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 "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_VdwCSTab_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: None
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecNone_VdwCSTab_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     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     int              nvdwtype;
  92     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  93     int              *vdwtype;
  94     real             *vdwparam;
  95     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  96     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  97     __m128i          vfitab;
  98     __m128i          ifour       = _mm_set1_epi32(4);
  99     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 100     real             *vftab;
 101     __m256d          dummy_mask,cutoff_mask;
 102     __m128           tmpmask0,tmpmask1;
 103     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 104     __m256d          one     = _mm256_set1_pd(1.0);
 105     __m256d          two     = _mm256_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     vftab            = kernel_data->table_vdw->data;
 122     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 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
 163         /* Reset potential sums */
 164         vvdwsum          = _mm256_setzero_pd();
 165
 166         /* Start inner kernel loop */
 167         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 168         {
 169
 170             /* Get j neighbor index, and coordinate index */
 171             jnrA             = jjnr[jidx];
 172             jnrB             = jjnr[jidx+1];
 173             jnrC             = jjnr[jidx+2];
 174             jnrD             = jjnr[jidx+3];
 175             j_coord_offsetA  = DIM*jnrA;
 176             j_coord_offsetB  = DIM*jnrB;
 177             j_coord_offsetC  = DIM*jnrC;
 178             j_coord_offsetD  = DIM*jnrD;
 179
 180             /* load j atom coordinates */
 181             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 182                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 183                                                  &jx0,&jy0,&jz0);
 184
 185             /* Calculate displacement vector */
 186             dx00             = _mm256_sub_pd(ix0,jx0);
 187             dy00             = _mm256_sub_pd(iy0,jy0);
 188             dz00             = _mm256_sub_pd(iz0,jz0);
 189
 190             /* Calculate squared distance and things based on it */
 191             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 192
 193             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 194
 195             /* Load parameters for j particles */
 196             vdwjidx0A        = 2*vdwtype[jnrA+0];
 197             vdwjidx0B        = 2*vdwtype[jnrB+0];
 198             vdwjidx0C        = 2*vdwtype[jnrC+0];
 199             vdwjidx0D        = 2*vdwtype[jnrD+0];
 200
 201             /**************************
 202              * CALCULATE INTERACTIONS *
 203              **************************/
 204
 205             r00              = _mm256_mul_pd(rsq00,rinv00);
 206
 207             /* Compute parameters for interactions between i and j atoms */
 208             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 209                                             vdwioffsetptr0+vdwjidx0B,
 210                                             vdwioffsetptr0+vdwjidx0C,
 211                                             vdwioffsetptr0+vdwjidx0D,
 212                                             &c6_00,&c12_00);
 213
 214             /* Calculate table index by multiplying r with table scale and truncate to integer */
 215             rt               = _mm256_mul_pd(r00,vftabscale);
 216             vfitab           = _mm256_cvttpd_epi32(rt);
 217             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 218             vfitab           = _mm_slli_epi32(vfitab,3);
 219
 220             /* CUBIC SPLINE TABLE DISPERSION */
 221             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 222             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 223             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 224             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 225             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 226             Heps             = _mm256_mul_pd(vfeps,H);
 227             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 228             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 229             vvdw6            = _mm256_mul_pd(c6_00,VV);
 230             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 231             fvdw6            = _mm256_mul_pd(c6_00,FF);
 232
 233             /* CUBIC SPLINE TABLE REPULSION */
 234             vfitab           = _mm_add_epi32(vfitab,ifour);
 235             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 236             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 237             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 238             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 239             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 240             Heps             = _mm256_mul_pd(vfeps,H);
 241             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 242             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 243             vvdw12           = _mm256_mul_pd(c12_00,VV);
 244             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 245             fvdw12           = _mm256_mul_pd(c12_00,FF);
 246             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 247             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 248
 249             /* Update potential sum for this i atom from the interaction with this j atom. */
 250             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 251
 252             fscal            = fvdw;
 253
 254             /* Calculate temporary vectorial force */
 255             tx               = _mm256_mul_pd(fscal,dx00);
 256             ty               = _mm256_mul_pd(fscal,dy00);
 257             tz               = _mm256_mul_pd(fscal,dz00);
 258
 259             /* Update vectorial force */
 260             fix0             = _mm256_add_pd(fix0,tx);
 261             fiy0             = _mm256_add_pd(fiy0,ty);
 262             fiz0             = _mm256_add_pd(fiz0,tz);
 263
 264             fjptrA             = f+j_coord_offsetA;
 265             fjptrB             = f+j_coord_offsetB;
 266             fjptrC             = f+j_coord_offsetC;
 267             fjptrD             = f+j_coord_offsetD;
 268             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 269
 270             /* Inner loop uses 56 flops */
 271         }
 272
 273         if(jidx<j_index_end)
 274         {
 275
 276             /* Get j neighbor index, and coordinate index */
 277             jnrlistA         = jjnr[jidx];
 278             jnrlistB         = jjnr[jidx+1];
 279             jnrlistC         = jjnr[jidx+2];
 280             jnrlistD         = jjnr[jidx+3];
 281             /* Sign of each element will be negative for non-real atoms.
 282              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 283              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 284              */
 285             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 286
 287             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 288             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 289             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 290
 291             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 292             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 293             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 294             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 295             j_coord_offsetA  = DIM*jnrA;
 296             j_coord_offsetB  = DIM*jnrB;
 297             j_coord_offsetC  = DIM*jnrC;
 298             j_coord_offsetD  = DIM*jnrD;
 299
 300             /* load j atom coordinates */
 301             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 302                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 303                                                  &jx0,&jy0,&jz0);
 304
 305             /* Calculate displacement vector */
 306             dx00             = _mm256_sub_pd(ix0,jx0);
 307             dy00             = _mm256_sub_pd(iy0,jy0);
 308             dz00             = _mm256_sub_pd(iz0,jz0);
 309
 310             /* Calculate squared distance and things based on it */
 311             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 312
 313             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 314
 315             /* Load parameters for j particles */
 316             vdwjidx0A        = 2*vdwtype[jnrA+0];
 317             vdwjidx0B        = 2*vdwtype[jnrB+0];
 318             vdwjidx0C        = 2*vdwtype[jnrC+0];
 319             vdwjidx0D        = 2*vdwtype[jnrD+0];
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             r00              = _mm256_mul_pd(rsq00,rinv00);
 326             r00              = _mm256_andnot_pd(dummy_mask,r00);
 327
 328             /* Compute parameters for interactions between i and j atoms */
 329             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 330                                             vdwioffsetptr0+vdwjidx0B,
 331                                             vdwioffsetptr0+vdwjidx0C,
 332                                             vdwioffsetptr0+vdwjidx0D,
 333                                             &c6_00,&c12_00);
 334
 335             /* Calculate table index by multiplying r with table scale and truncate to integer */
 336             rt               = _mm256_mul_pd(r00,vftabscale);
 337             vfitab           = _mm256_cvttpd_epi32(rt);
 338             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 339             vfitab           = _mm_slli_epi32(vfitab,3);
 340
 341             /* CUBIC SPLINE TABLE DISPERSION */
 342             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 343             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 344             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 345             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 346             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 347             Heps             = _mm256_mul_pd(vfeps,H);
 348             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 349             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 350             vvdw6            = _mm256_mul_pd(c6_00,VV);
 351             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 352             fvdw6            = _mm256_mul_pd(c6_00,FF);
 353
 354             /* CUBIC SPLINE TABLE REPULSION */
 355             vfitab           = _mm_add_epi32(vfitab,ifour);
 356             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 357             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 358             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 359             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 360             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 361             Heps             = _mm256_mul_pd(vfeps,H);
 362             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 363             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 364             vvdw12           = _mm256_mul_pd(c12_00,VV);
 365             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 366             fvdw12           = _mm256_mul_pd(c12_00,FF);
 367             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 368             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 372             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 373
 374             fscal            = fvdw;
 375
 376             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 377
 378             /* Calculate temporary vectorial force */
 379             tx               = _mm256_mul_pd(fscal,dx00);
 380             ty               = _mm256_mul_pd(fscal,dy00);
 381             tz               = _mm256_mul_pd(fscal,dz00);
 382
 383             /* Update vectorial force */
 384             fix0             = _mm256_add_pd(fix0,tx);
 385             fiy0             = _mm256_add_pd(fiy0,ty);
 386             fiz0             = _mm256_add_pd(fiz0,tz);
 387
 388             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 389             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 390             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 391             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 392             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 393
 394             /* Inner loop uses 57 flops */
 395         }
 396
 397         /* End of innermost loop */
 398
 399         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 400                                                  f+i_coord_offset,fshift+i_shift_offset);
 401
 402         ggid                        = gid[iidx];
 403         /* Update potential energies */
 404         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 405
 406         /* Increment number of inner iterations */
 407         inneriter                  += j_index_end - j_index_start;
 408
 409         /* Outer loop uses 7 flops */
 410     }
 411
 412     /* Increment number of outer iterations */
 413     outeriter        += nri;
 414
 415     /* Update outer/inner flops */
 416
 417     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*57);
 418 }
 419 /*
 420  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_double
 421  * Electrostatics interaction: None
 422  * VdW interaction:            CubicSplineTable
 423  * Geometry:                   Particle-Particle
 424  * Calculate force/pot:        Force
 425  */
 426 void
 427 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_double
 428                     (t_nblist                    * gmx_restrict       nlist,
 429                      rvec                        * gmx_restrict          xx,
 430                      rvec                        * gmx_restrict          ff,
 431                      t_forcerec                  * gmx_restrict          fr,
 432                      t_mdatoms                   * gmx_restrict     mdatoms,
 433                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 434                      t_nrnb                      * gmx_restrict        nrnb)
 435 {
 436     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 437      * just 0 for non-waters.
 438      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 439      * jnr indices corresponding to data put in the four positions in the SIMD register.
 440      */
 441     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 442     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 443     int              jnrA,jnrB,jnrC,jnrD;
 444     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 445     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 446     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 447     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 448     real             rcutoff_scalar;
 449     real             *shiftvec,*fshift,*x,*f;
 450     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 451     real             scratch[4*DIM];
 452     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 453     real *           vdwioffsetptr0;
 454     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 455     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 456     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 457     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 458     int              nvdwtype;
 459     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 460     int              *vdwtype;
 461     real             *vdwparam;
 462     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 463     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 464     __m128i          vfitab;
 465     __m128i          ifour       = _mm_set1_epi32(4);
 466     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 467     real             *vftab;
 468     __m256d          dummy_mask,cutoff_mask;
 469     __m128           tmpmask0,tmpmask1;
 470     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 471     __m256d          one     = _mm256_set1_pd(1.0);
 472     __m256d          two     = _mm256_set1_pd(2.0);
 473     x                = xx[0];
 474     f                = ff[0];
 475
 476     nri              = nlist->nri;
 477     iinr             = nlist->iinr;
 478     jindex           = nlist->jindex;
 479     jjnr             = nlist->jjnr;
 480     shiftidx         = nlist->shift;
 481     gid              = nlist->gid;
 482     shiftvec         = fr->shift_vec[0];
 483     fshift           = fr->fshift[0];
 484     nvdwtype         = fr->ntype;
 485     vdwparam         = fr->nbfp;
 486     vdwtype          = mdatoms->typeA;
 487
 488     vftab            = kernel_data->table_vdw->data;
 489     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 490
 491     /* Avoid stupid compiler warnings */
 492     jnrA = jnrB = jnrC = jnrD = 0;
 493     j_coord_offsetA = 0;
 494     j_coord_offsetB = 0;
 495     j_coord_offsetC = 0;
 496     j_coord_offsetD = 0;
 497
 498     outeriter        = 0;
 499     inneriter        = 0;
 500
 501     for(iidx=0;iidx<4*DIM;iidx++)
 502     {
 503         scratch[iidx] = 0.0;
 504     }
 505
 506     /* Start outer loop over neighborlists */
 507     for(iidx=0; iidx<nri; iidx++)
 508     {
 509         /* Load shift vector for this list */
 510         i_shift_offset   = DIM*shiftidx[iidx];
 511
 512         /* Load limits for loop over neighbors */
 513         j_index_start    = jindex[iidx];
 514         j_index_end      = jindex[iidx+1];
 515
 516         /* Get outer coordinate index */
 517         inr              = iinr[iidx];
 518         i_coord_offset   = DIM*inr;
 519
 520         /* Load i particle coords and add shift vector */
 521         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 522
 523         fix0             = _mm256_setzero_pd();
 524         fiy0             = _mm256_setzero_pd();
 525         fiz0             = _mm256_setzero_pd();
 526
 527         /* Load parameters for i particles */
 528         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 529
 530         /* Start inner kernel loop */
 531         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 532         {
 533
 534             /* Get j neighbor index, and coordinate index */
 535             jnrA             = jjnr[jidx];
 536             jnrB             = jjnr[jidx+1];
 537             jnrC             = jjnr[jidx+2];
 538             jnrD             = jjnr[jidx+3];
 539             j_coord_offsetA  = DIM*jnrA;
 540             j_coord_offsetB  = DIM*jnrB;
 541             j_coord_offsetC  = DIM*jnrC;
 542             j_coord_offsetD  = DIM*jnrD;
 543
 544             /* load j atom coordinates */
 545             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 546                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 547                                                  &jx0,&jy0,&jz0);
 548
 549             /* Calculate displacement vector */
 550             dx00             = _mm256_sub_pd(ix0,jx0);
 551             dy00             = _mm256_sub_pd(iy0,jy0);
 552             dz00             = _mm256_sub_pd(iz0,jz0);
 553
 554             /* Calculate squared distance and things based on it */
 555             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 556
 557             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 558
 559             /* Load parameters for j particles */
 560             vdwjidx0A        = 2*vdwtype[jnrA+0];
 561             vdwjidx0B        = 2*vdwtype[jnrB+0];
 562             vdwjidx0C        = 2*vdwtype[jnrC+0];
 563             vdwjidx0D        = 2*vdwtype[jnrD+0];
 564
 565             /**************************
 566              * CALCULATE INTERACTIONS *
 567              **************************/
 568
 569             r00              = _mm256_mul_pd(rsq00,rinv00);
 570
 571             /* Compute parameters for interactions between i and j atoms */
 572             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 573                                             vdwioffsetptr0+vdwjidx0B,
 574                                             vdwioffsetptr0+vdwjidx0C,
 575                                             vdwioffsetptr0+vdwjidx0D,
 576                                             &c6_00,&c12_00);
 577
 578             /* Calculate table index by multiplying r with table scale and truncate to integer */
 579             rt               = _mm256_mul_pd(r00,vftabscale);
 580             vfitab           = _mm256_cvttpd_epi32(rt);
 581             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 582             vfitab           = _mm_slli_epi32(vfitab,3);
 583
 584             /* CUBIC SPLINE TABLE DISPERSION */
 585             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 586             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 587             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 588             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 589             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 590             Heps             = _mm256_mul_pd(vfeps,H);
 591             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 592             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 593             fvdw6            = _mm256_mul_pd(c6_00,FF);
 594
 595             /* CUBIC SPLINE TABLE REPULSION */
 596             vfitab           = _mm_add_epi32(vfitab,ifour);
 597             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 598             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 599             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 600             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 601             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 602             Heps             = _mm256_mul_pd(vfeps,H);
 603             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 604             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 605             fvdw12           = _mm256_mul_pd(c12_00,FF);
 606             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 607
 608             fscal            = fvdw;
 609
 610             /* Calculate temporary vectorial force */
 611             tx               = _mm256_mul_pd(fscal,dx00);
 612             ty               = _mm256_mul_pd(fscal,dy00);
 613             tz               = _mm256_mul_pd(fscal,dz00);
 614
 615             /* Update vectorial force */
 616             fix0             = _mm256_add_pd(fix0,tx);
 617             fiy0             = _mm256_add_pd(fiy0,ty);
 618             fiz0             = _mm256_add_pd(fiz0,tz);
 619
 620             fjptrA             = f+j_coord_offsetA;
 621             fjptrB             = f+j_coord_offsetB;
 622             fjptrC             = f+j_coord_offsetC;
 623             fjptrD             = f+j_coord_offsetD;
 624             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 625
 626             /* Inner loop uses 48 flops */
 627         }
 628
 629         if(jidx<j_index_end)
 630         {
 631
 632             /* Get j neighbor index, and coordinate index */
 633             jnrlistA         = jjnr[jidx];
 634             jnrlistB         = jjnr[jidx+1];
 635             jnrlistC         = jjnr[jidx+2];
 636             jnrlistD         = jjnr[jidx+3];
 637             /* Sign of each element will be negative for non-real atoms.
 638              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 639              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 640              */
 641             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 642
 643             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 644             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 645             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 646
 647             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 648             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 649             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 650             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 651             j_coord_offsetA  = DIM*jnrA;
 652             j_coord_offsetB  = DIM*jnrB;
 653             j_coord_offsetC  = DIM*jnrC;
 654             j_coord_offsetD  = DIM*jnrD;
 655
 656             /* load j atom coordinates */
 657             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 658                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 659                                                  &jx0,&jy0,&jz0);
 660
 661             /* Calculate displacement vector */
 662             dx00             = _mm256_sub_pd(ix0,jx0);
 663             dy00             = _mm256_sub_pd(iy0,jy0);
 664             dz00             = _mm256_sub_pd(iz0,jz0);
 665
 666             /* Calculate squared distance and things based on it */
 667             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 668
 669             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 670
 671             /* Load parameters for j particles */
 672             vdwjidx0A        = 2*vdwtype[jnrA+0];
 673             vdwjidx0B        = 2*vdwtype[jnrB+0];
 674             vdwjidx0C        = 2*vdwtype[jnrC+0];
 675             vdwjidx0D        = 2*vdwtype[jnrD+0];
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             r00              = _mm256_mul_pd(rsq00,rinv00);
 682             r00              = _mm256_andnot_pd(dummy_mask,r00);
 683
 684             /* Compute parameters for interactions between i and j atoms */
 685             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 686                                             vdwioffsetptr0+vdwjidx0B,
 687                                             vdwioffsetptr0+vdwjidx0C,
 688                                             vdwioffsetptr0+vdwjidx0D,
 689                                             &c6_00,&c12_00);
 690
 691             /* Calculate table index by multiplying r with table scale and truncate to integer */
 692             rt               = _mm256_mul_pd(r00,vftabscale);
 693             vfitab           = _mm256_cvttpd_epi32(rt);
 694             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 695             vfitab           = _mm_slli_epi32(vfitab,3);
 696
 697             /* CUBIC SPLINE TABLE DISPERSION */
 698             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 699             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 700             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 701             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 702             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 703             Heps             = _mm256_mul_pd(vfeps,H);
 704             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 705             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 706             fvdw6            = _mm256_mul_pd(c6_00,FF);
 707
 708             /* CUBIC SPLINE TABLE REPULSION */
 709             vfitab           = _mm_add_epi32(vfitab,ifour);
 710             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 711             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 712             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 713             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 714             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 715             Heps             = _mm256_mul_pd(vfeps,H);
 716             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 717             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 718             fvdw12           = _mm256_mul_pd(c12_00,FF);
 719             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 720
 721             fscal            = fvdw;
 722
 723             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 724
 725             /* Calculate temporary vectorial force */
 726             tx               = _mm256_mul_pd(fscal,dx00);
 727             ty               = _mm256_mul_pd(fscal,dy00);
 728             tz               = _mm256_mul_pd(fscal,dz00);
 729
 730             /* Update vectorial force */
 731             fix0             = _mm256_add_pd(fix0,tx);
 732             fiy0             = _mm256_add_pd(fiy0,ty);
 733             fiz0             = _mm256_add_pd(fiz0,tz);
 734
 735             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 736             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 737             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 738             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 739             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 740
 741             /* Inner loop uses 49 flops */
 742         }
 743
 744         /* End of innermost loop */
 745
 746         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 747                                                  f+i_coord_offset,fshift+i_shift_offset);
 748
 749         /* Increment number of inner iterations */
 750         inneriter                  += j_index_end - j_index_start;
 751
 752         /* Outer loop uses 6 flops */
 753     }
 754
 755     /* Increment number of outer iterations */
 756     outeriter        += nri;
 757
 758     /* Update outer/inner flops */
 759
 760     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*49);
 761 }