src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCSTab_VdwLJ_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_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: CubicSplineTable
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCSTab_VdwLJ_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_elec->data;
 124     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->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,2);
 228
 229             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 230             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 231             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 232             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 233             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 234             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 235             Heps             = _mm256_mul_pd(vfeps,H);
 236             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 237             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 238             velec            = _mm256_mul_pd(qq00,VV);
 239             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 240             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 241
 242             /* LENNARD-JONES DISPERSION/REPULSION */
 243
 244             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 245             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 246             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 247             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 248             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 249
 250             /* Update potential sum for this i atom from the interaction with this j atom. */
 251             velecsum         = _mm256_add_pd(velecsum,velec);
 252             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 253
 254             fscal            = _mm256_add_pd(felec,fvdw);
 255
 256             /* Calculate temporary vectorial force */
 257             tx               = _mm256_mul_pd(fscal,dx00);
 258             ty               = _mm256_mul_pd(fscal,dy00);
 259             tz               = _mm256_mul_pd(fscal,dz00);
 260
 261             /* Update vectorial force */
 262             fix0             = _mm256_add_pd(fix0,tx);
 263             fiy0             = _mm256_add_pd(fiy0,ty);
 264             fiz0             = _mm256_add_pd(fiz0,tz);
 265
 266             fjptrA             = f+j_coord_offsetA;
 267             fjptrB             = f+j_coord_offsetB;
 268             fjptrC             = f+j_coord_offsetC;
 269             fjptrD             = f+j_coord_offsetD;
 270             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 271
 272             /* Inner loop uses 56 flops */
 273         }
 274
 275         if(jidx<j_index_end)
 276         {
 277
 278             /* Get j neighbor index, and coordinate index */
 279             jnrlistA         = jjnr[jidx];
 280             jnrlistB         = jjnr[jidx+1];
 281             jnrlistC         = jjnr[jidx+2];
 282             jnrlistD         = jjnr[jidx+3];
 283             /* Sign of each element will be negative for non-real atoms.
 284              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 285              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 286              */
 287             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 288
 289             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 290             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 291             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 292
 293             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 294             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 295             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 296             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 297             j_coord_offsetA  = DIM*jnrA;
 298             j_coord_offsetB  = DIM*jnrB;
 299             j_coord_offsetC  = DIM*jnrC;
 300             j_coord_offsetD  = DIM*jnrD;
 301
 302             /* load j atom coordinates */
 303             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 304                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 305                                                  &jx0,&jy0,&jz0);
 306
 307             /* Calculate displacement vector */
 308             dx00             = _mm256_sub_pd(ix0,jx0);
 309             dy00             = _mm256_sub_pd(iy0,jy0);
 310             dz00             = _mm256_sub_pd(iz0,jz0);
 311
 312             /* Calculate squared distance and things based on it */
 313             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 314
 315             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 316
 317             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 318
 319             /* Load parameters for j particles */
 320             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 321                                                                  charge+jnrC+0,charge+jnrD+0);
 322             vdwjidx0A        = 2*vdwtype[jnrA+0];
 323             vdwjidx0B        = 2*vdwtype[jnrB+0];
 324             vdwjidx0C        = 2*vdwtype[jnrC+0];
 325             vdwjidx0D        = 2*vdwtype[jnrD+0];
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             r00              = _mm256_mul_pd(rsq00,rinv00);
 332             r00              = _mm256_andnot_pd(dummy_mask,r00);
 333
 334             /* Compute parameters for interactions between i and j atoms */
 335             qq00             = _mm256_mul_pd(iq0,jq0);
 336             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 337                                             vdwioffsetptr0+vdwjidx0B,
 338                                             vdwioffsetptr0+vdwjidx0C,
 339                                             vdwioffsetptr0+vdwjidx0D,
 340                                             &c6_00,&c12_00);
 341
 342             /* Calculate table index by multiplying r with table scale and truncate to integer */
 343             rt               = _mm256_mul_pd(r00,vftabscale);
 344             vfitab           = _mm256_cvttpd_epi32(rt);
 345             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 346             vfitab           = _mm_slli_epi32(vfitab,2);
 347
 348             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 349             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 350             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 351             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 352             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 353             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 354             Heps             = _mm256_mul_pd(vfeps,H);
 355             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 356             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 357             velec            = _mm256_mul_pd(qq00,VV);
 358             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 359             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 360
 361             /* LENNARD-JONES DISPERSION/REPULSION */
 362
 363             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 364             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 365             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 366             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 367             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm256_andnot_pd(dummy_mask,velec);
 371             velecsum         = _mm256_add_pd(velecsum,velec);
 372             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 373             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 374
 375             fscal            = _mm256_add_pd(felec,fvdw);
 376
 377             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 378
 379             /* Calculate temporary vectorial force */
 380             tx               = _mm256_mul_pd(fscal,dx00);
 381             ty               = _mm256_mul_pd(fscal,dy00);
 382             tz               = _mm256_mul_pd(fscal,dz00);
 383
 384             /* Update vectorial force */
 385             fix0             = _mm256_add_pd(fix0,tx);
 386             fiy0             = _mm256_add_pd(fiy0,ty);
 387             fiz0             = _mm256_add_pd(fiz0,tz);
 388
 389             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 390             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 391             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 392             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 393             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 394
 395             /* Inner loop uses 57 flops */
 396         }
 397
 398         /* End of innermost loop */
 399
 400         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 401                                                  f+i_coord_offset,fshift+i_shift_offset);
 402
 403         ggid                        = gid[iidx];
 404         /* Update potential energies */
 405         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 406         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 407
 408         /* Increment number of inner iterations */
 409         inneriter                  += j_index_end - j_index_start;
 410
 411         /* Outer loop uses 9 flops */
 412     }
 413
 414     /* Increment number of outer iterations */
 415     outeriter        += nri;
 416
 417     /* Update outer/inner flops */
 418
 419     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*57);
 420 }
 421 /*
 422  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 423  * Electrostatics interaction: CubicSplineTable
 424  * VdW interaction:            LennardJones
 425  * Geometry:                   Particle-Particle
 426  * Calculate force/pot:        Force
 427  */
 428 void
 429 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 430                     (t_nblist                    * gmx_restrict       nlist,
 431                      rvec                        * gmx_restrict          xx,
 432                      rvec                        * gmx_restrict          ff,
 433                      t_forcerec                  * gmx_restrict          fr,
 434                      t_mdatoms                   * gmx_restrict     mdatoms,
 435                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 436                      t_nrnb                      * gmx_restrict        nrnb)
 437 {
 438     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 439      * just 0 for non-waters.
 440      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 441      * jnr indices corresponding to data put in the four positions in the SIMD register.
 442      */
 443     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 444     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 445     int              jnrA,jnrB,jnrC,jnrD;
 446     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 447     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 448     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 449     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 450     real             rcutoff_scalar;
 451     real             *shiftvec,*fshift,*x,*f;
 452     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 453     real             scratch[4*DIM];
 454     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 455     real *           vdwioffsetptr0;
 456     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 457     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 458     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 459     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 460     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 461     real             *charge;
 462     int              nvdwtype;
 463     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 464     int              *vdwtype;
 465     real             *vdwparam;
 466     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 467     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 468     __m128i          vfitab;
 469     __m128i          ifour       = _mm_set1_epi32(4);
 470     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 471     real             *vftab;
 472     __m256d          dummy_mask,cutoff_mask;
 473     __m128           tmpmask0,tmpmask1;
 474     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 475     __m256d          one     = _mm256_set1_pd(1.0);
 476     __m256d          two     = _mm256_set1_pd(2.0);
 477     x                = xx[0];
 478     f                = ff[0];
 479
 480     nri              = nlist->nri;
 481     iinr             = nlist->iinr;
 482     jindex           = nlist->jindex;
 483     jjnr             = nlist->jjnr;
 484     shiftidx         = nlist->shift;
 485     gid              = nlist->gid;
 486     shiftvec         = fr->shift_vec[0];
 487     fshift           = fr->fshift[0];
 488     facel            = _mm256_set1_pd(fr->epsfac);
 489     charge           = mdatoms->chargeA;
 490     nvdwtype         = fr->ntype;
 491     vdwparam         = fr->nbfp;
 492     vdwtype          = mdatoms->typeA;
 493
 494     vftab            = kernel_data->table_elec->data;
 495     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 496
 497     /* Avoid stupid compiler warnings */
 498     jnrA = jnrB = jnrC = jnrD = 0;
 499     j_coord_offsetA = 0;
 500     j_coord_offsetB = 0;
 501     j_coord_offsetC = 0;
 502     j_coord_offsetD = 0;
 503
 504     outeriter        = 0;
 505     inneriter        = 0;
 506
 507     for(iidx=0;iidx<4*DIM;iidx++)
 508     {
 509         scratch[iidx] = 0.0;
 510     }
 511
 512     /* Start outer loop over neighborlists */
 513     for(iidx=0; iidx<nri; iidx++)
 514     {
 515         /* Load shift vector for this list */
 516         i_shift_offset   = DIM*shiftidx[iidx];
 517
 518         /* Load limits for loop over neighbors */
 519         j_index_start    = jindex[iidx];
 520         j_index_end      = jindex[iidx+1];
 521
 522         /* Get outer coordinate index */
 523         inr              = iinr[iidx];
 524         i_coord_offset   = DIM*inr;
 525
 526         /* Load i particle coords and add shift vector */
 527         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 528
 529         fix0             = _mm256_setzero_pd();
 530         fiy0             = _mm256_setzero_pd();
 531         fiz0             = _mm256_setzero_pd();
 532
 533         /* Load parameters for i particles */
 534         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 535         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 536
 537         /* Start inner kernel loop */
 538         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 539         {
 540
 541             /* Get j neighbor index, and coordinate index */
 542             jnrA             = jjnr[jidx];
 543             jnrB             = jjnr[jidx+1];
 544             jnrC             = jjnr[jidx+2];
 545             jnrD             = jjnr[jidx+3];
 546             j_coord_offsetA  = DIM*jnrA;
 547             j_coord_offsetB  = DIM*jnrB;
 548             j_coord_offsetC  = DIM*jnrC;
 549             j_coord_offsetD  = DIM*jnrD;
 550
 551             /* load j atom coordinates */
 552             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 553                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 554                                                  &jx0,&jy0,&jz0);
 555
 556             /* Calculate displacement vector */
 557             dx00             = _mm256_sub_pd(ix0,jx0);
 558             dy00             = _mm256_sub_pd(iy0,jy0);
 559             dz00             = _mm256_sub_pd(iz0,jz0);
 560
 561             /* Calculate squared distance and things based on it */
 562             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 563
 564             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 565
 566             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 567
 568             /* Load parameters for j particles */
 569             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 570                                                                  charge+jnrC+0,charge+jnrD+0);
 571             vdwjidx0A        = 2*vdwtype[jnrA+0];
 572             vdwjidx0B        = 2*vdwtype[jnrB+0];
 573             vdwjidx0C        = 2*vdwtype[jnrC+0];
 574             vdwjidx0D        = 2*vdwtype[jnrD+0];
 575
 576             /**************************
 577              * CALCULATE INTERACTIONS *
 578              **************************/
 579
 580             r00              = _mm256_mul_pd(rsq00,rinv00);
 581
 582             /* Compute parameters for interactions between i and j atoms */
 583             qq00             = _mm256_mul_pd(iq0,jq0);
 584             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 585                                             vdwioffsetptr0+vdwjidx0B,
 586                                             vdwioffsetptr0+vdwjidx0C,
 587                                             vdwioffsetptr0+vdwjidx0D,
 588                                             &c6_00,&c12_00);
 589
 590             /* Calculate table index by multiplying r with table scale and truncate to integer */
 591             rt               = _mm256_mul_pd(r00,vftabscale);
 592             vfitab           = _mm256_cvttpd_epi32(rt);
 593             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 594             vfitab           = _mm_slli_epi32(vfitab,2);
 595
 596             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 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             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 606
 607             /* LENNARD-JONES DISPERSION/REPULSION */
 608
 609             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 610             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 611
 612             fscal            = _mm256_add_pd(felec,fvdw);
 613
 614             /* Calculate temporary vectorial force */
 615             tx               = _mm256_mul_pd(fscal,dx00);
 616             ty               = _mm256_mul_pd(fscal,dy00);
 617             tz               = _mm256_mul_pd(fscal,dz00);
 618
 619             /* Update vectorial force */
 620             fix0             = _mm256_add_pd(fix0,tx);
 621             fiy0             = _mm256_add_pd(fiy0,ty);
 622             fiz0             = _mm256_add_pd(fiz0,tz);
 623
 624             fjptrA             = f+j_coord_offsetA;
 625             fjptrB             = f+j_coord_offsetB;
 626             fjptrC             = f+j_coord_offsetC;
 627             fjptrD             = f+j_coord_offsetD;
 628             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 629
 630             /* Inner loop uses 47 flops */
 631         }
 632
 633         if(jidx<j_index_end)
 634         {
 635
 636             /* Get j neighbor index, and coordinate index */
 637             jnrlistA         = jjnr[jidx];
 638             jnrlistB         = jjnr[jidx+1];
 639             jnrlistC         = jjnr[jidx+2];
 640             jnrlistD         = jjnr[jidx+3];
 641             /* Sign of each element will be negative for non-real atoms.
 642              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 643              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 644              */
 645             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 646
 647             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 648             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 649             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 650
 651             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 652             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 653             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 654             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 655             j_coord_offsetA  = DIM*jnrA;
 656             j_coord_offsetB  = DIM*jnrB;
 657             j_coord_offsetC  = DIM*jnrC;
 658             j_coord_offsetD  = DIM*jnrD;
 659
 660             /* load j atom coordinates */
 661             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 662                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 663                                                  &jx0,&jy0,&jz0);
 664
 665             /* Calculate displacement vector */
 666             dx00             = _mm256_sub_pd(ix0,jx0);
 667             dy00             = _mm256_sub_pd(iy0,jy0);
 668             dz00             = _mm256_sub_pd(iz0,jz0);
 669
 670             /* Calculate squared distance and things based on it */
 671             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 672
 673             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 674
 675             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 676
 677             /* Load parameters for j particles */
 678             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 679                                                                  charge+jnrC+0,charge+jnrD+0);
 680             vdwjidx0A        = 2*vdwtype[jnrA+0];
 681             vdwjidx0B        = 2*vdwtype[jnrB+0];
 682             vdwjidx0C        = 2*vdwtype[jnrC+0];
 683             vdwjidx0D        = 2*vdwtype[jnrD+0];
 684
 685             /**************************
 686              * CALCULATE INTERACTIONS *
 687              **************************/
 688
 689             r00              = _mm256_mul_pd(rsq00,rinv00);
 690             r00              = _mm256_andnot_pd(dummy_mask,r00);
 691
 692             /* Compute parameters for interactions between i and j atoms */
 693             qq00             = _mm256_mul_pd(iq0,jq0);
 694             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 695                                             vdwioffsetptr0+vdwjidx0B,
 696                                             vdwioffsetptr0+vdwjidx0C,
 697                                             vdwioffsetptr0+vdwjidx0D,
 698                                             &c6_00,&c12_00);
 699
 700             /* Calculate table index by multiplying r with table scale and truncate to integer */
 701             rt               = _mm256_mul_pd(r00,vftabscale);
 702             vfitab           = _mm256_cvttpd_epi32(rt);
 703             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 704             vfitab           = _mm_slli_epi32(vfitab,2);
 705
 706             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 707             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 708             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 709             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 710             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 711             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 712             Heps             = _mm256_mul_pd(vfeps,H);
 713             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 714             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 715             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 716
 717             /* LENNARD-JONES DISPERSION/REPULSION */
 718
 719             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 720             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 721
 722             fscal            = _mm256_add_pd(felec,fvdw);
 723
 724             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 725
 726             /* Calculate temporary vectorial force */
 727             tx               = _mm256_mul_pd(fscal,dx00);
 728             ty               = _mm256_mul_pd(fscal,dy00);
 729             tz               = _mm256_mul_pd(fscal,dz00);
 730
 731             /* Update vectorial force */
 732             fix0             = _mm256_add_pd(fix0,tx);
 733             fiy0             = _mm256_add_pd(fiy0,ty);
 734             fiz0             = _mm256_add_pd(fiz0,tz);
 735
 736             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 737             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 738             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 739             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 740             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 741
 742             /* Inner loop uses 48 flops */
 743         }
 744
 745         /* End of innermost loop */
 746
 747         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 748                                                  f+i_coord_offset,fshift+i_shift_offset);
 749
 750         /* Increment number of inner iterations */
 751         inneriter                  += j_index_end - j_index_start;
 752
 753         /* Outer loop uses 7 flops */
 754     }
 755
 756     /* Increment number of outer iterations */
 757     outeriter        += nri;
 758
 759     /* Update outer/inner flops */
 760
 761     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*48);
 762 }