src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecNone_VdwLJEw_GeomP1P1_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_ElecNone_VdwLJEw_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: None
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecNone_VdwLJEw_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     real *           vdwgridioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     int              nvdwtype;
  90     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  94     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  95     __m256d           c6grid_00;
  96     real             *vdwgridparam;
  97     __m256d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  98     __m256d           one_half  = _mm256_set1_pd(0.5);
  99     __m256d           minus_one = _mm256_set1_pd(-1.0);
 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     nvdwtype         = fr->ntype;
 117     vdwparam         = fr->nbfp;
 118     vdwtype          = mdatoms->typeA;
 119     vdwgridparam     = fr->ljpme_c6grid;
 120     sh_lj_ewald      = _mm256_set1_pd(fr->ic->sh_lj_ewald);
 121     ewclj            = _mm256_set1_pd(fr->ic->ewaldcoeff_lj);
 122     ewclj2           = _mm256_mul_pd(minus_one,_mm256_mul_pd(ewclj,ewclj));
 123
 124     /* Avoid stupid compiler warnings */
 125     jnrA = jnrB = jnrC = jnrD = 0;
 126     j_coord_offsetA = 0;
 127     j_coord_offsetB = 0;
 128     j_coord_offsetC = 0;
 129     j_coord_offsetD = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm256_setzero_pd();
 157         fiy0             = _mm256_setzero_pd();
 158         fiz0             = _mm256_setzero_pd();
 159
 160         /* Load parameters for i particles */
 161         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 162         vdwgridioffsetptr0 = vdwgridparam+2*nvdwtype*vdwtype[inr+0];
 163
 164         /* Reset potential sums */
 165         vvdwsum          = _mm256_setzero_pd();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178             j_coord_offsetC  = DIM*jnrC;
 179             j_coord_offsetD  = DIM*jnrD;
 180
 181             /* load j atom coordinates */
 182             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 183                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 184                                                  &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm256_sub_pd(ix0,jx0);
 188             dy00             = _mm256_sub_pd(iy0,jy0);
 189             dz00             = _mm256_sub_pd(iz0,jz0);
 190
 191             /* Calculate squared distance and things based on it */
 192             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 193
 194             rinv00           = avx256_invsqrt_d(rsq00);
 195
 196             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 197
 198             /* Load parameters for j particles */
 199             vdwjidx0A        = 2*vdwtype[jnrA+0];
 200             vdwjidx0B        = 2*vdwtype[jnrB+0];
 201             vdwjidx0C        = 2*vdwtype[jnrC+0];
 202             vdwjidx0D        = 2*vdwtype[jnrD+0];
 203
 204             /**************************
 205              * CALCULATE INTERACTIONS *
 206              **************************/
 207
 208             r00              = _mm256_mul_pd(rsq00,rinv00);
 209
 210             /* Compute parameters for interactions between i and j atoms */
 211             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 212                                             vdwioffsetptr0+vdwjidx0B,
 213                                             vdwioffsetptr0+vdwjidx0C,
 214                                             vdwioffsetptr0+vdwjidx0D,
 215                                             &c6_00,&c12_00);
 216
 217             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 218                                                                   vdwgridioffsetptr0+vdwjidx0B,
 219                                                                   vdwgridioffsetptr0+vdwjidx0C,
 220                                                                   vdwgridioffsetptr0+vdwjidx0D);
 221
 222             /* Analytical LJ-PME */
 223             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 224             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 225             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 226             exponent         = avx256_exp_d(ewcljrsq);
 227             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 228             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 229             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 230             vvdw6            = _mm256_mul_pd(_mm256_sub_pd(c6_00,_mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly))),rinvsix);
 231             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 232             vvdw             = _mm256_sub_pd(_mm256_mul_pd(vvdw12,one_twelfth),_mm256_mul_pd(vvdw6,one_sixth));
 233             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 234             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,_mm256_sub_pd(vvdw6,_mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6)))),rinvsq00);
 235
 236             /* Update potential sum for this i atom from the interaction with this j atom. */
 237             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 238
 239             fscal            = fvdw;
 240
 241             /* Calculate temporary vectorial force */
 242             tx               = _mm256_mul_pd(fscal,dx00);
 243             ty               = _mm256_mul_pd(fscal,dy00);
 244             tz               = _mm256_mul_pd(fscal,dz00);
 245
 246             /* Update vectorial force */
 247             fix0             = _mm256_add_pd(fix0,tx);
 248             fiy0             = _mm256_add_pd(fiy0,ty);
 249             fiz0             = _mm256_add_pd(fiz0,tz);
 250
 251             fjptrA             = f+j_coord_offsetA;
 252             fjptrB             = f+j_coord_offsetB;
 253             fjptrC             = f+j_coord_offsetC;
 254             fjptrD             = f+j_coord_offsetD;
 255             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 256
 257             /* Inner loop uses 54 flops */
 258         }
 259
 260         if(jidx<j_index_end)
 261         {
 262
 263             /* Get j neighbor index, and coordinate index */
 264             jnrlistA         = jjnr[jidx];
 265             jnrlistB         = jjnr[jidx+1];
 266             jnrlistC         = jjnr[jidx+2];
 267             jnrlistD         = jjnr[jidx+3];
 268             /* Sign of each element will be negative for non-real atoms.
 269              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 270              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 271              */
 272             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 273
 274             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 275             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 276             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 277
 278             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 279             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 280             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 281             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 282             j_coord_offsetA  = DIM*jnrA;
 283             j_coord_offsetB  = DIM*jnrB;
 284             j_coord_offsetC  = DIM*jnrC;
 285             j_coord_offsetD  = DIM*jnrD;
 286
 287             /* load j atom coordinates */
 288             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 289                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 290                                                  &jx0,&jy0,&jz0);
 291
 292             /* Calculate displacement vector */
 293             dx00             = _mm256_sub_pd(ix0,jx0);
 294             dy00             = _mm256_sub_pd(iy0,jy0);
 295             dz00             = _mm256_sub_pd(iz0,jz0);
 296
 297             /* Calculate squared distance and things based on it */
 298             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 299
 300             rinv00           = avx256_invsqrt_d(rsq00);
 301
 302             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 303
 304             /* Load parameters for j particles */
 305             vdwjidx0A        = 2*vdwtype[jnrA+0];
 306             vdwjidx0B        = 2*vdwtype[jnrB+0];
 307             vdwjidx0C        = 2*vdwtype[jnrC+0];
 308             vdwjidx0D        = 2*vdwtype[jnrD+0];
 309
 310             /**************************
 311              * CALCULATE INTERACTIONS *
 312              **************************/
 313
 314             r00              = _mm256_mul_pd(rsq00,rinv00);
 315             r00              = _mm256_andnot_pd(dummy_mask,r00);
 316
 317             /* Compute parameters for interactions between i and j atoms */
 318             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 319                                             vdwioffsetptr0+vdwjidx0B,
 320                                             vdwioffsetptr0+vdwjidx0C,
 321                                             vdwioffsetptr0+vdwjidx0D,
 322                                             &c6_00,&c12_00);
 323
 324             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 325                                                                   vdwgridioffsetptr0+vdwjidx0B,
 326                                                                   vdwgridioffsetptr0+vdwjidx0C,
 327                                                                   vdwgridioffsetptr0+vdwjidx0D);
 328
 329             /* Analytical LJ-PME */
 330             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 331             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 332             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 333             exponent         = avx256_exp_d(ewcljrsq);
 334             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 335             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 336             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 337             vvdw6            = _mm256_mul_pd(_mm256_sub_pd(c6_00,_mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly))),rinvsix);
 338             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 339             vvdw             = _mm256_sub_pd(_mm256_mul_pd(vvdw12,one_twelfth),_mm256_mul_pd(vvdw6,one_sixth));
 340             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 341             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,_mm256_sub_pd(vvdw6,_mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6)))),rinvsq00);
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 345             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 346
 347             fscal            = fvdw;
 348
 349             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 350
 351             /* Calculate temporary vectorial force */
 352             tx               = _mm256_mul_pd(fscal,dx00);
 353             ty               = _mm256_mul_pd(fscal,dy00);
 354             tz               = _mm256_mul_pd(fscal,dz00);
 355
 356             /* Update vectorial force */
 357             fix0             = _mm256_add_pd(fix0,tx);
 358             fiy0             = _mm256_add_pd(fiy0,ty);
 359             fiz0             = _mm256_add_pd(fiz0,tz);
 360
 361             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 362             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 363             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 364             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 365             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 366
 367             /* Inner loop uses 55 flops */
 368         }
 369
 370         /* End of innermost loop */
 371
 372         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 373                                                  f+i_coord_offset,fshift+i_shift_offset);
 374
 375         ggid                        = gid[iidx];
 376         /* Update potential energies */
 377         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 378
 379         /* Increment number of inner iterations */
 380         inneriter                  += j_index_end - j_index_start;
 381
 382         /* Outer loop uses 7 flops */
 383     }
 384
 385     /* Increment number of outer iterations */
 386     outeriter        += nri;
 387
 388     /* Update outer/inner flops */
 389
 390     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*55);
 391 }
 392 /*
 393  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_256_double
 394  * Electrostatics interaction: None
 395  * VdW interaction:            LJEwald
 396  * Geometry:                   Particle-Particle
 397  * Calculate force/pot:        Force
 398  */
 399 void
 400 nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_256_double
 401                     (t_nblist                    * gmx_restrict       nlist,
 402                      rvec                        * gmx_restrict          xx,
 403                      rvec                        * gmx_restrict          ff,
 404                      struct t_forcerec           * gmx_restrict          fr,
 405                      t_mdatoms                   * gmx_restrict     mdatoms,
 406                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 407                      t_nrnb                      * gmx_restrict        nrnb)
 408 {
 409     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 410      * just 0 for non-waters.
 411      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 412      * jnr indices corresponding to data put in the four positions in the SIMD register.
 413      */
 414     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 415     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 416     int              jnrA,jnrB,jnrC,jnrD;
 417     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 418     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 419     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 420     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 421     real             rcutoff_scalar;
 422     real             *shiftvec,*fshift,*x,*f;
 423     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 424     real             scratch[4*DIM];
 425     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 426     real *           vdwioffsetptr0;
 427     real *           vdwgridioffsetptr0;
 428     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 429     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 430     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 431     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 432     int              nvdwtype;
 433     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 434     int              *vdwtype;
 435     real             *vdwparam;
 436     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 437     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 438     __m256d           c6grid_00;
 439     real             *vdwgridparam;
 440     __m256d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 441     __m256d           one_half  = _mm256_set1_pd(0.5);
 442     __m256d           minus_one = _mm256_set1_pd(-1.0);
 443     __m256d          dummy_mask,cutoff_mask;
 444     __m128           tmpmask0,tmpmask1;
 445     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 446     __m256d          one     = _mm256_set1_pd(1.0);
 447     __m256d          two     = _mm256_set1_pd(2.0);
 448     x                = xx[0];
 449     f                = ff[0];
 450
 451     nri              = nlist->nri;
 452     iinr             = nlist->iinr;
 453     jindex           = nlist->jindex;
 454     jjnr             = nlist->jjnr;
 455     shiftidx         = nlist->shift;
 456     gid              = nlist->gid;
 457     shiftvec         = fr->shift_vec[0];
 458     fshift           = fr->fshift[0];
 459     nvdwtype         = fr->ntype;
 460     vdwparam         = fr->nbfp;
 461     vdwtype          = mdatoms->typeA;
 462     vdwgridparam     = fr->ljpme_c6grid;
 463     sh_lj_ewald      = _mm256_set1_pd(fr->ic->sh_lj_ewald);
 464     ewclj            = _mm256_set1_pd(fr->ic->ewaldcoeff_lj);
 465     ewclj2           = _mm256_mul_pd(minus_one,_mm256_mul_pd(ewclj,ewclj));
 466
 467     /* Avoid stupid compiler warnings */
 468     jnrA = jnrB = jnrC = jnrD = 0;
 469     j_coord_offsetA = 0;
 470     j_coord_offsetB = 0;
 471     j_coord_offsetC = 0;
 472     j_coord_offsetD = 0;
 473
 474     outeriter        = 0;
 475     inneriter        = 0;
 476
 477     for(iidx=0;iidx<4*DIM;iidx++)
 478     {
 479         scratch[iidx] = 0.0;
 480     }
 481
 482     /* Start outer loop over neighborlists */
 483     for(iidx=0; iidx<nri; iidx++)
 484     {
 485         /* Load shift vector for this list */
 486         i_shift_offset   = DIM*shiftidx[iidx];
 487
 488         /* Load limits for loop over neighbors */
 489         j_index_start    = jindex[iidx];
 490         j_index_end      = jindex[iidx+1];
 491
 492         /* Get outer coordinate index */
 493         inr              = iinr[iidx];
 494         i_coord_offset   = DIM*inr;
 495
 496         /* Load i particle coords and add shift vector */
 497         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 498
 499         fix0             = _mm256_setzero_pd();
 500         fiy0             = _mm256_setzero_pd();
 501         fiz0             = _mm256_setzero_pd();
 502
 503         /* Load parameters for i particles */
 504         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 505         vdwgridioffsetptr0 = vdwgridparam+2*nvdwtype*vdwtype[inr+0];
 506
 507         /* Start inner kernel loop */
 508         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 509         {
 510
 511             /* Get j neighbor index, and coordinate index */
 512             jnrA             = jjnr[jidx];
 513             jnrB             = jjnr[jidx+1];
 514             jnrC             = jjnr[jidx+2];
 515             jnrD             = jjnr[jidx+3];
 516             j_coord_offsetA  = DIM*jnrA;
 517             j_coord_offsetB  = DIM*jnrB;
 518             j_coord_offsetC  = DIM*jnrC;
 519             j_coord_offsetD  = DIM*jnrD;
 520
 521             /* load j atom coordinates */
 522             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 523                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 524                                                  &jx0,&jy0,&jz0);
 525
 526             /* Calculate displacement vector */
 527             dx00             = _mm256_sub_pd(ix0,jx0);
 528             dy00             = _mm256_sub_pd(iy0,jy0);
 529             dz00             = _mm256_sub_pd(iz0,jz0);
 530
 531             /* Calculate squared distance and things based on it */
 532             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 533
 534             rinv00           = avx256_invsqrt_d(rsq00);
 535
 536             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 537
 538             /* Load parameters for j particles */
 539             vdwjidx0A        = 2*vdwtype[jnrA+0];
 540             vdwjidx0B        = 2*vdwtype[jnrB+0];
 541             vdwjidx0C        = 2*vdwtype[jnrC+0];
 542             vdwjidx0D        = 2*vdwtype[jnrD+0];
 543
 544             /**************************
 545              * CALCULATE INTERACTIONS *
 546              **************************/
 547
 548             r00              = _mm256_mul_pd(rsq00,rinv00);
 549
 550             /* Compute parameters for interactions between i and j atoms */
 551             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 552                                             vdwioffsetptr0+vdwjidx0B,
 553                                             vdwioffsetptr0+vdwjidx0C,
 554                                             vdwioffsetptr0+vdwjidx0D,
 555                                             &c6_00,&c12_00);
 556
 557             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 558                                                                   vdwgridioffsetptr0+vdwjidx0B,
 559                                                                   vdwgridioffsetptr0+vdwjidx0C,
 560                                                                   vdwgridioffsetptr0+vdwjidx0D);
 561
 562             /* Analytical LJ-PME */
 563             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 564             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 565             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 566             exponent         = avx256_exp_d(ewcljrsq);
 567             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 568             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 569             /* f6A = 6 * C6grid * (1 - poly) */
 570             f6A              = _mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly));
 571             /* f6B = C6grid * exponent * beta^6 */
 572             f6B              = _mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6));
 573             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 574             fvdw              = _mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),_mm256_sub_pd(c6_00,f6A)),rinvsix),f6B),rinvsq00);
 575
 576             fscal            = fvdw;
 577
 578             /* Calculate temporary vectorial force */
 579             tx               = _mm256_mul_pd(fscal,dx00);
 580             ty               = _mm256_mul_pd(fscal,dy00);
 581             tz               = _mm256_mul_pd(fscal,dz00);
 582
 583             /* Update vectorial force */
 584             fix0             = _mm256_add_pd(fix0,tx);
 585             fiy0             = _mm256_add_pd(fiy0,ty);
 586             fiz0             = _mm256_add_pd(fiz0,tz);
 587
 588             fjptrA             = f+j_coord_offsetA;
 589             fjptrB             = f+j_coord_offsetB;
 590             fjptrC             = f+j_coord_offsetC;
 591             fjptrD             = f+j_coord_offsetD;
 592             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 593
 594             /* Inner loop uses 46 flops */
 595         }
 596
 597         if(jidx<j_index_end)
 598         {
 599
 600             /* Get j neighbor index, and coordinate index */
 601             jnrlistA         = jjnr[jidx];
 602             jnrlistB         = jjnr[jidx+1];
 603             jnrlistC         = jjnr[jidx+2];
 604             jnrlistD         = jjnr[jidx+3];
 605             /* Sign of each element will be negative for non-real atoms.
 606              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 607              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 608              */
 609             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 610
 611             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 612             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 613             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 614
 615             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 616             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 617             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 618             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 619             j_coord_offsetA  = DIM*jnrA;
 620             j_coord_offsetB  = DIM*jnrB;
 621             j_coord_offsetC  = DIM*jnrC;
 622             j_coord_offsetD  = DIM*jnrD;
 623
 624             /* load j atom coordinates */
 625             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 626                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 627                                                  &jx0,&jy0,&jz0);
 628
 629             /* Calculate displacement vector */
 630             dx00             = _mm256_sub_pd(ix0,jx0);
 631             dy00             = _mm256_sub_pd(iy0,jy0);
 632             dz00             = _mm256_sub_pd(iz0,jz0);
 633
 634             /* Calculate squared distance and things based on it */
 635             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 636
 637             rinv00           = avx256_invsqrt_d(rsq00);
 638
 639             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 640
 641             /* Load parameters for j particles */
 642             vdwjidx0A        = 2*vdwtype[jnrA+0];
 643             vdwjidx0B        = 2*vdwtype[jnrB+0];
 644             vdwjidx0C        = 2*vdwtype[jnrC+0];
 645             vdwjidx0D        = 2*vdwtype[jnrD+0];
 646
 647             /**************************
 648              * CALCULATE INTERACTIONS *
 649              **************************/
 650
 651             r00              = _mm256_mul_pd(rsq00,rinv00);
 652             r00              = _mm256_andnot_pd(dummy_mask,r00);
 653
 654             /* Compute parameters for interactions between i and j atoms */
 655             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 656                                             vdwioffsetptr0+vdwjidx0B,
 657                                             vdwioffsetptr0+vdwjidx0C,
 658                                             vdwioffsetptr0+vdwjidx0D,
 659                                             &c6_00,&c12_00);
 660
 661             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 662                                                                   vdwgridioffsetptr0+vdwjidx0B,
 663                                                                   vdwgridioffsetptr0+vdwjidx0C,
 664                                                                   vdwgridioffsetptr0+vdwjidx0D);
 665
 666             /* Analytical LJ-PME */
 667             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 668             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 669             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 670             exponent         = avx256_exp_d(ewcljrsq);
 671             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 672             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 673             /* f6A = 6 * C6grid * (1 - poly) */
 674             f6A              = _mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly));
 675             /* f6B = C6grid * exponent * beta^6 */
 676             f6B              = _mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6));
 677             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 678             fvdw              = _mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),_mm256_sub_pd(c6_00,f6A)),rinvsix),f6B),rinvsq00);
 679
 680             fscal            = fvdw;
 681
 682             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 683
 684             /* Calculate temporary vectorial force */
 685             tx               = _mm256_mul_pd(fscal,dx00);
 686             ty               = _mm256_mul_pd(fscal,dy00);
 687             tz               = _mm256_mul_pd(fscal,dz00);
 688
 689             /* Update vectorial force */
 690             fix0             = _mm256_add_pd(fix0,tx);
 691             fiy0             = _mm256_add_pd(fiy0,ty);
 692             fiz0             = _mm256_add_pd(fiz0,tz);
 693
 694             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 695             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 696             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 697             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 698             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 699
 700             /* Inner loop uses 47 flops */
 701         }
 702
 703         /* End of innermost loop */
 704
 705         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 706                                                  f+i_coord_offset,fshift+i_shift_offset);
 707
 708         /* Increment number of inner iterations */
 709         inneriter                  += j_index_end - j_index_start;
 710
 711         /* Outer loop uses 6 flops */
 712     }
 713
 714     /* Increment number of outer iterations */
 715     outeriter        += nri;
 716
 717     /* Update outer/inner flops */
 718
 719     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*47);
 720 }