src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_avx_128_fma_double
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_avx_128_fma_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwioffset3;
  86     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  87     int              vdwjidx0A,vdwjidx0B;
  88     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     int              vdwjidx3A,vdwjidx3B;
  94     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  95     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  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          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  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          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 102     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 103     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 104     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 105     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 106     real             *charge;
 107     int              nvdwtype;
 108     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 109     int              *vdwtype;
 110     real             *vdwparam;
 111     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 112     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 113     __m128d          dummy_mask,cutoff_mask;
 114     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 115     __m128d          one     = _mm_set1_pd(1.0);
 116     __m128d          two     = _mm_set1_pd(2.0);
 117     x                = xx[0];
 118     f                = ff[0];
 119
 120     nri              = nlist->nri;
 121     iinr             = nlist->iinr;
 122     jindex           = nlist->jindex;
 123     jjnr             = nlist->jjnr;
 124     shiftidx         = nlist->shift;
 125     gid              = nlist->gid;
 126     shiftvec         = fr->shift_vec[0];
 127     fshift           = fr->fshift[0];
 128     facel            = _mm_set1_pd(fr->ic->epsfac);
 129     charge           = mdatoms->chargeA;
 130     nvdwtype         = fr->ntype;
 131     vdwparam         = fr->nbfp;
 132     vdwtype          = mdatoms->typeA;
 133
 134     /* Setup water-specific parameters */
 135     inr              = nlist->iinr[0];
 136     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 137     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 138     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 139     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 140
 141     jq1              = _mm_set1_pd(charge[inr+1]);
 142     jq2              = _mm_set1_pd(charge[inr+2]);
 143     jq3              = _mm_set1_pd(charge[inr+3]);
 144     vdwjidx0A        = 2*vdwtype[inr+0];
 145     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 146     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 147     qq11             = _mm_mul_pd(iq1,jq1);
 148     qq12             = _mm_mul_pd(iq1,jq2);
 149     qq13             = _mm_mul_pd(iq1,jq3);
 150     qq21             = _mm_mul_pd(iq2,jq1);
 151     qq22             = _mm_mul_pd(iq2,jq2);
 152     qq23             = _mm_mul_pd(iq2,jq3);
 153     qq31             = _mm_mul_pd(iq3,jq1);
 154     qq32             = _mm_mul_pd(iq3,jq2);
 155     qq33             = _mm_mul_pd(iq3,jq3);
 156
 157     /* Avoid stupid compiler warnings */
 158     jnrA = jnrB = 0;
 159     j_coord_offsetA = 0;
 160     j_coord_offsetB = 0;
 161
 162     outeriter        = 0;
 163     inneriter        = 0;
 164
 165     /* Start outer loop over neighborlists */
 166     for(iidx=0; iidx<nri; iidx++)
 167     {
 168         /* Load shift vector for this list */
 169         i_shift_offset   = DIM*shiftidx[iidx];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 181                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 182
 183         fix0             = _mm_setzero_pd();
 184         fiy0             = _mm_setzero_pd();
 185         fiz0             = _mm_setzero_pd();
 186         fix1             = _mm_setzero_pd();
 187         fiy1             = _mm_setzero_pd();
 188         fiz1             = _mm_setzero_pd();
 189         fix2             = _mm_setzero_pd();
 190         fiy2             = _mm_setzero_pd();
 191         fiz2             = _mm_setzero_pd();
 192         fix3             = _mm_setzero_pd();
 193         fiy3             = _mm_setzero_pd();
 194         fiz3             = _mm_setzero_pd();
 195
 196         /* Reset potential sums */
 197         velecsum         = _mm_setzero_pd();
 198         vvdwsum          = _mm_setzero_pd();
 199
 200         /* Start inner kernel loop */
 201         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 202         {
 203
 204             /* Get j neighbor index, and coordinate index */
 205             jnrA             = jjnr[jidx];
 206             jnrB             = jjnr[jidx+1];
 207             j_coord_offsetA  = DIM*jnrA;
 208             j_coord_offsetB  = DIM*jnrB;
 209
 210             /* load j atom coordinates */
 211             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 212                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 213                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 214
 215             /* Calculate displacement vector */
 216             dx00             = _mm_sub_pd(ix0,jx0);
 217             dy00             = _mm_sub_pd(iy0,jy0);
 218             dz00             = _mm_sub_pd(iz0,jz0);
 219             dx11             = _mm_sub_pd(ix1,jx1);
 220             dy11             = _mm_sub_pd(iy1,jy1);
 221             dz11             = _mm_sub_pd(iz1,jz1);
 222             dx12             = _mm_sub_pd(ix1,jx2);
 223             dy12             = _mm_sub_pd(iy1,jy2);
 224             dz12             = _mm_sub_pd(iz1,jz2);
 225             dx13             = _mm_sub_pd(ix1,jx3);
 226             dy13             = _mm_sub_pd(iy1,jy3);
 227             dz13             = _mm_sub_pd(iz1,jz3);
 228             dx21             = _mm_sub_pd(ix2,jx1);
 229             dy21             = _mm_sub_pd(iy2,jy1);
 230             dz21             = _mm_sub_pd(iz2,jz1);
 231             dx22             = _mm_sub_pd(ix2,jx2);
 232             dy22             = _mm_sub_pd(iy2,jy2);
 233             dz22             = _mm_sub_pd(iz2,jz2);
 234             dx23             = _mm_sub_pd(ix2,jx3);
 235             dy23             = _mm_sub_pd(iy2,jy3);
 236             dz23             = _mm_sub_pd(iz2,jz3);
 237             dx31             = _mm_sub_pd(ix3,jx1);
 238             dy31             = _mm_sub_pd(iy3,jy1);
 239             dz31             = _mm_sub_pd(iz3,jz1);
 240             dx32             = _mm_sub_pd(ix3,jx2);
 241             dy32             = _mm_sub_pd(iy3,jy2);
 242             dz32             = _mm_sub_pd(iz3,jz2);
 243             dx33             = _mm_sub_pd(ix3,jx3);
 244             dy33             = _mm_sub_pd(iy3,jy3);
 245             dz33             = _mm_sub_pd(iz3,jz3);
 246
 247             /* Calculate squared distance and things based on it */
 248             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 249             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 250             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 251             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 252             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 253             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 254             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 255             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 256             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 257             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 258
 259             rinv11           = avx128fma_invsqrt_d(rsq11);
 260             rinv12           = avx128fma_invsqrt_d(rsq12);
 261             rinv13           = avx128fma_invsqrt_d(rsq13);
 262             rinv21           = avx128fma_invsqrt_d(rsq21);
 263             rinv22           = avx128fma_invsqrt_d(rsq22);
 264             rinv23           = avx128fma_invsqrt_d(rsq23);
 265             rinv31           = avx128fma_invsqrt_d(rsq31);
 266             rinv32           = avx128fma_invsqrt_d(rsq32);
 267             rinv33           = avx128fma_invsqrt_d(rsq33);
 268
 269             rinvsq00         = avx128fma_inv_d(rsq00);
 270             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 271             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 272             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 273             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 274             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 275             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 276             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 277             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 278             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 279
 280             fjx0             = _mm_setzero_pd();
 281             fjy0             = _mm_setzero_pd();
 282             fjz0             = _mm_setzero_pd();
 283             fjx1             = _mm_setzero_pd();
 284             fjy1             = _mm_setzero_pd();
 285             fjz1             = _mm_setzero_pd();
 286             fjx2             = _mm_setzero_pd();
 287             fjy2             = _mm_setzero_pd();
 288             fjz2             = _mm_setzero_pd();
 289             fjx3             = _mm_setzero_pd();
 290             fjy3             = _mm_setzero_pd();
 291             fjz3             = _mm_setzero_pd();
 292
 293             /**************************
 294              * CALCULATE INTERACTIONS *
 295              **************************/
 296
 297             /* LENNARD-JONES DISPERSION/REPULSION */
 298
 299             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 300             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 301             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 302             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 303             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 307
 308             fscal            = fvdw;
 309
 310             /* Update vectorial force */
 311             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 312             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 313             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 314
 315             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 316             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 317             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             /* COULOMB ELECTROSTATICS */
 324             velec            = _mm_mul_pd(qq11,rinv11);
 325             felec            = _mm_mul_pd(velec,rinvsq11);
 326
 327             /* Update potential sum for this i atom from the interaction with this j atom. */
 328             velecsum         = _mm_add_pd(velecsum,velec);
 329
 330             fscal            = felec;
 331
 332             /* Update vectorial force */
 333             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 334             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 335             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 336
 337             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 338             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 339             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 340
 341             /**************************
 342              * CALCULATE INTERACTIONS *
 343              **************************/
 344
 345             /* COULOMB ELECTROSTATICS */
 346             velec            = _mm_mul_pd(qq12,rinv12);
 347             felec            = _mm_mul_pd(velec,rinvsq12);
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             velecsum         = _mm_add_pd(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             /* Update vectorial force */
 355             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 356             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 357             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 358
 359             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 360             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 361             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             /* COULOMB ELECTROSTATICS */
 368             velec            = _mm_mul_pd(qq13,rinv13);
 369             felec            = _mm_mul_pd(velec,rinvsq13);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velecsum         = _mm_add_pd(velecsum,velec);
 373
 374             fscal            = felec;
 375
 376             /* Update vectorial force */
 377             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 378             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 379             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 380
 381             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 382             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 383             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 384
 385             /**************************
 386              * CALCULATE INTERACTIONS *
 387              **************************/
 388
 389             /* COULOMB ELECTROSTATICS */
 390             velec            = _mm_mul_pd(qq21,rinv21);
 391             felec            = _mm_mul_pd(velec,rinvsq21);
 392
 393             /* Update potential sum for this i atom from the interaction with this j atom. */
 394             velecsum         = _mm_add_pd(velecsum,velec);
 395
 396             fscal            = felec;
 397
 398             /* Update vectorial force */
 399             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 400             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 401             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 402
 403             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 404             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 405             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             /* COULOMB ELECTROSTATICS */
 412             velec            = _mm_mul_pd(qq22,rinv22);
 413             felec            = _mm_mul_pd(velec,rinvsq22);
 414
 415             /* Update potential sum for this i atom from the interaction with this j atom. */
 416             velecsum         = _mm_add_pd(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             /* Update vectorial force */
 421             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 422             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 423             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 424
 425             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 426             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 427             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 428
 429             /**************************
 430              * CALCULATE INTERACTIONS *
 431              **************************/
 432
 433             /* COULOMB ELECTROSTATICS */
 434             velec            = _mm_mul_pd(qq23,rinv23);
 435             felec            = _mm_mul_pd(velec,rinvsq23);
 436
 437             /* Update potential sum for this i atom from the interaction with this j atom. */
 438             velecsum         = _mm_add_pd(velecsum,velec);
 439
 440             fscal            = felec;
 441
 442             /* Update vectorial force */
 443             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 444             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 445             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 446
 447             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 448             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 449             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 450
 451             /**************************
 452              * CALCULATE INTERACTIONS *
 453              **************************/
 454
 455             /* COULOMB ELECTROSTATICS */
 456             velec            = _mm_mul_pd(qq31,rinv31);
 457             felec            = _mm_mul_pd(velec,rinvsq31);
 458
 459             /* Update potential sum for this i atom from the interaction with this j atom. */
 460             velecsum         = _mm_add_pd(velecsum,velec);
 461
 462             fscal            = felec;
 463
 464             /* Update vectorial force */
 465             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 466             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 467             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 468
 469             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 470             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 471             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 472
 473             /**************************
 474              * CALCULATE INTERACTIONS *
 475              **************************/
 476
 477             /* COULOMB ELECTROSTATICS */
 478             velec            = _mm_mul_pd(qq32,rinv32);
 479             felec            = _mm_mul_pd(velec,rinvsq32);
 480
 481             /* Update potential sum for this i atom from the interaction with this j atom. */
 482             velecsum         = _mm_add_pd(velecsum,velec);
 483
 484             fscal            = felec;
 485
 486             /* Update vectorial force */
 487             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 488             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 489             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 490
 491             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 492             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 493             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 494
 495             /**************************
 496              * CALCULATE INTERACTIONS *
 497              **************************/
 498
 499             /* COULOMB ELECTROSTATICS */
 500             velec            = _mm_mul_pd(qq33,rinv33);
 501             felec            = _mm_mul_pd(velec,rinvsq33);
 502
 503             /* Update potential sum for this i atom from the interaction with this j atom. */
 504             velecsum         = _mm_add_pd(velecsum,velec);
 505
 506             fscal            = felec;
 507
 508             /* Update vectorial force */
 509             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 510             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 511             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 512
 513             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 514             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 515             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 516
 517             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 518
 519             /* Inner loop uses 317 flops */
 520         }
 521
 522         if(jidx<j_index_end)
 523         {
 524
 525             jnrA             = jjnr[jidx];
 526             j_coord_offsetA  = DIM*jnrA;
 527
 528             /* load j atom coordinates */
 529             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 530                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 531                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 532
 533             /* Calculate displacement vector */
 534             dx00             = _mm_sub_pd(ix0,jx0);
 535             dy00             = _mm_sub_pd(iy0,jy0);
 536             dz00             = _mm_sub_pd(iz0,jz0);
 537             dx11             = _mm_sub_pd(ix1,jx1);
 538             dy11             = _mm_sub_pd(iy1,jy1);
 539             dz11             = _mm_sub_pd(iz1,jz1);
 540             dx12             = _mm_sub_pd(ix1,jx2);
 541             dy12             = _mm_sub_pd(iy1,jy2);
 542             dz12             = _mm_sub_pd(iz1,jz2);
 543             dx13             = _mm_sub_pd(ix1,jx3);
 544             dy13             = _mm_sub_pd(iy1,jy3);
 545             dz13             = _mm_sub_pd(iz1,jz3);
 546             dx21             = _mm_sub_pd(ix2,jx1);
 547             dy21             = _mm_sub_pd(iy2,jy1);
 548             dz21             = _mm_sub_pd(iz2,jz1);
 549             dx22             = _mm_sub_pd(ix2,jx2);
 550             dy22             = _mm_sub_pd(iy2,jy2);
 551             dz22             = _mm_sub_pd(iz2,jz2);
 552             dx23             = _mm_sub_pd(ix2,jx3);
 553             dy23             = _mm_sub_pd(iy2,jy3);
 554             dz23             = _mm_sub_pd(iz2,jz3);
 555             dx31             = _mm_sub_pd(ix3,jx1);
 556             dy31             = _mm_sub_pd(iy3,jy1);
 557             dz31             = _mm_sub_pd(iz3,jz1);
 558             dx32             = _mm_sub_pd(ix3,jx2);
 559             dy32             = _mm_sub_pd(iy3,jy2);
 560             dz32             = _mm_sub_pd(iz3,jz2);
 561             dx33             = _mm_sub_pd(ix3,jx3);
 562             dy33             = _mm_sub_pd(iy3,jy3);
 563             dz33             = _mm_sub_pd(iz3,jz3);
 564
 565             /* Calculate squared distance and things based on it */
 566             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 567             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 568             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 569             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 570             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 571             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 572             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 573             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 574             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 575             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 576
 577             rinv11           = avx128fma_invsqrt_d(rsq11);
 578             rinv12           = avx128fma_invsqrt_d(rsq12);
 579             rinv13           = avx128fma_invsqrt_d(rsq13);
 580             rinv21           = avx128fma_invsqrt_d(rsq21);
 581             rinv22           = avx128fma_invsqrt_d(rsq22);
 582             rinv23           = avx128fma_invsqrt_d(rsq23);
 583             rinv31           = avx128fma_invsqrt_d(rsq31);
 584             rinv32           = avx128fma_invsqrt_d(rsq32);
 585             rinv33           = avx128fma_invsqrt_d(rsq33);
 586
 587             rinvsq00         = avx128fma_inv_d(rsq00);
 588             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 589             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 590             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 591             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 592             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 593             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 594             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 595             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 596             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 597
 598             fjx0             = _mm_setzero_pd();
 599             fjy0             = _mm_setzero_pd();
 600             fjz0             = _mm_setzero_pd();
 601             fjx1             = _mm_setzero_pd();
 602             fjy1             = _mm_setzero_pd();
 603             fjz1             = _mm_setzero_pd();
 604             fjx2             = _mm_setzero_pd();
 605             fjy2             = _mm_setzero_pd();
 606             fjz2             = _mm_setzero_pd();
 607             fjx3             = _mm_setzero_pd();
 608             fjy3             = _mm_setzero_pd();
 609             fjz3             = _mm_setzero_pd();
 610
 611             /**************************
 612              * CALCULATE INTERACTIONS *
 613              **************************/
 614
 615             /* LENNARD-JONES DISPERSION/REPULSION */
 616
 617             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 618             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 619             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 620             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 621             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 622
 623             /* Update potential sum for this i atom from the interaction with this j atom. */
 624             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 625             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 626
 627             fscal            = fvdw;
 628
 629             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 630
 631             /* Update vectorial force */
 632             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 633             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 634             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 635
 636             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 637             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 638             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 639
 640             /**************************
 641              * CALCULATE INTERACTIONS *
 642              **************************/
 643
 644             /* COULOMB ELECTROSTATICS */
 645             velec            = _mm_mul_pd(qq11,rinv11);
 646             felec            = _mm_mul_pd(velec,rinvsq11);
 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(dx11,fscal,fix1);
 658             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 659             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 660
 661             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 662             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 663             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 664
 665             /**************************
 666              * CALCULATE INTERACTIONS *
 667              **************************/
 668
 669             /* COULOMB ELECTROSTATICS */
 670             velec            = _mm_mul_pd(qq12,rinv12);
 671             felec            = _mm_mul_pd(velec,rinvsq12);
 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(dx12,fscal,fix1);
 683             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 684             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 685
 686             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 687             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 688             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 689
 690             /**************************
 691              * CALCULATE INTERACTIONS *
 692              **************************/
 693
 694             /* COULOMB ELECTROSTATICS */
 695             velec            = _mm_mul_pd(qq13,rinv13);
 696             felec            = _mm_mul_pd(velec,rinvsq13);
 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(dx13,fscal,fix1);
 708             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 709             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 710
 711             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 712             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 713             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 714
 715             /**************************
 716              * CALCULATE INTERACTIONS *
 717              **************************/
 718
 719             /* COULOMB ELECTROSTATICS */
 720             velec            = _mm_mul_pd(qq21,rinv21);
 721             felec            = _mm_mul_pd(velec,rinvsq21);
 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(dx21,fscal,fix2);
 733             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 734             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 735
 736             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 737             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 738             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 739
 740             /**************************
 741              * CALCULATE INTERACTIONS *
 742              **************************/
 743
 744             /* COULOMB ELECTROSTATICS */
 745             velec            = _mm_mul_pd(qq22,rinv22);
 746             felec            = _mm_mul_pd(velec,rinvsq22);
 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(dx22,fscal,fix2);
 758             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 759             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 760
 761             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 762             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 763             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             /* COULOMB ELECTROSTATICS */
 770             velec            = _mm_mul_pd(qq23,rinv23);
 771             felec            = _mm_mul_pd(velec,rinvsq23);
 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(dx23,fscal,fix2);
 783             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 784             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 785
 786             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 787             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 788             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 789
 790             /**************************
 791              * CALCULATE INTERACTIONS *
 792              **************************/
 793
 794             /* COULOMB ELECTROSTATICS */
 795             velec            = _mm_mul_pd(qq31,rinv31);
 796             felec            = _mm_mul_pd(velec,rinvsq31);
 797
 798             /* Update potential sum for this i atom from the interaction with this j atom. */
 799             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 800             velecsum         = _mm_add_pd(velecsum,velec);
 801
 802             fscal            = felec;
 803
 804             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 805
 806             /* Update vectorial force */
 807             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 808             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 809             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 810
 811             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 812             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 813             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 814
 815             /**************************
 816              * CALCULATE INTERACTIONS *
 817              **************************/
 818
 819             /* COULOMB ELECTROSTATICS */
 820             velec            = _mm_mul_pd(qq32,rinv32);
 821             felec            = _mm_mul_pd(velec,rinvsq32);
 822
 823             /* Update potential sum for this i atom from the interaction with this j atom. */
 824             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 825             velecsum         = _mm_add_pd(velecsum,velec);
 826
 827             fscal            = felec;
 828
 829             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 830
 831             /* Update vectorial force */
 832             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 833             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 834             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 835
 836             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 837             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 838             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 839
 840             /**************************
 841              * CALCULATE INTERACTIONS *
 842              **************************/
 843
 844             /* COULOMB ELECTROSTATICS */
 845             velec            = _mm_mul_pd(qq33,rinv33);
 846             felec            = _mm_mul_pd(velec,rinvsq33);
 847
 848             /* Update potential sum for this i atom from the interaction with this j atom. */
 849             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 850             velecsum         = _mm_add_pd(velecsum,velec);
 851
 852             fscal            = felec;
 853
 854             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 855
 856             /* Update vectorial force */
 857             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 858             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 859             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 860
 861             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 862             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 863             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 864
 865             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 866
 867             /* Inner loop uses 317 flops */
 868         }
 869
 870         /* End of innermost loop */
 871
 872         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 873                                               f+i_coord_offset,fshift+i_shift_offset);
 874
 875         ggid                        = gid[iidx];
 876         /* Update potential energies */
 877         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 878         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 879
 880         /* Increment number of inner iterations */
 881         inneriter                  += j_index_end - j_index_start;
 882
 883         /* Outer loop uses 26 flops */
 884     }
 885
 886     /* Increment number of outer iterations */
 887     outeriter        += nri;
 888
 889     /* Update outer/inner flops */
 890
 891     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*317);
 892 }
 893 /*
 894  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_double
 895  * Electrostatics interaction: Coulomb
 896  * VdW interaction:            LennardJones
 897  * Geometry:                   Water4-Water4
 898  * Calculate force/pot:        Force
 899  */
 900 void
 901 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_double
 902                     (t_nblist                    * gmx_restrict       nlist,
 903                      rvec                        * gmx_restrict          xx,
 904                      rvec                        * gmx_restrict          ff,
 905                      struct t_forcerec           * gmx_restrict          fr,
 906                      t_mdatoms                   * gmx_restrict     mdatoms,
 907                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 908                      t_nrnb                      * gmx_restrict        nrnb)
 909 {
 910     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 911      * just 0 for non-waters.
 912      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 913      * jnr indices corresponding to data put in the four positions in the SIMD register.
 914      */
 915     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 916     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 917     int              jnrA,jnrB;
 918     int              j_coord_offsetA,j_coord_offsetB;
 919     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 920     real             rcutoff_scalar;
 921     real             *shiftvec,*fshift,*x,*f;
 922     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 923     int              vdwioffset0;
 924     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 925     int              vdwioffset1;
 926     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 927     int              vdwioffset2;
 928     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 929     int              vdwioffset3;
 930     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 931     int              vdwjidx0A,vdwjidx0B;
 932     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 933     int              vdwjidx1A,vdwjidx1B;
 934     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 935     int              vdwjidx2A,vdwjidx2B;
 936     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 937     int              vdwjidx3A,vdwjidx3B;
 938     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 939     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 940     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 941     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 942     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 943     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 944     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 945     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 946     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 947     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 948     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 949     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 950     real             *charge;
 951     int              nvdwtype;
 952     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 953     int              *vdwtype;
 954     real             *vdwparam;
 955     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 956     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 957     __m128d          dummy_mask,cutoff_mask;
 958     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 959     __m128d          one     = _mm_set1_pd(1.0);
 960     __m128d          two     = _mm_set1_pd(2.0);
 961     x                = xx[0];
 962     f                = ff[0];
 963
 964     nri              = nlist->nri;
 965     iinr             = nlist->iinr;
 966     jindex           = nlist->jindex;
 967     jjnr             = nlist->jjnr;
 968     shiftidx         = nlist->shift;
 969     gid              = nlist->gid;
 970     shiftvec         = fr->shift_vec[0];
 971     fshift           = fr->fshift[0];
 972     facel            = _mm_set1_pd(fr->ic->epsfac);
 973     charge           = mdatoms->chargeA;
 974     nvdwtype         = fr->ntype;
 975     vdwparam         = fr->nbfp;
 976     vdwtype          = mdatoms->typeA;
 977
 978     /* Setup water-specific parameters */
 979     inr              = nlist->iinr[0];
 980     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 981     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 982     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 983     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 984
 985     jq1              = _mm_set1_pd(charge[inr+1]);
 986     jq2              = _mm_set1_pd(charge[inr+2]);
 987     jq3              = _mm_set1_pd(charge[inr+3]);
 988     vdwjidx0A        = 2*vdwtype[inr+0];
 989     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 990     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 991     qq11             = _mm_mul_pd(iq1,jq1);
 992     qq12             = _mm_mul_pd(iq1,jq2);
 993     qq13             = _mm_mul_pd(iq1,jq3);
 994     qq21             = _mm_mul_pd(iq2,jq1);
 995     qq22             = _mm_mul_pd(iq2,jq2);
 996     qq23             = _mm_mul_pd(iq2,jq3);
 997     qq31             = _mm_mul_pd(iq3,jq1);
 998     qq32             = _mm_mul_pd(iq3,jq2);
 999     qq33             = _mm_mul_pd(iq3,jq3);
1000
1001     /* Avoid stupid compiler warnings */
1002     jnrA = jnrB = 0;
1003     j_coord_offsetA = 0;
1004     j_coord_offsetB = 0;
1005
1006     outeriter        = 0;
1007     inneriter        = 0;
1008
1009     /* Start outer loop over neighborlists */
1010     for(iidx=0; iidx<nri; iidx++)
1011     {
1012         /* Load shift vector for this list */
1013         i_shift_offset   = DIM*shiftidx[iidx];
1014
1015         /* Load limits for loop over neighbors */
1016         j_index_start    = jindex[iidx];
1017         j_index_end      = jindex[iidx+1];
1018
1019         /* Get outer coordinate index */
1020         inr              = iinr[iidx];
1021         i_coord_offset   = DIM*inr;
1022
1023         /* Load i particle coords and add shift vector */
1024         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1025                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1026
1027         fix0             = _mm_setzero_pd();
1028         fiy0             = _mm_setzero_pd();
1029         fiz0             = _mm_setzero_pd();
1030         fix1             = _mm_setzero_pd();
1031         fiy1             = _mm_setzero_pd();
1032         fiz1             = _mm_setzero_pd();
1033         fix2             = _mm_setzero_pd();
1034         fiy2             = _mm_setzero_pd();
1035         fiz2             = _mm_setzero_pd();
1036         fix3             = _mm_setzero_pd();
1037         fiy3             = _mm_setzero_pd();
1038         fiz3             = _mm_setzero_pd();
1039
1040         /* Start inner kernel loop */
1041         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1042         {
1043
1044             /* Get j neighbor index, and coordinate index */
1045             jnrA             = jjnr[jidx];
1046             jnrB             = jjnr[jidx+1];
1047             j_coord_offsetA  = DIM*jnrA;
1048             j_coord_offsetB  = DIM*jnrB;
1049
1050             /* load j atom coordinates */
1051             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1052                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1053                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1054
1055             /* Calculate displacement vector */
1056             dx00             = _mm_sub_pd(ix0,jx0);
1057             dy00             = _mm_sub_pd(iy0,jy0);
1058             dz00             = _mm_sub_pd(iz0,jz0);
1059             dx11             = _mm_sub_pd(ix1,jx1);
1060             dy11             = _mm_sub_pd(iy1,jy1);
1061             dz11             = _mm_sub_pd(iz1,jz1);
1062             dx12             = _mm_sub_pd(ix1,jx2);
1063             dy12             = _mm_sub_pd(iy1,jy2);
1064             dz12             = _mm_sub_pd(iz1,jz2);
1065             dx13             = _mm_sub_pd(ix1,jx3);
1066             dy13             = _mm_sub_pd(iy1,jy3);
1067             dz13             = _mm_sub_pd(iz1,jz3);
1068             dx21             = _mm_sub_pd(ix2,jx1);
1069             dy21             = _mm_sub_pd(iy2,jy1);
1070             dz21             = _mm_sub_pd(iz2,jz1);
1071             dx22             = _mm_sub_pd(ix2,jx2);
1072             dy22             = _mm_sub_pd(iy2,jy2);
1073             dz22             = _mm_sub_pd(iz2,jz2);
1074             dx23             = _mm_sub_pd(ix2,jx3);
1075             dy23             = _mm_sub_pd(iy2,jy3);
1076             dz23             = _mm_sub_pd(iz2,jz3);
1077             dx31             = _mm_sub_pd(ix3,jx1);
1078             dy31             = _mm_sub_pd(iy3,jy1);
1079             dz31             = _mm_sub_pd(iz3,jz1);
1080             dx32             = _mm_sub_pd(ix3,jx2);
1081             dy32             = _mm_sub_pd(iy3,jy2);
1082             dz32             = _mm_sub_pd(iz3,jz2);
1083             dx33             = _mm_sub_pd(ix3,jx3);
1084             dy33             = _mm_sub_pd(iy3,jy3);
1085             dz33             = _mm_sub_pd(iz3,jz3);
1086
1087             /* Calculate squared distance and things based on it */
1088             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1089             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1090             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1091             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1092             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1093             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1094             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1095             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1096             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1097             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1098
1099             rinv11           = avx128fma_invsqrt_d(rsq11);
1100             rinv12           = avx128fma_invsqrt_d(rsq12);
1101             rinv13           = avx128fma_invsqrt_d(rsq13);
1102             rinv21           = avx128fma_invsqrt_d(rsq21);
1103             rinv22           = avx128fma_invsqrt_d(rsq22);
1104             rinv23           = avx128fma_invsqrt_d(rsq23);
1105             rinv31           = avx128fma_invsqrt_d(rsq31);
1106             rinv32           = avx128fma_invsqrt_d(rsq32);
1107             rinv33           = avx128fma_invsqrt_d(rsq33);
1108
1109             rinvsq00         = avx128fma_inv_d(rsq00);
1110             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1111             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1112             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1113             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1114             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1115             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1116             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1117             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1118             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1119
1120             fjx0             = _mm_setzero_pd();
1121             fjy0             = _mm_setzero_pd();
1122             fjz0             = _mm_setzero_pd();
1123             fjx1             = _mm_setzero_pd();
1124             fjy1             = _mm_setzero_pd();
1125             fjz1             = _mm_setzero_pd();
1126             fjx2             = _mm_setzero_pd();
1127             fjy2             = _mm_setzero_pd();
1128             fjz2             = _mm_setzero_pd();
1129             fjx3             = _mm_setzero_pd();
1130             fjy3             = _mm_setzero_pd();
1131             fjz3             = _mm_setzero_pd();
1132
1133             /**************************
1134              * CALCULATE INTERACTIONS *
1135              **************************/
1136
1137             /* LENNARD-JONES DISPERSION/REPULSION */
1138
1139             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1140             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1141
1142             fscal            = fvdw;
1143
1144             /* Update vectorial force */
1145             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1146             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1147             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1148
1149             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1150             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1151             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1152
1153             /**************************
1154              * CALCULATE INTERACTIONS *
1155              **************************/
1156
1157             /* COULOMB ELECTROSTATICS */
1158             velec            = _mm_mul_pd(qq11,rinv11);
1159             felec            = _mm_mul_pd(velec,rinvsq11);
1160
1161             fscal            = felec;
1162
1163             /* Update vectorial force */
1164             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1165             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1166             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1167
1168             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1169             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1170             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1171
1172             /**************************
1173              * CALCULATE INTERACTIONS *
1174              **************************/
1175
1176             /* COULOMB ELECTROSTATICS */
1177             velec            = _mm_mul_pd(qq12,rinv12);
1178             felec            = _mm_mul_pd(velec,rinvsq12);
1179
1180             fscal            = felec;
1181
1182             /* Update vectorial force */
1183             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1184             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1185             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1186
1187             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1188             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1189             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1190
1191             /**************************
1192              * CALCULATE INTERACTIONS *
1193              **************************/
1194
1195             /* COULOMB ELECTROSTATICS */
1196             velec            = _mm_mul_pd(qq13,rinv13);
1197             felec            = _mm_mul_pd(velec,rinvsq13);
1198
1199             fscal            = felec;
1200
1201             /* Update vectorial force */
1202             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1203             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1204             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1205
1206             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1207             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1208             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1209
1210             /**************************
1211              * CALCULATE INTERACTIONS *
1212              **************************/
1213
1214             /* COULOMB ELECTROSTATICS */
1215             velec            = _mm_mul_pd(qq21,rinv21);
1216             felec            = _mm_mul_pd(velec,rinvsq21);
1217
1218             fscal            = felec;
1219
1220             /* Update vectorial force */
1221             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1222             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1223             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1224
1225             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1226             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1227             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1228
1229             /**************************
1230              * CALCULATE INTERACTIONS *
1231              **************************/
1232
1233             /* COULOMB ELECTROSTATICS */
1234             velec            = _mm_mul_pd(qq22,rinv22);
1235             felec            = _mm_mul_pd(velec,rinvsq22);
1236
1237             fscal            = felec;
1238
1239             /* Update vectorial force */
1240             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1241             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1242             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1243
1244             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1245             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1246             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1247
1248             /**************************
1249              * CALCULATE INTERACTIONS *
1250              **************************/
1251
1252             /* COULOMB ELECTROSTATICS */
1253             velec            = _mm_mul_pd(qq23,rinv23);
1254             felec            = _mm_mul_pd(velec,rinvsq23);
1255
1256             fscal            = felec;
1257
1258             /* Update vectorial force */
1259             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1260             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1261             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1262
1263             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1264             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1265             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1266
1267             /**************************
1268              * CALCULATE INTERACTIONS *
1269              **************************/
1270
1271             /* COULOMB ELECTROSTATICS */
1272             velec            = _mm_mul_pd(qq31,rinv31);
1273             felec            = _mm_mul_pd(velec,rinvsq31);
1274
1275             fscal            = felec;
1276
1277             /* Update vectorial force */
1278             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1279             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1280             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1281
1282             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1283             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1284             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1285
1286             /**************************
1287              * CALCULATE INTERACTIONS *
1288              **************************/
1289
1290             /* COULOMB ELECTROSTATICS */
1291             velec            = _mm_mul_pd(qq32,rinv32);
1292             felec            = _mm_mul_pd(velec,rinvsq32);
1293
1294             fscal            = felec;
1295
1296             /* Update vectorial force */
1297             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1298             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1299             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1300
1301             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1302             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1303             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1304
1305             /**************************
1306              * CALCULATE INTERACTIONS *
1307              **************************/
1308
1309             /* COULOMB ELECTROSTATICS */
1310             velec            = _mm_mul_pd(qq33,rinv33);
1311             felec            = _mm_mul_pd(velec,rinvsq33);
1312
1313             fscal            = felec;
1314
1315             /* Update vectorial force */
1316             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1317             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1318             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1319
1320             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1321             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1322             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1323
1324             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1325
1326             /* Inner loop uses 303 flops */
1327         }
1328
1329         if(jidx<j_index_end)
1330         {
1331
1332             jnrA             = jjnr[jidx];
1333             j_coord_offsetA  = DIM*jnrA;
1334
1335             /* load j atom coordinates */
1336             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1337                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1338                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1339
1340             /* Calculate displacement vector */
1341             dx00             = _mm_sub_pd(ix0,jx0);
1342             dy00             = _mm_sub_pd(iy0,jy0);
1343             dz00             = _mm_sub_pd(iz0,jz0);
1344             dx11             = _mm_sub_pd(ix1,jx1);
1345             dy11             = _mm_sub_pd(iy1,jy1);
1346             dz11             = _mm_sub_pd(iz1,jz1);
1347             dx12             = _mm_sub_pd(ix1,jx2);
1348             dy12             = _mm_sub_pd(iy1,jy2);
1349             dz12             = _mm_sub_pd(iz1,jz2);
1350             dx13             = _mm_sub_pd(ix1,jx3);
1351             dy13             = _mm_sub_pd(iy1,jy3);
1352             dz13             = _mm_sub_pd(iz1,jz3);
1353             dx21             = _mm_sub_pd(ix2,jx1);
1354             dy21             = _mm_sub_pd(iy2,jy1);
1355             dz21             = _mm_sub_pd(iz2,jz1);
1356             dx22             = _mm_sub_pd(ix2,jx2);
1357             dy22             = _mm_sub_pd(iy2,jy2);
1358             dz22             = _mm_sub_pd(iz2,jz2);
1359             dx23             = _mm_sub_pd(ix2,jx3);
1360             dy23             = _mm_sub_pd(iy2,jy3);
1361             dz23             = _mm_sub_pd(iz2,jz3);
1362             dx31             = _mm_sub_pd(ix3,jx1);
1363             dy31             = _mm_sub_pd(iy3,jy1);
1364             dz31             = _mm_sub_pd(iz3,jz1);
1365             dx32             = _mm_sub_pd(ix3,jx2);
1366             dy32             = _mm_sub_pd(iy3,jy2);
1367             dz32             = _mm_sub_pd(iz3,jz2);
1368             dx33             = _mm_sub_pd(ix3,jx3);
1369             dy33             = _mm_sub_pd(iy3,jy3);
1370             dz33             = _mm_sub_pd(iz3,jz3);
1371
1372             /* Calculate squared distance and things based on it */
1373             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1374             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1375             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1376             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1377             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1378             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1379             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1380             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1381             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1382             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1383
1384             rinv11           = avx128fma_invsqrt_d(rsq11);
1385             rinv12           = avx128fma_invsqrt_d(rsq12);
1386             rinv13           = avx128fma_invsqrt_d(rsq13);
1387             rinv21           = avx128fma_invsqrt_d(rsq21);
1388             rinv22           = avx128fma_invsqrt_d(rsq22);
1389             rinv23           = avx128fma_invsqrt_d(rsq23);
1390             rinv31           = avx128fma_invsqrt_d(rsq31);
1391             rinv32           = avx128fma_invsqrt_d(rsq32);
1392             rinv33           = avx128fma_invsqrt_d(rsq33);
1393
1394             rinvsq00         = avx128fma_inv_d(rsq00);
1395             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1396             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1397             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1398             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1399             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1400             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1401             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1402             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1403             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1404
1405             fjx0             = _mm_setzero_pd();
1406             fjy0             = _mm_setzero_pd();
1407             fjz0             = _mm_setzero_pd();
1408             fjx1             = _mm_setzero_pd();
1409             fjy1             = _mm_setzero_pd();
1410             fjz1             = _mm_setzero_pd();
1411             fjx2             = _mm_setzero_pd();
1412             fjy2             = _mm_setzero_pd();
1413             fjz2             = _mm_setzero_pd();
1414             fjx3             = _mm_setzero_pd();
1415             fjy3             = _mm_setzero_pd();
1416             fjz3             = _mm_setzero_pd();
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             /* LENNARD-JONES DISPERSION/REPULSION */
1423
1424             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1425             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1426
1427             fscal            = fvdw;
1428
1429             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1430
1431             /* Update vectorial force */
1432             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1433             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1434             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1435
1436             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1437             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1438             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1439
1440             /**************************
1441              * CALCULATE INTERACTIONS *
1442              **************************/
1443
1444             /* COULOMB ELECTROSTATICS */
1445             velec            = _mm_mul_pd(qq11,rinv11);
1446             felec            = _mm_mul_pd(velec,rinvsq11);
1447
1448             fscal            = felec;
1449
1450             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1451
1452             /* Update vectorial force */
1453             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1454             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1455             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1456
1457             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1458             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1459             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1460
1461             /**************************
1462              * CALCULATE INTERACTIONS *
1463              **************************/
1464
1465             /* COULOMB ELECTROSTATICS */
1466             velec            = _mm_mul_pd(qq12,rinv12);
1467             felec            = _mm_mul_pd(velec,rinvsq12);
1468
1469             fscal            = felec;
1470
1471             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1472
1473             /* Update vectorial force */
1474             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1475             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1476             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1477
1478             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1479             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1480             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1481
1482             /**************************
1483              * CALCULATE INTERACTIONS *
1484              **************************/
1485
1486             /* COULOMB ELECTROSTATICS */
1487             velec            = _mm_mul_pd(qq13,rinv13);
1488             felec            = _mm_mul_pd(velec,rinvsq13);
1489
1490             fscal            = felec;
1491
1492             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1493
1494             /* Update vectorial force */
1495             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1496             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1497             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1498
1499             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1500             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1501             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1502
1503             /**************************
1504              * CALCULATE INTERACTIONS *
1505              **************************/
1506
1507             /* COULOMB ELECTROSTATICS */
1508             velec            = _mm_mul_pd(qq21,rinv21);
1509             felec            = _mm_mul_pd(velec,rinvsq21);
1510
1511             fscal            = felec;
1512
1513             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1514
1515             /* Update vectorial force */
1516             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1517             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1518             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1519
1520             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1521             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1522             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1523
1524             /**************************
1525              * CALCULATE INTERACTIONS *
1526              **************************/
1527
1528             /* COULOMB ELECTROSTATICS */
1529             velec            = _mm_mul_pd(qq22,rinv22);
1530             felec            = _mm_mul_pd(velec,rinvsq22);
1531
1532             fscal            = felec;
1533
1534             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1535
1536             /* Update vectorial force */
1537             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1538             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1539             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1540
1541             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1542             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1543             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1544
1545             /**************************
1546              * CALCULATE INTERACTIONS *
1547              **************************/
1548
1549             /* COULOMB ELECTROSTATICS */
1550             velec            = _mm_mul_pd(qq23,rinv23);
1551             felec            = _mm_mul_pd(velec,rinvsq23);
1552
1553             fscal            = felec;
1554
1555             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1556
1557             /* Update vectorial force */
1558             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1559             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1560             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1561
1562             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1563             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1564             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1565
1566             /**************************
1567              * CALCULATE INTERACTIONS *
1568              **************************/
1569
1570             /* COULOMB ELECTROSTATICS */
1571             velec            = _mm_mul_pd(qq31,rinv31);
1572             felec            = _mm_mul_pd(velec,rinvsq31);
1573
1574             fscal            = felec;
1575
1576             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1577
1578             /* Update vectorial force */
1579             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1580             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1581             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1582
1583             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1584             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1585             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1586
1587             /**************************
1588              * CALCULATE INTERACTIONS *
1589              **************************/
1590
1591             /* COULOMB ELECTROSTATICS */
1592             velec            = _mm_mul_pd(qq32,rinv32);
1593             felec            = _mm_mul_pd(velec,rinvsq32);
1594
1595             fscal            = felec;
1596
1597             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1598
1599             /* Update vectorial force */
1600             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1601             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1602             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1603
1604             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1605             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1606             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1607
1608             /**************************
1609              * CALCULATE INTERACTIONS *
1610              **************************/
1611
1612             /* COULOMB ELECTROSTATICS */
1613             velec            = _mm_mul_pd(qq33,rinv33);
1614             felec            = _mm_mul_pd(velec,rinvsq33);
1615
1616             fscal            = felec;
1617
1618             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1619
1620             /* Update vectorial force */
1621             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1622             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1623             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1624
1625             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1626             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1627             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1628
1629             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1630
1631             /* Inner loop uses 303 flops */
1632         }
1633
1634         /* End of innermost loop */
1635
1636         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1637                                               f+i_coord_offset,fshift+i_shift_offset);
1638
1639         /* Increment number of inner iterations */
1640         inneriter                  += j_index_end - j_index_start;
1641
1642         /* Outer loop uses 24 flops */
1643     }
1644
1645     /* Increment number of outer iterations */
1646     outeriter        += nri;
1647
1648     /* Update outer/inner flops */
1649
1650     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1651 }