src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCoul_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 "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "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_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_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     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  90     real             *charge;
  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     facel            = _mm256_set1_pd(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_vdw->data;
 124     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = jnrC = jnrD = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130     j_coord_offsetC = 0;
 131     j_coord_offsetD = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     for(iidx=0;iidx<4*DIM;iidx++)
 137     {
 138         scratch[iidx] = 0.0;
 139     }
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 157
 158         fix0             = _mm256_setzero_pd();
 159         fiy0             = _mm256_setzero_pd();
 160         fiz0             = _mm256_setzero_pd();
 161
 162         /* Load parameters for i particles */
 163         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 164         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 165
 166         /* Reset potential sums */
 167         velecsum         = _mm256_setzero_pd();
 168         vvdwsum          = _mm256_setzero_pd();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             jnrC             = jjnr[jidx+2];
 178             jnrD             = jjnr[jidx+3];
 179             j_coord_offsetA  = DIM*jnrA;
 180             j_coord_offsetB  = DIM*jnrB;
 181             j_coord_offsetC  = DIM*jnrC;
 182             j_coord_offsetD  = DIM*jnrD;
 183
 184             /* load j atom coordinates */
 185             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 186                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 187                                                  &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _mm256_sub_pd(ix0,jx0);
 191             dy00             = _mm256_sub_pd(iy0,jy0);
 192             dz00             = _mm256_sub_pd(iz0,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 196
 197             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 198
 199             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 200
 201             /* Load parameters for j particles */
 202             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 203                                                                  charge+jnrC+0,charge+jnrD+0);
 204             vdwjidx0A        = 2*vdwtype[jnrA+0];
 205             vdwjidx0B        = 2*vdwtype[jnrB+0];
 206             vdwjidx0C        = 2*vdwtype[jnrC+0];
 207             vdwjidx0D        = 2*vdwtype[jnrD+0];
 208
 209             /**************************
 210              * CALCULATE INTERACTIONS *
 211              **************************/
 212
 213             r00              = _mm256_mul_pd(rsq00,rinv00);
 214
 215             /* Compute parameters for interactions between i and j atoms */
 216             qq00             = _mm256_mul_pd(iq0,jq0);
 217             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 218                                             vdwioffsetptr0+vdwjidx0B,
 219                                             vdwioffsetptr0+vdwjidx0C,
 220                                             vdwioffsetptr0+vdwjidx0D,
 221                                             &c6_00,&c12_00);
 222
 223             /* Calculate table index by multiplying r with table scale and truncate to integer */
 224             rt               = _mm256_mul_pd(r00,vftabscale);
 225             vfitab           = _mm256_cvttpd_epi32(rt);
 226             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 227             vfitab           = _mm_slli_epi32(vfitab,3);
 228
 229             /* COULOMB ELECTROSTATICS */
 230             velec            = _mm256_mul_pd(qq00,rinv00);
 231             felec            = _mm256_mul_pd(velec,rinvsq00);
 232
 233             /* CUBIC SPLINE TABLE DISPERSION */
 234             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 235             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 236             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 237             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 238             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 239             Heps             = _mm256_mul_pd(vfeps,H);
 240             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 241             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 242             vvdw6            = _mm256_mul_pd(c6_00,VV);
 243             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 244             fvdw6            = _mm256_mul_pd(c6_00,FF);
 245
 246             /* CUBIC SPLINE TABLE REPULSION */
 247             vfitab           = _mm_add_epi32(vfitab,ifour);
 248             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 249             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 250             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 251             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 252             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 253             Heps             = _mm256_mul_pd(vfeps,H);
 254             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 255             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 256             vvdw12           = _mm256_mul_pd(c12_00,VV);
 257             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 258             fvdw12           = _mm256_mul_pd(c12_00,FF);
 259             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 260             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 261
 262             /* Update potential sum for this i atom from the interaction with this j atom. */
 263             velecsum         = _mm256_add_pd(velecsum,velec);
 264             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 265
 266             fscal            = _mm256_add_pd(felec,fvdw);
 267
 268             /* Calculate temporary vectorial force */
 269             tx               = _mm256_mul_pd(fscal,dx00);
 270             ty               = _mm256_mul_pd(fscal,dy00);
 271             tz               = _mm256_mul_pd(fscal,dz00);
 272
 273             /* Update vectorial force */
 274             fix0             = _mm256_add_pd(fix0,tx);
 275             fiy0             = _mm256_add_pd(fiy0,ty);
 276             fiz0             = _mm256_add_pd(fiz0,tz);
 277
 278             fjptrA             = f+j_coord_offsetA;
 279             fjptrB             = f+j_coord_offsetB;
 280             fjptrC             = f+j_coord_offsetC;
 281             fjptrD             = f+j_coord_offsetD;
 282             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 283
 284             /* Inner loop uses 62 flops */
 285         }
 286
 287         if(jidx<j_index_end)
 288         {
 289
 290             /* Get j neighbor index, and coordinate index */
 291             jnrlistA         = jjnr[jidx];
 292             jnrlistB         = jjnr[jidx+1];
 293             jnrlistC         = jjnr[jidx+2];
 294             jnrlistD         = jjnr[jidx+3];
 295             /* Sign of each element will be negative for non-real atoms.
 296              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 297              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 298              */
 299             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 300
 301             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 302             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 303             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 304
 305             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 306             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 307             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 308             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 309             j_coord_offsetA  = DIM*jnrA;
 310             j_coord_offsetB  = DIM*jnrB;
 311             j_coord_offsetC  = DIM*jnrC;
 312             j_coord_offsetD  = DIM*jnrD;
 313
 314             /* load j atom coordinates */
 315             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 316                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 317                                                  &jx0,&jy0,&jz0);
 318
 319             /* Calculate displacement vector */
 320             dx00             = _mm256_sub_pd(ix0,jx0);
 321             dy00             = _mm256_sub_pd(iy0,jy0);
 322             dz00             = _mm256_sub_pd(iz0,jz0);
 323
 324             /* Calculate squared distance and things based on it */
 325             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 326
 327             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 328
 329             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 330
 331             /* Load parameters for j particles */
 332             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 333                                                                  charge+jnrC+0,charge+jnrD+0);
 334             vdwjidx0A        = 2*vdwtype[jnrA+0];
 335             vdwjidx0B        = 2*vdwtype[jnrB+0];
 336             vdwjidx0C        = 2*vdwtype[jnrC+0];
 337             vdwjidx0D        = 2*vdwtype[jnrD+0];
 338
 339             /**************************
 340              * CALCULATE INTERACTIONS *
 341              **************************/
 342
 343             r00              = _mm256_mul_pd(rsq00,rinv00);
 344             r00              = _mm256_andnot_pd(dummy_mask,r00);
 345
 346             /* Compute parameters for interactions between i and j atoms */
 347             qq00             = _mm256_mul_pd(iq0,jq0);
 348             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 349                                             vdwioffsetptr0+vdwjidx0B,
 350                                             vdwioffsetptr0+vdwjidx0C,
 351                                             vdwioffsetptr0+vdwjidx0D,
 352                                             &c6_00,&c12_00);
 353
 354             /* Calculate table index by multiplying r with table scale and truncate to integer */
 355             rt               = _mm256_mul_pd(r00,vftabscale);
 356             vfitab           = _mm256_cvttpd_epi32(rt);
 357             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 358             vfitab           = _mm_slli_epi32(vfitab,3);
 359
 360             /* COULOMB ELECTROSTATICS */
 361             velec            = _mm256_mul_pd(qq00,rinv00);
 362             felec            = _mm256_mul_pd(velec,rinvsq00);
 363
 364             /* CUBIC SPLINE TABLE DISPERSION */
 365             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 366             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 367             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 368             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 369             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 370             Heps             = _mm256_mul_pd(vfeps,H);
 371             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 372             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 373             vvdw6            = _mm256_mul_pd(c6_00,VV);
 374             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 375             fvdw6            = _mm256_mul_pd(c6_00,FF);
 376
 377             /* CUBIC SPLINE TABLE REPULSION */
 378             vfitab           = _mm_add_epi32(vfitab,ifour);
 379             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 380             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 381             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 382             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 383             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 384             Heps             = _mm256_mul_pd(vfeps,H);
 385             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 386             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 387             vvdw12           = _mm256_mul_pd(c12_00,VV);
 388             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 389             fvdw12           = _mm256_mul_pd(c12_00,FF);
 390             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 391             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 392
 393             /* Update potential sum for this i atom from the interaction with this j atom. */
 394             velec            = _mm256_andnot_pd(dummy_mask,velec);
 395             velecsum         = _mm256_add_pd(velecsum,velec);
 396             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 397             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 398
 399             fscal            = _mm256_add_pd(felec,fvdw);
 400
 401             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 402
 403             /* Calculate temporary vectorial force */
 404             tx               = _mm256_mul_pd(fscal,dx00);
 405             ty               = _mm256_mul_pd(fscal,dy00);
 406             tz               = _mm256_mul_pd(fscal,dz00);
 407
 408             /* Update vectorial force */
 409             fix0             = _mm256_add_pd(fix0,tx);
 410             fiy0             = _mm256_add_pd(fiy0,ty);
 411             fiz0             = _mm256_add_pd(fiz0,tz);
 412
 413             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 414             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 415             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 416             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 417             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 418
 419             /* Inner loop uses 63 flops */
 420         }
 421
 422         /* End of innermost loop */
 423
 424         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 425                                                  f+i_coord_offset,fshift+i_shift_offset);
 426
 427         ggid                        = gid[iidx];
 428         /* Update potential energies */
 429         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 430         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 431
 432         /* Increment number of inner iterations */
 433         inneriter                  += j_index_end - j_index_start;
 434
 435         /* Outer loop uses 9 flops */
 436     }
 437
 438     /* Increment number of outer iterations */
 439     outeriter        += nri;
 440
 441     /* Update outer/inner flops */
 442
 443     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*63);
 444 }
 445 /*
 446  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_256_double
 447  * Electrostatics interaction: Coulomb
 448  * VdW interaction:            CubicSplineTable
 449  * Geometry:                   Particle-Particle
 450  * Calculate force/pot:        Force
 451  */
 452 void
 453 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_256_double
 454                     (t_nblist                    * gmx_restrict       nlist,
 455                      rvec                        * gmx_restrict          xx,
 456                      rvec                        * gmx_restrict          ff,
 457                      t_forcerec                  * gmx_restrict          fr,
 458                      t_mdatoms                   * gmx_restrict     mdatoms,
 459                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 460                      t_nrnb                      * gmx_restrict        nrnb)
 461 {
 462     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 463      * just 0 for non-waters.
 464      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 465      * jnr indices corresponding to data put in the four positions in the SIMD register.
 466      */
 467     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 468     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 469     int              jnrA,jnrB,jnrC,jnrD;
 470     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 471     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 472     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 473     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 474     real             rcutoff_scalar;
 475     real             *shiftvec,*fshift,*x,*f;
 476     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 477     real             scratch[4*DIM];
 478     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 479     real *           vdwioffsetptr0;
 480     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 481     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 482     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 483     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 484     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 485     real             *charge;
 486     int              nvdwtype;
 487     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 488     int              *vdwtype;
 489     real             *vdwparam;
 490     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 491     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 492     __m128i          vfitab;
 493     __m128i          ifour       = _mm_set1_epi32(4);
 494     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 495     real             *vftab;
 496     __m256d          dummy_mask,cutoff_mask;
 497     __m128           tmpmask0,tmpmask1;
 498     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 499     __m256d          one     = _mm256_set1_pd(1.0);
 500     __m256d          two     = _mm256_set1_pd(2.0);
 501     x                = xx[0];
 502     f                = ff[0];
 503
 504     nri              = nlist->nri;
 505     iinr             = nlist->iinr;
 506     jindex           = nlist->jindex;
 507     jjnr             = nlist->jjnr;
 508     shiftidx         = nlist->shift;
 509     gid              = nlist->gid;
 510     shiftvec         = fr->shift_vec[0];
 511     fshift           = fr->fshift[0];
 512     facel            = _mm256_set1_pd(fr->epsfac);
 513     charge           = mdatoms->chargeA;
 514     nvdwtype         = fr->ntype;
 515     vdwparam         = fr->nbfp;
 516     vdwtype          = mdatoms->typeA;
 517
 518     vftab            = kernel_data->table_vdw->data;
 519     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 520
 521     /* Avoid stupid compiler warnings */
 522     jnrA = jnrB = jnrC = jnrD = 0;
 523     j_coord_offsetA = 0;
 524     j_coord_offsetB = 0;
 525     j_coord_offsetC = 0;
 526     j_coord_offsetD = 0;
 527
 528     outeriter        = 0;
 529     inneriter        = 0;
 530
 531     for(iidx=0;iidx<4*DIM;iidx++)
 532     {
 533         scratch[iidx] = 0.0;
 534     }
 535
 536     /* Start outer loop over neighborlists */
 537     for(iidx=0; iidx<nri; iidx++)
 538     {
 539         /* Load shift vector for this list */
 540         i_shift_offset   = DIM*shiftidx[iidx];
 541
 542         /* Load limits for loop over neighbors */
 543         j_index_start    = jindex[iidx];
 544         j_index_end      = jindex[iidx+1];
 545
 546         /* Get outer coordinate index */
 547         inr              = iinr[iidx];
 548         i_coord_offset   = DIM*inr;
 549
 550         /* Load i particle coords and add shift vector */
 551         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 552
 553         fix0             = _mm256_setzero_pd();
 554         fiy0             = _mm256_setzero_pd();
 555         fiz0             = _mm256_setzero_pd();
 556
 557         /* Load parameters for i particles */
 558         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 559         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 560
 561         /* Start inner kernel loop */
 562         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 563         {
 564
 565             /* Get j neighbor index, and coordinate index */
 566             jnrA             = jjnr[jidx];
 567             jnrB             = jjnr[jidx+1];
 568             jnrC             = jjnr[jidx+2];
 569             jnrD             = jjnr[jidx+3];
 570             j_coord_offsetA  = DIM*jnrA;
 571             j_coord_offsetB  = DIM*jnrB;
 572             j_coord_offsetC  = DIM*jnrC;
 573             j_coord_offsetD  = DIM*jnrD;
 574
 575             /* load j atom coordinates */
 576             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 577                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 578                                                  &jx0,&jy0,&jz0);
 579
 580             /* Calculate displacement vector */
 581             dx00             = _mm256_sub_pd(ix0,jx0);
 582             dy00             = _mm256_sub_pd(iy0,jy0);
 583             dz00             = _mm256_sub_pd(iz0,jz0);
 584
 585             /* Calculate squared distance and things based on it */
 586             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 587
 588             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 589
 590             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 591
 592             /* Load parameters for j particles */
 593             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 594                                                                  charge+jnrC+0,charge+jnrD+0);
 595             vdwjidx0A        = 2*vdwtype[jnrA+0];
 596             vdwjidx0B        = 2*vdwtype[jnrB+0];
 597             vdwjidx0C        = 2*vdwtype[jnrC+0];
 598             vdwjidx0D        = 2*vdwtype[jnrD+0];
 599
 600             /**************************
 601              * CALCULATE INTERACTIONS *
 602              **************************/
 603
 604             r00              = _mm256_mul_pd(rsq00,rinv00);
 605
 606             /* Compute parameters for interactions between i and j atoms */
 607             qq00             = _mm256_mul_pd(iq0,jq0);
 608             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 609                                             vdwioffsetptr0+vdwjidx0B,
 610                                             vdwioffsetptr0+vdwjidx0C,
 611                                             vdwioffsetptr0+vdwjidx0D,
 612                                             &c6_00,&c12_00);
 613
 614             /* Calculate table index by multiplying r with table scale and truncate to integer */
 615             rt               = _mm256_mul_pd(r00,vftabscale);
 616             vfitab           = _mm256_cvttpd_epi32(rt);
 617             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 618             vfitab           = _mm_slli_epi32(vfitab,3);
 619
 620             /* COULOMB ELECTROSTATICS */
 621             velec            = _mm256_mul_pd(qq00,rinv00);
 622             felec            = _mm256_mul_pd(velec,rinvsq00);
 623
 624             /* CUBIC SPLINE TABLE DISPERSION */
 625             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 626             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 627             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 628             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 629             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 630             Heps             = _mm256_mul_pd(vfeps,H);
 631             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 632             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 633             fvdw6            = _mm256_mul_pd(c6_00,FF);
 634
 635             /* CUBIC SPLINE TABLE REPULSION */
 636             vfitab           = _mm_add_epi32(vfitab,ifour);
 637             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 638             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 639             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 640             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 641             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 642             Heps             = _mm256_mul_pd(vfeps,H);
 643             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 644             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 645             fvdw12           = _mm256_mul_pd(c12_00,FF);
 646             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 647
 648             fscal            = _mm256_add_pd(felec,fvdw);
 649
 650             /* Calculate temporary vectorial force */
 651             tx               = _mm256_mul_pd(fscal,dx00);
 652             ty               = _mm256_mul_pd(fscal,dy00);
 653             tz               = _mm256_mul_pd(fscal,dz00);
 654
 655             /* Update vectorial force */
 656             fix0             = _mm256_add_pd(fix0,tx);
 657             fiy0             = _mm256_add_pd(fiy0,ty);
 658             fiz0             = _mm256_add_pd(fiz0,tz);
 659
 660             fjptrA             = f+j_coord_offsetA;
 661             fjptrB             = f+j_coord_offsetB;
 662             fjptrC             = f+j_coord_offsetC;
 663             fjptrD             = f+j_coord_offsetD;
 664             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 665
 666             /* Inner loop uses 53 flops */
 667         }
 668
 669         if(jidx<j_index_end)
 670         {
 671
 672             /* Get j neighbor index, and coordinate index */
 673             jnrlistA         = jjnr[jidx];
 674             jnrlistB         = jjnr[jidx+1];
 675             jnrlistC         = jjnr[jidx+2];
 676             jnrlistD         = jjnr[jidx+3];
 677             /* Sign of each element will be negative for non-real atoms.
 678              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 679              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 680              */
 681             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 682
 683             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 684             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 685             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 686
 687             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 688             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 689             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 690             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 691             j_coord_offsetA  = DIM*jnrA;
 692             j_coord_offsetB  = DIM*jnrB;
 693             j_coord_offsetC  = DIM*jnrC;
 694             j_coord_offsetD  = DIM*jnrD;
 695
 696             /* load j atom coordinates */
 697             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 698                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 699                                                  &jx0,&jy0,&jz0);
 700
 701             /* Calculate displacement vector */
 702             dx00             = _mm256_sub_pd(ix0,jx0);
 703             dy00             = _mm256_sub_pd(iy0,jy0);
 704             dz00             = _mm256_sub_pd(iz0,jz0);
 705
 706             /* Calculate squared distance and things based on it */
 707             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 708
 709             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 710
 711             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 712
 713             /* Load parameters for j particles */
 714             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 715                                                                  charge+jnrC+0,charge+jnrD+0);
 716             vdwjidx0A        = 2*vdwtype[jnrA+0];
 717             vdwjidx0B        = 2*vdwtype[jnrB+0];
 718             vdwjidx0C        = 2*vdwtype[jnrC+0];
 719             vdwjidx0D        = 2*vdwtype[jnrD+0];
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             r00              = _mm256_mul_pd(rsq00,rinv00);
 726             r00              = _mm256_andnot_pd(dummy_mask,r00);
 727
 728             /* Compute parameters for interactions between i and j atoms */
 729             qq00             = _mm256_mul_pd(iq0,jq0);
 730             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 731                                             vdwioffsetptr0+vdwjidx0B,
 732                                             vdwioffsetptr0+vdwjidx0C,
 733                                             vdwioffsetptr0+vdwjidx0D,
 734                                             &c6_00,&c12_00);
 735
 736             /* Calculate table index by multiplying r with table scale and truncate to integer */
 737             rt               = _mm256_mul_pd(r00,vftabscale);
 738             vfitab           = _mm256_cvttpd_epi32(rt);
 739             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 740             vfitab           = _mm_slli_epi32(vfitab,3);
 741
 742             /* COULOMB ELECTROSTATICS */
 743             velec            = _mm256_mul_pd(qq00,rinv00);
 744             felec            = _mm256_mul_pd(velec,rinvsq00);
 745
 746             /* CUBIC SPLINE TABLE DISPERSION */
 747             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 748             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 749             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 750             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 751             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 752             Heps             = _mm256_mul_pd(vfeps,H);
 753             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 754             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 755             fvdw6            = _mm256_mul_pd(c6_00,FF);
 756
 757             /* CUBIC SPLINE TABLE REPULSION */
 758             vfitab           = _mm_add_epi32(vfitab,ifour);
 759             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 760             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 761             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 762             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 763             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 764             Heps             = _mm256_mul_pd(vfeps,H);
 765             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 766             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 767             fvdw12           = _mm256_mul_pd(c12_00,FF);
 768             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 769
 770             fscal            = _mm256_add_pd(felec,fvdw);
 771
 772             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 773
 774             /* Calculate temporary vectorial force */
 775             tx               = _mm256_mul_pd(fscal,dx00);
 776             ty               = _mm256_mul_pd(fscal,dy00);
 777             tz               = _mm256_mul_pd(fscal,dz00);
 778
 779             /* Update vectorial force */
 780             fix0             = _mm256_add_pd(fix0,tx);
 781             fiy0             = _mm256_add_pd(fiy0,ty);
 782             fiz0             = _mm256_add_pd(fiz0,tz);
 783
 784             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 785             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 786             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 787             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 788             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 789
 790             /* Inner loop uses 54 flops */
 791         }
 792
 793         /* End of innermost loop */
 794
 795         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 796                                                  f+i_coord_offset,fshift+i_shift_offset);
 797
 798         /* Increment number of inner iterations */
 799         inneriter                  += j_index_end - j_index_start;
 800
 801         /* Outer loop uses 7 flops */
 802     }
 803
 804     /* Increment number of outer iterations */
 805     outeriter        += nri;
 806
 807     /* Update outer/inner flops */
 808
 809     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*54);
 810 }