift,*x,*f;
 813     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 814     int              vdwioffset1;
 815     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 816     int              vdwioffset2;
 817     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 818     int              vdwioffset3;
 819     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 820     int              vdwjidx1A,vdwjidx1B;
 821     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 822     int              vdwjidx2A,vdwjidx2B;
 823     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 824     int              vdwjidx3A,vdwjidx3B;
 825     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 826     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 827     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 828     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 829     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 830     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 831     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 832     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 833     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 834     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 835     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 836     real             *charge;
 837     __m128d          dummy_mask,cutoff_mask;
 838     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 839     __m128d          one     = _mm_set1_pd(1.0);
 840     __m128d          two     = _mm_set1_pd(2.0);
 841     x                = xx[0];
 842     f                = ff[0];
 843
 844     nri              = nlist->nri;
 845     iinr             = nlist->iinr;
 846     jindex           = nlist->jindex;
 847     jjnr             = nlist->jjnr;
 848     shiftidx         = nlist->shift;
 849     gid              = nlist->gid;
 850     shiftvec         = fr->shift_vec[0];
 851     fshift           = fr->fshift[0];
 852     facel            = _mm_set1_pd(fr->epsfac);
 853     charge           = mdatoms->chargeA;
 854
 855     /* Setup water-specific parameters */
 856     inr              = nlist->iinr[0];
 857     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 858     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 859     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 860
 861     jq1              = _mm_set1_pd(charge[inr+1]);
 862     jq2              = _mm_set1_pd(charge[inr+2]);
 863     jq3              = _mm_set1_pd(charge[inr+3]);
 864     qq11             = _mm_mul_pd(iq1,jq1);
 865     qq12             = _mm_mul_pd(iq1,jq2);
 866     qq13             = _mm_mul_pd(iq1,jq3);
 867     qq21             = _mm_mul_pd(iq2,jq1);
 868     qq22             = _mm_mul_pd(iq2,jq2);
 869     qq23             = _mm_mul_pd(iq2,jq3);
 870     qq31             = _mm_mul_pd(iq3,jq1);
 871     qq32             = _mm_mul_pd(iq3,jq2);
 872     qq33             = _mm_mul_pd(iq3,jq3);
 873
 874     /* Avoid stupid compiler warnings */
 875     jnrA = jnrB = 0;
 876     j_coord_offsetA = 0;
 877     j_coord_offsetB = 0;
 878
 879     outeriter        = 0;
 880     inneriter        = 0;
 881
 882     /* Start outer loop over neighborlists */
 883     for(iidx=0; iidx<nri; iidx++)
 884     {
 885         /* Load shift vector for this list */
 886         i_shift_offset   = DIM*shiftidx[iidx];
 887
 888         /* Load limits for loop over neighbors */
 889         j_index_start    = jindex[iidx];
 890         j_index_end      = jindex[iidx+1];
 891
 892         /* Get outer coordinate index */
 893         inr              = iinr[iidx];
 894         i_coord_offset   = DIM*inr;
 895
 896         /* Load i particle coords and add shift vector */
 897         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 898                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 899
 900         fix1             = _mm_setzero_pd();
 901         fiy1             = _mm_setzero_pd();
 902         fiz1             = _mm_setzero_pd();
 903         fix2             = _mm_setzero_pd();
 904         fiy2             = _mm_setzero_pd();
 905         fiz2             = _mm_setzero_pd();
 906         fix3             = _mm_setzero_pd();
 907         fiy3             = _mm_setzero_pd();
 908         fiz3             = _mm_setzero_pd();
 909
 910         /* Start inner kernel loop */
 911         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 912         {
 913
 914             /* Get j neighbor index, and coordinate index */
 915             jnrA             = jjnr[jidx];
 916             jnrB             = jjnr[jidx+1];
 917             j_coord_offsetA  = DIM*jnrA;
 918             j_coord_offsetB  = DIM*jnrB;
 919
 920             /* load j atom coordinates */
 921             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 922                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 923
 924             /* Calculate displacement vector */
 925             dx11             = _mm_sub_pd(ix1,jx1);
 926             dy11             = _mm_sub_pd(iy1,jy1);
 927             dz11             = _mm_sub_pd(iz1,jz1);
 928             dx12             = _mm_sub_pd(ix1,jx2);
 929             dy12             = _mm_sub_pd(iy1,jy2);
 930             dz12             = _mm_sub_pd(iz1,jz2);
 931             dx13             = _mm_sub_pd(ix1,jx3);
 932             dy13             = _mm_sub_pd(iy1,jy3);
 933             dz13             = _mm_sub_pd(iz1,jz3);
 934             dx21             = _mm_sub_pd(ix2,jx1);
 935             dy21             = _mm_sub_pd(iy2,jy1);
 936             dz21             = _mm_sub_pd(iz2,jz1);
 937             dx22             = _mm_sub_pd(ix2,jx2);
 938             dy22             = _mm_sub_pd(iy2,jy2);
 939             dz22             = _mm_sub_pd(iz2,jz2);
 940             dx23             = _mm_sub_pd(ix2,jx3);
 941             dy23             = _mm_sub_pd(iy2,jy3);
 942             dz23             = _mm_sub_pd(iz2,jz3);
 943             dx31             = _mm_sub_pd(ix3,jx1);
 944             dy31             = _mm_sub_pd(iy3,jy1);
 945             dz31             = _mm_sub_pd(iz3,jz1);
 946             dx32             = _mm_sub_pd(ix3,jx2);
 947             dy32             = _mm_sub_pd(iy3,jy2);
 948             dz32             = _mm_sub_pd(iz3,jz2);
 949             dx33             = _mm_sub_pd(ix3,jx3);
 950             dy33             = _mm_sub_pd(iy3,jy3);
 951             dz33             = _mm_sub_pd(iz3,jz3);
 952
 953             /* Calculate squared distance and things based on it */
 954             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 955             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 956             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 957             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 958             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 959             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 960             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 961             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 962             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 963
 964             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 965             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 966             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 967             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 968             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 969             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 970             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 971             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 972             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 973
 974             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 975             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 976             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 977             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 978             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 979             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 980             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 981             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 982             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 983
 984             fjx1             = _mm_setzero_pd();
 985             fjy1             = _mm_setzero_pd();
 986             fjz1             = _mm_setzero_pd();
 987             fjx2             = _mm_setzero_pd();
 988             fjy2             = _mm_setzero_pd();
 989             fjz2             = _mm_setzero_pd();
 990             fjx3             = _mm_setzero_pd();
 991             fjy3             = _mm_setzero_pd();
 992             fjz3             = _mm_setzero_pd();
 993
 994             /**************************
 995              * CALCULATE INTERACTIONS *
 996              **************************/
 997
 998             /* COULOMB ELECTROSTATICS */
 999             velec            = _mm_mul_pd(qq11,rinv11);
1000             felec            = _mm_mul_pd(velec,rinvsq11);
1001
1002             fscal            = felec;
1003
1004             /* Update vectorial force */
1005             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1006             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1007             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1008
1009             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1010             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1011             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1012
1013             /**************************
1014              * CALCULATE INTERACTIONS *
1015              **************************/
1016
1017             /* COULOMB ELECTROSTATICS */
1018             velec            = _mm_mul_pd(qq12,rinv12);
1019             felec            = _mm_mul_pd(velec,rinvsq12);
1020
1021             fscal            = felec;
1022
1023             /* Update vectorial force */
1024             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1025             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1026             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1027
1028             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1029             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1030             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1031
1032             /**************************
1033              * CALCULATE INTERACTIONS *
1034              **************************/
1035
1036             /* COULOMB ELECTROSTATICS */
1037             velec            = _mm_mul_pd(qq13,rinv13);
1038             felec            = _mm_mul_pd(velec,rinvsq13);
1039
1040             fscal            = felec;
1041
1042             /* Update vectorial force */
1043             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1044             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1045             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1046
1047             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1048             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1049             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1050
1051             /**************************
1052              * CALCULATE INTERACTIONS *
1053              **************************/
1054
1055             /* COULOMB ELECTROSTATICS */
1056             velec            = _mm_mul_pd(qq21,rinv21);
1057             felec            = _mm_mul_pd(velec,rinvsq21);
1058
1059             fscal            = felec;
1060
1061             /* Update vectorial force */
1062             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1063             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1064             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1065
1066             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1067             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1068             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1069
1070             /**************************
1071              * CALCULATE INTERACTIONS *
1072              **************************/
1073
1074             /* COULOMB ELECTROSTATICS */
1075             velec            = _mm_mul_pd(qq22,rinv22);
1076             felec            = _mm_mul_pd(velec,rinvsq22);
1077
1078             fscal            = felec;
1079
1080             /* Update vectorial force */
1081             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1082             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1083             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1084
1085             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1086             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1087             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1088
1089             /**************************
1090              * CALCULATE INTERACTIONS *
1091              **************************/
1092
1093             /* COULOMB ELECTROSTATICS */
1094             velec            = _mm_mul_pd(qq23,rinv23);
1095             felec            = _mm_mul_pd(velec,rinvsq23);
1096
1097             fscal            = felec;
1098
1099             /* Update vectorial force */
1100             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1101             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1102             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1103
1104             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1105             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1106             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1107
1108             /**************************
1109              * CALCULATE INTERACTIONS *
1110              **************************/
1111
1112             /* COULOMB ELECTROSTATICS */
1113             velec            = _mm_mul_pd(qq31,rinv31);
1114             felec            = _mm_mul_pd(velec,rinvsq31);
1115
1116             fscal            = felec;
1117
1118             /* Update vectorial force */
1119             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1120             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1121             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1122
1123             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1124             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1125             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1126
1127             /**************************
1128              * CALCULATE INTERACTIONS *
1129              **************************/
1130
1131             /* COULOMB ELECTROSTATICS */
1132             velec            = _mm_mul_pd(qq32,rinv32);
1133             felec            = _mm_mul_pd(velec,rinvsq32);
1134
1135             fscal            = felec;
1136
1137             /* Update vectorial force */
1138             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1139             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1140             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1141
1142             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1143             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1144             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1145
1146             /**************************
1147              * CALCULATE INTERACTIONS *
1148              **************************/
1149
1150             /* COULOMB ELECTROSTATICS */
1151             velec            = _mm_mul_pd(qq33,rinv33);
1152             felec            = _mm_mul_pd(velec,rinvsq33);
1153
1154             fscal            = felec;
1155
1156             /* Update vectorial force */
1157             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1158             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1159             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1160
1161             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1162             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1163             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1164
1165             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);
1166
1167             /* Inner loop uses 270 flops */
1168         }
1169
1170         if(jidx<j_index_end)
1171         {
1172
1173             jnrA             = jjnr[jidx];
1174             j_coord_offsetA  = DIM*jnrA;
1175
1176             /* load j atom coordinates */
1177             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1178                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1179
1180             /* Calculate displacement vector */
1181             dx11             = _mm_sub_pd(ix1,jx1);
1182             dy11             = _mm_sub_pd(iy1,jy1);
1183             dz11             = _mm_sub_pd(iz1,jz1);
1184             dx12             = _mm_sub_pd(ix1,jx2);
1185             dy12             = _mm_sub_pd(iy1,jy2);
1186             dz12             = _mm_sub_pd(iz1,jz2);
1187             dx13             = _mm_sub_pd(ix1,jx3);
1188             dy13             = _mm_sub_pd(iy1,jy3);
1189             dz13             = _mm_sub_pd(iz1,jz3);
1190             dx21             = _mm_sub_pd(ix2,jx1);
1191             dy21             = _mm_sub_pd(iy2,jy1);
1192             dz21             = _mm_sub_pd(iz2,jz1);
1193             dx22             = _mm_sub_pd(ix2,jx2);
1194             dy22             = _mm_sub_pd(iy2,jy2);
1195             dz22             = _mm_sub_pd(iz2,jz2);
1196             dx23             = _mm_sub_pd(ix2,jx3);
1197             dy23             = _mm_sub_pd(iy2,jy3);
1198             dz23             = _mm_sub_pd(iz2,jz3);
1199             dx31             = _mm_sub_pd(ix3,jx1);
1200             dy31             = _mm_sub_pd(iy3,jy1);
1201             dz31             = _mm_sub_pd(iz3,jz1);
1202             dx32             = _mm_sub_pd(ix3,jx2);
1203             dy32             = _mm_sub_pd(iy3,jy2);
1204             dz32             = _mm_sub_pd(iz3,jz2);
1205             dx33             = _mm_sub_pd(ix3,jx3);
1206             dy33             = _mm_sub_pd(iy3,jy3);
1207             dz33             = _mm_sub_pd(iz3,jz3);
1208
1209             /* Calculate squared distance and things based on it */
1210             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1211             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1212             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1213             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1214             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1215             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1216             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1217             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1218             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1219
1220             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1221             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1222             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1223             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1224             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1225             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1226             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1227             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1228             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1229
1230             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1231             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1232             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1233             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1234             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1235             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1236             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1237             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1238             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1239
1240             fjx1             = _mm_setzero_pd();
1241             fjy1             = _mm_setzero_pd();
1242             fjz1             = _mm_setzero_pd();
1243             fjx2             = _mm_setzero_pd();
1244             fjy2             = _mm_setzero_pd();
1245             fjz2             = _mm_setzero_pd();
1246             fjx3             = _mm_setzero_pd();
1247             fjy3             = _mm_setzero_pd();
1248             fjz3             = _mm_setzero_pd();
1249
1250             /**************************
1251              * CALCULATE INTERACTIONS *
1252              **************************/
1253
1254             /* COULOMB ELECTROSTATICS */
1255             velec            = _mm_mul_pd(qq11,rinv11);
1256             felec            = _mm_mul_pd(velec,rinvsq11);
1257
1258             fscal            = felec;
1259
1260             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1261
1262             /* Update vectorial force */
1263             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1264             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1265             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1266
1267             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1268             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1269             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1270
1271             /**************************
1272              * CALCULATE INTERACTIONS *
1273              **************************/
1274
1275             /* COULOMB ELECTROSTATICS */
1276             velec            = _mm_mul_pd(qq12,rinv12);
1277             felec            = _mm_mul_pd(velec,rinvsq12);
1278
1279             fscal            = felec;
1280
1281             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1282
1283             /* Update vectorial force */
1284             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1285             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1286             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1287
1288             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1289             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1290             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             /* COULOMB ELECTROSTATICS */
1297             velec            = _mm_mul_pd(qq13,rinv13);
1298             felec            = _mm_mul_pd(velec,rinvsq13);
1299
1300             fscal            = felec;
1301
1302             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1303
1304             /* Update vectorial force */
1305             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1306             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1307             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1308
1309             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1310             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1311             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1312
1313             /**************************
1314              * CALCULATE INTERACTIONS *
1315              **************************/
1316
1317             /* COULOMB ELECTROSTATICS */
1318             velec            = _mm_mul_pd(qq21,rinv21);
1319             felec            = _mm_mul_pd(velec,rinvsq21);
1320
1321             fscal            = felec;
1322
1323             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1324
1325             /* Update vectorial force */
1326             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1327             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1328             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1329
1330             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1331             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1332             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1333
1334             /**************************
1335              * CALCULATE INTERACTIONS *
1336              **************************/
1337
1338             /* COULOMB ELECTROSTATICS */
1339             velec            = _mm_mul_pd(qq22,rinv22);
1340             felec            = _mm_mul_pd(velec,rinvsq22);
1341
1342             fscal            = felec;
1343
1344             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1345
1346             /* Update vectorial force */
1347             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1348             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1349             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1350
1351             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1352             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1353             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1354
1355             /**************************
1356              * CALCULATE INTERACTIONS *
1357              **************************/
1358
1359             /* COULOMB ELECTROSTATICS */
1360             velec            = _mm_mul_pd(qq23,rinv23);
1361             felec            = _mm_mul_pd(velec,rinvsq23);
1362
1363             fscal            = felec;
1364
1365             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1366
1367             /* Update vectorial force */
1368             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1369             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1370             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1371
1372             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1373             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1374             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1375
1376             /**************************
1377              * CALCULATE INTERACTIONS *
1378              **************************/
1379
1380             /* COULOMB ELECTROSTATICS */
1381             velec            = _mm_mul_pd(qq31,rinv31);
1382             felec            = _mm_mul_pd(velec,rinvsq31);
1383
1384             fscal            = felec;
1385
1386             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1387
1388             /* Update vectorial force */
1389             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1390             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1391             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1392
1393             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1394             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1395             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1396
1397             /**************************
1398              * CALCULATE INTERACTIONS *
1399              **************************/
1400
1401             /* COULOMB ELECTROSTATICS */
1402             velec            = _mm_mul_pd(qq32,rinv32);
1403             felec            = _mm_mul_pd(velec,rinvsq32);
1404
1405             fscal            = felec;
1406
1407             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1408
1409             /* Update vectorial force */
1410             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1411             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1412             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1413
1414             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1415             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1416             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             /* COULOMB ELECTROSTATICS */
1423             velec            = _mm_mul_pd(qq33,rinv33);
1424             felec            = _mm_mul_pd(velec,rinvsq33);
1425
1426             fscal            = felec;
1427
1428             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1429
1430             /* Update vectorial force */
1431             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1432             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1433             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1434
1435             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1436             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1437             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1438
1439             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1440
1441             /* Inner loop uses 270 flops */
1442         }
1443
1444         /* End of innermost loop */
1445
1446         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1447                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1448
1449         /* Increment number of inner iterations */
1450         inneriter                  += j_index_end - j_index_start;
1451
1452         /* Outer loop uses 18 flops */
1453     }
1454
1455     /* Increment number of outer iterations */
1456     outeriter        += nri;
1457
1458     /* Update outer/inner flops */
1459
1460     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1461 }