src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecNone_VdwLJEw_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEw_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: None
  55  * VdW interaction:            LJEwald
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecNone_VdwLJEw_GeomP1P1_VF_avx_256_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  79     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  84     real             scratch[4*DIM];
  85     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     real *           vdwgridioffsetptr0;
  88     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  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     int              nvdwtype;
  93     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  97     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  98     __m256d           c6grid_00;
  99     real             *vdwgridparam;
 100     __m256d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 101     __m256d           one_half  = _mm256_set1_pd(0.5);
 102     __m256d           minus_one = _mm256_set1_pd(-1.0);
 103     __m256d          dummy_mask,cutoff_mask;
 104     __m128           tmpmask0,tmpmask1;
 105     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 106     __m256d          one     = _mm256_set1_pd(1.0);
 107     __m256d          two     = _mm256_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122     vdwgridparam     = fr->ljpme_c6grid;
 123     sh_lj_ewald      = _mm256_set1_pd(fr->ic->sh_lj_ewald);
 124     ewclj            = _mm256_set1_pd(fr->ewaldcoeff_lj);
 125     ewclj2           = _mm256_mul_pd(minus_one,_mm256_mul_pd(ewclj,ewclj));
 126
 127     /* Avoid stupid compiler warnings */
 128     jnrA = jnrB = jnrC = jnrD = 0;
 129     j_coord_offsetA = 0;
 130     j_coord_offsetB = 0;
 131     j_coord_offsetC = 0;
 132     j_coord_offsetD = 0;
 133
 134     outeriter        = 0;
 135     inneriter        = 0;
 136
 137     for(iidx=0;iidx<4*DIM;iidx++)
 138     {
 139         scratch[iidx] = 0.0;
 140     }
 141
 142     /* Start outer loop over neighborlists */
 143     for(iidx=0; iidx<nri; iidx++)
 144     {
 145         /* Load shift vector for this list */
 146         i_shift_offset   = DIM*shiftidx[iidx];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 158
 159         fix0             = _mm256_setzero_pd();
 160         fiy0             = _mm256_setzero_pd();
 161         fiz0             = _mm256_setzero_pd();
 162
 163         /* Load parameters for i particles */
 164         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 165         vdwgridioffsetptr0 = vdwgridparam+2*nvdwtype*vdwtype[inr+0];
 166
 167         /* Reset potential sums */
 168         vvdwsum          = _mm256_setzero_pd();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             jnrC             = jjnr[jidx+2];
 178             jnrD             = jjnr[jidx+3];
 179             j_coord_offsetA  = DIM*jnrA;
 180             j_coord_offsetB  = DIM*jnrB;
 181             j_coord_offsetC  = DIM*jnrC;
 182             j_coord_offsetD  = DIM*jnrD;
 183
 184             /* load j atom coordinates */
 185             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 186                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 187                                                  &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _mm256_sub_pd(ix0,jx0);
 191             dy00             = _mm256_sub_pd(iy0,jy0);
 192             dz00             = _mm256_sub_pd(iz0,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 196
 197             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 198
 199             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 200
 201             /* Load parameters for j particles */
 202             vdwjidx0A        = 2*vdwtype[jnrA+0];
 203             vdwjidx0B        = 2*vdwtype[jnrB+0];
 204             vdwjidx0C        = 2*vdwtype[jnrC+0];
 205             vdwjidx0D        = 2*vdwtype[jnrD+0];
 206
 207             /**************************
 208              * CALCULATE INTERACTIONS *
 209              **************************/
 210
 211             r00              = _mm256_mul_pd(rsq00,rinv00);
 212
 213             /* Compute parameters for interactions between i and j atoms */
 214             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 215                                             vdwioffsetptr0+vdwjidx0B,
 216                                             vdwioffsetptr0+vdwjidx0C,
 217                                             vdwioffsetptr0+vdwjidx0D,
 218                                             &c6_00,&c12_00);
 219
 220             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 221                                                                   vdwgridioffsetptr0+vdwjidx0B,
 222                                                                   vdwgridioffsetptr0+vdwjidx0C,
 223                                                                   vdwgridioffsetptr0+vdwjidx0D);
 224
 225             /* Analytical LJ-PME */
 226             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 227             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 228             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 229             exponent         = gmx_simd_exp_d(ewcljrsq);
 230             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 231             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 232             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 233             vvdw6            = _mm256_mul_pd(_mm256_sub_pd(c6_00,_mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly))),rinvsix);
 234             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 235             vvdw             = _mm256_sub_pd(_mm256_mul_pd(vvdw12,one_twelfth),_mm256_mul_pd(vvdw6,one_sixth));
 236             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 237             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);
 238
 239             /* Update potential sum for this i atom from the interaction with this j atom. */
 240             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 241
 242             fscal            = fvdw;
 243
 244             /* Calculate temporary vectorial force */
 245             tx               = _mm256_mul_pd(fscal,dx00);
 246             ty               = _mm256_mul_pd(fscal,dy00);
 247             tz               = _mm256_mul_pd(fscal,dz00);
 248
 249             /* Update vectorial force */
 250             fix0             = _mm256_add_pd(fix0,tx);
 251             fiy0             = _mm256_add_pd(fiy0,ty);
 252             fiz0             = _mm256_add_pd(fiz0,tz);
 253
 254             fjptrA             = f+j_coord_offsetA;
 255             fjptrB             = f+j_coord_offsetB;
 256             fjptrC             = f+j_coord_offsetC;
 257             fjptrD             = f+j_coord_offsetD;
 258             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 259
 260             /* Inner loop uses 54 flops */
 261         }
 262
 263         if(jidx<j_index_end)
 264         {
 265
 266             /* Get j neighbor index, and coordinate index */
 267             jnrlistA         = jjnr[jidx];
 268             jnrlistB         = jjnr[jidx+1];
 269             jnrlistC         = jjnr[jidx+2];
 270             jnrlistD         = jjnr[jidx+3];
 271             /* Sign of each element will be negative for non-real atoms.
 272              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 273              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 274              */
 275             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 276
 277             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 278             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 279             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 280
 281             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 282             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 283             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 284             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 285             j_coord_offsetA  = DIM*jnrA;
 286             j_coord_offsetB  = DIM*jnrB;
 287             j_coord_offsetC  = DIM*jnrC;
 288             j_coord_offsetD  = DIM*jnrD;
 289
 290             /* load j atom coordinates */
 291             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 292                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 293                                                  &jx0,&jy0,&jz0);
 294
 295             /* Calculate displacement vector */
 296             dx00             = _mm256_sub_pd(ix0,jx0);
 297             dy00             = _mm256_sub_pd(iy0,jy0);
 298             dz00             = _mm256_sub_pd(iz0,jz0);
 299
 300             /* Calculate squared distance and things based on it */
 301             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 302
 303             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 304
 305             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 306
 307             /* Load parameters for j particles */
 308             vdwjidx0A        = 2*vdwtype[jnrA+0];
 309             vdwjidx0B        = 2*vdwtype[jnrB+0];
 310             vdwjidx0C        = 2*vdwtype[jnrC+0];
 311             vdwjidx0D        = 2*vdwtype[jnrD+0];
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             r00              = _mm256_mul_pd(rsq00,rinv00);
 318             r00              = _mm256_andnot_pd(dummy_mask,r00);
 319
 320             /* Compute parameters for interactions between i and j atoms */
 321             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 322                                             vdwioffsetptr0+vdwjidx0B,
 323                                             vdwioffsetptr0+vdwjidx0C,
 324                                             vdwioffsetptr0+vdwjidx0D,
 325                                             &c6_00,&c12_00);
 326
 327             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 328                                                                   vdwgridioffsetptr0+vdwjidx0B,
 329                                                                   vdwgridioffsetptr0+vdwjidx0C,
 330                                                                   vdwgridioffsetptr0+vdwjidx0D);
 331
 332             /* Analytical LJ-PME */
 333             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 334             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 335             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 336             exponent         = gmx_simd_exp_d(ewcljrsq);
 337             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 338             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 339             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 340             vvdw6            = _mm256_mul_pd(_mm256_sub_pd(c6_00,_mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly))),rinvsix);
 341             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 342             vvdw             = _mm256_sub_pd(_mm256_mul_pd(vvdw12,one_twelfth),_mm256_mul_pd(vvdw6,one_sixth));
 343             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 344             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);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 348             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 349
 350             fscal            = fvdw;
 351
 352             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 353
 354             /* Calculate temporary vectorial force */
 355             tx               = _mm256_mul_pd(fscal,dx00);
 356             ty               = _mm256_mul_pd(fscal,dy00);
 357             tz               = _mm256_mul_pd(fscal,dz00);
 358
 359             /* Update vectorial force */
 360             fix0             = _mm256_add_pd(fix0,tx);
 361             fiy0             = _mm256_add_pd(fiy0,ty);
 362             fiz0             = _mm256_add_pd(fiz0,tz);
 363
 364             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 365             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 366             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 367             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 368             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 369
 370             /* Inner loop uses 55 flops */
 371         }
 372
 373         /* End of innermost loop */
 374
 375         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 376                                                  f+i_coord_offset,fshift+i_shift_offset);
 377
 378         ggid                        = gid[iidx];
 379         /* Update potential energies */
 380         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 381
 382         /* Increment number of inner iterations */
 383         inneriter                  += j_index_end - j_index_start;
 384
 385         /* Outer loop uses 7 flops */
 386     }
 387
 388     /* Increment number of outer iterations */
 389     outeriter        += nri;
 390
 391     /* Update outer/inner flops */
 392
 393     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*55);
 394 }
 395 /*
 396  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_256_double
 397  * Electrostatics interaction: None
 398  * VdW interaction:            LJEwald
 399  * Geometry:                   Particle-Particle
 400  * Calculate force/pot:        Force
 401  */
 402 void
 403 nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_256_double
 404                     (t_nblist                    * gmx_restrict       nlist,
 405                      rvec                        * gmx_restrict          xx,
 406                      rvec                        * gmx_restrict          ff,
 407                      t_forcerec                  * gmx_restrict          fr,
 408                      t_mdatoms                   * gmx_restrict     mdatoms,
 409                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 410                      t_nrnb                      * gmx_restrict        nrnb)
 411 {
 412     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 413      * just 0 for non-waters.
 414      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 415      * jnr indices corresponding to data put in the four positions in the SIMD register.
 416      */
 417     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 418     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 419     int              jnrA,jnrB,jnrC,jnrD;
 420     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 421     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 422     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 423     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 424     real             rcutoff_scalar;
 425     real             *shiftvec,*fshift,*x,*f;
 426     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 427     real             scratch[4*DIM];
 428     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 429     real *           vdwioffsetptr0;
 430     real *           vdwgridioffsetptr0;
 431     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 432     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 433     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 434     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 435     int              nvdwtype;
 436     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 437     int              *vdwtype;
 438     real             *vdwparam;
 439     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 440     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 441     __m256d           c6grid_00;
 442     real             *vdwgridparam;
 443     __m256d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 444     __m256d           one_half  = _mm256_set1_pd(0.5);
 445     __m256d           minus_one = _mm256_set1_pd(-1.0);
 446     __m256d          dummy_mask,cutoff_mask;
 447     __m128           tmpmask0,tmpmask1;
 448     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 449     __m256d          one     = _mm256_set1_pd(1.0);
 450     __m256d          two     = _mm256_set1_pd(2.0);
 451     x                = xx[0];
 452     f                = ff[0];
 453
 454     nri              = nlist->nri;
 455     iinr             = nlist->iinr;
 456     jindex           = nlist->jindex;
 457     jjnr             = nlist->jjnr;
 458     shiftidx         = nlist->shift;
 459     gid              = nlist->gid;
 460     shiftvec         = fr->shift_vec[0];
 461     fshift           = fr->fshift[0];
 462     nvdwtype         = fr->ntype;
 463     vdwparam         = fr->nbfp;
 464     vdwtype          = mdatoms->typeA;
 465     vdwgridparam     = fr->ljpme_c6grid;
 466     sh_lj_ewald      = _mm256_set1_pd(fr->ic->sh_lj_ewald);
 467     ewclj            = _mm256_set1_pd(fr->ewaldcoeff_lj);
 468     ewclj2           = _mm256_mul_pd(minus_one,_mm256_mul_pd(ewclj,ewclj));
 469
 470     /* Avoid stupid compiler warnings */
 471     jnrA = jnrB = jnrC = jnrD = 0;
 472     j_coord_offsetA = 0;
 473     j_coord_offsetB = 0;
 474     j_coord_offsetC = 0;
 475     j_coord_offsetD = 0;
 476
 477     outeriter        = 0;
 478     inneriter        = 0;
 479
 480     for(iidx=0;iidx<4*DIM;iidx++)
 481     {
 482         scratch[iidx] = 0.0;
 483     }
 484
 485     /* Start outer loop over neighborlists */
 486     for(iidx=0; iidx<nri; iidx++)
 487     {
 488         /* Load shift vector for this list */
 489         i_shift_offset   = DIM*shiftidx[iidx];
 490
 491         /* Load limits for loop over neighbors */
 492         j_index_start    = jindex[iidx];
 493         j_index_end      = jindex[iidx+1];
 494
 495         /* Get outer coordinate index */
 496         inr              = iinr[iidx];
 497         i_coord_offset   = DIM*inr;
 498
 499         /* Load i particle coords and add shift vector */
 500         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 501
 502         fix0             = _mm256_setzero_pd();
 503         fiy0             = _mm256_setzero_pd();
 504         fiz0             = _mm256_setzero_pd();
 505
 506         /* Load parameters for i particles */
 507         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 508         vdwgridioffsetptr0 = vdwgridparam+2*nvdwtype*vdwtype[inr+0];
 509
 510         /* Start inner kernel loop */
 511         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 512         {
 513
 514             /* Get j neighbor index, and coordinate index */
 515             jnrA             = jjnr[jidx];
 516             jnrB             = jjnr[jidx+1];
 517             jnrC             = jjnr[jidx+2];
 518             jnrD             = jjnr[jidx+3];
 519             j_coord_offsetA  = DIM*jnrA;
 520             j_coord_offsetB  = DIM*jnrB;
 521             j_coord_offsetC  = DIM*jnrC;
 522             j_coord_offsetD  = DIM*jnrD;
 523
 524             /* load j atom coordinates */
 525             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 526                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 527                                                  &jx0,&jy0,&jz0);
 528
 529             /* Calculate displacement vector */
 530             dx00             = _mm256_sub_pd(ix0,jx0);
 531             dy00             = _mm256_sub_pd(iy0,jy0);
 532             dz00             = _mm256_sub_pd(iz0,jz0);
 533
 534             /* Calculate squared distance and things based on it */
 535             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 536
 537             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 538
 539             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 540
 541             /* Load parameters for j particles */
 542             vdwjidx0A        = 2*vdwtype[jnrA+0];
 543             vdwjidx0B        = 2*vdwtype[jnrB+0];
 544             vdwjidx0C        = 2*vdwtype[jnrC+0];
 545             vdwjidx0D        = 2*vdwtype[jnrD+0];
 546
 547             /**************************
 548              * CALCULATE INTERACTIONS *
 549              **************************/
 550
 551             r00              = _mm256_mul_pd(rsq00,rinv00);
 552
 553             /* Compute parameters for interactions between i and j atoms */
 554             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 555                                             vdwioffsetptr0+vdwjidx0B,
 556                                             vdwioffsetptr0+vdwjidx0C,
 557                                             vdwioffsetptr0+vdwjidx0D,
 558                                             &c6_00,&c12_00);
 559
 560             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 561                                                                   vdwgridioffsetptr0+vdwjidx0B,
 562                                                                   vdwgridioffsetptr0+vdwjidx0C,
 563                                                                   vdwgridioffsetptr0+vdwjidx0D);
 564
 565             /* Analytical LJ-PME */
 566             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 567             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 568             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 569             exponent         = gmx_simd_exp_d(ewcljrsq);
 570             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 571             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 572             /* f6A = 6 * C6grid * (1 - poly) */
 573             f6A              = _mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly));
 574             /* f6B = C6grid * exponent * beta^6 */
 575             f6B              = _mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6));
 576             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 577             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);
 578
 579             fscal            = fvdw;
 580
 581             /* Calculate temporary vectorial force */
 582             tx               = _mm256_mul_pd(fscal,dx00);
 583             ty               = _mm256_mul_pd(fscal,dy00);
 584             tz               = _mm256_mul_pd(fscal,dz00);
 585
 586             /* Update vectorial force */
 587             fix0             = _mm256_add_pd(fix0,tx);
 588             fiy0             = _mm256_add_pd(fiy0,ty);
 589             fiz0             = _mm256_add_pd(fiz0,tz);
 590
 591             fjptrA             = f+j_coord_offsetA;
 592             fjptrB             = f+j_coord_offsetB;
 593             fjptrC             = f+j_coord_offsetC;
 594             fjptrD             = f+j_coord_offsetD;
 595             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 596
 597             /* Inner loop uses 46 flops */
 598         }
 599
 600         if(jidx<j_index_end)
 601         {
 602
 603             /* Get j neighbor index, and coordinate index */
 604             jnrlistA         = jjnr[jidx];
 605             jnrlistB         = jjnr[jidx+1];
 606             jnrlistC         = jjnr[jidx+2];
 607             jnrlistD         = jjnr[jidx+3];
 608             /* Sign of each element will be negative for non-real atoms.
 609              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 610              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 611              */
 612             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 613
 614             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 615             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 616             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 617
 618             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 619             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 620             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 621             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 622             j_coord_offsetA  = DIM*jnrA;
 623             j_coord_offsetB  = DIM*jnrB;
 624             j_coord_offsetC  = DIM*jnrC;
 625             j_coord_offsetD  = DIM*jnrD;
 626
 627             /* load j atom coordinates */
 628             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 629                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 630                                                  &jx0,&jy0,&jz0);
 631
 632             /* Calculate displacement vector */
 633             dx00             = _mm256_sub_pd(ix0,jx0);
 634             dy00             = _mm256_sub_pd(iy0,jy0);
 635             dz00             = _mm256_sub_pd(iz0,jz0);
 636
 637             /* Calculate squared distance and things based on it */
 638             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 639
 640             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 641
 642             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 643
 644             /* Load parameters for j particles */
 645             vdwjidx0A        = 2*vdwtype[jnrA+0];
 646             vdwjidx0B        = 2*vdwtype[jnrB+0];
 647             vdwjidx0C        = 2*vdwtype[jnrC+0];
 648             vdwjidx0D        = 2*vdwtype[jnrD+0];
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             r00              = _mm256_mul_pd(rsq00,rinv00);
 655             r00              = _mm256_andnot_pd(dummy_mask,r00);
 656
 657             /* Compute parameters for interactions between i and j atoms */
 658             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 659                                             vdwioffsetptr0+vdwjidx0B,
 660                                             vdwioffsetptr0+vdwjidx0C,
 661                                             vdwioffsetptr0+vdwjidx0D,
 662                                             &c6_00,&c12_00);
 663
 664             c6grid_00       = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
 665                                                                   vdwgridioffsetptr0+vdwjidx0B,
 666                                                                   vdwgridioffsetptr0+vdwjidx0C,
 667                                                                   vdwgridioffsetptr0+vdwjidx0D);
 668
 669             /* Analytical LJ-PME */
 670             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 671             ewcljrsq         = _mm256_mul_pd(ewclj2,rsq00);
 672             ewclj6           = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
 673             exponent         = gmx_simd_exp_d(ewcljrsq);
 674             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 675             poly             = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
 676             /* f6A = 6 * C6grid * (1 - poly) */
 677             f6A              = _mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly));
 678             /* f6B = C6grid * exponent * beta^6 */
 679             f6B              = _mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6));
 680             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 681             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);
 682
 683             fscal            = fvdw;
 684
 685             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 686
 687             /* Calculate temporary vectorial force */
 688             tx               = _mm256_mul_pd(fscal,dx00);
 689             ty               = _mm256_mul_pd(fscal,dy00);
 690             tz               = _mm256_mul_pd(fscal,dz00);
 691
 692             /* Update vectorial force */
 693             fix0             = _mm256_add_pd(fix0,tx);
 694             fiy0             = _mm256_add_pd(fiy0,ty);
 695             fiz0             = _mm256_add_pd(fiz0,tz);
 696
 697             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 698             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 699             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 700             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 701             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 702
 703             /* Inner loop uses 47 flops */
 704         }
 705
 706         /* End of innermost loop */
 707
 708         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 709                                                  f+i_coord_offset,fshift+i_shift_offset);
 710
 711         /* Increment number of inner iterations */
 712         inneriter                  += j_index_end - j_index_start;
 713
 714         /* Outer loop uses 6 flops */
 715     }
 716
 717     /* Increment number of outer iterations */
 718     outeriter        += nri;
 719
 720     /* Update outer/inner flops */
 721
 722     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*47);
 723 }