src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecNone_VdwCSTab_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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: None
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  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     __m128i          vfitab;
  96     __m128i          ifour       = _mm_set1_epi32(4);
  97     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  98     real             *vftab;
  99     __m256d          dummy_mask,cutoff_mask;
 100     __m128           tmpmask0,tmpmask1;
 101     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 102     __m256d          one     = _mm256_set1_pd(1.0);
 103     __m256d          two     = _mm256_set1_pd(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     nvdwtype         = fr->ntype;
 116     vdwparam         = fr->nbfp;
 117     vdwtype          = mdatoms->typeA;
 118
 119     vftab            = kernel_data->table_vdw->data;
 120     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm256_setzero_pd();
 155         fiy0             = _mm256_setzero_pd();
 156         fiz0             = _mm256_setzero_pd();
 157
 158         /* Load parameters for i particles */
 159         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 160
 161         /* Reset potential sums */
 162         vvdwsum          = _mm256_setzero_pd();
 163
 164         /* Start inner kernel loop */
 165         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 166         {
 167
 168             /* Get j neighbor index, and coordinate index */
 169             jnrA             = jjnr[jidx];
 170             jnrB             = jjnr[jidx+1];
 171             jnrC             = jjnr[jidx+2];
 172             jnrD             = jjnr[jidx+3];
 173             j_coord_offsetA  = DIM*jnrA;
 174             j_coord_offsetB  = DIM*jnrB;
 175             j_coord_offsetC  = DIM*jnrC;
 176             j_coord_offsetD  = DIM*jnrD;
 177
 178             /* load j atom coordinates */
 179             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 180                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 181                                                  &jx0,&jy0,&jz0);
 182
 183             /* Calculate displacement vector */
 184             dx00             = _mm256_sub_pd(ix0,jx0);
 185             dy00             = _mm256_sub_pd(iy0,jy0);
 186             dz00             = _mm256_sub_pd(iz0,jz0);
 187
 188             /* Calculate squared distance and things based on it */
 189             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 190
 191             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 192
 193             /* Load parameters for j particles */
 194             vdwjidx0A        = 2*vdwtype[jnrA+0];
 195             vdwjidx0B        = 2*vdwtype[jnrB+0];
 196             vdwjidx0C        = 2*vdwtype[jnrC+0];
 197             vdwjidx0D        = 2*vdwtype[jnrD+0];
 198
 199             /**************************
 200              * CALCULATE INTERACTIONS *
 201              **************************/
 202
 203             r00              = _mm256_mul_pd(rsq00,rinv00);
 204
 205             /* Compute parameters for interactions between i and j atoms */
 206             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 207                                             vdwioffsetptr0+vdwjidx0B,
 208                                             vdwioffsetptr0+vdwjidx0C,
 209                                             vdwioffsetptr0+vdwjidx0D,
 210                                             &c6_00,&c12_00);
 211
 212             /* Calculate table index by multiplying r with table scale and truncate to integer */
 213             rt               = _mm256_mul_pd(r00,vftabscale);
 214             vfitab           = _mm256_cvttpd_epi32(rt);
 215             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 216             vfitab           = _mm_slli_epi32(vfitab,3);
 217
 218             /* CUBIC SPLINE TABLE DISPERSION */
 219             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 220             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 221             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 222             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 223             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 224             Heps             = _mm256_mul_pd(vfeps,H);
 225             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 226             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 227             vvdw6            = _mm256_mul_pd(c6_00,VV);
 228             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 229             fvdw6            = _mm256_mul_pd(c6_00,FF);
 230
 231             /* CUBIC SPLINE TABLE REPULSION */
 232             vfitab           = _mm_add_epi32(vfitab,ifour);
 233             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 234             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 235             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 236             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 237             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 238             Heps             = _mm256_mul_pd(vfeps,H);
 239             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 240             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 241             vvdw12           = _mm256_mul_pd(c12_00,VV);
 242             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 243             fvdw12           = _mm256_mul_pd(c12_00,FF);
 244             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 245             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 246
 247             /* Update potential sum for this i atom from the interaction with this j atom. */
 248             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 249
 250             fscal            = fvdw;
 251
 252             /* Calculate temporary vectorial force */
 253             tx               = _mm256_mul_pd(fscal,dx00);
 254             ty               = _mm256_mul_pd(fscal,dy00);
 255             tz               = _mm256_mul_pd(fscal,dz00);
 256
 257             /* Update vectorial force */
 258             fix0             = _mm256_add_pd(fix0,tx);
 259             fiy0             = _mm256_add_pd(fiy0,ty);
 260             fiz0             = _mm256_add_pd(fiz0,tz);
 261
 262             fjptrA             = f+j_coord_offsetA;
 263             fjptrB             = f+j_coord_offsetB;
 264             fjptrC             = f+j_coord_offsetC;
 265             fjptrD             = f+j_coord_offsetD;
 266             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 267
 268             /* Inner loop uses 56 flops */
 269         }
 270
 271         if(jidx<j_index_end)
 272         {
 273
 274             /* Get j neighbor index, and coordinate index */
 275             jnrlistA         = jjnr[jidx];
 276             jnrlistB         = jjnr[jidx+1];
 277             jnrlistC         = jjnr[jidx+2];
 278             jnrlistD         = jjnr[jidx+3];
 279             /* Sign of each element will be negative for non-real atoms.
 280              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 281              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 282              */
 283             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 284
 285             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 286             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 287             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 288
 289             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 290             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 291             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 292             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 293             j_coord_offsetA  = DIM*jnrA;
 294             j_coord_offsetB  = DIM*jnrB;
 295             j_coord_offsetC  = DIM*jnrC;
 296             j_coord_offsetD  = DIM*jnrD;
 297
 298             /* load j atom coordinates */
 299             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 300                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 301                                                  &jx0,&jy0,&jz0);
 302
 303             /* Calculate displacement vector */
 304             dx00             = _mm256_sub_pd(ix0,jx0);
 305             dy00             = _mm256_sub_pd(iy0,jy0);
 306             dz00             = _mm256_sub_pd(iz0,jz0);
 307
 308             /* Calculate squared distance and things based on it */
 309             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 310
 311             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 312
 313             /* Load parameters for j particles */
 314             vdwjidx0A        = 2*vdwtype[jnrA+0];
 315             vdwjidx0B        = 2*vdwtype[jnrB+0];
 316             vdwjidx0C        = 2*vdwtype[jnrC+0];
 317             vdwjidx0D        = 2*vdwtype[jnrD+0];
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             r00              = _mm256_mul_pd(rsq00,rinv00);
 324             r00              = _mm256_andnot_pd(dummy_mask,r00);
 325
 326             /* Compute parameters for interactions between i and j atoms */
 327             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 328                                             vdwioffsetptr0+vdwjidx0B,
 329                                             vdwioffsetptr0+vdwjidx0C,
 330                                             vdwioffsetptr0+vdwjidx0D,
 331                                             &c6_00,&c12_00);
 332
 333             /* Calculate table index by multiplying r with table scale and truncate to integer */
 334             rt               = _mm256_mul_pd(r00,vftabscale);
 335             vfitab           = _mm256_cvttpd_epi32(rt);
 336             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 337             vfitab           = _mm_slli_epi32(vfitab,3);
 338
 339             /* CUBIC SPLINE TABLE DISPERSION */
 340             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 341             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 342             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 343             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 344             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 345             Heps             = _mm256_mul_pd(vfeps,H);
 346             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 347             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 348             vvdw6            = _mm256_mul_pd(c6_00,VV);
 349             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 350             fvdw6            = _mm256_mul_pd(c6_00,FF);
 351
 352             /* CUBIC SPLINE TABLE REPULSION */
 353             vfitab           = _mm_add_epi32(vfitab,ifour);
 354             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 355             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 356             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 357             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 358             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 359             Heps             = _mm256_mul_pd(vfeps,H);
 360             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 361             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 362             vvdw12           = _mm256_mul_pd(c12_00,VV);
 363             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 364             fvdw12           = _mm256_mul_pd(c12_00,FF);
 365             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 366             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 370             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 371
 372             fscal            = fvdw;
 373
 374             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 375
 376             /* Calculate temporary vectorial force */
 377             tx               = _mm256_mul_pd(fscal,dx00);
 378             ty               = _mm256_mul_pd(fscal,dy00);
 379             tz               = _mm256_mul_pd(fscal,dz00);
 380
 381             /* Update vectorial force */
 382             fix0             = _mm256_add_pd(fix0,tx);
 383             fiy0             = _mm256_add_pd(fiy0,ty);
 384             fiz0             = _mm256_add_pd(fiz0,tz);
 385
 386             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 387             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 388             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 389             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 390             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 391
 392             /* Inner loop uses 57 flops */
 393         }
 394
 395         /* End of innermost loop */
 396
 397         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 398                                                  f+i_coord_offset,fshift+i_shift_offset);
 399
 400         ggid                        = gid[iidx];
 401         /* Update potential energies */
 402         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 403
 404         /* Increment number of inner iterations */
 405         inneriter                  += j_index_end - j_index_start;
 406
 407         /* Outer loop uses 7 flops */
 408     }
 409
 410     /* Increment number of outer iterations */
 411     outeriter        += nri;
 412
 413     /* Update outer/inner flops */
 414
 415     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*57);
 416 }
 417 /*
 418  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_double
 419  * Electrostatics interaction: None
 420  * VdW interaction:            CubicSplineTable
 421  * Geometry:                   Particle-Particle
 422  * Calculate force/pot:        Force
 423  */
 424 void
 425 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_double
 426                     (t_nblist                    * gmx_restrict       nlist,
 427                      rvec                        * gmx_restrict          xx,
 428                      rvec                        * gmx_restrict          ff,
 429                      t_forcerec                  * gmx_restrict          fr,
 430                      t_mdatoms                   * gmx_restrict     mdatoms,
 431                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 432                      t_nrnb                      * gmx_restrict        nrnb)
 433 {
 434     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 435      * just 0 for non-waters.
 436      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 437      * jnr indices corresponding to data put in the four positions in the SIMD register.
 438      */
 439     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 440     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 441     int              jnrA,jnrB,jnrC,jnrD;
 442     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 443     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 444     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 445     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 446     real             rcutoff_scalar;
 447     real             *shiftvec,*fshift,*x,*f;
 448     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 449     real             scratch[4*DIM];
 450     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 451     real *           vdwioffsetptr0;
 452     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 453     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 454     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 455     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 456     int              nvdwtype;
 457     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 458     int              *vdwtype;
 459     real             *vdwparam;
 460     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 461     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 462     __m128i          vfitab;
 463     __m128i          ifour       = _mm_set1_epi32(4);
 464     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 465     real             *vftab;
 466     __m256d          dummy_mask,cutoff_mask;
 467     __m128           tmpmask0,tmpmask1;
 468     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 469     __m256d          one     = _mm256_set1_pd(1.0);
 470     __m256d          two     = _mm256_set1_pd(2.0);
 471     x                = xx[0];
 472     f                = ff[0];
 473
 474     nri              = nlist->nri;
 475     iinr             = nlist->iinr;
 476     jindex           = nlist->jindex;
 477     jjnr             = nlist->jjnr;
 478     shiftidx         = nlist->shift;
 479     gid              = nlist->gid;
 480     shiftvec         = fr->shift_vec[0];
 481     fshift           = fr->fshift[0];
 482     nvdwtype         = fr->ntype;
 483     vdwparam         = fr->nbfp;
 484     vdwtype          = mdatoms->typeA;
 485
 486     vftab            = kernel_data->table_vdw->data;
 487     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 488
 489     /* Avoid stupid compiler warnings */
 490     jnrA = jnrB = jnrC = jnrD = 0;
 491     j_coord_offsetA = 0;
 492     j_coord_offsetB = 0;
 493     j_coord_offsetC = 0;
 494     j_coord_offsetD = 0;
 495
 496     outeriter        = 0;
 497     inneriter        = 0;
 498
 499     for(iidx=0;iidx<4*DIM;iidx++)
 500     {
 501         scratch[iidx] = 0.0;
 502     }
 503
 504     /* Start outer loop over neighborlists */
 505     for(iidx=0; iidx<nri; iidx++)
 506     {
 507         /* Load shift vector for this list */
 508         i_shift_offset   = DIM*shiftidx[iidx];
 509
 510         /* Load limits for loop over neighbors */
 511         j_index_start    = jindex[iidx];
 512         j_index_end      = jindex[iidx+1];
 513
 514         /* Get outer coordinate index */
 515         inr              = iinr[iidx];
 516         i_coord_offset   = DIM*inr;
 517
 518         /* Load i particle coords and add shift vector */
 519         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 520
 521         fix0             = _mm256_setzero_pd();
 522         fiy0             = _mm256_setzero_pd();
 523         fiz0             = _mm256_setzero_pd();
 524
 525         /* Load parameters for i particles */
 526         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 527
 528         /* Start inner kernel loop */
 529         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 530         {
 531
 532             /* Get j neighbor index, and coordinate index */
 533             jnrA             = jjnr[jidx];
 534             jnrB             = jjnr[jidx+1];
 535             jnrC             = jjnr[jidx+2];
 536             jnrD             = jjnr[jidx+3];
 537             j_coord_offsetA  = DIM*jnrA;
 538             j_coord_offsetB  = DIM*jnrB;
 539             j_coord_offsetC  = DIM*jnrC;
 540             j_coord_offsetD  = DIM*jnrD;
 541
 542             /* load j atom coordinates */
 543             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 544                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 545                                                  &jx0,&jy0,&jz0);
 546
 547             /* Calculate displacement vector */
 548             dx00             = _mm256_sub_pd(ix0,jx0);
 549             dy00             = _mm256_sub_pd(iy0,jy0);
 550             dz00             = _mm256_sub_pd(iz0,jz0);
 551
 552             /* Calculate squared distance and things based on it */
 553             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 554
 555             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 556
 557             /* Load parameters for j particles */
 558             vdwjidx0A        = 2*vdwtype[jnrA+0];
 559             vdwjidx0B        = 2*vdwtype[jnrB+0];
 560             vdwjidx0C        = 2*vdwtype[jnrC+0];
 561             vdwjidx0D        = 2*vdwtype[jnrD+0];
 562
 563             /**************************
 564              * CALCULATE INTERACTIONS *
 565              **************************/
 566
 567             r00              = _mm256_mul_pd(rsq00,rinv00);
 568
 569             /* Compute parameters for interactions between i and j atoms */
 570             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 571                                             vdwioffsetptr0+vdwjidx0B,
 572                                             vdwioffsetptr0+vdwjidx0C,
 573                                             vdwioffsetptr0+vdwjidx0D,
 574                                             &c6_00,&c12_00);
 575
 576             /* Calculate table index by multiplying r with table scale and truncate to integer */
 577             rt               = _mm256_mul_pd(r00,vftabscale);
 578             vfitab           = _mm256_cvttpd_epi32(rt);
 579             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 580             vfitab           = _mm_slli_epi32(vfitab,3);
 581
 582             /* CUBIC SPLINE TABLE DISPERSION */
 583             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 584             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 585             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 586             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 587             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 588             Heps             = _mm256_mul_pd(vfeps,H);
 589             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 590             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 591             fvdw6            = _mm256_mul_pd(c6_00,FF);
 592
 593             /* CUBIC SPLINE TABLE REPULSION */
 594             vfitab           = _mm_add_epi32(vfitab,ifour);
 595             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 596             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 597             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 598             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 599             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 600             Heps             = _mm256_mul_pd(vfeps,H);
 601             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 602             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 603             fvdw12           = _mm256_mul_pd(c12_00,FF);
 604             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 605
 606             fscal            = fvdw;
 607
 608             /* Calculate temporary vectorial force */
 609             tx               = _mm256_mul_pd(fscal,dx00);
 610             ty               = _mm256_mul_pd(fscal,dy00);
 611             tz               = _mm256_mul_pd(fscal,dz00);
 612
 613             /* Update vectorial force */
 614             fix0             = _mm256_add_pd(fix0,tx);
 615             fiy0             = _mm256_add_pd(fiy0,ty);
 616             fiz0             = _mm256_add_pd(fiz0,tz);
 617
 618             fjptrA             = f+j_coord_offsetA;
 619             fjptrB             = f+j_coord_offsetB;
 620             fjptrC             = f+j_coord_offsetC;
 621             fjptrD             = f+j_coord_offsetD;
 622             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 623
 624             /* Inner loop uses 48 flops */
 625         }
 626
 627         if(jidx<j_index_end)
 628         {
 629
 630             /* Get j neighbor index, and coordinate index */
 631             jnrlistA         = jjnr[jidx];
 632             jnrlistB         = jjnr[jidx+1];
 633             jnrlistC         = jjnr[jidx+2];
 634             jnrlistD         = jjnr[jidx+3];
 635             /* Sign of each element will be negative for non-real atoms.
 636              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 637              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 638              */
 639             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 640
 641             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 642             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 643             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 644
 645             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 646             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 647             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 648             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 649             j_coord_offsetA  = DIM*jnrA;
 650             j_coord_offsetB  = DIM*jnrB;
 651             j_coord_offsetC  = DIM*jnrC;
 652             j_coord_offsetD  = DIM*jnrD;
 653
 654             /* load j atom coordinates */
 655             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 656                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 657                                                  &jx0,&jy0,&jz0);
 658
 659             /* Calculate displacement vector */
 660             dx00             = _mm256_sub_pd(ix0,jx0);
 661             dy00             = _mm256_sub_pd(iy0,jy0);
 662             dz00             = _mm256_sub_pd(iz0,jz0);
 663
 664             /* Calculate squared distance and things based on it */
 665             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 666
 667             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 668
 669             /* Load parameters for j particles */
 670             vdwjidx0A        = 2*vdwtype[jnrA+0];
 671             vdwjidx0B        = 2*vdwtype[jnrB+0];
 672             vdwjidx0C        = 2*vdwtype[jnrC+0];
 673             vdwjidx0D        = 2*vdwtype[jnrD+0];
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             r00              = _mm256_mul_pd(rsq00,rinv00);
 680             r00              = _mm256_andnot_pd(dummy_mask,r00);
 681
 682             /* Compute parameters for interactions between i and j atoms */
 683             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 684                                             vdwioffsetptr0+vdwjidx0B,
 685                                             vdwioffsetptr0+vdwjidx0C,
 686                                             vdwioffsetptr0+vdwjidx0D,
 687                                             &c6_00,&c12_00);
 688
 689             /* Calculate table index by multiplying r with table scale and truncate to integer */
 690             rt               = _mm256_mul_pd(r00,vftabscale);
 691             vfitab           = _mm256_cvttpd_epi32(rt);
 692             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 693             vfitab           = _mm_slli_epi32(vfitab,3);
 694
 695             /* CUBIC SPLINE TABLE DISPERSION */
 696             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 697             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 698             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 699             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 700             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 701             Heps             = _mm256_mul_pd(vfeps,H);
 702             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 703             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 704             fvdw6            = _mm256_mul_pd(c6_00,FF);
 705
 706             /* CUBIC SPLINE TABLE REPULSION */
 707             vfitab           = _mm_add_epi32(vfitab,ifour);
 708             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 709             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 710             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 711             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 712             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 713             Heps             = _mm256_mul_pd(vfeps,H);
 714             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 715             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 716             fvdw12           = _mm256_mul_pd(c12_00,FF);
 717             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 718
 719             fscal            = 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 49 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 6 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_VDW_F,outeriter*6 + inneriter*49);
 759 }