src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B;
  75     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B;
  77     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128d          dummy_mask,cutoff_mask;
  96     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  97     __m128d          one     = _mm_set1_pd(1.0);
  98     __m128d          two     = _mm_set1_pd(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_pd(fr->epsfac);
 111     charge           = mdatoms->chargeA;
 112     nvdwtype         = fr->ntype;
 113     vdwparam         = fr->nbfp;
 114     vdwtype          = mdatoms->typeA;
 115
 116     /* Setup water-specific parameters */
 117     inr              = nlist->iinr[0];
 118     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 119     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 120     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 121     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 122
 123     jq0              = _mm_set1_pd(charge[inr+0]);
 124     jq1              = _mm_set1_pd(charge[inr+1]);
 125     jq2              = _mm_set1_pd(charge[inr+2]);
 126     vdwjidx0A        = 2*vdwtype[inr+0];
 127     qq00             = _mm_mul_pd(iq0,jq0);
 128     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 129     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 130     qq01             = _mm_mul_pd(iq0,jq1);
 131     qq02             = _mm_mul_pd(iq0,jq2);
 132     qq10             = _mm_mul_pd(iq1,jq0);
 133     qq11             = _mm_mul_pd(iq1,jq1);
 134     qq12             = _mm_mul_pd(iq1,jq2);
 135     qq20             = _mm_mul_pd(iq2,jq0);
 136     qq21             = _mm_mul_pd(iq2,jq1);
 137     qq22             = _mm_mul_pd(iq2,jq2);
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 163                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 164
 165         fix0             = _mm_setzero_pd();
 166         fiy0             = _mm_setzero_pd();
 167         fiz0             = _mm_setzero_pd();
 168         fix1             = _mm_setzero_pd();
 169         fiy1             = _mm_setzero_pd();
 170         fiz1             = _mm_setzero_pd();
 171         fix2             = _mm_setzero_pd();
 172         fiy2             = _mm_setzero_pd();
 173         fiz2             = _mm_setzero_pd();
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_pd();
 177         vvdwsum          = _mm_setzero_pd();
 178
 179         /* Start inner kernel loop */
 180         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 181         {
 182
 183             /* Get j neighbor index, and coordinate index */
 184             jnrA             = jjnr[jidx];
 185             jnrB             = jjnr[jidx+1];
 186             j_coord_offsetA  = DIM*jnrA;
 187             j_coord_offsetB  = DIM*jnrB;
 188
 189             /* load j atom coordinates */
 190             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 192
 193             /* Calculate displacement vector */
 194             dx00             = _mm_sub_pd(ix0,jx0);
 195             dy00             = _mm_sub_pd(iy0,jy0);
 196             dz00             = _mm_sub_pd(iz0,jz0);
 197             dx01             = _mm_sub_pd(ix0,jx1);
 198             dy01             = _mm_sub_pd(iy0,jy1);
 199             dz01             = _mm_sub_pd(iz0,jz1);
 200             dx02             = _mm_sub_pd(ix0,jx2);
 201             dy02             = _mm_sub_pd(iy0,jy2);
 202             dz02             = _mm_sub_pd(iz0,jz2);
 203             dx10             = _mm_sub_pd(ix1,jx0);
 204             dy10             = _mm_sub_pd(iy1,jy0);
 205             dz10             = _mm_sub_pd(iz1,jz0);
 206             dx11             = _mm_sub_pd(ix1,jx1);
 207             dy11             = _mm_sub_pd(iy1,jy1);
 208             dz11             = _mm_sub_pd(iz1,jz1);
 209             dx12             = _mm_sub_pd(ix1,jx2);
 210             dy12             = _mm_sub_pd(iy1,jy2);
 211             dz12             = _mm_sub_pd(iz1,jz2);
 212             dx20             = _mm_sub_pd(ix2,jx0);
 213             dy20             = _mm_sub_pd(iy2,jy0);
 214             dz20             = _mm_sub_pd(iz2,jz0);
 215             dx21             = _mm_sub_pd(ix2,jx1);
 216             dy21             = _mm_sub_pd(iy2,jy1);
 217             dz21             = _mm_sub_pd(iz2,jz1);
 218             dx22             = _mm_sub_pd(ix2,jx2);
 219             dy22             = _mm_sub_pd(iy2,jy2);
 220             dz22             = _mm_sub_pd(iz2,jz2);
 221
 222             /* Calculate squared distance and things based on it */
 223             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 224             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 225             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 226             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 227             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 228             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 229             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 230             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 231             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 232
 233             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 234             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 235             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 236             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 237             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 238             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 239             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 240             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 241             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 242
 243             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 244             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 245             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 246             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 247             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 248             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 249             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 250             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 251             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 252
 253             fjx0             = _mm_setzero_pd();
 254             fjy0             = _mm_setzero_pd();
 255             fjz0             = _mm_setzero_pd();
 256             fjx1             = _mm_setzero_pd();
 257             fjy1             = _mm_setzero_pd();
 258             fjz1             = _mm_setzero_pd();
 259             fjx2             = _mm_setzero_pd();
 260             fjy2             = _mm_setzero_pd();
 261             fjz2             = _mm_setzero_pd();
 262
 263             /**************************
 264              * CALCULATE INTERACTIONS *
 265              **************************/
 266
 267             /* COULOMB ELECTROSTATICS */
 268             velec            = _mm_mul_pd(qq00,rinv00);
 269             felec            = _mm_mul_pd(velec,rinvsq00);
 270
 271             /* LENNARD-JONES DISPERSION/REPULSION */
 272
 273             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 274             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 275             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 276             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 277             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 278
 279             /* Update potential sum for this i atom from the interaction with this j atom. */
 280             velecsum         = _mm_add_pd(velecsum,velec);
 281             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 282
 283             fscal            = _mm_add_pd(felec,fvdw);
 284
 285             /* Update vectorial force */
 286             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 287             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 288             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 289
 290             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 291             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 292             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             /* COULOMB ELECTROSTATICS */
 299             velec            = _mm_mul_pd(qq01,rinv01);
 300             felec            = _mm_mul_pd(velec,rinvsq01);
 301
 302             /* Update potential sum for this i atom from the interaction with this j atom. */
 303             velecsum         = _mm_add_pd(velecsum,velec);
 304
 305             fscal            = felec;
 306
 307             /* Update vectorial force */
 308             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 309             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 310             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 311
 312             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 313             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 314             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             /* COULOMB ELECTROSTATICS */
 321             velec            = _mm_mul_pd(qq02,rinv02);
 322             felec            = _mm_mul_pd(velec,rinvsq02);
 323
 324             /* Update potential sum for this i atom from the interaction with this j atom. */
 325             velecsum         = _mm_add_pd(velecsum,velec);
 326
 327             fscal            = felec;
 328
 329             /* Update vectorial force */
 330             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 331             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 332             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 333
 334             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 335             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 336             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             /* COULOMB ELECTROSTATICS */
 343             velec            = _mm_mul_pd(qq10,rinv10);
 344             felec            = _mm_mul_pd(velec,rinvsq10);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velecsum         = _mm_add_pd(velecsum,velec);
 348
 349             fscal            = felec;
 350
 351             /* Update vectorial force */
 352             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 353             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 354             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 355
 356             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 357             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 358             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 359
 360             /**************************
 361              * CALCULATE INTERACTIONS *
 362              **************************/
 363
 364             /* COULOMB ELECTROSTATICS */
 365             velec            = _mm_mul_pd(qq11,rinv11);
 366             felec            = _mm_mul_pd(velec,rinvsq11);
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velecsum         = _mm_add_pd(velecsum,velec);
 370
 371             fscal            = felec;
 372
 373             /* Update vectorial force */
 374             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 375             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 376             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 377
 378             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 379             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 380             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             /* COULOMB ELECTROSTATICS */
 387             velec            = _mm_mul_pd(qq12,rinv12);
 388             felec            = _mm_mul_pd(velec,rinvsq12);
 389
 390             /* Update potential sum for this i atom from the interaction with this j atom. */
 391             velecsum         = _mm_add_pd(velecsum,velec);
 392
 393             fscal            = felec;
 394
 395             /* Update vectorial force */
 396             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 397             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 398             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 399
 400             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 401             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 402             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 403
 404             /**************************
 405              * CALCULATE INTERACTIONS *
 406              **************************/
 407
 408             /* COULOMB ELECTROSTATICS */
 409             velec            = _mm_mul_pd(qq20,rinv20);
 410             felec            = _mm_mul_pd(velec,rinvsq20);
 411
 412             /* Update potential sum for this i atom from the interaction with this j atom. */
 413             velecsum         = _mm_add_pd(velecsum,velec);
 414
 415             fscal            = felec;
 416
 417             /* Update vectorial force */
 418             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 419             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 420             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 421
 422             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 423             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 424             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 425
 426             /**************************
 427              * CALCULATE INTERACTIONS *
 428              **************************/
 429
 430             /* COULOMB ELECTROSTATICS */
 431             velec            = _mm_mul_pd(qq21,rinv21);
 432             felec            = _mm_mul_pd(velec,rinvsq21);
 433
 434             /* Update potential sum for this i atom from the interaction with this j atom. */
 435             velecsum         = _mm_add_pd(velecsum,velec);
 436
 437             fscal            = felec;
 438
 439             /* Update vectorial force */
 440             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 441             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 442             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 443
 444             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 445             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 446             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 447
 448             /**************************
 449              * CALCULATE INTERACTIONS *
 450              **************************/
 451
 452             /* COULOMB ELECTROSTATICS */
 453             velec            = _mm_mul_pd(qq22,rinv22);
 454             felec            = _mm_mul_pd(velec,rinvsq22);
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             velecsum         = _mm_add_pd(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             /* Update vectorial force */
 462             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 463             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 464             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 465
 466             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 467             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 468             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 469
 470             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 471
 472             /* Inner loop uses 291 flops */
 473         }
 474
 475         if(jidx<j_index_end)
 476         {
 477
 478             jnrA             = jjnr[jidx];
 479             j_coord_offsetA  = DIM*jnrA;
 480
 481             /* load j atom coordinates */
 482             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 483                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 484
 485             /* Calculate displacement vector */
 486             dx00             = _mm_sub_pd(ix0,jx0);
 487             dy00             = _mm_sub_pd(iy0,jy0);
 488             dz00             = _mm_sub_pd(iz0,jz0);
 489             dx01             = _mm_sub_pd(ix0,jx1);
 490             dy01             = _mm_sub_pd(iy0,jy1);
 491             dz01             = _mm_sub_pd(iz0,jz1);
 492             dx02             = _mm_sub_pd(ix0,jx2);
 493             dy02             = _mm_sub_pd(iy0,jy2);
 494             dz02             = _mm_sub_pd(iz0,jz2);
 495             dx10             = _mm_sub_pd(ix1,jx0);
 496             dy10             = _mm_sub_pd(iy1,jy0);
 497             dz10             = _mm_sub_pd(iz1,jz0);
 498             dx11             = _mm_sub_pd(ix1,jx1);
 499             dy11             = _mm_sub_pd(iy1,jy1);
 500             dz11             = _mm_sub_pd(iz1,jz1);
 501             dx12             = _mm_sub_pd(ix1,jx2);
 502             dy12             = _mm_sub_pd(iy1,jy2);
 503             dz12             = _mm_sub_pd(iz1,jz2);
 504             dx20             = _mm_sub_pd(ix2,jx0);
 505             dy20             = _mm_sub_pd(iy2,jy0);
 506             dz20             = _mm_sub_pd(iz2,jz0);
 507             dx21             = _mm_sub_pd(ix2,jx1);
 508             dy21             = _mm_sub_pd(iy2,jy1);
 509             dz21             = _mm_sub_pd(iz2,jz1);
 510             dx22             = _mm_sub_pd(ix2,jx2);
 511             dy22             = _mm_sub_pd(iy2,jy2);
 512             dz22             = _mm_sub_pd(iz2,jz2);
 513
 514             /* Calculate squared distance and things based on it */
 515             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 516             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 517             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 518             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 519             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 520             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 521             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 522             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 523             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 524
 525             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 526             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 527             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 528             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 529             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 530             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 531             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 532             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 533             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 534
 535             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 536             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 537             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 538             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 539             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 540             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 541             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 542             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 543             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 544
 545             fjx0             = _mm_setzero_pd();
 546             fjy0             = _mm_setzero_pd();
 547             fjz0             = _mm_setzero_pd();
 548             fjx1             = _mm_setzero_pd();
 549             fjy1             = _mm_setzero_pd();
 550             fjz1             = _mm_setzero_pd();
 551             fjx2             = _mm_setzero_pd();
 552             fjy2             = _mm_setzero_pd();
 553             fjz2             = _mm_setzero_pd();
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             /* COULOMB ELECTROSTATICS */
 560             velec            = _mm_mul_pd(qq00,rinv00);
 561             felec            = _mm_mul_pd(velec,rinvsq00);
 562
 563             /* LENNARD-JONES DISPERSION/REPULSION */
 564
 565             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 566             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 567             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 568             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 569             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 570
 571             /* Update potential sum for this i atom from the interaction with this j atom. */
 572             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 573             velecsum         = _mm_add_pd(velecsum,velec);
 574             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 575             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 576
 577             fscal            = _mm_add_pd(felec,fvdw);
 578
 579             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 580
 581             /* Update vectorial force */
 582             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 583             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 584             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 585
 586             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 587             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 588             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 589
 590             /**************************
 591              * CALCULATE INTERACTIONS *
 592              **************************/
 593
 594             /* COULOMB ELECTROSTATICS */
 595             velec            = _mm_mul_pd(qq01,rinv01);
 596             felec            = _mm_mul_pd(velec,rinvsq01);
 597
 598             /* Update potential sum for this i atom from the interaction with this j atom. */
 599             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 600             velecsum         = _mm_add_pd(velecsum,velec);
 601
 602             fscal            = felec;
 603
 604             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 605
 606             /* Update vectorial force */
 607             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 608             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 609             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 610
 611             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 612             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 613             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 614
 615             /**************************
 616              * CALCULATE INTERACTIONS *
 617              **************************/
 618
 619             /* COULOMB ELECTROSTATICS */
 620             velec            = _mm_mul_pd(qq02,rinv02);
 621             felec            = _mm_mul_pd(velec,rinvsq02);
 622
 623             /* Update potential sum for this i atom from the interaction with this j atom. */
 624             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 625             velecsum         = _mm_add_pd(velecsum,velec);
 626
 627             fscal            = felec;
 628
 629             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 630
 631             /* Update vectorial force */
 632             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 633             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 634             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 635
 636             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 637             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 638             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 639
 640             /**************************
 641              * CALCULATE INTERACTIONS *
 642              **************************/
 643
 644             /* COULOMB ELECTROSTATICS */
 645             velec            = _mm_mul_pd(qq10,rinv10);
 646             felec            = _mm_mul_pd(velec,rinvsq10);
 647
 648             /* Update potential sum for this i atom from the interaction with this j atom. */
 649             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 650             velecsum         = _mm_add_pd(velecsum,velec);
 651
 652             fscal            = felec;
 653
 654             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 655
 656             /* Update vectorial force */
 657             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 658             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 659             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 660
 661             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 662             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 663             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 664
 665             /**************************
 666              * CALCULATE INTERACTIONS *
 667              **************************/
 668
 669             /* COULOMB ELECTROSTATICS */
 670             velec            = _mm_mul_pd(qq11,rinv11);
 671             felec            = _mm_mul_pd(velec,rinvsq11);
 672
 673             /* Update potential sum for this i atom from the interaction with this j atom. */
 674             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 675             velecsum         = _mm_add_pd(velecsum,velec);
 676
 677             fscal            = felec;
 678
 679             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 680
 681             /* Update vectorial force */
 682             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 683             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 684             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 685
 686             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 687             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 688             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 689
 690             /**************************
 691              * CALCULATE INTERACTIONS *
 692              **************************/
 693
 694             /* COULOMB ELECTROSTATICS */
 695             velec            = _mm_mul_pd(qq12,rinv12);
 696             felec            = _mm_mul_pd(velec,rinvsq12);
 697
 698             /* Update potential sum for this i atom from the interaction with this j atom. */
 699             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 700             velecsum         = _mm_add_pd(velecsum,velec);
 701
 702             fscal            = felec;
 703
 704             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 705
 706             /* Update vectorial force */
 707             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 708             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 709             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 710
 711             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 712             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 713             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 714
 715             /**************************
 716              * CALCULATE INTERACTIONS *
 717              **************************/
 718
 719             /* COULOMB ELECTROSTATICS */
 720             velec            = _mm_mul_pd(qq20,rinv20);
 721             felec            = _mm_mul_pd(velec,rinvsq20);
 722
 723             /* Update potential sum for this i atom from the interaction with this j atom. */
 724             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 725             velecsum         = _mm_add_pd(velecsum,velec);
 726
 727             fscal            = felec;
 728
 729             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 730
 731             /* Update vectorial force */
 732             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 733             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 734             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 735
 736             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 737             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 738             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 739
 740             /**************************
 741              * CALCULATE INTERACTIONS *
 742              **************************/
 743
 744             /* COULOMB ELECTROSTATICS */
 745             velec            = _mm_mul_pd(qq21,rinv21);
 746             felec            = _mm_mul_pd(velec,rinvsq21);
 747
 748             /* Update potential sum for this i atom from the interaction with this j atom. */
 749             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 750             velecsum         = _mm_add_pd(velecsum,velec);
 751
 752             fscal            = felec;
 753
 754             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 755
 756             /* Update vectorial force */
 757             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 758             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 759             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 760
 761             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 762             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 763             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             /* COULOMB ELECTROSTATICS */
 770             velec            = _mm_mul_pd(qq22,rinv22);
 771             felec            = _mm_mul_pd(velec,rinvsq22);
 772
 773             /* Update potential sum for this i atom from the interaction with this j atom. */
 774             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 775             velecsum         = _mm_add_pd(velecsum,velec);
 776
 777             fscal            = felec;
 778
 779             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 780
 781             /* Update vectorial force */
 782             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 783             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 784             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 785
 786             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 787             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 788             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 789
 790             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 791
 792             /* Inner loop uses 291 flops */
 793         }
 794
 795         /* End of innermost loop */
 796
 797         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 798                                               f+i_coord_offset,fshift+i_shift_offset);
 799
 800         ggid                        = gid[iidx];
 801         /* Update potential energies */
 802         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 803         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 804
 805         /* Increment number of inner iterations */
 806         inneriter                  += j_index_end - j_index_start;
 807
 808         /* Outer loop uses 20 flops */
 809     }
 810
 811     /* Increment number of outer iterations */
 812     outeriter        += nri;
 813
 814     /* Update outer/inner flops */
 815
 816     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*291);
 817 }
 818 /*
 819  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_avx_128_fma_double
 820  * Electrostatics interaction: Coulomb
 821  * VdW interaction:            LennardJones
 822  * Geometry:                   Water3-Water3
 823  * Calculate force/pot:        Force
 824  */
 825 void
 826 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_avx_128_fma_double
 827                     (t_nblist * gmx_restrict                nlist,
 828                      rvec * gmx_restrict                    xx,
 829                      rvec * gmx_restrict                    ff,
 830                      t_forcerec * gmx_restrict              fr,
 831                      t_mdatoms * gmx_restrict               mdatoms,
 832                      nb_kernel_data_t * gmx_restrict        kernel_data,
 833                      t_nrnb * gmx_restrict                  nrnb)
 834 {
 835     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 836      * just 0 for non-waters.
 837      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 838      * jnr indices corresponding to data put in the four positions in the SIMD register.
 839      */
 840     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 841     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 842     int              jnrA,jnrB;
 843     int              j_coord_offsetA,j_coord_offsetB;
 844     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 845     real             rcutoff_scalar;
 846     real             *shiftvec,*fshift,*x,*f;
 847     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 848     int              vdwioffset0;
 849     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 850     int              vdwioffset1;
 851     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 852     int              vdwioffset2;
 853     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 854     int              vdwjidx0A,vdwjidx0B;
 855     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 856     int              vdwjidx1A,vdwjidx1B;
 857     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 858     int              vdwjidx2A,vdwjidx2B;
 859     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 860     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 861     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 862     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 863     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 864     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 865     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 866     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 867     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 868     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 869     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 870     real             *charge;
 871     int              nvdwtype;
 872     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 873     int              *vdwtype;
 874     real             *vdwparam;
 875     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 876     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 877     __m128d          dummy_mask,cutoff_mask;
 878     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 879     __m128d          one     = _mm_set1_pd(1.0);
 880     __m128d          two     = _mm_set1_pd(2.0);
 881     x                = xx[0];
 882     f                = ff[0];
 883
 884     nri              = nlist->nri;
 885     iinr             = nlist->iinr;
 886     jindex           = nlist->jindex;
 887     jjnr             = nlist->jjnr;
 888     shiftidx         = nlist->shift;
 889     gid              = nlist->gid;
 890     shiftvec         = fr->shift_vec[0];
 891     fshift           = fr->fshift[0];
 892     facel            = _mm_set1_pd(fr->epsfac);
 893     charge           = mdatoms->chargeA;
 894     nvdwtype         = fr->ntype;
 895     vdwparam         = fr->nbfp;
 896     vdwtype          = mdatoms->typeA;
 897
 898     /* Setup water-specific parameters */
 899     inr              = nlist->iinr[0];
 900     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 901     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 902     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 903     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 904
 905     jq0              = _mm_set1_pd(charge[inr+0]);
 906     jq1              = _mm_set1_pd(charge[inr+1]);
 907     jq2              = _mm_set1_pd(charge[inr+2]);
 908     vdwjidx0A        = 2*vdwtype[inr+0];
 909     qq00             = _mm_mul_pd(iq0,jq0);
 910     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 911     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 912     qq01             = _mm_mul_pd(iq0,jq1);
 913     qq02             = _mm_mul_pd(iq0,jq2);
 914     qq10             = _mm_mul_pd(iq1,jq0);
 915     qq11             = _mm_mul_pd(iq1,jq1);
 916     qq12             = _mm_mul_pd(iq1,jq2);
 917     qq20             = _mm_mul_pd(iq2,jq0);
 918     qq21             = _mm_mul_pd(iq2,jq1);
 919     qq22             = _mm_mul_pd(iq2,jq2);
 920
 921     /* Avoid stupid compiler warnings */
 922     jnrA = jnrB = 0;
 923     j_coord_offsetA = 0;
 924     j_coord_offsetB = 0;
 925
 926     outeriter        = 0;
 927     inneriter        = 0;
 928
 929     /* Start outer loop over neighborlists */
 930     for(iidx=0; iidx<nri; iidx++)
 931     {
 932         /* Load shift vector for this list */
 933         i_shift_offset   = DIM*shiftidx[iidx];
 934
 935         /* Load limits for loop over neighbors */
 936         j_index_start    = jindex[iidx];
 937         j_index_end      = jindex[iidx+1];
 938
 939         /* Get outer coordinate index */
 940         inr              = iinr[iidx];
 941         i_coord_offset   = DIM*inr;
 942
 943         /* Load i particle coords and add shift vector */
 944         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 945                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 946
 947         fix0             = _mm_setzero_pd();
 948         fiy0             = _mm_setzero_pd();
 949         fiz0             = _mm_setzero_pd();
 950         fix1             = _mm_setzero_pd();
 951         fiy1             = _mm_setzero_pd();
 952         fiz1             = _mm_setzero_pd();
 953         fix2             = _mm_setzero_pd();
 954         fiy2             = _mm_setzero_pd();
 955         fiz2             = _mm_setzero_pd();
 956
 957         /* Start inner kernel loop */
 958         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 959         {
 960
 961             /* Get j neighbor index, and coordinate index */
 962             jnrA             = jjnr[jidx];
 963             jnrB             = jjnr[jidx+1];
 964             j_coord_offsetA  = DIM*jnrA;
 965             j_coord_offsetB  = DIM*jnrB;
 966
 967             /* load j atom coordinates */
 968             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 969                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 970
 971             /* Calculate displacement vector */
 972             dx00             = _mm_sub_pd(ix0,jx0);
 973             dy00             = _mm_sub_pd(iy0,jy0);
 974             dz00             = _mm_sub_pd(iz0,jz0);
 975             dx01             = _mm_sub_pd(ix0,jx1);
 976             dy01             = _mm_sub_pd(iy0,jy1);
 977             dz01             = _mm_sub_pd(iz0,jz1);
 978             dx02             = _mm_sub_pd(ix0,jx2);
 979             dy02             = _mm_sub_pd(iy0,jy2);
 980             dz02             = _mm_sub_pd(iz0,jz2);
 981             dx10             = _mm_sub_pd(ix1,jx0);
 982             dy10             = _mm_sub_pd(iy1,jy0);
 983             dz10             = _mm_sub_pd(iz1,jz0);
 984             dx11             = _mm_sub_pd(ix1,jx1);
 985             dy11             = _mm_sub_pd(iy1,jy1);
 986             dz11             = _mm_sub_pd(iz1,jz1);
 987             dx12             = _mm_sub_pd(ix1,jx2);
 988             dy12             = _mm_sub_pd(iy1,jy2);
 989             dz12             = _mm_sub_pd(iz1,jz2);
 990             dx20             = _mm_sub_pd(ix2,jx0);
 991             dy20             = _mm_sub_pd(iy2,jy0);
 992             dz20             = _mm_sub_pd(iz2,jz0);
 993             dx21             = _mm_sub_pd(ix2,jx1);
 994             dy21             = _mm_sub_pd(iy2,jy1);
 995             dz21             = _mm_sub_pd(iz2,jz1);
 996             dx22             = _mm_sub_pd(ix2,jx2);
 997             dy22             = _mm_sub_pd(iy2,jy2);
 998             dz22             = _mm_sub_pd(iz2,jz2);
 999
1000             /* Calculate squared distance and things based on it */
1001             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1002             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1003             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1004             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1005             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1006             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1007             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1008             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1009             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1010
1011             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1012             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1013             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1014             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1015             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1016             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1017             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1018             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1019             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1020
1021             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1022             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1023             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1024             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1025             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1026             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1027             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1028             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1029             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1030
1031             fjx0             = _mm_setzero_pd();
1032             fjy0             = _mm_setzero_pd();
1033             fjz0             = _mm_setzero_pd();
1034             fjx1             = _mm_setzero_pd();
1035             fjy1             = _mm_setzero_pd();
1036             fjz1             = _mm_setzero_pd();
1037             fjx2             = _mm_setzero_pd();
1038             fjy2             = _mm_setzero_pd();
1039             fjz2             = _mm_setzero_pd();
1040
1041             /**************************
1042              * CALCULATE INTERACTIONS *
1043              **************************/
1044
1045             /* COULOMB ELECTROSTATICS */
1046             velec            = _mm_mul_pd(qq00,rinv00);
1047             felec            = _mm_mul_pd(velec,rinvsq00);
1048
1049             /* LENNARD-JONES DISPERSION/REPULSION */
1050
1051             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1052             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1053
1054             fscal            = _mm_add_pd(felec,fvdw);
1055
1056             /* Update vectorial force */
1057             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1058             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1059             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1060
1061             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1062             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1063             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1064
1065             /**************************
1066              * CALCULATE INTERACTIONS *
1067              **************************/
1068
1069             /* COULOMB ELECTROSTATICS */
1070             velec            = _mm_mul_pd(qq01,rinv01);
1071             felec            = _mm_mul_pd(velec,rinvsq01);
1072
1073             fscal            = felec;
1074
1075             /* Update vectorial force */
1076             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1077             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1078             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1079
1080             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1081             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1082             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1083
1084             /**************************
1085              * CALCULATE INTERACTIONS *
1086              **************************/
1087
1088             /* COULOMB ELECTROSTATICS */
1089             velec            = _mm_mul_pd(qq02,rinv02);
1090             felec            = _mm_mul_pd(velec,rinvsq02);
1091
1092             fscal            = felec;
1093
1094             /* Update vectorial force */
1095             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1096             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1097             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1098
1099             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1100             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1101             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1102
1103             /**************************
1104              * CALCULATE INTERACTIONS *
1105              **************************/
1106
1107             /* COULOMB ELECTROSTATICS */
1108             velec            = _mm_mul_pd(qq10,rinv10);
1109             felec            = _mm_mul_pd(velec,rinvsq10);
1110
1111             fscal            = felec;
1112
1113             /* Update vectorial force */
1114             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1115             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1116             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1117
1118             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1119             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1120             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1121
1122             /**************************
1123              * CALCULATE INTERACTIONS *
1124              **************************/
1125
1126             /* COULOMB ELECTROSTATICS */
1127             velec            = _mm_mul_pd(qq11,rinv11);
1128             felec            = _mm_mul_pd(velec,rinvsq11);
1129
1130             fscal            = felec;
1131
1132             /* Update vectorial force */
1133             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1134             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1135             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1136
1137             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1138             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1139             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1140
1141             /**************************
1142              * CALCULATE INTERACTIONS *
1143              **************************/
1144
1145             /* COULOMB ELECTROSTATICS */
1146             velec            = _mm_mul_pd(qq12,rinv12);
1147             felec            = _mm_mul_pd(velec,rinvsq12);
1148
1149             fscal            = felec;
1150
1151             /* Update vectorial force */
1152             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1153             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1154             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1155
1156             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1157             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1158             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1159
1160             /**************************
1161              * CALCULATE INTERACTIONS *
1162              **************************/
1163
1164             /* COULOMB ELECTROSTATICS */
1165             velec            = _mm_mul_pd(qq20,rinv20);
1166             felec            = _mm_mul_pd(velec,rinvsq20);
1167
1168             fscal            = felec;
1169
1170             /* Update vectorial force */
1171             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1172             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1173             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1174
1175             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1176             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1177             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1178
1179             /**************************
1180              * CALCULATE INTERACTIONS *
1181              **************************/
1182
1183             /* COULOMB ELECTROSTATICS */
1184             velec            = _mm_mul_pd(qq21,rinv21);
1185             felec            = _mm_mul_pd(velec,rinvsq21);
1186
1187             fscal            = felec;
1188
1189             /* Update vectorial force */
1190             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1191             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1192             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1193
1194             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1195             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1196             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1197
1198             /**************************
1199              * CALCULATE INTERACTIONS *
1200              **************************/
1201
1202             /* COULOMB ELECTROSTATICS */
1203             velec            = _mm_mul_pd(qq22,rinv22);
1204             felec            = _mm_mul_pd(velec,rinvsq22);
1205
1206             fscal            = felec;
1207
1208             /* Update vectorial force */
1209             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1210             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1211             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1212
1213             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1214             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1215             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1216
1217             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1218
1219             /* Inner loop uses 277 flops */
1220         }
1221
1222         if(jidx<j_index_end)
1223         {
1224
1225             jnrA             = jjnr[jidx];
1226             j_coord_offsetA  = DIM*jnrA;
1227
1228             /* load j atom coordinates */
1229             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1230                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1231
1232             /* Calculate displacement vector */
1233             dx00             = _mm_sub_pd(ix0,jx0);
1234             dy00             = _mm_sub_pd(iy0,jy0);
1235             dz00             = _mm_sub_pd(iz0,jz0);
1236             dx01             = _mm_sub_pd(ix0,jx1);
1237             dy01             = _mm_sub_pd(iy0,jy1);
1238             dz01             = _mm_sub_pd(iz0,jz1);
1239             dx02             = _mm_sub_pd(ix0,jx2);
1240             dy02             = _mm_sub_pd(iy0,jy2);
1241             dz02             = _mm_sub_pd(iz0,jz2);
1242             dx10             = _mm_sub_pd(ix1,jx0);
1243             dy10             = _mm_sub_pd(iy1,jy0);
1244             dz10             = _mm_sub_pd(iz1,jz0);
1245             dx11             = _mm_sub_pd(ix1,jx1);
1246             dy11             = _mm_sub_pd(iy1,jy1);
1247             dz11             = _mm_sub_pd(iz1,jz1);
1248             dx12             = _mm_sub_pd(ix1,jx2);
1249             dy12             = _mm_sub_pd(iy1,jy2);
1250             dz12             = _mm_sub_pd(iz1,jz2);
1251             dx20             = _mm_sub_pd(ix2,jx0);
1252             dy20             = _mm_sub_pd(iy2,jy0);
1253             dz20             = _mm_sub_pd(iz2,jz0);
1254             dx21             = _mm_sub_pd(ix2,jx1);
1255             dy21             = _mm_sub_pd(iy2,jy1);
1256             dz21             = _mm_sub_pd(iz2,jz1);
1257             dx22             = _mm_sub_pd(ix2,jx2);
1258             dy22             = _mm_sub_pd(iy2,jy2);
1259             dz22             = _mm_sub_pd(iz2,jz2);
1260
1261             /* Calculate squared distance and things based on it */
1262             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1263             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1264             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1265             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1266             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1267             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1268             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1269             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1270             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1271
1272             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1273             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1274             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1275             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1276             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1277             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1278             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1279             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1280             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1281
1282             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1283             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1284             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1285             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1286             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1287             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1288             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1289             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1290             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1291
1292             fjx0             = _mm_setzero_pd();
1293             fjy0             = _mm_setzero_pd();
1294             fjz0             = _mm_setzero_pd();
1295             fjx1             = _mm_setzero_pd();
1296             fjy1             = _mm_setzero_pd();
1297             fjz1             = _mm_setzero_pd();
1298             fjx2             = _mm_setzero_pd();
1299             fjy2             = _mm_setzero_pd();
1300             fjz2             = _mm_setzero_pd();
1301
1302             /**************************
1303              * CALCULATE INTERACTIONS *
1304              **************************/
1305
1306             /* COULOMB ELECTROSTATICS */
1307             velec            = _mm_mul_pd(qq00,rinv00);
1308             felec            = _mm_mul_pd(velec,rinvsq00);
1309
1310             /* LENNARD-JONES DISPERSION/REPULSION */
1311
1312             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1313             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1314
1315             fscal            = _mm_add_pd(felec,fvdw);
1316
1317             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1318
1319             /* Update vectorial force */
1320             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1321             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1322             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1323
1324             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1325             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1326             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1327
1328             /**************************
1329              * CALCULATE INTERACTIONS *
1330              **************************/
1331
1332             /* COULOMB ELECTROSTATICS */
1333             velec            = _mm_mul_pd(qq01,rinv01);
1334             felec            = _mm_mul_pd(velec,rinvsq01);
1335
1336             fscal            = felec;
1337
1338             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1339
1340             /* Update vectorial force */
1341             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1342             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1343             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1344
1345             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1346             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1347             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1348
1349             /**************************
1350              * CALCULATE INTERACTIONS *
1351              **************************/
1352
1353             /* COULOMB ELECTROSTATICS */
1354             velec            = _mm_mul_pd(qq02,rinv02);
1355             felec            = _mm_mul_pd(velec,rinvsq02);
1356
1357             fscal            = felec;
1358
1359             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1360
1361             /* Update vectorial force */
1362             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1363             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1364             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1365
1366             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1367             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1368             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1369
1370             /**************************
1371              * CALCULATE INTERACTIONS *
1372              **************************/
1373
1374             /* COULOMB ELECTROSTATICS */
1375             velec            = _mm_mul_pd(qq10,rinv10);
1376             felec            = _mm_mul_pd(velec,rinvsq10);
1377
1378             fscal            = felec;
1379
1380             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1381
1382             /* Update vectorial force */
1383             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1384             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1385             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1386
1387             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1388             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1389             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1390
1391             /**************************
1392              * CALCULATE INTERACTIONS *
1393              **************************/
1394
1395             /* COULOMB ELECTROSTATICS */
1396             velec            = _mm_mul_pd(qq11,rinv11);
1397             felec            = _mm_mul_pd(velec,rinvsq11);
1398
1399             fscal            = felec;
1400
1401             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1402
1403             /* Update vectorial force */
1404             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1405             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1406             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1407
1408             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1409             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1410             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             /* COULOMB ELECTROSTATICS */
1417             velec            = _mm_mul_pd(qq12,rinv12);
1418             felec            = _mm_mul_pd(velec,rinvsq12);
1419
1420             fscal            = felec;
1421
1422             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1423
1424             /* Update vectorial force */
1425             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1426             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1427             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1428
1429             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1430             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1431             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1432
1433             /**************************
1434              * CALCULATE INTERACTIONS *
1435              **************************/
1436
1437             /* COULOMB ELECTROSTATICS */
1438             velec            = _mm_mul_pd(qq20,rinv20);
1439             felec            = _mm_mul_pd(velec,rinvsq20);
1440
1441             fscal            = felec;
1442
1443             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1444
1445             /* Update vectorial force */
1446             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1447             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1448             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1449
1450             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1451             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1452             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1453
1454             /**************************
1455              * CALCULATE INTERACTIONS *
1456              **************************/
1457
1458             /* COULOMB ELECTROSTATICS */
1459             velec            = _mm_mul_pd(qq21,rinv21);
1460             felec            = _mm_mul_pd(velec,rinvsq21);
1461
1462             fscal            = felec;
1463
1464             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1465
1466             /* Update vectorial force */
1467             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1468             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1469             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1470
1471             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1472             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1473             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1474
1475             /**************************
1476              * CALCULATE INTERACTIONS *
1477              **************************/
1478
1479             /* COULOMB ELECTROSTATICS */
1480             velec            = _mm_mul_pd(qq22,rinv22);
1481             felec            = _mm_mul_pd(velec,rinvsq22);
1482
1483             fscal            = felec;
1484
1485             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1486
1487             /* Update vectorial force */
1488             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1489             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1490             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1491
1492             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1493             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1494             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1495
1496             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1497
1498             /* Inner loop uses 277 flops */
1499         }
1500
1501         /* End of innermost loop */
1502
1503         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1504                                               f+i_coord_offset,fshift+i_shift_offset);
1505
1506         /* Increment number of inner iterations */
1507         inneriter                  += j_index_end - j_index_start;
1508
1509         /* Outer loop uses 18 flops */
1510     }
1511
1512     /* Increment number of outer iterations */
1513     outeriter        += nri;
1514
1515     /* Update outer/inner flops */
1516
1517     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1518 }