src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_VdwLJ_GeomP1P1_avx_256_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_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_ElecEw_VdwLJ_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJ_GeomP1P1_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     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  95     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  96     __m128i          ewitab;
  97     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  98     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  99     real             *ewtab;
 100     __m256d          dummy_mask,cutoff_mask;
 101     __m128           tmpmask0,tmpmask1;
 102     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 103     __m256d          one     = _mm256_set1_pd(1.0);
 104     __m256d          two     = _mm256_set1_pd(2.0);
 105     x                = xx[0];
 106     f                = ff[0];
 107
 108     nri              = nlist->nri;
 109     iinr             = nlist->iinr;
 110     jindex           = nlist->jindex;
 111     jjnr             = nlist->jjnr;
 112     shiftidx         = nlist->shift;
 113     gid              = nlist->gid;
 114     shiftvec         = fr->shift_vec[0];
 115     fshift           = fr->fshift[0];
 116     facel            = _mm256_set1_pd(fr->ic->epsfac);
 117     charge           = mdatoms->chargeA;
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 123     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 124     beta2            = _mm256_mul_pd(beta,beta);
 125     beta3            = _mm256_mul_pd(beta,beta2);
 126
 127     ewtab            = fr->ic->tabq_coul_FDV0;
 128     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 129     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 130
 131     /* Avoid stupid compiler warnings */
 132     jnrA = jnrB = jnrC = jnrD = 0;
 133     j_coord_offsetA = 0;
 134     j_coord_offsetB = 0;
 135     j_coord_offsetC = 0;
 136     j_coord_offsetD = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     for(iidx=0;iidx<4*DIM;iidx++)
 142     {
 143         scratch[iidx] = 0.0;
 144     }
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 162
 163         fix0             = _mm256_setzero_pd();
 164         fiy0             = _mm256_setzero_pd();
 165         fiz0             = _mm256_setzero_pd();
 166
 167         /* Load parameters for i particles */
 168         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 169         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 170
 171         /* Reset potential sums */
 172         velecsum         = _mm256_setzero_pd();
 173         vvdwsum          = _mm256_setzero_pd();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188
 189             /* load j atom coordinates */
 190             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 192                                                  &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm256_sub_pd(ix0,jx0);
 196             dy00             = _mm256_sub_pd(iy0,jy0);
 197             dz00             = _mm256_sub_pd(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 201
 202             rinv00           = avx256_invsqrt_d(rsq00);
 203
 204             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 208                                                                  charge+jnrC+0,charge+jnrD+0);
 209             vdwjidx0A        = 2*vdwtype[jnrA+0];
 210             vdwjidx0B        = 2*vdwtype[jnrB+0];
 211             vdwjidx0C        = 2*vdwtype[jnrC+0];
 212             vdwjidx0D        = 2*vdwtype[jnrD+0];
 213
 214             /**************************
 215              * CALCULATE INTERACTIONS *
 216              **************************/
 217
 218             r00              = _mm256_mul_pd(rsq00,rinv00);
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             qq00             = _mm256_mul_pd(iq0,jq0);
 222             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 223                                             vdwioffsetptr0+vdwjidx0B,
 224                                             vdwioffsetptr0+vdwjidx0C,
 225                                             vdwioffsetptr0+vdwjidx0D,
 226                                             &c6_00,&c12_00);
 227
 228             /* EWALD ELECTROSTATICS */
 229
 230             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 231             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 232             ewitab           = _mm256_cvttpd_epi32(ewrt);
 233             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 234             ewitab           = _mm_slli_epi32(ewitab,2);
 235             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 236             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 237             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 238             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 239             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 240             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 241             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 242             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 243             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 244
 245             /* LENNARD-JONES DISPERSION/REPULSION */
 246
 247             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 248             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 249             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 250             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 251             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 252
 253             /* Update potential sum for this i atom from the interaction with this j atom. */
 254             velecsum         = _mm256_add_pd(velecsum,velec);
 255             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 256
 257             fscal            = _mm256_add_pd(felec,fvdw);
 258
 259             /* Calculate temporary vectorial force */
 260             tx               = _mm256_mul_pd(fscal,dx00);
 261             ty               = _mm256_mul_pd(fscal,dy00);
 262             tz               = _mm256_mul_pd(fscal,dz00);
 263
 264             /* Update vectorial force */
 265             fix0             = _mm256_add_pd(fix0,tx);
 266             fiy0             = _mm256_add_pd(fiy0,ty);
 267             fiz0             = _mm256_add_pd(fiz0,tz);
 268
 269             fjptrA             = f+j_coord_offsetA;
 270             fjptrB             = f+j_coord_offsetB;
 271             fjptrC             = f+j_coord_offsetC;
 272             fjptrD             = f+j_coord_offsetD;
 273             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 274
 275             /* Inner loop uses 53 flops */
 276         }
 277
 278         if(jidx<j_index_end)
 279         {
 280
 281             /* Get j neighbor index, and coordinate index */
 282             jnrlistA         = jjnr[jidx];
 283             jnrlistB         = jjnr[jidx+1];
 284             jnrlistC         = jjnr[jidx+2];
 285             jnrlistD         = jjnr[jidx+3];
 286             /* Sign of each element will be negative for non-real atoms.
 287              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 288              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 289              */
 290             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 291
 292             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 293             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 294             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 295
 296             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 297             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 298             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 299             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 300             j_coord_offsetA  = DIM*jnrA;
 301             j_coord_offsetB  = DIM*jnrB;
 302             j_coord_offsetC  = DIM*jnrC;
 303             j_coord_offsetD  = DIM*jnrD;
 304
 305             /* load j atom coordinates */
 306             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 307                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 308                                                  &jx0,&jy0,&jz0);
 309
 310             /* Calculate displacement vector */
 311             dx00             = _mm256_sub_pd(ix0,jx0);
 312             dy00             = _mm256_sub_pd(iy0,jy0);
 313             dz00             = _mm256_sub_pd(iz0,jz0);
 314
 315             /* Calculate squared distance and things based on it */
 316             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 317
 318             rinv00           = avx256_invsqrt_d(rsq00);
 319
 320             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 321
 322             /* Load parameters for j particles */
 323             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 324                                                                  charge+jnrC+0,charge+jnrD+0);
 325             vdwjidx0A        = 2*vdwtype[jnrA+0];
 326             vdwjidx0B        = 2*vdwtype[jnrB+0];
 327             vdwjidx0C        = 2*vdwtype[jnrC+0];
 328             vdwjidx0D        = 2*vdwtype[jnrD+0];
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             r00              = _mm256_mul_pd(rsq00,rinv00);
 335             r00              = _mm256_andnot_pd(dummy_mask,r00);
 336
 337             /* Compute parameters for interactions between i and j atoms */
 338             qq00             = _mm256_mul_pd(iq0,jq0);
 339             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 340                                             vdwioffsetptr0+vdwjidx0B,
 341                                             vdwioffsetptr0+vdwjidx0C,
 342                                             vdwioffsetptr0+vdwjidx0D,
 343                                             &c6_00,&c12_00);
 344
 345             /* EWALD ELECTROSTATICS */
 346
 347             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 348             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 349             ewitab           = _mm256_cvttpd_epi32(ewrt);
 350             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 351             ewitab           = _mm_slli_epi32(ewitab,2);
 352             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 353             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 354             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 355             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 356             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 357             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 358             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 359             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 360             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 361
 362             /* LENNARD-JONES DISPERSION/REPULSION */
 363
 364             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 365             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 366             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 367             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 368             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velec            = _mm256_andnot_pd(dummy_mask,velec);
 372             velecsum         = _mm256_add_pd(velecsum,velec);
 373             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 374             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 375
 376             fscal            = _mm256_add_pd(felec,fvdw);
 377
 378             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 379
 380             /* Calculate temporary vectorial force */
 381             tx               = _mm256_mul_pd(fscal,dx00);
 382             ty               = _mm256_mul_pd(fscal,dy00);
 383             tz               = _mm256_mul_pd(fscal,dz00);
 384
 385             /* Update vectorial force */
 386             fix0             = _mm256_add_pd(fix0,tx);
 387             fiy0             = _mm256_add_pd(fiy0,ty);
 388             fiz0             = _mm256_add_pd(fiz0,tz);
 389
 390             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 391             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 392             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 393             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 394             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 395
 396             /* Inner loop uses 54 flops */
 397         }
 398
 399         /* End of innermost loop */
 400
 401         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 402                                                  f+i_coord_offset,fshift+i_shift_offset);
 403
 404         ggid                        = gid[iidx];
 405         /* Update potential energies */
 406         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 407         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 408
 409         /* Increment number of inner iterations */
 410         inneriter                  += j_index_end - j_index_start;
 411
 412         /* Outer loop uses 9 flops */
 413     }
 414
 415     /* Increment number of outer iterations */
 416     outeriter        += nri;
 417
 418     /* Update outer/inner flops */
 419
 420     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 421 }
 422 /*
 423  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 424  * Electrostatics interaction: Ewald
 425  * VdW interaction:            LennardJones
 426  * Geometry:                   Particle-Particle
 427  * Calculate force/pot:        Force
 428  */
 429 void
 430 nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 431                     (t_nblist                    * gmx_restrict       nlist,
 432                      rvec                        * gmx_restrict          xx,
 433                      rvec                        * gmx_restrict          ff,
 434                      struct t_forcerec           * gmx_restrict          fr,
 435                      t_mdatoms                   * gmx_restrict     mdatoms,
 436                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 437                      t_nrnb                      * gmx_restrict        nrnb)
 438 {
 439     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 440      * just 0 for non-waters.
 441      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 442      * jnr indices corresponding to data put in the four positions in the SIMD register.
 443      */
 444     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 445     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 446     int              jnrA,jnrB,jnrC,jnrD;
 447     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 448     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 449     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 450     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 451     real             rcutoff_scalar;
 452     real             *shiftvec,*fshift,*x,*f;
 453     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 454     real             scratch[4*DIM];
 455     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 456     real *           vdwioffsetptr0;
 457     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 458     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 459     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 460     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 461     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 462     real             *charge;
 463     int              nvdwtype;
 464     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 465     int              *vdwtype;
 466     real             *vdwparam;
 467     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 468     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 469     __m128i          ewitab;
 470     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 471     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 472     real             *ewtab;
 473     __m256d          dummy_mask,cutoff_mask;
 474     __m128           tmpmask0,tmpmask1;
 475     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 476     __m256d          one     = _mm256_set1_pd(1.0);
 477     __m256d          two     = _mm256_set1_pd(2.0);
 478     x                = xx[0];
 479     f                = ff[0];
 480
 481     nri              = nlist->nri;
 482     iinr             = nlist->iinr;
 483     jindex           = nlist->jindex;
 484     jjnr             = nlist->jjnr;
 485     shiftidx         = nlist->shift;
 486     gid              = nlist->gid;
 487     shiftvec         = fr->shift_vec[0];
 488     fshift           = fr->fshift[0];
 489     facel            = _mm256_set1_pd(fr->ic->epsfac);
 490     charge           = mdatoms->chargeA;
 491     nvdwtype         = fr->ntype;
 492     vdwparam         = fr->nbfp;
 493     vdwtype          = mdatoms->typeA;
 494
 495     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 496     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 497     beta2            = _mm256_mul_pd(beta,beta);
 498     beta3            = _mm256_mul_pd(beta,beta2);
 499
 500     ewtab            = fr->ic->tabq_coul_F;
 501     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 502     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 503
 504     /* Avoid stupid compiler warnings */
 505     jnrA = jnrB = jnrC = jnrD = 0;
 506     j_coord_offsetA = 0;
 507     j_coord_offsetB = 0;
 508     j_coord_offsetC = 0;
 509     j_coord_offsetD = 0;
 510
 511     outeriter        = 0;
 512     inneriter        = 0;
 513
 514     for(iidx=0;iidx<4*DIM;iidx++)
 515     {
 516         scratch[iidx] = 0.0;
 517     }
 518
 519     /* Start outer loop over neighborlists */
 520     for(iidx=0; iidx<nri; iidx++)
 521     {
 522         /* Load shift vector for this list */
 523         i_shift_offset   = DIM*shiftidx[iidx];
 524
 525         /* Load limits for loop over neighbors */
 526         j_index_start    = jindex[iidx];
 527         j_index_end      = jindex[iidx+1];
 528
 529         /* Get outer coordinate index */
 530         inr              = iinr[iidx];
 531         i_coord_offset   = DIM*inr;
 532
 533         /* Load i particle coords and add shift vector */
 534         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 535
 536         fix0             = _mm256_setzero_pd();
 537         fiy0             = _mm256_setzero_pd();
 538         fiz0             = _mm256_setzero_pd();
 539
 540         /* Load parameters for i particles */
 541         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 542         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 543
 544         /* Start inner kernel loop */
 545         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 546         {
 547
 548             /* Get j neighbor index, and coordinate index */
 549             jnrA             = jjnr[jidx];
 550             jnrB             = jjnr[jidx+1];
 551             jnrC             = jjnr[jidx+2];
 552             jnrD             = jjnr[jidx+3];
 553             j_coord_offsetA  = DIM*jnrA;
 554             j_coord_offsetB  = DIM*jnrB;
 555             j_coord_offsetC  = DIM*jnrC;
 556             j_coord_offsetD  = DIM*jnrD;
 557
 558             /* load j atom coordinates */
 559             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 560                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 561                                                  &jx0,&jy0,&jz0);
 562
 563             /* Calculate displacement vector */
 564             dx00             = _mm256_sub_pd(ix0,jx0);
 565             dy00             = _mm256_sub_pd(iy0,jy0);
 566             dz00             = _mm256_sub_pd(iz0,jz0);
 567
 568             /* Calculate squared distance and things based on it */
 569             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 570
 571             rinv00           = avx256_invsqrt_d(rsq00);
 572
 573             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 574
 575             /* Load parameters for j particles */
 576             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 577                                                                  charge+jnrC+0,charge+jnrD+0);
 578             vdwjidx0A        = 2*vdwtype[jnrA+0];
 579             vdwjidx0B        = 2*vdwtype[jnrB+0];
 580             vdwjidx0C        = 2*vdwtype[jnrC+0];
 581             vdwjidx0D        = 2*vdwtype[jnrD+0];
 582
 583             /**************************
 584              * CALCULATE INTERACTIONS *
 585              **************************/
 586
 587             r00              = _mm256_mul_pd(rsq00,rinv00);
 588
 589             /* Compute parameters for interactions between i and j atoms */
 590             qq00             = _mm256_mul_pd(iq0,jq0);
 591             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 592                                             vdwioffsetptr0+vdwjidx0B,
 593                                             vdwioffsetptr0+vdwjidx0C,
 594                                             vdwioffsetptr0+vdwjidx0D,
 595                                             &c6_00,&c12_00);
 596
 597             /* EWALD ELECTROSTATICS */
 598
 599             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 600             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 601             ewitab           = _mm256_cvttpd_epi32(ewrt);
 602             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 603             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 604                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 605                                             &ewtabF,&ewtabFn);
 606             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 607             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 608
 609             /* LENNARD-JONES DISPERSION/REPULSION */
 610
 611             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 612             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 613
 614             fscal            = _mm256_add_pd(felec,fvdw);
 615
 616             /* Calculate temporary vectorial force */
 617             tx               = _mm256_mul_pd(fscal,dx00);
 618             ty               = _mm256_mul_pd(fscal,dy00);
 619             tz               = _mm256_mul_pd(fscal,dz00);
 620
 621             /* Update vectorial force */
 622             fix0             = _mm256_add_pd(fix0,tx);
 623             fiy0             = _mm256_add_pd(fiy0,ty);
 624             fiz0             = _mm256_add_pd(fiz0,tz);
 625
 626             fjptrA             = f+j_coord_offsetA;
 627             fjptrB             = f+j_coord_offsetB;
 628             fjptrC             = f+j_coord_offsetC;
 629             fjptrD             = f+j_coord_offsetD;
 630             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 631
 632             /* Inner loop uses 43 flops */
 633         }
 634
 635         if(jidx<j_index_end)
 636         {
 637
 638             /* Get j neighbor index, and coordinate index */
 639             jnrlistA         = jjnr[jidx];
 640             jnrlistB         = jjnr[jidx+1];
 641             jnrlistC         = jjnr[jidx+2];
 642             jnrlistD         = jjnr[jidx+3];
 643             /* Sign of each element will be negative for non-real atoms.
 644              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 645              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 646              */
 647             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 648
 649             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 650             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 651             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 652
 653             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 654             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 655             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 656             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 657             j_coord_offsetA  = DIM*jnrA;
 658             j_coord_offsetB  = DIM*jnrB;
 659             j_coord_offsetC  = DIM*jnrC;
 660             j_coord_offsetD  = DIM*jnrD;
 661
 662             /* load j atom coordinates */
 663             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 664                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 665                                                  &jx0,&jy0,&jz0);
 666
 667             /* Calculate displacement vector */
 668             dx00             = _mm256_sub_pd(ix0,jx0);
 669             dy00             = _mm256_sub_pd(iy0,jy0);
 670             dz00             = _mm256_sub_pd(iz0,jz0);
 671
 672             /* Calculate squared distance and things based on it */
 673             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 674
 675             rinv00           = avx256_invsqrt_d(rsq00);
 676
 677             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 678
 679             /* Load parameters for j particles */
 680             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 681                                                                  charge+jnrC+0,charge+jnrD+0);
 682             vdwjidx0A        = 2*vdwtype[jnrA+0];
 683             vdwjidx0B        = 2*vdwtype[jnrB+0];
 684             vdwjidx0C        = 2*vdwtype[jnrC+0];
 685             vdwjidx0D        = 2*vdwtype[jnrD+0];
 686
 687             /**************************
 688              * CALCULATE INTERACTIONS *
 689              **************************/
 690
 691             r00              = _mm256_mul_pd(rsq00,rinv00);
 692             r00              = _mm256_andnot_pd(dummy_mask,r00);
 693
 694             /* Compute parameters for interactions between i and j atoms */
 695             qq00             = _mm256_mul_pd(iq0,jq0);
 696             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 697                                             vdwioffsetptr0+vdwjidx0B,
 698                                             vdwioffsetptr0+vdwjidx0C,
 699                                             vdwioffsetptr0+vdwjidx0D,
 700                                             &c6_00,&c12_00);
 701
 702             /* EWALD ELECTROSTATICS */
 703
 704             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 705             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 706             ewitab           = _mm256_cvttpd_epi32(ewrt);
 707             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 708             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 709                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 710                                             &ewtabF,&ewtabFn);
 711             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 712             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 713
 714             /* LENNARD-JONES DISPERSION/REPULSION */
 715
 716             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 717             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 718
 719             fscal            = _mm256_add_pd(felec,fvdw);
 720
 721             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 722
 723             /* Calculate temporary vectorial force */
 724             tx               = _mm256_mul_pd(fscal,dx00);
 725             ty               = _mm256_mul_pd(fscal,dy00);
 726             tz               = _mm256_mul_pd(fscal,dz00);
 727
 728             /* Update vectorial force */
 729             fix0             = _mm256_add_pd(fix0,tx);
 730             fiy0             = _mm256_add_pd(fiy0,ty);
 731             fiz0             = _mm256_add_pd(fiz0,tz);
 732
 733             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 734             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 735             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 736             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 737             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 738
 739             /* Inner loop uses 44 flops */
 740         }
 741
 742         /* End of innermost loop */
 743
 744         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 745                                                  f+i_coord_offset,fshift+i_shift_offset);
 746
 747         /* Increment number of inner iterations */
 748         inneriter                  += j_index_end - j_index_start;
 749
 750         /* Outer loop uses 7 flops */
 751     }
 752
 753     /* Increment number of outer iterations */
 754     outeriter        += nri;
 755
 756     /* Update outer/inner flops */
 757
 758     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*44);
 759 }