src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_avx_256_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: CubicSplineTable
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_256_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     real *           vdwioffsetptr0;
  84     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  95     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  96     __m128i          vfitab;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  99     real             *vftab;
 100     __m256d          dummy_mask,cutoff_mask;
 101     __m128           tmpmask0,tmpmask1;
 102     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 103     __m256d          one     = _mm256_set1_pd(1.0);
 104     __m256d          two     = _mm256_set1_pd(2.0);
 105     x                = xx[0];
 106     f                = ff[0];
 107
 108     nri              = nlist->nri;
 109     iinr             = nlist->iinr;
 110     jindex           = nlist->jindex;
 111     jjnr             = nlist->jjnr;
 112     shiftidx         = nlist->shift;
 113     gid              = nlist->gid;
 114     shiftvec         = fr->shift_vec[0];
 115     fshift           = fr->fshift[0];
 116     facel            = _mm256_set1_pd(fr->ic->epsfac);
 117     charge           = mdatoms->chargeA;
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     vftab            = kernel_data->table_elec->data;
 123     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 124
 125     /* Avoid stupid compiler warnings */
 126     jnrA = jnrB = jnrC = jnrD = 0;
 127     j_coord_offsetA = 0;
 128     j_coord_offsetB = 0;
 129     j_coord_offsetC = 0;
 130     j_coord_offsetD = 0;
 131
 132     outeriter        = 0;
 133     inneriter        = 0;
 134
 135     for(iidx=0;iidx<4*DIM;iidx++)
 136     {
 137         scratch[iidx] = 0.0;
 138     }
 139
 140     /* Start outer loop over neighborlists */
 141     for(iidx=0; iidx<nri; iidx++)
 142     {
 143         /* Load shift vector for this list */
 144         i_shift_offset   = DIM*shiftidx[iidx];
 145
 146         /* Load limits for loop over neighbors */
 147         j_index_start    = jindex[iidx];
 148         j_index_end      = jindex[iidx+1];
 149
 150         /* Get outer coordinate index */
 151         inr              = iinr[iidx];
 152         i_coord_offset   = DIM*inr;
 153
 154         /* Load i particle coords and add shift vector */
 155         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 156
 157         fix0             = _mm256_setzero_pd();
 158         fiy0             = _mm256_setzero_pd();
 159         fiz0             = _mm256_setzero_pd();
 160
 161         /* Load parameters for i particles */
 162         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 163         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 164
 165         /* Reset potential sums */
 166         velecsum         = _mm256_setzero_pd();
 167         vvdwsum          = _mm256_setzero_pd();
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 171         {
 172
 173             /* Get j neighbor index, and coordinate index */
 174             jnrA             = jjnr[jidx];
 175             jnrB             = jjnr[jidx+1];
 176             jnrC             = jjnr[jidx+2];
 177             jnrD             = jjnr[jidx+3];
 178             j_coord_offsetA  = DIM*jnrA;
 179             j_coord_offsetB  = DIM*jnrB;
 180             j_coord_offsetC  = DIM*jnrC;
 181             j_coord_offsetD  = DIM*jnrD;
 182
 183             /* load j atom coordinates */
 184             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 185                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 186                                                  &jx0,&jy0,&jz0);
 187
 188             /* Calculate displacement vector */
 189             dx00             = _mm256_sub_pd(ix0,jx0);
 190             dy00             = _mm256_sub_pd(iy0,jy0);
 191             dz00             = _mm256_sub_pd(iz0,jz0);
 192
 193             /* Calculate squared distance and things based on it */
 194             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 195
 196             rinv00           = avx256_invsqrt_d(rsq00);
 197
 198             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 199
 200             /* Load parameters for j particles */
 201             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 202                                                                  charge+jnrC+0,charge+jnrD+0);
 203             vdwjidx0A        = 2*vdwtype[jnrA+0];
 204             vdwjidx0B        = 2*vdwtype[jnrB+0];
 205             vdwjidx0C        = 2*vdwtype[jnrC+0];
 206             vdwjidx0D        = 2*vdwtype[jnrD+0];
 207
 208             /**************************
 209              * CALCULATE INTERACTIONS *
 210              **************************/
 211
 212             r00              = _mm256_mul_pd(rsq00,rinv00);
 213
 214             /* Compute parameters for interactions between i and j atoms */
 215             qq00             = _mm256_mul_pd(iq0,jq0);
 216             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 217                                             vdwioffsetptr0+vdwjidx0B,
 218                                             vdwioffsetptr0+vdwjidx0C,
 219                                             vdwioffsetptr0+vdwjidx0D,
 220                                             &c6_00,&c12_00);
 221
 222             /* Calculate table index by multiplying r with table scale and truncate to integer */
 223             rt               = _mm256_mul_pd(r00,vftabscale);
 224             vfitab           = _mm256_cvttpd_epi32(rt);
 225             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 226             vfitab           = _mm_slli_epi32(vfitab,2);
 227
 228             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 229             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 230             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 231             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 232             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 233             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 234             Heps             = _mm256_mul_pd(vfeps,H);
 235             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 236             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 237             velec            = _mm256_mul_pd(qq00,VV);
 238             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 239             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 240
 241             /* LENNARD-JONES DISPERSION/REPULSION */
 242
 243             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 244             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 245             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 246             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 247             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 248
 249             /* Update potential sum for this i atom from the interaction with this j atom. */
 250             velecsum         = _mm256_add_pd(velecsum,velec);
 251             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 252
 253             fscal            = _mm256_add_pd(felec,fvdw);
 254
 255             /* Calculate temporary vectorial force */
 256             tx               = _mm256_mul_pd(fscal,dx00);
 257             ty               = _mm256_mul_pd(fscal,dy00);
 258             tz               = _mm256_mul_pd(fscal,dz00);
 259
 260             /* Update vectorial force */
 261             fix0             = _mm256_add_pd(fix0,tx);
 262             fiy0             = _mm256_add_pd(fiy0,ty);
 263             fiz0             = _mm256_add_pd(fiz0,tz);
 264
 265             fjptrA             = f+j_coord_offsetA;
 266             fjptrB             = f+j_coord_offsetB;
 267             fjptrC             = f+j_coord_offsetC;
 268             fjptrD             = f+j_coord_offsetD;
 269             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 270
 271             /* Inner loop uses 56 flops */
 272         }
 273
 274         if(jidx<j_index_end)
 275         {
 276
 277             /* Get j neighbor index, and coordinate index */
 278             jnrlistA         = jjnr[jidx];
 279             jnrlistB         = jjnr[jidx+1];
 280             jnrlistC         = jjnr[jidx+2];
 281             jnrlistD         = jjnr[jidx+3];
 282             /* Sign of each element will be negative for non-real atoms.
 283              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 284              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 285              */
 286             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 287
 288             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 289             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 290             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 291
 292             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 293             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 294             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 295             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 296             j_coord_offsetA  = DIM*jnrA;
 297             j_coord_offsetB  = DIM*jnrB;
 298             j_coord_offsetC  = DIM*jnrC;
 299             j_coord_offsetD  = DIM*jnrD;
 300
 301             /* load j atom coordinates */
 302             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 303                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 304                                                  &jx0,&jy0,&jz0);
 305
 306             /* Calculate displacement vector */
 307             dx00             = _mm256_sub_pd(ix0,jx0);
 308             dy00             = _mm256_sub_pd(iy0,jy0);
 309             dz00             = _mm256_sub_pd(iz0,jz0);
 310
 311             /* Calculate squared distance and things based on it */
 312             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 313
 314             rinv00           = avx256_invsqrt_d(rsq00);
 315
 316             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 317
 318             /* Load parameters for j particles */
 319             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 320                                                                  charge+jnrC+0,charge+jnrD+0);
 321             vdwjidx0A        = 2*vdwtype[jnrA+0];
 322             vdwjidx0B        = 2*vdwtype[jnrB+0];
 323             vdwjidx0C        = 2*vdwtype[jnrC+0];
 324             vdwjidx0D        = 2*vdwtype[jnrD+0];
 325
 326             /**************************
 327              * CALCULATE INTERACTIONS *
 328              **************************/
 329
 330             r00              = _mm256_mul_pd(rsq00,rinv00);
 331             r00              = _mm256_andnot_pd(dummy_mask,r00);
 332
 333             /* Compute parameters for interactions between i and j atoms */
 334             qq00             = _mm256_mul_pd(iq0,jq0);
 335             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 336                                             vdwioffsetptr0+vdwjidx0B,
 337                                             vdwioffsetptr0+vdwjidx0C,
 338                                             vdwioffsetptr0+vdwjidx0D,
 339                                             &c6_00,&c12_00);
 340
 341             /* Calculate table index by multiplying r with table scale and truncate to integer */
 342             rt               = _mm256_mul_pd(r00,vftabscale);
 343             vfitab           = _mm256_cvttpd_epi32(rt);
 344             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 345             vfitab           = _mm_slli_epi32(vfitab,2);
 346
 347             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 348             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 349             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 350             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 351             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 352             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 353             Heps             = _mm256_mul_pd(vfeps,H);
 354             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 355             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 356             velec            = _mm256_mul_pd(qq00,VV);
 357             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 358             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 359
 360             /* LENNARD-JONES DISPERSION/REPULSION */
 361
 362             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 363             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 364             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 365             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 366             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velec            = _mm256_andnot_pd(dummy_mask,velec);
 370             velecsum         = _mm256_add_pd(velecsum,velec);
 371             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 372             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 373
 374             fscal            = _mm256_add_pd(felec,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(velecsum,kernel_data->energygrp_elec+ggid);
 405         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 406
 407         /* Increment number of inner iterations */
 408         inneriter                  += j_index_end - j_index_start;
 409
 410         /* Outer loop uses 9 flops */
 411     }
 412
 413     /* Increment number of outer iterations */
 414     outeriter        += nri;
 415
 416     /* Update outer/inner flops */
 417
 418     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*57);
 419 }
 420 /*
 421  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 422  * Electrostatics interaction: CubicSplineTable
 423  * VdW interaction:            LennardJones
 424  * Geometry:                   Particle-Particle
 425  * Calculate force/pot:        Force
 426  */
 427 void
 428 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 429                     (t_nblist                    * gmx_restrict       nlist,
 430                      rvec                        * gmx_restrict          xx,
 431                      rvec                        * gmx_restrict          ff,
 432                      struct t_forcerec           * gmx_restrict          fr,
 433                      t_mdatoms                   * gmx_restrict     mdatoms,
 434                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 435                      t_nrnb                      * gmx_restrict        nrnb)
 436 {
 437     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 438      * just 0 for non-waters.
 439      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 440      * jnr indices corresponding to data put in the four positions in the SIMD register.
 441      */
 442     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 443     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 444     int              jnrA,jnrB,jnrC,jnrD;
 445     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 446     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 447     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 448     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 449     real             rcutoff_scalar;
 450     real             *shiftvec,*fshift,*x,*f;
 451     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 452     real             scratch[4*DIM];
 453     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 454     real *           vdwioffsetptr0;
 455     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 456     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 457     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 458     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 459     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 460     real             *charge;
 461     int              nvdwtype;
 462     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 463     int              *vdwtype;
 464     real             *vdwparam;
 465     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 466     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 467     __m128i          vfitab;
 468     __m128i          ifour       = _mm_set1_epi32(4);
 469     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 470     real             *vftab;
 471     __m256d          dummy_mask,cutoff_mask;
 472     __m128           tmpmask0,tmpmask1;
 473     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 474     __m256d          one     = _mm256_set1_pd(1.0);
 475     __m256d          two     = _mm256_set1_pd(2.0);
 476     x                = xx[0];
 477     f                = ff[0];
 478
 479     nri              = nlist->nri;
 480     iinr             = nlist->iinr;
 481     jindex           = nlist->jindex;
 482     jjnr             = nlist->jjnr;
 483     shiftidx         = nlist->shift;
 484     gid              = nlist->gid;
 485     shiftvec         = fr->shift_vec[0];
 486     fshift           = fr->fshift[0];
 487     facel            = _mm256_set1_pd(fr->ic->epsfac);
 488     charge           = mdatoms->chargeA;
 489     nvdwtype         = fr->ntype;
 490     vdwparam         = fr->nbfp;
 491     vdwtype          = mdatoms->typeA;
 492
 493     vftab            = kernel_data->table_elec->data;
 494     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 495
 496     /* Avoid stupid compiler warnings */
 497     jnrA = jnrB = jnrC = jnrD = 0;
 498     j_coord_offsetA = 0;
 499     j_coord_offsetB = 0;
 500     j_coord_offsetC = 0;
 501     j_coord_offsetD = 0;
 502
 503     outeriter        = 0;
 504     inneriter        = 0;
 505
 506     for(iidx=0;iidx<4*DIM;iidx++)
 507     {
 508         scratch[iidx] = 0.0;
 509     }
 510
 511     /* Start outer loop over neighborlists */
 512     for(iidx=0; iidx<nri; iidx++)
 513     {
 514         /* Load shift vector for this list */
 515         i_shift_offset   = DIM*shiftidx[iidx];
 516
 517         /* Load limits for loop over neighbors */
 518         j_index_start    = jindex[iidx];
 519         j_index_end      = jindex[iidx+1];
 520
 521         /* Get outer coordinate index */
 522         inr              = iinr[iidx];
 523         i_coord_offset   = DIM*inr;
 524
 525         /* Load i particle coords and add shift vector */
 526         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 527
 528         fix0             = _mm256_setzero_pd();
 529         fiy0             = _mm256_setzero_pd();
 530         fiz0             = _mm256_setzero_pd();
 531
 532         /* Load parameters for i particles */
 533         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 534         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 535
 536         /* Start inner kernel loop */
 537         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 538         {
 539
 540             /* Get j neighbor index, and coordinate index */
 541             jnrA             = jjnr[jidx];
 542             jnrB             = jjnr[jidx+1];
 543             jnrC             = jjnr[jidx+2];
 544             jnrD             = jjnr[jidx+3];
 545             j_coord_offsetA  = DIM*jnrA;
 546             j_coord_offsetB  = DIM*jnrB;
 547             j_coord_offsetC  = DIM*jnrC;
 548             j_coord_offsetD  = DIM*jnrD;
 549
 550             /* load j atom coordinates */
 551             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 552                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 553                                                  &jx0,&jy0,&jz0);
 554
 555             /* Calculate displacement vector */
 556             dx00             = _mm256_sub_pd(ix0,jx0);
 557             dy00             = _mm256_sub_pd(iy0,jy0);
 558             dz00             = _mm256_sub_pd(iz0,jz0);
 559
 560             /* Calculate squared distance and things based on it */
 561             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 562
 563             rinv00           = avx256_invsqrt_d(rsq00);
 564
 565             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 566
 567             /* Load parameters for j particles */
 568             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 569                                                                  charge+jnrC+0,charge+jnrD+0);
 570             vdwjidx0A        = 2*vdwtype[jnrA+0];
 571             vdwjidx0B        = 2*vdwtype[jnrB+0];
 572             vdwjidx0C        = 2*vdwtype[jnrC+0];
 573             vdwjidx0D        = 2*vdwtype[jnrD+0];
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             r00              = _mm256_mul_pd(rsq00,rinv00);
 580
 581             /* Compute parameters for interactions between i and j atoms */
 582             qq00             = _mm256_mul_pd(iq0,jq0);
 583             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 584                                             vdwioffsetptr0+vdwjidx0B,
 585                                             vdwioffsetptr0+vdwjidx0C,
 586                                             vdwioffsetptr0+vdwjidx0D,
 587                                             &c6_00,&c12_00);
 588
 589             /* Calculate table index by multiplying r with table scale and truncate to integer */
 590             rt               = _mm256_mul_pd(r00,vftabscale);
 591             vfitab           = _mm256_cvttpd_epi32(rt);
 592             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 593             vfitab           = _mm_slli_epi32(vfitab,2);
 594
 595             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 596             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 597             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 598             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 599             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 600             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 601             Heps             = _mm256_mul_pd(vfeps,H);
 602             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 603             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 604             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 605
 606             /* LENNARD-JONES DISPERSION/REPULSION */
 607
 608             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 609             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 610
 611             fscal            = _mm256_add_pd(felec,fvdw);
 612
 613             /* Calculate temporary vectorial force */
 614             tx               = _mm256_mul_pd(fscal,dx00);
 615             ty               = _mm256_mul_pd(fscal,dy00);
 616             tz               = _mm256_mul_pd(fscal,dz00);
 617
 618             /* Update vectorial force */
 619             fix0             = _mm256_add_pd(fix0,tx);
 620             fiy0             = _mm256_add_pd(fiy0,ty);
 621             fiz0             = _mm256_add_pd(fiz0,tz);
 622
 623             fjptrA             = f+j_coord_offsetA;
 624             fjptrB             = f+j_coord_offsetB;
 625             fjptrC             = f+j_coord_offsetC;
 626             fjptrD             = f+j_coord_offsetD;
 627             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 628
 629             /* Inner loop uses 47 flops */
 630         }
 631
 632         if(jidx<j_index_end)
 633         {
 634
 635             /* Get j neighbor index, and coordinate index */
 636             jnrlistA         = jjnr[jidx];
 637             jnrlistB         = jjnr[jidx+1];
 638             jnrlistC         = jjnr[jidx+2];
 639             jnrlistD         = jjnr[jidx+3];
 640             /* Sign of each element will be negative for non-real atoms.
 641              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 642              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 643              */
 644             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 645
 646             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 647             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 648             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 649
 650             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 651             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 652             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 653             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 654             j_coord_offsetA  = DIM*jnrA;
 655             j_coord_offsetB  = DIM*jnrB;
 656             j_coord_offsetC  = DIM*jnrC;
 657             j_coord_offsetD  = DIM*jnrD;
 658
 659             /* load j atom coordinates */
 660             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 661                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 662                                                  &jx0,&jy0,&jz0);
 663
 664             /* Calculate displacement vector */
 665             dx00             = _mm256_sub_pd(ix0,jx0);
 666             dy00             = _mm256_sub_pd(iy0,jy0);
 667             dz00             = _mm256_sub_pd(iz0,jz0);
 668
 669             /* Calculate squared distance and things based on it */
 670             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 671
 672             rinv00           = avx256_invsqrt_d(rsq00);
 673
 674             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 675
 676             /* Load parameters for j particles */
 677             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 678                                                                  charge+jnrC+0,charge+jnrD+0);
 679             vdwjidx0A        = 2*vdwtype[jnrA+0];
 680             vdwjidx0B        = 2*vdwtype[jnrB+0];
 681             vdwjidx0C        = 2*vdwtype[jnrC+0];
 682             vdwjidx0D        = 2*vdwtype[jnrD+0];
 683
 684             /**************************
 685              * CALCULATE INTERACTIONS *
 686              **************************/
 687
 688             r00              = _mm256_mul_pd(rsq00,rinv00);
 689             r00              = _mm256_andnot_pd(dummy_mask,r00);
 690
 691             /* Compute parameters for interactions between i and j atoms */
 692             qq00             = _mm256_mul_pd(iq0,jq0);
 693             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 694                                             vdwioffsetptr0+vdwjidx0B,
 695                                             vdwioffsetptr0+vdwjidx0C,
 696                                             vdwioffsetptr0+vdwjidx0D,
 697                                             &c6_00,&c12_00);
 698
 699             /* Calculate table index by multiplying r with table scale and truncate to integer */
 700             rt               = _mm256_mul_pd(r00,vftabscale);
 701             vfitab           = _mm256_cvttpd_epi32(rt);
 702             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 703             vfitab           = _mm_slli_epi32(vfitab,2);
 704
 705             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 706             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 707             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 708             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 709             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 710             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 711             Heps             = _mm256_mul_pd(vfeps,H);
 712             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 713             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 714             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 715
 716             /* LENNARD-JONES DISPERSION/REPULSION */
 717
 718             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 719             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 720
 721             fscal            = _mm256_add_pd(felec,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 48 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 7 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_ELEC_VDW_F,outeriter*7 + inneriter*48);
 761 }