src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_avx_256_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_256_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_256_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_VF_avx_256_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_VF_avx_256_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,C,D refer to j loop unrolling done with AVX, e.g. for the four 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,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     real *           vdwioffsetptr0;
  84     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     real *           vdwioffsetptr1;
  86     __m256d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     real *           vdwioffsetptr2;
  88     __m256d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m256d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  93     __m256d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  94     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  95     real             *charge;
  96     int              nvdwtype;
  97     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  98     int              *vdwtype;
  99     real             *vdwparam;
 100     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 101     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 102     __m128i          vfitab;
 103     __m128i          ifour       = _mm_set1_epi32(4);
 104     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 105     real             *vftab;
 106     __m256d          dummy_mask,cutoff_mask;
 107     __m128           tmpmask0,tmpmask1;
 108     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 109     __m256d          one     = _mm256_set1_pd(1.0);
 110     __m256d          two     = _mm256_set1_pd(2.0);
 111     x                = xx[0];
 112     f                = ff[0];
 113
 114     nri              = nlist->nri;
 115     iinr             = nlist->iinr;
 116     jindex           = nlist->jindex;
 117     jjnr             = nlist->jjnr;
 118     shiftidx         = nlist->shift;
 119     gid              = nlist->gid;
 120     shiftvec         = fr->shift_vec[0];
 121     fshift           = fr->fshift[0];
 122     facel            = _mm256_set1_pd(fr->ic->epsfac);
 123     charge           = mdatoms->chargeA;
 124     krf              = _mm256_set1_pd(fr->ic->k_rf);
 125     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 126     crf              = _mm256_set1_pd(fr->ic->c_rf);
 127     nvdwtype         = fr->ntype;
 128     vdwparam         = fr->nbfp;
 129     vdwtype          = mdatoms->typeA;
 130
 131     vftab            = kernel_data->table_vdw->data;
 132     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 133
 134     /* Setup water-specific parameters */
 135     inr              = nlist->iinr[0];
 136     iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 137     iq1              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
 138     iq2              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
 139     vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 140
 141     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 142     rcutoff_scalar   = fr->ic->rcoulomb;
 143     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 144     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 145
 146     /* Avoid stupid compiler warnings */
 147     jnrA = jnrB = jnrC = jnrD = 0;
 148     j_coord_offsetA = 0;
 149     j_coord_offsetB = 0;
 150     j_coord_offsetC = 0;
 151     j_coord_offsetD = 0;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     for(iidx=0;iidx<4*DIM;iidx++)
 157     {
 158         scratch[iidx] = 0.0;
 159     }
 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_mm256_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             = _mm256_setzero_pd();
 180         fiy0             = _mm256_setzero_pd();
 181         fiz0             = _mm256_setzero_pd();
 182         fix1             = _mm256_setzero_pd();
 183         fiy1             = _mm256_setzero_pd();
 184         fiz1             = _mm256_setzero_pd();
 185         fix2             = _mm256_setzero_pd();
 186         fiy2             = _mm256_setzero_pd();
 187         fiz2             = _mm256_setzero_pd();
 188
 189         /* Reset potential sums */
 190         velecsum         = _mm256_setzero_pd();
 191         vvdwsum          = _mm256_setzero_pd();
 192
 193         /* Start inner kernel loop */
 194         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 195         {
 196
 197             /* Get j neighbor index, and coordinate index */
 198             jnrA             = jjnr[jidx];
 199             jnrB             = jjnr[jidx+1];
 200             jnrC             = jjnr[jidx+2];
 201             jnrD             = jjnr[jidx+3];
 202             j_coord_offsetA  = DIM*jnrA;
 203             j_coord_offsetB  = DIM*jnrB;
 204             j_coord_offsetC  = DIM*jnrC;
 205             j_coord_offsetD  = DIM*jnrD;
 206
 207             /* load j atom coordinates */
 208             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 209                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 210                                                  &jx0,&jy0,&jz0);
 211
 212             /* Calculate displacement vector */
 213             dx00             = _mm256_sub_pd(ix0,jx0);
 214             dy00             = _mm256_sub_pd(iy0,jy0);
 215             dz00             = _mm256_sub_pd(iz0,jz0);
 216             dx10             = _mm256_sub_pd(ix1,jx0);
 217             dy10             = _mm256_sub_pd(iy1,jy0);
 218             dz10             = _mm256_sub_pd(iz1,jz0);
 219             dx20             = _mm256_sub_pd(ix2,jx0);
 220             dy20             = _mm256_sub_pd(iy2,jy0);
 221             dz20             = _mm256_sub_pd(iz2,jz0);
 222
 223             /* Calculate squared distance and things based on it */
 224             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 225             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
 226             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
 227
 228             rinv00           = avx256_invsqrt_d(rsq00);
 229             rinv10           = avx256_invsqrt_d(rsq10);
 230             rinv20           = avx256_invsqrt_d(rsq20);
 231
 232             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 233             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
 234             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
 235
 236             /* Load parameters for j particles */
 237             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 238                                                                  charge+jnrC+0,charge+jnrD+0);
 239             vdwjidx0A        = 2*vdwtype[jnrA+0];
 240             vdwjidx0B        = 2*vdwtype[jnrB+0];
 241             vdwjidx0C        = 2*vdwtype[jnrC+0];
 242             vdwjidx0D        = 2*vdwtype[jnrD+0];
 243
 244             fjx0             = _mm256_setzero_pd();
 245             fjy0             = _mm256_setzero_pd();
 246             fjz0             = _mm256_setzero_pd();
 247
 248             /**************************
 249              * CALCULATE INTERACTIONS *
 250              **************************/
 251
 252             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 253             {
 254
 255             r00              = _mm256_mul_pd(rsq00,rinv00);
 256
 257             /* Compute parameters for interactions between i and j atoms */
 258             qq00             = _mm256_mul_pd(iq0,jq0);
 259             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 260                                             vdwioffsetptr0+vdwjidx0B,
 261                                             vdwioffsetptr0+vdwjidx0C,
 262                                             vdwioffsetptr0+vdwjidx0D,
 263                                             &c6_00,&c12_00);
 264
 265             /* Calculate table index by multiplying r with table scale and truncate to integer */
 266             rt               = _mm256_mul_pd(r00,vftabscale);
 267             vfitab           = _mm256_cvttpd_epi32(rt);
 268             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 269             vfitab           = _mm_slli_epi32(vfitab,3);
 270
 271             /* REACTION-FIELD ELECTROSTATICS */
 272             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 273             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 274
 275             /* CUBIC SPLINE TABLE DISPERSION */
 276             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 277             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 278             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 279             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 280             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 281             Heps             = _mm256_mul_pd(vfeps,H);
 282             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 283             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 284             vvdw6            = _mm256_mul_pd(c6_00,VV);
 285             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 286             fvdw6            = _mm256_mul_pd(c6_00,FF);
 287
 288             /* CUBIC SPLINE TABLE REPULSION */
 289             vfitab           = _mm_add_epi32(vfitab,ifour);
 290             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 291             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 292             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 293             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 294             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 295             Heps             = _mm256_mul_pd(vfeps,H);
 296             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 297             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 298             vvdw12           = _mm256_mul_pd(c12_00,VV);
 299             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 300             fvdw12           = _mm256_mul_pd(c12_00,FF);
 301             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 302             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 303
 304             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 305
 306             /* Update potential sum for this i atom from the interaction with this j atom. */
 307             velec            = _mm256_and_pd(velec,cutoff_mask);
 308             velecsum         = _mm256_add_pd(velecsum,velec);
 309             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 310             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 311
 312             fscal            = _mm256_add_pd(felec,fvdw);
 313
 314             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 315
 316             /* Calculate temporary vectorial force */
 317             tx               = _mm256_mul_pd(fscal,dx00);
 318             ty               = _mm256_mul_pd(fscal,dy00);
 319             tz               = _mm256_mul_pd(fscal,dz00);
 320
 321             /* Update vectorial force */
 322             fix0             = _mm256_add_pd(fix0,tx);
 323             fiy0             = _mm256_add_pd(fiy0,ty);
 324             fiz0             = _mm256_add_pd(fiz0,tz);
 325
 326             fjx0             = _mm256_add_pd(fjx0,tx);
 327             fjy0             = _mm256_add_pd(fjy0,ty);
 328             fjz0             = _mm256_add_pd(fjz0,tz);
 329
 330             }
 331
 332             /**************************
 333              * CALCULATE INTERACTIONS *
 334              **************************/
 335
 336             if (gmx_mm256_any_lt(rsq10,rcutoff2))
 337             {
 338
 339             /* Compute parameters for interactions between i and j atoms */
 340             qq10             = _mm256_mul_pd(iq1,jq0);
 341
 342             /* REACTION-FIELD ELECTROSTATICS */
 343             velec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
 344             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
 345
 346             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
 347
 348             /* Update potential sum for this i atom from the interaction with this j atom. */
 349             velec            = _mm256_and_pd(velec,cutoff_mask);
 350             velecsum         = _mm256_add_pd(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 355
 356             /* Calculate temporary vectorial force */
 357             tx               = _mm256_mul_pd(fscal,dx10);
 358             ty               = _mm256_mul_pd(fscal,dy10);
 359             tz               = _mm256_mul_pd(fscal,dz10);
 360
 361             /* Update vectorial force */
 362             fix1             = _mm256_add_pd(fix1,tx);
 363             fiy1             = _mm256_add_pd(fiy1,ty);
 364             fiz1             = _mm256_add_pd(fiz1,tz);
 365
 366             fjx0             = _mm256_add_pd(fjx0,tx);
 367             fjy0             = _mm256_add_pd(fjy0,ty);
 368             fjz0             = _mm256_add_pd(fjz0,tz);
 369
 370             }
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             if (gmx_mm256_any_lt(rsq20,rcutoff2))
 377             {
 378
 379             /* Compute parameters for interactions between i and j atoms */
 380             qq20             = _mm256_mul_pd(iq2,jq0);
 381
 382             /* REACTION-FIELD ELECTROSTATICS */
 383             velec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
 384             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
 385
 386             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
 387
 388             /* Update potential sum for this i atom from the interaction with this j atom. */
 389             velec            = _mm256_and_pd(velec,cutoff_mask);
 390             velecsum         = _mm256_add_pd(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 395
 396             /* Calculate temporary vectorial force */
 397             tx               = _mm256_mul_pd(fscal,dx20);
 398             ty               = _mm256_mul_pd(fscal,dy20);
 399             tz               = _mm256_mul_pd(fscal,dz20);
 400
 401             /* Update vectorial force */
 402             fix2             = _mm256_add_pd(fix2,tx);
 403             fiy2             = _mm256_add_pd(fiy2,ty);
 404             fiz2             = _mm256_add_pd(fiz2,tz);
 405
 406             fjx0             = _mm256_add_pd(fjx0,tx);
 407             fjy0             = _mm256_add_pd(fjy0,ty);
 408             fjz0             = _mm256_add_pd(fjz0,tz);
 409
 410             }
 411
 412             fjptrA             = f+j_coord_offsetA;
 413             fjptrB             = f+j_coord_offsetB;
 414             fjptrC             = f+j_coord_offsetC;
 415             fjptrD             = f+j_coord_offsetD;
 416
 417             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 418
 419             /* Inner loop uses 147 flops */
 420         }
 421
 422         if(jidx<j_index_end)
 423         {
 424
 425             /* Get j neighbor index, and coordinate index */
 426             jnrlistA         = jjnr[jidx];
 427             jnrlistB         = jjnr[jidx+1];
 428             jnrlistC         = jjnr[jidx+2];
 429             jnrlistD         = jjnr[jidx+3];
 430             /* Sign of each element will be negative for non-real atoms.
 431              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 432              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 433              */
 434             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 435
 436             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 437             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 438             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 439
 440             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 441             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 442             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 443             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 444             j_coord_offsetA  = DIM*jnrA;
 445             j_coord_offsetB  = DIM*jnrB;
 446             j_coord_offsetC  = DIM*jnrC;
 447             j_coord_offsetD  = DIM*jnrD;
 448
 449             /* load j atom coordinates */
 450             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 451                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 452                                                  &jx0,&jy0,&jz0);
 453
 454             /* Calculate displacement vector */
 455             dx00             = _mm256_sub_pd(ix0,jx0);
 456             dy00             = _mm256_sub_pd(iy0,jy0);
 457             dz00             = _mm256_sub_pd(iz0,jz0);
 458             dx10             = _mm256_sub_pd(ix1,jx0);
 459             dy10             = _mm256_sub_pd(iy1,jy0);
 460             dz10             = _mm256_sub_pd(iz1,jz0);
 461             dx20             = _mm256_sub_pd(ix2,jx0);
 462             dy20             = _mm256_sub_pd(iy2,jy0);
 463             dz20             = _mm256_sub_pd(iz2,jz0);
 464
 465             /* Calculate squared distance and things based on it */
 466             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 467             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
 468             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
 469
 470             rinv00           = avx256_invsqrt_d(rsq00);
 471             rinv10           = avx256_invsqrt_d(rsq10);
 472             rinv20           = avx256_invsqrt_d(rsq20);
 473
 474             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 475             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
 476             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
 477
 478             /* Load parameters for j particles */
 479             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 480                                                                  charge+jnrC+0,charge+jnrD+0);
 481             vdwjidx0A        = 2*vdwtype[jnrA+0];
 482             vdwjidx0B        = 2*vdwtype[jnrB+0];
 483             vdwjidx0C        = 2*vdwtype[jnrC+0];
 484             vdwjidx0D        = 2*vdwtype[jnrD+0];
 485
 486             fjx0             = _mm256_setzero_pd();
 487             fjy0             = _mm256_setzero_pd();
 488             fjz0             = _mm256_setzero_pd();
 489
 490             /**************************
 491              * CALCULATE INTERACTIONS *
 492              **************************/
 493
 494             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 495             {
 496
 497             r00              = _mm256_mul_pd(rsq00,rinv00);
 498             r00              = _mm256_andnot_pd(dummy_mask,r00);
 499
 500             /* Compute parameters for interactions between i and j atoms */
 501             qq00             = _mm256_mul_pd(iq0,jq0);
 502             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 503                                             vdwioffsetptr0+vdwjidx0B,
 504                                             vdwioffsetptr0+vdwjidx0C,
 505                                             vdwioffsetptr0+vdwjidx0D,
 506                                             &c6_00,&c12_00);
 507
 508             /* Calculate table index by multiplying r with table scale and truncate to integer */
 509             rt               = _mm256_mul_pd(r00,vftabscale);
 510             vfitab           = _mm256_cvttpd_epi32(rt);
 511             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 512             vfitab           = _mm_slli_epi32(vfitab,3);
 513
 514             /* REACTION-FIELD ELECTROSTATICS */
 515             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 516             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 517
 518             /* CUBIC SPLINE TABLE DISPERSION */
 519             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 520             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 521             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 522             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 523             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 524             Heps             = _mm256_mul_pd(vfeps,H);
 525             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 526             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 527             vvdw6            = _mm256_mul_pd(c6_00,VV);
 528             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 529             fvdw6            = _mm256_mul_pd(c6_00,FF);
 530
 531             /* CUBIC SPLINE TABLE REPULSION */
 532             vfitab           = _mm_add_epi32(vfitab,ifour);
 533             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 534             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 535             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 536             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 537             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 538             Heps             = _mm256_mul_pd(vfeps,H);
 539             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 540             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 541             vvdw12           = _mm256_mul_pd(c12_00,VV);
 542             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 543             fvdw12           = _mm256_mul_pd(c12_00,FF);
 544             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 545             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 546
 547             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 548
 549             /* Update potential sum for this i atom from the interaction with this j atom. */
 550             velec            = _mm256_and_pd(velec,cutoff_mask);
 551             velec            = _mm256_andnot_pd(dummy_mask,velec);
 552             velecsum         = _mm256_add_pd(velecsum,velec);
 553             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 554             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 555             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 556
 557             fscal            = _mm256_add_pd(felec,fvdw);
 558
 559             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 560
 561             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 562
 563             /* Calculate temporary vectorial force */
 564             tx               = _mm256_mul_pd(fscal,dx00);
 565             ty               = _mm256_mul_pd(fscal,dy00);
 566             tz               = _mm256_mul_pd(fscal,dz00);
 567
 568             /* Update vectorial force */
 569             fix0             = _mm256_add_pd(fix0,tx);
 570             fiy0             = _mm256_add_pd(fiy0,ty);
 571             fiz0             = _mm256_add_pd(fiz0,tz);
 572
 573             fjx0             = _mm256_add_pd(fjx0,tx);
 574             fjy0             = _mm256_add_pd(fjy0,ty);
 575             fjz0             = _mm256_add_pd(fjz0,tz);
 576
 577             }
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             if (gmx_mm256_any_lt(rsq10,rcutoff2))
 584             {
 585
 586             /* Compute parameters for interactions between i and j atoms */
 587             qq10             = _mm256_mul_pd(iq1,jq0);
 588
 589             /* REACTION-FIELD ELECTROSTATICS */
 590             velec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
 591             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
 592
 593             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
 594
 595             /* Update potential sum for this i atom from the interaction with this j atom. */
 596             velec            = _mm256_and_pd(velec,cutoff_mask);
 597             velec            = _mm256_andnot_pd(dummy_mask,velec);
 598             velecsum         = _mm256_add_pd(velecsum,velec);
 599
 600             fscal            = felec;
 601
 602             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 603
 604             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 605
 606             /* Calculate temporary vectorial force */
 607             tx               = _mm256_mul_pd(fscal,dx10);
 608             ty               = _mm256_mul_pd(fscal,dy10);
 609             tz               = _mm256_mul_pd(fscal,dz10);
 610
 611             /* Update vectorial force */
 612             fix1             = _mm256_add_pd(fix1,tx);
 613             fiy1             = _mm256_add_pd(fiy1,ty);
 614             fiz1             = _mm256_add_pd(fiz1,tz);
 615
 616             fjx0             = _mm256_add_pd(fjx0,tx);
 617             fjy0             = _mm256_add_pd(fjy0,ty);
 618             fjz0             = _mm256_add_pd(fjz0,tz);
 619
 620             }
 621
 622             /**************************
 623              * CALCULATE INTERACTIONS *
 624              **************************/
 625
 626             if (gmx_mm256_any_lt(rsq20,rcutoff2))
 627             {
 628
 629             /* Compute parameters for interactions between i and j atoms */
 630             qq20             = _mm256_mul_pd(iq2,jq0);
 631
 632             /* REACTION-FIELD ELECTROSTATICS */
 633             velec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
 634             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
 635
 636             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
 637
 638             /* Update potential sum for this i atom from the interaction with this j atom. */
 639             velec            = _mm256_and_pd(velec,cutoff_mask);
 640             velec            = _mm256_andnot_pd(dummy_mask,velec);
 641             velecsum         = _mm256_add_pd(velecsum,velec);
 642
 643             fscal            = felec;
 644
 645             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 646
 647             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 648
 649             /* Calculate temporary vectorial force */
 650             tx               = _mm256_mul_pd(fscal,dx20);
 651             ty               = _mm256_mul_pd(fscal,dy20);
 652             tz               = _mm256_mul_pd(fscal,dz20);
 653
 654             /* Update vectorial force */
 655             fix2             = _mm256_add_pd(fix2,tx);
 656             fiy2             = _mm256_add_pd(fiy2,ty);
 657             fiz2             = _mm256_add_pd(fiz2,tz);
 658
 659             fjx0             = _mm256_add_pd(fjx0,tx);
 660             fjy0             = _mm256_add_pd(fjy0,ty);
 661             fjz0             = _mm256_add_pd(fjz0,tz);
 662
 663             }
 664
 665             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 666             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 667             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 668             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 669
 670             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 671
 672             /* Inner loop uses 148 flops */
 673         }
 674
 675         /* End of innermost loop */
 676
 677         gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 678                                                  f+i_coord_offset,fshift+i_shift_offset);
 679
 680         ggid                        = gid[iidx];
 681         /* Update potential energies */
 682         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 683         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 684
 685         /* Increment number of inner iterations */
 686         inneriter                  += j_index_end - j_index_start;
 687
 688         /* Outer loop uses 20 flops */
 689     }
 690
 691     /* Increment number of outer iterations */
 692     outeriter        += nri;
 693
 694     /* Update outer/inner flops */
 695
 696     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*148);
 697 }
 698 /*
 699  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_F_avx_256_double
 700  * Electrostatics interaction: ReactionField
 701  * VdW interaction:            CubicSplineTable
 702  * Geometry:                   Water3-Particle
 703  * Calculate force/pot:        Force
 704  */
 705 void
 706 nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_F_avx_256_double
 707                     (t_nblist                    * gmx_restrict       nlist,
 708                      rvec                        * gmx_restrict          xx,
 709                      rvec                        * gmx_restrict          ff,
 710                      struct t_forcerec           * gmx_restrict          fr,
 711                      t_mdatoms                   * gmx_restrict     mdatoms,
 712                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 713                      t_nrnb                      * gmx_restrict        nrnb)
 714 {
 715     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 716      * just 0 for non-waters.
 717      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 718      * jnr indices corresponding to data put in the four positions in the SIMD register.
 719      */
 720     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 721     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 722     int              jnrA,jnrB,jnrC,jnrD;
 723     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 724     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 725     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 726     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 727     real             rcutoff_scalar;
 728     real             *shiftvec,*fshift,*x,*f;
 729     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 730     real             scratch[4*DIM];
 731     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 732     real *           vdwioffsetptr0;
 733     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 734     real *           vdwioffsetptr1;
 735     __m256d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 736     real *           vdwioffsetptr2;
 737     __m256d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 738     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 739     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 740     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 741     __m256d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 742     __m256d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 743     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 744     real             *charge;
 745     int              nvdwtype;
 746     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 747     int              *vdwtype;
 748     real             *vdwparam;
 749     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 750     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 751     __m128i          vfitab;
 752     __m128i          ifour       = _mm_set1_epi32(4);
 753     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 754     real             *vftab;
 755     __m256d          dummy_mask,cutoff_mask;
 756     __m128           tmpmask0,tmpmask1;
 757     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 758     __m256d          one     = _mm256_set1_pd(1.0);
 759     __m256d          two     = _mm256_set1_pd(2.0);
 760     x                = xx[0];
 761     f                = ff[0];
 762
 763     nri              = nlist->nri;
 764     iinr             = nlist->iinr;
 765     jindex           = nlist->jindex;
 766     jjnr             = nlist->jjnr;
 767     shiftidx         = nlist->shift;
 768     gid              = nlist->gid;
 769     shiftvec         = fr->shift_vec[0];
 770     fshift           = fr->fshift[0];
 771     facel            = _mm256_set1_pd(fr->ic->epsfac);
 772     charge           = mdatoms->chargeA;
 773     krf              = _mm256_set1_pd(fr->ic->k_rf);
 774     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 775     crf              = _mm256_set1_pd(fr->ic->c_rf);
 776     nvdwtype         = fr->ntype;
 777     vdwparam         = fr->nbfp;
 778     vdwtype          = mdatoms->typeA;
 779
 780     vftab            = kernel_data->table_vdw->data;
 781     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 782
 783     /* Setup water-specific parameters */
 784     inr              = nlist->iinr[0];
 785     iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 786     iq1              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
 787     iq2              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
 788     vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 789
 790     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 791     rcutoff_scalar   = fr->ic->rcoulomb;
 792     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 793     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 794
 795     /* Avoid stupid compiler warnings */
 796     jnrA = jnrB = jnrC = jnrD = 0;
 797     j_coord_offsetA = 0;
 798     j_coord_offsetB = 0;
 799     j_coord_offsetC = 0;
 800     j_coord_offsetD = 0;
 801
 802     outeriter        = 0;
 803     inneriter        = 0;
 804
 805     for(iidx=0;iidx<4*DIM;iidx++)
 806     {
 807         scratch[iidx] = 0.0;
 808     }
 809
 810     /* Start outer loop over neighborlists */
 811     for(iidx=0; iidx<nri; iidx++)
 812     {
 813         /* Load shift vector for this list */
 814         i_shift_offset   = DIM*shiftidx[iidx];
 815
 816         /* Load limits for loop over neighbors */
 817         j_index_start    = jindex[iidx];
 818         j_index_end      = jindex[iidx+1];
 819
 820         /* Get outer coordinate index */
 821         inr              = iinr[iidx];
 822         i_coord_offset   = DIM*inr;
 823
 824         /* Load i particle coords and add shift vector */
 825         gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 826                                                     &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 827
 828         fix0             = _mm256_setzero_pd();
 829         fiy0             = _mm256_setzero_pd();
 830         fiz0             = _mm256_setzero_pd();
 831         fix1             = _mm256_setzero_pd();
 832         fiy1             = _mm256_setzero_pd();
 833         fiz1             = _mm256_setzero_pd();
 834         fix2             = _mm256_setzero_pd();
 835         fiy2             = _mm256_setzero_pd();
 836         fiz2             = _mm256_setzero_pd();
 837
 838         /* Start inner kernel loop */
 839         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 840         {
 841
 842             /* Get j neighbor index, and coordinate index */
 843             jnrA             = jjnr[jidx];
 844             jnrB             = jjnr[jidx+1];
 845             jnrC             = jjnr[jidx+2];
 846             jnrD             = jjnr[jidx+3];
 847             j_coord_offsetA  = DIM*jnrA;
 848             j_coord_offsetB  = DIM*jnrB;
 849             j_coord_offsetC  = DIM*jnrC;
 850             j_coord_offsetD  = DIM*jnrD;
 851
 852             /* load j atom coordinates */
 853             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 854                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 855                                                  &jx0,&jy0,&jz0);
 856
 857             /* Calculate displacement vector */
 858             dx00             = _mm256_sub_pd(ix0,jx0);
 859             dy00             = _mm256_sub_pd(iy0,jy0);
 860             dz00             = _mm256_sub_pd(iz0,jz0);
 861             dx10             = _mm256_sub_pd(ix1,jx0);
 862             dy10             = _mm256_sub_pd(iy1,jy0);
 863             dz10             = _mm256_sub_pd(iz1,jz0);
 864             dx20             = _mm256_sub_pd(ix2,jx0);
 865             dy20             = _mm256_sub_pd(iy2,jy0);
 866             dz20             = _mm256_sub_pd(iz2,jz0);
 867
 868             /* Calculate squared distance and things based on it */
 869             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 870             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
 871             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
 872
 873             rinv00           = avx256_invsqrt_d(rsq00);
 874             rinv10           = avx256_invsqrt_d(rsq10);
 875             rinv20           = avx256_invsqrt_d(rsq20);
 876
 877             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 878             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
 879             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
 880
 881             /* Load parameters for j particles */
 882             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 883                                                                  charge+jnrC+0,charge+jnrD+0);
 884             vdwjidx0A        = 2*vdwtype[jnrA+0];
 885             vdwjidx0B        = 2*vdwtype[jnrB+0];
 886             vdwjidx0C        = 2*vdwtype[jnrC+0];
 887             vdwjidx0D        = 2*vdwtype[jnrD+0];
 888
 889             fjx0             = _mm256_setzero_pd();
 890             fjy0             = _mm256_setzero_pd();
 891             fjz0             = _mm256_setzero_pd();
 892
 893             /**************************
 894              * CALCULATE INTERACTIONS *
 895              **************************/
 896
 897             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 898             {
 899
 900             r00              = _mm256_mul_pd(rsq00,rinv00);
 901
 902             /* Compute parameters for interactions between i and j atoms */
 903             qq00             = _mm256_mul_pd(iq0,jq0);
 904             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 905                                             vdwioffsetptr0+vdwjidx0B,
 906                                             vdwioffsetptr0+vdwjidx0C,
 907                                             vdwioffsetptr0+vdwjidx0D,
 908                                             &c6_00,&c12_00);
 909
 910             /* Calculate table index by multiplying r with table scale and truncate to integer */
 911             rt               = _mm256_mul_pd(r00,vftabscale);
 912             vfitab           = _mm256_cvttpd_epi32(rt);
 913             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 914             vfitab           = _mm_slli_epi32(vfitab,3);
 915
 916             /* REACTION-FIELD ELECTROSTATICS */
 917             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 918
 919             /* CUBIC SPLINE TABLE DISPERSION */
 920             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 921             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 922             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 923             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 924             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 925             Heps             = _mm256_mul_pd(vfeps,H);
 926             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 927             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 928             fvdw6            = _mm256_mul_pd(c6_00,FF);
 929
 930             /* CUBIC SPLINE TABLE REPULSION */
 931             vfitab           = _mm_add_epi32(vfitab,ifour);
 932             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 933             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 934             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 935             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 936             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 937             Heps             = _mm256_mul_pd(vfeps,H);
 938             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 939             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 940             fvdw12           = _mm256_mul_pd(c12_00,FF);
 941             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 942
 943             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 944
 945             fscal            = _mm256_add_pd(felec,fvdw);
 946
 947             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 948
 949             /* Calculate temporary vectorial force */
 950             tx               = _mm256_mul_pd(fscal,dx00);
 951             ty               = _mm256_mul_pd(fscal,dy00);
 952             tz               = _mm256_mul_pd(fscal,dz00);
 953
 954             /* Update vectorial force */
 955             fix0             = _mm256_add_pd(fix0,tx);
 956             fiy0             = _mm256_add_pd(fiy0,ty);
 957             fiz0             = _mm256_add_pd(fiz0,tz);
 958
 959             fjx0             = _mm256_add_pd(fjx0,tx);
 960             fjy0             = _mm256_add_pd(fjy0,ty);
 961             fjz0             = _mm256_add_pd(fjz0,tz);
 962
 963             }
 964
 965             /**************************
 966              * CALCULATE INTERACTIONS *
 967              **************************/
 968
 969             if (gmx_mm256_any_lt(rsq10,rcutoff2))
 970             {
 971
 972             /* Compute parameters for interactions between i and j atoms */
 973             qq10             = _mm256_mul_pd(iq1,jq0);
 974
 975             /* REACTION-FIELD ELECTROSTATICS */
 976             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
 977
 978             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
 979
 980             fscal            = felec;
 981
 982             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 983
 984             /* Calculate temporary vectorial force */
 985             tx               = _mm256_mul_pd(fscal,dx10);
 986             ty               = _mm256_mul_pd(fscal,dy10);
 987             tz               = _mm256_mul_pd(fscal,dz10);
 988
 989             /* Update vectorial force */
 990             fix1             = _mm256_add_pd(fix1,tx);
 991             fiy1             = _mm256_add_pd(fiy1,ty);
 992             fiz1             = _mm256_add_pd(fiz1,tz);
 993
 994             fjx0             = _mm256_add_pd(fjx0,tx);
 995             fjy0             = _mm256_add_pd(fjy0,ty);
 996             fjz0             = _mm256_add_pd(fjz0,tz);
 997
 998             }
 999
1000             /**************************
1001              * CALCULATE INTERACTIONS *
1002              **************************/
1003
1004             if (gmx_mm256_any_lt(rsq20,rcutoff2))
1005             {
1006
1007             /* Compute parameters for interactions between i and j atoms */
1008             qq20             = _mm256_mul_pd(iq2,jq0);
1009
1010             /* REACTION-FIELD ELECTROSTATICS */
1011             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
1012
1013             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1014
1015             fscal            = felec;
1016
1017             fscal            = _mm256_and_pd(fscal,cutoff_mask);
1018
1019             /* Calculate temporary vectorial force */
1020             tx               = _mm256_mul_pd(fscal,dx20);
1021             ty               = _mm256_mul_pd(fscal,dy20);
1022             tz               = _mm256_mul_pd(fscal,dz20);
1023
1024             /* Update vectorial force */
1025             fix2             = _mm256_add_pd(fix2,tx);
1026             fiy2             = _mm256_add_pd(fiy2,ty);
1027             fiz2             = _mm256_add_pd(fiz2,tz);
1028
1029             fjx0             = _mm256_add_pd(fjx0,tx);
1030             fjy0             = _mm256_add_pd(fjy0,ty);
1031             fjz0             = _mm256_add_pd(fjz0,tz);
1032
1033             }
1034
1035             fjptrA             = f+j_coord_offsetA;
1036             fjptrB             = f+j_coord_offsetB;
1037             fjptrC             = f+j_coord_offsetC;
1038             fjptrD             = f+j_coord_offsetD;
1039
1040             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1041
1042             /* Inner loop uses 120 flops */
1043         }
1044
1045         if(jidx<j_index_end)
1046         {
1047
1048             /* Get j neighbor index, and coordinate index */
1049             jnrlistA         = jjnr[jidx];
1050             jnrlistB         = jjnr[jidx+1];
1051             jnrlistC         = jjnr[jidx+2];
1052             jnrlistD         = jjnr[jidx+3];
1053             /* Sign of each element will be negative for non-real atoms.
1054              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1055              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
1056              */
1057             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1058
1059             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
1060             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
1061             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
1062
1063             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1064             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1065             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1066             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1067             j_coord_offsetA  = DIM*jnrA;
1068             j_coord_offsetB  = DIM*jnrB;
1069             j_coord_offsetC  = DIM*jnrC;
1070             j_coord_offsetD  = DIM*jnrD;
1071
1072             /* load j atom coordinates */
1073             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1074                                                  x+j_coord_offsetC,x+j_coord_offsetD,
1075                                                  &jx0,&jy0,&jz0);
1076
1077             /* Calculate displacement vector */
1078             dx00             = _mm256_sub_pd(ix0,jx0);
1079             dy00             = _mm256_sub_pd(iy0,jy0);
1080             dz00             = _mm256_sub_pd(iz0,jz0);
1081             dx10             = _mm256_sub_pd(ix1,jx0);
1082             dy10             = _mm256_sub_pd(iy1,jy0);
1083             dz10             = _mm256_sub_pd(iz1,jz0);
1084             dx20             = _mm256_sub_pd(ix2,jx0);
1085             dy20             = _mm256_sub_pd(iy2,jy0);
1086             dz20             = _mm256_sub_pd(iz2,jz0);
1087
1088             /* Calculate squared distance and things based on it */
1089             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
1090             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
1091             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
1092
1093             rinv00           = avx256_invsqrt_d(rsq00);
1094             rinv10           = avx256_invsqrt_d(rsq10);
1095             rinv20           = avx256_invsqrt_d(rsq20);
1096
1097             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
1098             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
1099             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
1100
1101             /* Load parameters for j particles */
1102             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
1103                                                                  charge+jnrC+0,charge+jnrD+0);
1104             vdwjidx0A        = 2*vdwtype[jnrA+0];
1105             vdwjidx0B        = 2*vdwtype[jnrB+0];
1106             vdwjidx0C        = 2*vdwtype[jnrC+0];
1107             vdwjidx0D        = 2*vdwtype[jnrD+0];
1108
1109             fjx0             = _mm256_setzero_pd();
1110             fjy0             = _mm256_setzero_pd();
1111             fjz0             = _mm256_setzero_pd();
1112
1113             /**************************
1114              * CALCULATE INTERACTIONS *
1115              **************************/
1116
1117             if (gmx_mm256_any_lt(rsq00,rcutoff2))
1118             {
1119
1120             r00              = _mm256_mul_pd(rsq00,rinv00);
1121             r00              = _mm256_andnot_pd(dummy_mask,r00);
1122
1123             /* Compute parameters for interactions between i and j atoms */
1124             qq00             = _mm256_mul_pd(iq0,jq0);
1125             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
1126                                             vdwioffsetptr0+vdwjidx0B,
1127                                             vdwioffsetptr0+vdwjidx0C,
1128                                             vdwioffsetptr0+vdwjidx0D,
1129                                             &c6_00,&c12_00);
1130
1131             /* Calculate table index by multiplying r with table scale and truncate to integer */
1132             rt               = _mm256_mul_pd(r00,vftabscale);
1133             vfitab           = _mm256_cvttpd_epi32(rt);
1134             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
1135             vfitab           = _mm_slli_epi32(vfitab,3);
1136
1137             /* REACTION-FIELD ELECTROSTATICS */
1138             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
1139
1140             /* CUBIC SPLINE TABLE DISPERSION */
1141             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1142             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1143             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
1144             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
1145             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
1146             Heps             = _mm256_mul_pd(vfeps,H);
1147             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
1148             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
1149             fvdw6            = _mm256_mul_pd(c6_00,FF);
1150
1151             /* CUBIC SPLINE TABLE REPULSION */
1152             vfitab           = _mm_add_epi32(vfitab,ifour);
1153             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1154             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1155             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
1156             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
1157             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
1158             Heps             = _mm256_mul_pd(vfeps,H);
1159             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
1160             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
1161             fvdw12           = _mm256_mul_pd(c12_00,FF);
1162             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
1163
1164             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
1165
1166             fscal            = _mm256_add_pd(felec,fvdw);
1167
1168             fscal            = _mm256_and_pd(fscal,cutoff_mask);
1169
1170             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
1171
1172             /* Calculate temporary vectorial force */
1173             tx               = _mm256_mul_pd(fscal,dx00);
1174             ty               = _mm256_mul_pd(fscal,dy00);
1175             tz               = _mm256_mul_pd(fscal,dz00);
1176
1177             /* Update vectorial force */
1178             fix0             = _mm256_add_pd(fix0,tx);
1179             fiy0             = _mm256_add_pd(fiy0,ty);
1180             fiz0             = _mm256_add_pd(fiz0,tz);
1181
1182             fjx0             = _mm256_add_pd(fjx0,tx);
1183             fjy0             = _mm256_add_pd(fjy0,ty);
1184             fjz0             = _mm256_add_pd(fjz0,tz);
1185
1186             }
1187
1188             /**************************
1189              * CALCULATE INTERACTIONS *
1190              **************************/
1191
1192             if (gmx_mm256_any_lt(rsq10,rcutoff2))
1193             {
1194
1195             /* Compute parameters for interactions between i and j atoms */
1196             qq10             = _mm256_mul_pd(iq1,jq0);
1197
1198             /* REACTION-FIELD ELECTROSTATICS */
1199             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
1200
1201             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
1202
1203             fscal            = felec;
1204
1205             fscal            = _mm256_and_pd(fscal,cutoff_mask);
1206
1207             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
1208
1209             /* Calculate temporary vectorial force */
1210             tx               = _mm256_mul_pd(fscal,dx10);
1211             ty               = _mm256_mul_pd(fscal,dy10);
1212             tz               = _mm256_mul_pd(fscal,dz10);
1213
1214             /* Update vectorial force */
1215             fix1             = _mm256_add_pd(fix1,tx);
1216             fiy1             = _mm256_add_pd(fiy1,ty);
1217             fiz1             = _mm256_add_pd(fiz1,tz);
1218
1219             fjx0             = _mm256_add_pd(fjx0,tx);
1220             fjy0             = _mm256_add_pd(fjy0,ty);
1221             fjz0             = _mm256_add_pd(fjz0,tz);
1222
1223             }
1224
1225             /**************************
1226              * CALCULATE INTERACTIONS *
1227              **************************/
1228
1229             if (gmx_mm256_any_lt(rsq20,rcutoff2))
1230             {
1231
1232             /* Compute parameters for interactions between i and j atoms */
1233             qq20             = _mm256_mul_pd(iq2,jq0);
1234
1235             /* REACTION-FIELD ELECTROSTATICS */
1236             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
1237
1238             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1239
1240             fscal            = felec;
1241
1242             fscal            = _mm256_and_pd(fscal,cutoff_mask);
1243
1244             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
1245
1246             /* Calculate temporary vectorial force */
1247             tx               = _mm256_mul_pd(fscal,dx20);
1248             ty               = _mm256_mul_pd(fscal,dy20);
1249             tz               = _mm256_mul_pd(fscal,dz20);
1250
1251             /* Update vectorial force */
1252             fix2             = _mm256_add_pd(fix2,tx);
1253             fiy2             = _mm256_add_pd(fiy2,ty);
1254             fiz2             = _mm256_add_pd(fiz2,tz);
1255
1256             fjx0             = _mm256_add_pd(fjx0,tx);
1257             fjy0             = _mm256_add_pd(fjy0,ty);
1258             fjz0             = _mm256_add_pd(fjz0,tz);
1259
1260             }
1261
1262             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1263             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1264             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1265             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1266
1267             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1268
1269             /* Inner loop uses 121 flops */
1270         }
1271
1272         /* End of innermost loop */
1273
1274         gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1275                                                  f+i_coord_offset,fshift+i_shift_offset);
1276
1277         /* Increment number of inner iterations */
1278         inneriter                  += j_index_end - j_index_start;
1279
1280         /* Outer loop uses 18 flops */
1281     }
1282
1283     /* Increment number of outer iterations */
1284     outeriter        += nri;
1285
1286     /* Update outer/inner flops */
1287
1288     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*121);
1289 }