src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwjidx0A,vdwjidx0B;
  87     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  94     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  95     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     int              nvdwtype;
 104     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 105     int              *vdwtype;
 106     real             *vdwparam;
 107     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 108     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 109     __m128d          dummy_mask,cutoff_mask;
 110     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 111     __m128d          one     = _mm_set1_pd(1.0);
 112     __m128d          two     = _mm_set1_pd(2.0);
 113     x                = xx[0];
 114     f                = ff[0];
 115
 116     nri              = nlist->nri;
 117     iinr             = nlist->iinr;
 118     jindex           = nlist->jindex;
 119     jjnr             = nlist->jjnr;
 120     shiftidx         = nlist->shift;
 121     gid              = nlist->gid;
 122     shiftvec         = fr->shift_vec[0];
 123     fshift           = fr->fshift[0];
 124     facel            = _mm_set1_pd(fr->epsfac);
 125     charge           = mdatoms->chargeA;
 126     nvdwtype         = fr->ntype;
 127     vdwparam         = fr->nbfp;
 128     vdwtype          = mdatoms->typeA;
 129
 130     /* Setup water-specific parameters */
 131     inr              = nlist->iinr[0];
 132     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 133     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 134     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 135     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 136
 137     jq0              = _mm_set1_pd(charge[inr+0]);
 138     jq1              = _mm_set1_pd(charge[inr+1]);
 139     jq2              = _mm_set1_pd(charge[inr+2]);
 140     vdwjidx0A        = 2*vdwtype[inr+0];
 141     qq00             = _mm_mul_pd(iq0,jq0);
 142     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 143     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 144     qq01             = _mm_mul_pd(iq0,jq1);
 145     qq02             = _mm_mul_pd(iq0,jq2);
 146     qq10             = _mm_mul_pd(iq1,jq0);
 147     qq11             = _mm_mul_pd(iq1,jq1);
 148     qq12             = _mm_mul_pd(iq1,jq2);
 149     qq20             = _mm_mul_pd(iq2,jq0);
 150     qq21             = _mm_mul_pd(iq2,jq1);
 151     qq22             = _mm_mul_pd(iq2,jq2);
 152
 153     /* Avoid stupid compiler warnings */
 154     jnrA = jnrB = 0;
 155     j_coord_offsetA = 0;
 156     j_coord_offsetB = 0;
 157
 158     outeriter        = 0;
 159     inneriter        = 0;
 160
 161     /* Start outer loop over neighborlists */
 162     for(iidx=0; iidx<nri; iidx++)
 163     {
 164         /* Load shift vector for this list */
 165         i_shift_offset   = DIM*shiftidx[iidx];
 166
 167         /* Load limits for loop over neighbors */
 168         j_index_start    = jindex[iidx];
 169         j_index_end      = jindex[iidx+1];
 170
 171         /* Get outer coordinate index */
 172         inr              = iinr[iidx];
 173         i_coord_offset   = DIM*inr;
 174
 175         /* Load i particle coords and add shift vector */
 176         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 177                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 178
 179         fix0             = _mm_setzero_pd();
 180         fiy0             = _mm_setzero_pd();
 181         fiz0             = _mm_setzero_pd();
 182         fix1             = _mm_setzero_pd();
 183         fiy1             = _mm_setzero_pd();
 184         fiz1             = _mm_setzero_pd();
 185         fix2             = _mm_setzero_pd();
 186         fiy2             = _mm_setzero_pd();
 187         fiz2             = _mm_setzero_pd();
 188
 189         /* Reset potential sums */
 190         velecsum         = _mm_setzero_pd();
 191         vvdwsum          = _mm_setzero_pd();
 192
 193         /* Start inner kernel loop */
 194         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 195         {
 196
 197             /* Get j neighbor index, and coordinate index */
 198             jnrA             = jjnr[jidx];
 199             jnrB             = jjnr[jidx+1];
 200             j_coord_offsetA  = DIM*jnrA;
 201             j_coord_offsetB  = DIM*jnrB;
 202
 203             /* load j atom coordinates */
 204             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 205                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 206
 207             /* Calculate displacement vector */
 208             dx00             = _mm_sub_pd(ix0,jx0);
 209             dy00             = _mm_sub_pd(iy0,jy0);
 210             dz00             = _mm_sub_pd(iz0,jz0);
 211             dx01             = _mm_sub_pd(ix0,jx1);
 212             dy01             = _mm_sub_pd(iy0,jy1);
 213             dz01             = _mm_sub_pd(iz0,jz1);
 214             dx02             = _mm_sub_pd(ix0,jx2);
 215             dy02             = _mm_sub_pd(iy0,jy2);
 216             dz02             = _mm_sub_pd(iz0,jz2);
 217             dx10             = _mm_sub_pd(ix1,jx0);
 218             dy10             = _mm_sub_pd(iy1,jy0);
 219             dz10             = _mm_sub_pd(iz1,jz0);
 220             dx11             = _mm_sub_pd(ix1,jx1);
 221             dy11             = _mm_sub_pd(iy1,jy1);
 222             dz11             = _mm_sub_pd(iz1,jz1);
 223             dx12             = _mm_sub_pd(ix1,jx2);
 224             dy12             = _mm_sub_pd(iy1,jy2);
 225             dz12             = _mm_sub_pd(iz1,jz2);
 226             dx20             = _mm_sub_pd(ix2,jx0);
 227             dy20             = _mm_sub_pd(iy2,jy0);
 228             dz20             = _mm_sub_pd(iz2,jz0);
 229             dx21             = _mm_sub_pd(ix2,jx1);
 230             dy21             = _mm_sub_pd(iy2,jy1);
 231             dz21             = _mm_sub_pd(iz2,jz1);
 232             dx22             = _mm_sub_pd(ix2,jx2);
 233             dy22             = _mm_sub_pd(iy2,jy2);
 234             dz22             = _mm_sub_pd(iz2,jz2);
 235
 236             /* Calculate squared distance and things based on it */
 237             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 238             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 239             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 240             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 241             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 242             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 243             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 244             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 245             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 246
 247             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 248             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 249             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 250             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 251             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 252             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 253             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 254             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 255             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 256
 257             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 258             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 259             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 260             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 261             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 262             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 263             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 264             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 265             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 266
 267             fjx0             = _mm_setzero_pd();
 268             fjy0             = _mm_setzero_pd();
 269             fjz0             = _mm_setzero_pd();
 270             fjx1             = _mm_setzero_pd();
 271             fjy1             = _mm_setzero_pd();
 272             fjz1             = _mm_setzero_pd();
 273             fjx2             = _mm_setzero_pd();
 274             fjy2             = _mm_setzero_pd();
 275             fjz2             = _mm_setzero_pd();
 276
 277             /**************************
 278              * CALCULATE INTERACTIONS *
 279              **************************/
 280
 281             /* COULOMB ELECTROSTATICS */
 282             velec            = _mm_mul_pd(qq00,rinv00);
 283             felec            = _mm_mul_pd(velec,rinvsq00);
 284
 285             /* LENNARD-JONES DISPERSION/REPULSION */
 286
 287             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 288             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 289             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 290             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 291             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 292
 293             /* Update potential sum for this i atom from the interaction with this j atom. */
 294             velecsum         = _mm_add_pd(velecsum,velec);
 295             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 296
 297             fscal            = _mm_add_pd(felec,fvdw);
 298
 299             /* Update vectorial force */
 300             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 301             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 302             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 303
 304             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 305             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 306             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             /* COULOMB ELECTROSTATICS */
 313             velec            = _mm_mul_pd(qq01,rinv01);
 314             felec            = _mm_mul_pd(velec,rinvsq01);
 315
 316             /* Update potential sum for this i atom from the interaction with this j atom. */
 317             velecsum         = _mm_add_pd(velecsum,velec);
 318
 319             fscal            = felec;
 320
 321             /* Update vectorial force */
 322             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 323             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 324             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 325
 326             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 327             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 328             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             /* COULOMB ELECTROSTATICS */
 335             velec            = _mm_mul_pd(qq02,rinv02);
 336             felec            = _mm_mul_pd(velec,rinvsq02);
 337
 338             /* Update potential sum for this i atom from the interaction with this j atom. */
 339             velecsum         = _mm_add_pd(velecsum,velec);
 340
 341             fscal            = felec;
 342
 343             /* Update vectorial force */
 344             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 345             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 346             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 347
 348             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 349             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 350             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 351
 352             /**************************
 353              * CALCULATE INTERACTIONS *
 354              **************************/
 355
 356             /* COULOMB ELECTROSTATICS */
 357             velec            = _mm_mul_pd(qq10,rinv10);
 358             felec            = _mm_mul_pd(velec,rinvsq10);
 359
 360             /* Update potential sum for this i atom from the interaction with this j atom. */
 361             velecsum         = _mm_add_pd(velecsum,velec);
 362
 363             fscal            = felec;
 364
 365             /* Update vectorial force */
 366             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 367             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 368             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 369
 370             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 371             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 372             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 373
 374             /**************************
 375              * CALCULATE INTERACTIONS *
 376              **************************/
 377
 378             /* COULOMB ELECTROSTATICS */
 379             velec            = _mm_mul_pd(qq11,rinv11);
 380             felec            = _mm_mul_pd(velec,rinvsq11);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velecsum         = _mm_add_pd(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387             /* Update vectorial force */
 388             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 389             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 390             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 391
 392             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 393             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 394             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 395
 396             /**************************
 397              * CALCULATE INTERACTIONS *
 398              **************************/
 399
 400             /* COULOMB ELECTROSTATICS */
 401             velec            = _mm_mul_pd(qq12,rinv12);
 402             felec            = _mm_mul_pd(velec,rinvsq12);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velecsum         = _mm_add_pd(velecsum,velec);
 406
 407             fscal            = felec;
 408
 409             /* Update vectorial force */
 410             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 411             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 412             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 413
 414             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 415             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 416             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 417
 418             /**************************
 419              * CALCULATE INTERACTIONS *
 420              **************************/
 421
 422             /* COULOMB ELECTROSTATICS */
 423             velec            = _mm_mul_pd(qq20,rinv20);
 424             felec            = _mm_mul_pd(velec,rinvsq20);
 425
 426             /* Update potential sum for this i atom from the interaction with this j atom. */
 427             velecsum         = _mm_add_pd(velecsum,velec);
 428
 429             fscal            = felec;
 430
 431             /* Update vectorial force */
 432             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 433             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 434             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 435
 436             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 437             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 438             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 439
 440             /**************************
 441              * CALCULATE INTERACTIONS *
 442              **************************/
 443
 444             /* COULOMB ELECTROSTATICS */
 445             velec            = _mm_mul_pd(qq21,rinv21);
 446             felec            = _mm_mul_pd(velec,rinvsq21);
 447
 448             /* Update potential sum for this i atom from the interaction with this j atom. */
 449             velecsum         = _mm_add_pd(velecsum,velec);
 450
 451             fscal            = felec;
 452
 453             /* Update vectorial force */
 454             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 455             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 456             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 457
 458             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 459             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 460             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             /* COULOMB ELECTROSTATICS */
 467             velec            = _mm_mul_pd(qq22,rinv22);
 468             felec            = _mm_mul_pd(velec,rinvsq22);
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velecsum         = _mm_add_pd(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475             /* Update vectorial force */
 476             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 477             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 478             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 479
 480             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 481             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 482             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 483
 484             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 485
 486             /* Inner loop uses 291 flops */
 487         }
 488
 489         if(jidx<j_index_end)
 490         {
 491
 492             jnrA             = jjnr[jidx];
 493             j_coord_offsetA  = DIM*jnrA;
 494
 495             /* load j atom coordinates */
 496             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 497                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 498
 499             /* Calculate displacement vector */
 500             dx00             = _mm_sub_pd(ix0,jx0);
 501             dy00             = _mm_sub_pd(iy0,jy0);
 502             dz00             = _mm_sub_pd(iz0,jz0);
 503             dx01             = _mm_sub_pd(ix0,jx1);
 504             dy01             = _mm_sub_pd(iy0,jy1);
 505             dz01             = _mm_sub_pd(iz0,jz1);
 506             dx02             = _mm_sub_pd(ix0,jx2);
 507             dy02             = _mm_sub_pd(iy0,jy2);
 508             dz02             = _mm_sub_pd(iz0,jz2);
 509             dx10             = _mm_sub_pd(ix1,jx0);
 510             dy10             = _mm_sub_pd(iy1,jy0);
 511             dz10             = _mm_sub_pd(iz1,jz0);
 512             dx11             = _mm_sub_pd(ix1,jx1);
 513             dy11             = _mm_sub_pd(iy1,jy1);
 514             dz11             = _mm_sub_pd(iz1,jz1);
 515             dx12             = _mm_sub_pd(ix1,jx2);
 516             dy12             = _mm_sub_pd(iy1,jy2);
 517             dz12             = _mm_sub_pd(iz1,jz2);
 518             dx20             = _mm_sub_pd(ix2,jx0);
 519             dy20             = _mm_sub_pd(iy2,jy0);
 520             dz20             = _mm_sub_pd(iz2,jz0);
 521             dx21             = _mm_sub_pd(ix2,jx1);
 522             dy21             = _mm_sub_pd(iy2,jy1);
 523             dz21             = _mm_sub_pd(iz2,jz1);
 524             dx22             = _mm_sub_pd(ix2,jx2);
 525             dy22             = _mm_sub_pd(iy2,jy2);
 526             dz22             = _mm_sub_pd(iz2,jz2);
 527
 528             /* Calculate squared distance and things based on it */
 529             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 530             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 531             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 532             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 533             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 534             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 535             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 536             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 537             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 538
 539             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 540             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 541             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 542             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 543             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 544             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 545             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 546             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 547             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 548
 549             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 550             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 551             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 552             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 553             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 554             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 555             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 556             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 557             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 558
 559             fjx0             = _mm_setzero_pd();
 560             fjy0             = _mm_setzero_pd();
 561             fjz0             = _mm_setzero_pd();
 562             fjx1             = _mm_setzero_pd();
 563             fjy1             = _mm_setzero_pd();
 564             fjz1             = _mm_setzero_pd();
 565             fjx2             = _mm_setzero_pd();
 566             fjy2             = _mm_setzero_pd();
 567             fjz2             = _mm_setzero_pd();
 568
 569             /**************************
 570              * CALCULATE INTERACTIONS *
 571              **************************/
 572
 573             /* COULOMB ELECTROSTATICS */
 574             velec            = _mm_mul_pd(qq00,rinv00);
 575             felec            = _mm_mul_pd(velec,rinvsq00);
 576
 577             /* LENNARD-JONES DISPERSION/REPULSION */
 578
 579             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 580             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 581             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 582             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 583             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 584
 585             /* Update potential sum for this i atom from the interaction with this j atom. */
 586             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 587             velecsum         = _mm_add_pd(velecsum,velec);
 588             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 589             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 590
 591             fscal            = _mm_add_pd(felec,fvdw);
 592
 593             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 594
 595             /* Update vectorial force */
 596             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 597             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 598             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 599
 600             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 601             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 602             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 603
 604             /**************************
 605              * CALCULATE INTERACTIONS *
 606              **************************/
 607
 608             /* COULOMB ELECTROSTATICS */
 609             velec            = _mm_mul_pd(qq01,rinv01);
 610             felec            = _mm_mul_pd(velec,rinvsq01);
 611
 612             /* Update potential sum for this i atom from the interaction with this j atom. */
 613             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 614             velecsum         = _mm_add_pd(velecsum,velec);
 615
 616             fscal            = felec;
 617
 618             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 619
 620             /* Update vectorial force */
 621             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 622             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 623             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 624
 625             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 626             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 627             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 628
 629             /**************************
 630              * CALCULATE INTERACTIONS *
 631              **************************/
 632
 633             /* COULOMB ELECTROSTATICS */
 634             velec            = _mm_mul_pd(qq02,rinv02);
 635             felec            = _mm_mul_pd(velec,rinvsq02);
 636
 637             /* Update potential sum for this i atom from the interaction with this j atom. */
 638             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 639             velecsum         = _mm_add_pd(velecsum,velec);
 640
 641             fscal            = felec;
 642
 643             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 644
 645             /* Update vectorial force */
 646             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 647             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 648             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 649
 650             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 651             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 652             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 653
 654             /**************************
 655              * CALCULATE INTERACTIONS *
 656              **************************/
 657
 658             /* COULOMB ELECTROSTATICS */
 659             velec            = _mm_mul_pd(qq10,rinv10);
 660             felec            = _mm_mul_pd(velec,rinvsq10);
 661
 662             /* Update potential sum for this i atom from the interaction with this j atom. */
 663             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 664             velecsum         = _mm_add_pd(velecsum,velec);
 665
 666             fscal            = felec;
 667
 668             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 669
 670             /* Update vectorial force */
 671             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 672             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 673             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 674
 675             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 676             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 677             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 678
 679             /**************************
 680              * CALCULATE INTERACTIONS *
 681              **************************/
 682
 683             /* COULOMB ELECTROSTATICS */
 684             velec            = _mm_mul_pd(qq11,rinv11);
 685             felec            = _mm_mul_pd(velec,rinvsq11);
 686
 687             /* Update potential sum for this i atom from the interaction with this j atom. */
 688             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 689             velecsum         = _mm_add_pd(velecsum,velec);
 690
 691             fscal            = felec;
 692
 693             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 694
 695             /* Update vectorial force */
 696             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 697             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 698             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 699
 700             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 701             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 702             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 703
 704             /**************************
 705              * CALCULATE INTERACTIONS *
 706              **************************/
 707
 708             /* COULOMB ELECTROSTATICS */
 709             velec            = _mm_mul_pd(qq12,rinv12);
 710             felec            = _mm_mul_pd(velec,rinvsq12);
 711
 712             /* Update potential sum for this i atom from the interaction with this j atom. */
 713             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 714             velecsum         = _mm_add_pd(velecsum,velec);
 715
 716             fscal            = felec;
 717
 718             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 719
 720             /* Update vectorial force */
 721             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 722             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 723             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 724
 725             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 726             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 727             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 728
 729             /**************************
 730              * CALCULATE INTERACTIONS *
 731              **************************/
 732
 733             /* COULOMB ELECTROSTATICS */
 734             velec            = _mm_mul_pd(qq20,rinv20);
 735             felec            = _mm_mul_pd(velec,rinvsq20);
 736
 737             /* Update potential sum for this i atom from the interaction with this j atom. */
 738             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 739             velecsum         = _mm_add_pd(velecsum,velec);
 740
 741             fscal            = felec;
 742
 743             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 744
 745             /* Update vectorial force */
 746             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 747             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 748             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 749
 750             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 751             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 752             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             /* COULOMB ELECTROSTATICS */
 759             velec            = _mm_mul_pd(qq21,rinv21);
 760             felec            = _mm_mul_pd(velec,rinvsq21);
 761
 762             /* Update potential sum for this i atom from the interaction with this j atom. */
 763             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 764             velecsum         = _mm_add_pd(velecsum,velec);
 765
 766             fscal            = felec;
 767
 768             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 769
 770             /* Update vectorial force */
 771             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 772             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 773             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 774
 775             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 776             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 777             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 778
 779             /**************************
 780              * CALCULATE INTERACTIONS *
 781              **************************/
 782
 783             /* COULOMB ELECTROSTATICS */
 784             velec            = _mm_mul_pd(qq22,rinv22);
 785             felec            = _mm_mul_pd(velec,rinvsq22);
 786
 787             /* Update potential sum for this i atom from the interaction with this j atom. */
 788             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 789             velecsum         = _mm_add_pd(velecsum,velec);
 790
 791             fscal            = felec;
 792
 793             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 794
 795             /* Update vectorial force */
 796             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 797             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 798             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 799
 800             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 801             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 802             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 803
 804             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 805
 806             /* Inner loop uses 291 flops */
 807         }
 808
 809         /* End of innermost loop */
 810
 811         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 812                                               f+i_coord_offset,fshift+i_shift_offset);
 813
 814         ggid                        = gid[iidx];
 815         /* Update potential energies */
 816         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 817         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 818
 819         /* Increment number of inner iterations */
 820         inneriter                  += j_index_end - j_index_start;
 821
 822         /* Outer loop uses 20 flops */
 823     }
 824
 825     /* Increment number of outer iterations */
 826     outeriter        += nri;
 827
 828     /* Update outer/inner flops */
 829
 830     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*291);
 831 }
 832 /*
 833  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_avx_128_fma_double
 834  * Electrostatics interaction: Coulomb
 835  * VdW interaction:            LennardJones
 836  * Geometry:                   Water3-Water3
 837  * Calculate force/pot:        Force
 838  */
 839 void
 840 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_avx_128_fma_double
 841                     (t_nblist                    * gmx_restrict       nlist,
 842                      rvec                        * gmx_restrict          xx,
 843                      rvec                        * gmx_restrict          ff,
 844                      t_forcerec                  * gmx_restrict          fr,
 845                      t_mdatoms                   * gmx_restrict     mdatoms,
 846                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 847                      t_nrnb                      * gmx_restrict        nrnb)
 848 {
 849     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 850      * just 0 for non-waters.
 851      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 852      * jnr indices corresponding to data put in the four positions in the SIMD register.
 853      */
 854     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 855     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 856     int              jnrA,jnrB;
 857     int              j_coord_offsetA,j_coord_offsetB;
 858     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 859     real             rcutoff_scalar;
 860     real             *shiftvec,*fshift,*x,*f;
 861     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 862     int              vdwioffset0;
 863     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 864     int              vdwioffset1;
 865     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 866     int              vdwioffset2;
 867     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 868     int              vdwjidx0A,vdwjidx0B;
 869     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 870     int              vdwjidx1A,vdwjidx1B;
 871     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 872     int              vdwjidx2A,vdwjidx2B;
 873     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 874     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 875     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 876     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 877     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 878     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 879     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 880     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 881     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 882     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 883     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 884     real             *charge;
 885     int              nvdwtype;
 886     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 887     int              *vdwtype;
 888     real             *vdwparam;
 889     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 890     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 891     __m128d          dummy_mask,cutoff_mask;
 892     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 893     __m128d          one     = _mm_set1_pd(1.0);
 894     __m128d          two     = _mm_set1_pd(2.0);
 895     x                = xx[0];
 896     f                = ff[0];
 897
 898     nri              = nlist->nri;
 899     iinr             = nlist->iinr;
 900     jindex           = nlist->jindex;
 901     jjnr             = nlist->jjnr;
 902     shiftidx         = nlist->shift;
 903     gid              = nlist->gid;
 904     shiftvec         = fr->shift_vec[0];
 905     fshift           = fr->fshift[0];
 906     facel            = _mm_set1_pd(fr->epsfac);
 907     charge           = mdatoms->chargeA;
 908     nvdwtype         = fr->ntype;
 909     vdwparam         = fr->nbfp;
 910     vdwtype          = mdatoms->typeA;
 911
 912     /* Setup water-specific parameters */
 913     inr              = nlist->iinr[0];
 914     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 915     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 916     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 917     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 918
 919     jq0              = _mm_set1_pd(charge[inr+0]);
 920     jq1              = _mm_set1_pd(charge[inr+1]);
 921     jq2              = _mm_set1_pd(charge[inr+2]);
 922     vdwjidx0A        = 2*vdwtype[inr+0];
 923     qq00             = _mm_mul_pd(iq0,jq0);
 924     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 925     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 926     qq01             = _mm_mul_pd(iq0,jq1);
 927     qq02             = _mm_mul_pd(iq0,jq2);
 928     qq10             = _mm_mul_pd(iq1,jq0);
 929     qq11             = _mm_mul_pd(iq1,jq1);
 930     qq12             = _mm_mul_pd(iq1,jq2);
 931     qq20             = _mm_mul_pd(iq2,jq0);
 932     qq21             = _mm_mul_pd(iq2,jq1);
 933     qq22             = _mm_mul_pd(iq2,jq2);
 934
 935     /* Avoid stupid compiler warnings */
 936     jnrA = jnrB = 0;
 937     j_coord_offsetA = 0;
 938     j_coord_offsetB = 0;
 939
 940     outeriter        = 0;
 941     inneriter        = 0;
 942
 943     /* Start outer loop over neighborlists */
 944     for(iidx=0; iidx<nri; iidx++)
 945     {
 946         /* Load shift vector for this list */
 947         i_shift_offset   = DIM*shiftidx[iidx];
 948
 949         /* Load limits for loop over neighbors */
 950         j_index_start    = jindex[iidx];
 951         j_index_end      = jindex[iidx+1];
 952
 953         /* Get outer coordinate index */
 954         inr              = iinr[iidx];
 955         i_coord_offset   = DIM*inr;
 956
 957         /* Load i particle coords and add shift vector */
 958         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 959                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 960
 961         fix0             = _mm_setzero_pd();
 962         fiy0             = _mm_setzero_pd();
 963         fiz0             = _mm_setzero_pd();
 964         fix1             = _mm_setzero_pd();
 965         fiy1             = _mm_setzero_pd();
 966         fiz1             = _mm_setzero_pd();
 967         fix2             = _mm_setzero_pd();
 968         fiy2             = _mm_setzero_pd();
 969         fiz2             = _mm_setzero_pd();
 970
 971         /* Start inner kernel loop */
 972         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 973         {
 974
 975             /* Get j neighbor index, and coordinate index */
 976             jnrA             = jjnr[jidx];
 977             jnrB             = jjnr[jidx+1];
 978             j_coord_offsetA  = DIM*jnrA;
 979             j_coord_offsetB  = DIM*jnrB;
 980
 981             /* load j atom coordinates */
 982             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 983                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 984
 985             /* Calculate displacement vector */
 986             dx00             = _mm_sub_pd(ix0,jx0);
 987             dy00             = _mm_sub_pd(iy0,jy0);
 988             dz00             = _mm_sub_pd(iz0,jz0);
 989             dx01             = _mm_sub_pd(ix0,jx1);
 990             dy01             = _mm_sub_pd(iy0,jy1);
 991             dz01             = _mm_sub_pd(iz0,jz1);
 992             dx02             = _mm_sub_pd(ix0,jx2);
 993             dy02             = _mm_sub_pd(iy0,jy2);
 994             dz02             = _mm_sub_pd(iz0,jz2);
 995             dx10             = _mm_sub_pd(ix1,jx0);
 996             dy10             = _mm_sub_pd(iy1,jy0);
 997             dz10             = _mm_sub_pd(iz1,jz0);
 998             dx11             = _mm_sub_pd(ix1,jx1);
 999             dy11             = _mm_sub_pd(iy1,jy1);
1000             dz11             = _mm_sub_pd(iz1,jz1);
1001             dx12             = _mm_sub_pd(ix1,jx2);
1002             dy12             = _mm_sub_pd(iy1,jy2);
1003             dz12             = _mm_sub_pd(iz1,jz2);
1004             dx20             = _mm_sub_pd(ix2,jx0);
1005             dy20             = _mm_sub_pd(iy2,jy0);
1006             dz20             = _mm_sub_pd(iz2,jz0);
1007             dx21             = _mm_sub_pd(ix2,jx1);
1008             dy21             = _mm_sub_pd(iy2,jy1);
1009             dz21             = _mm_sub_pd(iz2,jz1);
1010             dx22             = _mm_sub_pd(ix2,jx2);
1011             dy22             = _mm_sub_pd(iy2,jy2);
1012             dz22             = _mm_sub_pd(iz2,jz2);
1013
1014             /* Calculate squared distance and things based on it */
1015             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1016             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1017             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1018             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1019             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1020             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1021             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1022             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1023             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1024
1025             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1026             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1027             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1028             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1029             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1030             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1031             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1032             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1033             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1034
1035             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1036             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1037             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1038             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1039             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1040             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1041             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1042             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1043             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1044
1045             fjx0             = _mm_setzero_pd();
1046             fjy0             = _mm_setzero_pd();
1047             fjz0             = _mm_setzero_pd();
1048             fjx1             = _mm_setzero_pd();
1049             fjy1             = _mm_setzero_pd();
1050             fjz1             = _mm_setzero_pd();
1051             fjx2             = _mm_setzero_pd();
1052             fjy2             = _mm_setzero_pd();
1053             fjz2             = _mm_setzero_pd();
1054
1055             /**************************
1056              * CALCULATE INTERACTIONS *
1057              **************************/
1058
1059             /* COULOMB ELECTROSTATICS */
1060             velec            = _mm_mul_pd(qq00,rinv00);
1061             felec            = _mm_mul_pd(velec,rinvsq00);
1062
1063             /* LENNARD-JONES DISPERSION/REPULSION */
1064
1065             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1066             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1067
1068             fscal            = _mm_add_pd(felec,fvdw);
1069
1070             /* Update vectorial force */
1071             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1072             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1073             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1074
1075             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1076             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1077             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1078
1079             /**************************
1080              * CALCULATE INTERACTIONS *
1081              **************************/
1082
1083             /* COULOMB ELECTROSTATICS */
1084             velec            = _mm_mul_pd(qq01,rinv01);
1085             felec            = _mm_mul_pd(velec,rinvsq01);
1086
1087             fscal            = felec;
1088
1089             /* Update vectorial force */
1090             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1091             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1092             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1093
1094             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1095             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1096             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1097
1098             /**************************
1099              * CALCULATE INTERACTIONS *
1100              **************************/
1101
1102             /* COULOMB ELECTROSTATICS */
1103             velec            = _mm_mul_pd(qq02,rinv02);
1104             felec            = _mm_mul_pd(velec,rinvsq02);
1105
1106             fscal            = felec;
1107
1108             /* Update vectorial force */
1109             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1110             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1111             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1112
1113             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1114             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1115             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1116
1117             /**************************
1118              * CALCULATE INTERACTIONS *
1119              **************************/
1120
1121             /* COULOMB ELECTROSTATICS */
1122             velec            = _mm_mul_pd(qq10,rinv10);
1123             felec            = _mm_mul_pd(velec,rinvsq10);
1124
1125             fscal            = felec;
1126
1127             /* Update vectorial force */
1128             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1129             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1130             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1131
1132             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1133             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1134             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1135
1136             /**************************
1137              * CALCULATE INTERACTIONS *
1138              **************************/
1139
1140             /* COULOMB ELECTROSTATICS */
1141             velec            = _mm_mul_pd(qq11,rinv11);
1142             felec            = _mm_mul_pd(velec,rinvsq11);
1143
1144             fscal            = felec;
1145
1146             /* Update vectorial force */
1147             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1148             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1149             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1150
1151             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1152             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1153             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1154
1155             /**************************
1156              * CALCULATE INTERACTIONS *
1157              **************************/
1158
1159             /* COULOMB ELECTROSTATICS */
1160             velec            = _mm_mul_pd(qq12,rinv12);
1161             felec            = _mm_mul_pd(velec,rinvsq12);
1162
1163             fscal            = felec;
1164
1165             /* Update vectorial force */
1166             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1167             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1168             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1169
1170             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1171             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1172             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1173
1174             /**************************
1175              * CALCULATE INTERACTIONS *
1176              **************************/
1177
1178             /* COULOMB ELECTROSTATICS */
1179             velec            = _mm_mul_pd(qq20,rinv20);
1180             felec            = _mm_mul_pd(velec,rinvsq20);
1181
1182             fscal            = felec;
1183
1184             /* Update vectorial force */
1185             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1186             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1187             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1188
1189             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1190             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1191             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1192
1193             /**************************
1194              * CALCULATE INTERACTIONS *
1195              **************************/
1196
1197             /* COULOMB ELECTROSTATICS */
1198             velec            = _mm_mul_pd(qq21,rinv21);
1199             felec            = _mm_mul_pd(velec,rinvsq21);
1200
1201             fscal            = felec;
1202
1203             /* Update vectorial force */
1204             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1205             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1206             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1207
1208             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1209             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1210             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1211
1212             /**************************
1213              * CALCULATE INTERACTIONS *
1214              **************************/
1215
1216             /* COULOMB ELECTROSTATICS */
1217             velec            = _mm_mul_pd(qq22,rinv22);
1218             felec            = _mm_mul_pd(velec,rinvsq22);
1219
1220             fscal            = felec;
1221
1222             /* Update vectorial force */
1223             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1224             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1225             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1226
1227             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1228             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1229             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1230
1231             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1232
1233             /* Inner loop uses 277 flops */
1234         }
1235
1236         if(jidx<j_index_end)
1237         {
1238
1239             jnrA             = jjnr[jidx];
1240             j_coord_offsetA  = DIM*jnrA;
1241
1242             /* load j atom coordinates */
1243             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1244                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1245
1246             /* Calculate displacement vector */
1247             dx00             = _mm_sub_pd(ix0,jx0);
1248             dy00             = _mm_sub_pd(iy0,jy0);
1249             dz00             = _mm_sub_pd(iz0,jz0);
1250             dx01             = _mm_sub_pd(ix0,jx1);
1251             dy01             = _mm_sub_pd(iy0,jy1);
1252             dz01             = _mm_sub_pd(iz0,jz1);
1253             dx02             = _mm_sub_pd(ix0,jx2);
1254             dy02             = _mm_sub_pd(iy0,jy2);
1255             dz02             = _mm_sub_pd(iz0,jz2);
1256             dx10             = _mm_sub_pd(ix1,jx0);
1257             dy10             = _mm_sub_pd(iy1,jy0);
1258             dz10             = _mm_sub_pd(iz1,jz0);
1259             dx11             = _mm_sub_pd(ix1,jx1);
1260             dy11             = _mm_sub_pd(iy1,jy1);
1261             dz11             = _mm_sub_pd(iz1,jz1);
1262             dx12             = _mm_sub_pd(ix1,jx2);
1263             dy12             = _mm_sub_pd(iy1,jy2);
1264             dz12             = _mm_sub_pd(iz1,jz2);
1265             dx20             = _mm_sub_pd(ix2,jx0);
1266             dy20             = _mm_sub_pd(iy2,jy0);
1267             dz20             = _mm_sub_pd(iz2,jz0);
1268             dx21             = _mm_sub_pd(ix2,jx1);
1269             dy21             = _mm_sub_pd(iy2,jy1);
1270             dz21             = _mm_sub_pd(iz2,jz1);
1271             dx22             = _mm_sub_pd(ix2,jx2);
1272             dy22             = _mm_sub_pd(iy2,jy2);
1273             dz22             = _mm_sub_pd(iz2,jz2);
1274
1275             /* Calculate squared distance and things based on it */
1276             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1277             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1278             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1279             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1280             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1281             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1282             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1283             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1284             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1285
1286             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1287             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1288             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1289             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1290             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1291             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1292             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1293             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1294             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1295
1296             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1297             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1298             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1299             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1300             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1301             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1302             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1303             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1304             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1305
1306             fjx0             = _mm_setzero_pd();
1307             fjy0             = _mm_setzero_pd();
1308             fjz0             = _mm_setzero_pd();
1309             fjx1             = _mm_setzero_pd();
1310             fjy1             = _mm_setzero_pd();
1311             fjz1             = _mm_setzero_pd();
1312             fjx2             = _mm_setzero_pd();
1313             fjy2             = _mm_setzero_pd();
1314             fjz2             = _mm_setzero_pd();
1315
1316             /**************************
1317              * CALCULATE INTERACTIONS *
1318              **************************/
1319
1320             /* COULOMB ELECTROSTATICS */
1321             velec            = _mm_mul_pd(qq00,rinv00);
1322             felec            = _mm_mul_pd(velec,rinvsq00);
1323
1324             /* LENNARD-JONES DISPERSION/REPULSION */
1325
1326             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1327             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1328
1329             fscal            = _mm_add_pd(felec,fvdw);
1330
1331             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1332
1333             /* Update vectorial force */
1334             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1335             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1336             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1337
1338             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1339             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1340             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1341
1342             /**************************
1343              * CALCULATE INTERACTIONS *
1344              **************************/
1345
1346             /* COULOMB ELECTROSTATICS */
1347             velec            = _mm_mul_pd(qq01,rinv01);
1348             felec            = _mm_mul_pd(velec,rinvsq01);
1349
1350             fscal            = felec;
1351
1352             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1353
1354             /* Update vectorial force */
1355             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1356             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1357             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1358
1359             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1360             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1361             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1362
1363             /**************************
1364              * CALCULATE INTERACTIONS *
1365              **************************/
1366
1367             /* COULOMB ELECTROSTATICS */
1368             velec            = _mm_mul_pd(qq02,rinv02);
1369             felec            = _mm_mul_pd(velec,rinvsq02);
1370
1371             fscal            = felec;
1372
1373             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1374
1375             /* Update vectorial force */
1376             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1377             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1378             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1379
1380             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1381             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1382             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             /* COULOMB ELECTROSTATICS */
1389             velec            = _mm_mul_pd(qq10,rinv10);
1390             felec            = _mm_mul_pd(velec,rinvsq10);
1391
1392             fscal            = felec;
1393
1394             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1395
1396             /* Update vectorial force */
1397             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1398             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1399             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1400
1401             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1402             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1403             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1404
1405             /**************************
1406              * CALCULATE INTERACTIONS *
1407              **************************/
1408
1409             /* COULOMB ELECTROSTATICS */
1410             velec            = _mm_mul_pd(qq11,rinv11);
1411             felec            = _mm_mul_pd(velec,rinvsq11);
1412
1413             fscal            = felec;
1414
1415             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1416
1417             /* Update vectorial force */
1418             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1419             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1420             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1421
1422             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1423             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1424             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1425
1426             /**************************
1427              * CALCULATE INTERACTIONS *
1428              **************************/
1429
1430             /* COULOMB ELECTROSTATICS */
1431             velec            = _mm_mul_pd(qq12,rinv12);
1432             felec            = _mm_mul_pd(velec,rinvsq12);
1433
1434             fscal            = felec;
1435
1436             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1437
1438             /* Update vectorial force */
1439             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1440             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1441             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1442
1443             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1444             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1445             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1446
1447             /**************************
1448              * CALCULATE INTERACTIONS *
1449              **************************/
1450
1451             /* COULOMB ELECTROSTATICS */
1452             velec            = _mm_mul_pd(qq20,rinv20);
1453             felec            = _mm_mul_pd(velec,rinvsq20);
1454
1455             fscal            = felec;
1456
1457             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1458
1459             /* Update vectorial force */
1460             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1461             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1462             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1463
1464             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1465             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1466             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1467
1468             /**************************
1469              * CALCULATE INTERACTIONS *
1470              **************************/
1471
1472             /* COULOMB ELECTROSTATICS */
1473             velec            = _mm_mul_pd(qq21,rinv21);
1474             felec            = _mm_mul_pd(velec,rinvsq21);
1475
1476             fscal            = felec;
1477
1478             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1479
1480             /* Update vectorial force */
1481             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1482             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1483             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1484
1485             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1486             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1487             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1488
1489             /**************************
1490              * CALCULATE INTERACTIONS *
1491              **************************/
1492
1493             /* COULOMB ELECTROSTATICS */
1494             velec            = _mm_mul_pd(qq22,rinv22);
1495             felec            = _mm_mul_pd(velec,rinvsq22);
1496
1497             fscal            = felec;
1498
1499             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1500
1501             /* Update vectorial force */
1502             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1503             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1504             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1505
1506             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1507             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1508             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1509
1510             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1511
1512             /* Inner loop uses 277 flops */
1513         }
1514
1515         /* End of innermost loop */
1516
1517         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1518                                               f+i_coord_offset,fshift+i_shift_offset);
1519
1520         /* Increment number of inner iterations */
1521         inneriter                  += j_index_end - j_index_start;
1522
1523         /* Outer loop uses 18 flops */
1524     }
1525
1526     /* Increment number of outer iterations */
1527     outeriter        += nri;
1528
1529     /* Update outer/inner flops */
1530
1531     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1532 }