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, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: CubicSplineTable
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_256_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  79     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  84     real             scratch[4*DIM];
  85     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  98     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  99     __m128i          vfitab;
 100     __m128i          ifour       = _mm_set1_epi32(4);
 101     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 102     real             *vftab;
 103     __m256d          dummy_mask,cutoff_mask;
 104     __m128           tmpmask0,tmpmask1;
 105     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 106     __m256d          one     = _mm256_set1_pd(1.0);
 107     __m256d          two     = _mm256_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm256_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     vftab            = kernel_data->table_elec->data;
 126     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 127
 128     /* Avoid stupid compiler warnings */
 129     jnrA = jnrB = jnrC = jnrD = 0;
 130     j_coord_offsetA = 0;
 131     j_coord_offsetB = 0;
 132     j_coord_offsetC = 0;
 133     j_coord_offsetD = 0;
 134
 135     outeriter        = 0;
 136     inneriter        = 0;
 137
 138     for(iidx=0;iidx<4*DIM;iidx++)
 139     {
 140         scratch[iidx] = 0.0;
 141     }
 142
 143     /* Start outer loop over neighborlists */
 144     for(iidx=0; iidx<nri; iidx++)
 145     {
 146         /* Load shift vector for this list */
 147         i_shift_offset   = DIM*shiftidx[iidx];
 148
 149         /* Load limits for loop over neighbors */
 150         j_index_start    = jindex[iidx];
 151         j_index_end      = jindex[iidx+1];
 152
 153         /* Get outer coordinate index */
 154         inr              = iinr[iidx];
 155         i_coord_offset   = DIM*inr;
 156
 157         /* Load i particle coords and add shift vector */
 158         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 159
 160         fix0             = _mm256_setzero_pd();
 161         fiy0             = _mm256_setzero_pd();
 162         fiz0             = _mm256_setzero_pd();
 163
 164         /* Load parameters for i particles */
 165         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 166         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 167
 168         /* Reset potential sums */
 169         velecsum         = _mm256_setzero_pd();
 170         vvdwsum          = _mm256_setzero_pd();
 171
 172         /* Start inner kernel loop */
 173         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 174         {
 175
 176             /* Get j neighbor index, and coordinate index */
 177             jnrA             = jjnr[jidx];
 178             jnrB             = jjnr[jidx+1];
 179             jnrC             = jjnr[jidx+2];
 180             jnrD             = jjnr[jidx+3];
 181             j_coord_offsetA  = DIM*jnrA;
 182             j_coord_offsetB  = DIM*jnrB;
 183             j_coord_offsetC  = DIM*jnrC;
 184             j_coord_offsetD  = DIM*jnrD;
 185
 186             /* load j atom coordinates */
 187             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 188                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 189                                                  &jx0,&jy0,&jz0);
 190
 191             /* Calculate displacement vector */
 192             dx00             = _mm256_sub_pd(ix0,jx0);
 193             dy00             = _mm256_sub_pd(iy0,jy0);
 194             dz00             = _mm256_sub_pd(iz0,jz0);
 195
 196             /* Calculate squared distance and things based on it */
 197             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 198
 199             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 200
 201             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 202
 203             /* Load parameters for j particles */
 204             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 205                                                                  charge+jnrC+0,charge+jnrD+0);
 206             vdwjidx0A        = 2*vdwtype[jnrA+0];
 207             vdwjidx0B        = 2*vdwtype[jnrB+0];
 208             vdwjidx0C        = 2*vdwtype[jnrC+0];
 209             vdwjidx0D        = 2*vdwtype[jnrD+0];
 210
 211             /**************************
 212              * CALCULATE INTERACTIONS *
 213              **************************/
 214
 215             r00              = _mm256_mul_pd(rsq00,rinv00);
 216
 217             /* Compute parameters for interactions between i and j atoms */
 218             qq00             = _mm256_mul_pd(iq0,jq0);
 219             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 220                                             vdwioffsetptr0+vdwjidx0B,
 221                                             vdwioffsetptr0+vdwjidx0C,
 222                                             vdwioffsetptr0+vdwjidx0D,
 223                                             &c6_00,&c12_00);
 224
 225             /* Calculate table index by multiplying r with table scale and truncate to integer */
 226             rt               = _mm256_mul_pd(r00,vftabscale);
 227             vfitab           = _mm256_cvttpd_epi32(rt);
 228             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 229             vfitab           = _mm_slli_epi32(vfitab,2);
 230
 231             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 232             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 233             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 234             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 235             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 236             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 237             Heps             = _mm256_mul_pd(vfeps,H);
 238             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 239             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 240             velec            = _mm256_mul_pd(qq00,VV);
 241             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 242             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 243
 244             /* LENNARD-JONES DISPERSION/REPULSION */
 245
 246             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 247             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 248             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 249             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 250             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 251
 252             /* Update potential sum for this i atom from the interaction with this j atom. */
 253             velecsum         = _mm256_add_pd(velecsum,velec);
 254             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 255
 256             fscal            = _mm256_add_pd(felec,fvdw);
 257
 258             /* Calculate temporary vectorial force */
 259             tx               = _mm256_mul_pd(fscal,dx00);
 260             ty               = _mm256_mul_pd(fscal,dy00);
 261             tz               = _mm256_mul_pd(fscal,dz00);
 262
 263             /* Update vectorial force */
 264             fix0             = _mm256_add_pd(fix0,tx);
 265             fiy0             = _mm256_add_pd(fiy0,ty);
 266             fiz0             = _mm256_add_pd(fiz0,tz);
 267
 268             fjptrA             = f+j_coord_offsetA;
 269             fjptrB             = f+j_coord_offsetB;
 270             fjptrC             = f+j_coord_offsetC;
 271             fjptrD             = f+j_coord_offsetD;
 272             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 273
 274             /* Inner loop uses 56 flops */
 275         }
 276
 277         if(jidx<j_index_end)
 278         {
 279
 280             /* Get j neighbor index, and coordinate index */
 281             jnrlistA         = jjnr[jidx];
 282             jnrlistB         = jjnr[jidx+1];
 283             jnrlistC         = jjnr[jidx+2];
 284             jnrlistD         = jjnr[jidx+3];
 285             /* Sign of each element will be negative for non-real atoms.
 286              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 287              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 288              */
 289             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 290
 291             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 292             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 293             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 294
 295             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 296             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 297             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 298             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 299             j_coord_offsetA  = DIM*jnrA;
 300             j_coord_offsetB  = DIM*jnrB;
 301             j_coord_offsetC  = DIM*jnrC;
 302             j_coord_offsetD  = DIM*jnrD;
 303
 304             /* load j atom coordinates */
 305             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 306                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 307                                                  &jx0,&jy0,&jz0);
 308
 309             /* Calculate displacement vector */
 310             dx00             = _mm256_sub_pd(ix0,jx0);
 311             dy00             = _mm256_sub_pd(iy0,jy0);
 312             dz00             = _mm256_sub_pd(iz0,jz0);
 313
 314             /* Calculate squared distance and things based on it */
 315             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 316
 317             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 318
 319             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 320
 321             /* Load parameters for j particles */
 322             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 323                                                                  charge+jnrC+0,charge+jnrD+0);
 324             vdwjidx0A        = 2*vdwtype[jnrA+0];
 325             vdwjidx0B        = 2*vdwtype[jnrB+0];
 326             vdwjidx0C        = 2*vdwtype[jnrC+0];
 327             vdwjidx0D        = 2*vdwtype[jnrD+0];
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             r00              = _mm256_mul_pd(rsq00,rinv00);
 334             r00              = _mm256_andnot_pd(dummy_mask,r00);
 335
 336             /* Compute parameters for interactions between i and j atoms */
 337             qq00             = _mm256_mul_pd(iq0,jq0);
 338             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 339                                             vdwioffsetptr0+vdwjidx0B,
 340                                             vdwioffsetptr0+vdwjidx0C,
 341                                             vdwioffsetptr0+vdwjidx0D,
 342                                             &c6_00,&c12_00);
 343
 344             /* Calculate table index by multiplying r with table scale and truncate to integer */
 345             rt               = _mm256_mul_pd(r00,vftabscale);
 346             vfitab           = _mm256_cvttpd_epi32(rt);
 347             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 348             vfitab           = _mm_slli_epi32(vfitab,2);
 349
 350             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 351             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 352             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 353             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 354             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 355             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 356             Heps             = _mm256_mul_pd(vfeps,H);
 357             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 358             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 359             velec            = _mm256_mul_pd(qq00,VV);
 360             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 361             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 362
 363             /* LENNARD-JONES DISPERSION/REPULSION */
 364
 365             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 366             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 367             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 368             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 369             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velec            = _mm256_andnot_pd(dummy_mask,velec);
 373             velecsum         = _mm256_add_pd(velecsum,velec);
 374             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 375             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 376
 377             fscal            = _mm256_add_pd(felec,fvdw);
 378
 379             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = _mm256_mul_pd(fscal,dx00);
 383             ty               = _mm256_mul_pd(fscal,dy00);
 384             tz               = _mm256_mul_pd(fscal,dz00);
 385
 386             /* Update vectorial force */
 387             fix0             = _mm256_add_pd(fix0,tx);
 388             fiy0             = _mm256_add_pd(fiy0,ty);
 389             fiz0             = _mm256_add_pd(fiz0,tz);
 390
 391             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 392             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 393             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 394             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 395             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 396
 397             /* Inner loop uses 57 flops */
 398         }
 399
 400         /* End of innermost loop */
 401
 402         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 403                                                  f+i_coord_offset,fshift+i_shift_offset);
 404
 405         ggid                        = gid[iidx];
 406         /* Update potential energies */
 407         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 408         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 409
 410         /* Increment number of inner iterations */
 411         inneriter                  += j_index_end - j_index_start;
 412
 413         /* Outer loop uses 9 flops */
 414     }
 415
 416     /* Increment number of outer iterations */
 417     outeriter        += nri;
 418
 419     /* Update outer/inner flops */
 420
 421     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*57);
 422 }
 423 /*
 424  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 425  * Electrostatics interaction: CubicSplineTable
 426  * VdW interaction:            LennardJones
 427  * Geometry:                   Particle-Particle
 428  * Calculate force/pot:        Force
 429  */
 430 void
 431 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 432                     (t_nblist                    * gmx_restrict       nlist,
 433                      rvec                        * gmx_restrict          xx,
 434                      rvec                        * gmx_restrict          ff,
 435                      t_forcerec                  * gmx_restrict          fr,
 436                      t_mdatoms                   * gmx_restrict     mdatoms,
 437                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 438                      t_nrnb                      * gmx_restrict        nrnb)
 439 {
 440     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 441      * just 0 for non-waters.
 442      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 443      * jnr indices corresponding to data put in the four positions in the SIMD register.
 444      */
 445     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 446     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 447     int              jnrA,jnrB,jnrC,jnrD;
 448     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 449     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 450     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 451     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 452     real             rcutoff_scalar;
 453     real             *shiftvec,*fshift,*x,*f;
 454     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 455     real             scratch[4*DIM];
 456     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 457     real *           vdwioffsetptr0;
 458     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 459     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 460     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 461     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 462     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 463     real             *charge;
 464     int              nvdwtype;
 465     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 466     int              *vdwtype;
 467     real             *vdwparam;
 468     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 469     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 470     __m128i          vfitab;
 471     __m128i          ifour       = _mm_set1_epi32(4);
 472     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 473     real             *vftab;
 474     __m256d          dummy_mask,cutoff_mask;
 475     __m128           tmpmask0,tmpmask1;
 476     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 477     __m256d          one     = _mm256_set1_pd(1.0);
 478     __m256d          two     = _mm256_set1_pd(2.0);
 479     x                = xx[0];
 480     f                = ff[0];
 481
 482     nri              = nlist->nri;
 483     iinr             = nlist->iinr;
 484     jindex           = nlist->jindex;
 485     jjnr             = nlist->jjnr;
 486     shiftidx         = nlist->shift;
 487     gid              = nlist->gid;
 488     shiftvec         = fr->shift_vec[0];
 489     fshift           = fr->fshift[0];
 490     facel            = _mm256_set1_pd(fr->epsfac);
 491     charge           = mdatoms->chargeA;
 492     nvdwtype         = fr->ntype;
 493     vdwparam         = fr->nbfp;
 494     vdwtype          = mdatoms->typeA;
 495
 496     vftab            = kernel_data->table_elec->data;
 497     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 498
 499     /* Avoid stupid compiler warnings */
 500     jnrA = jnrB = jnrC = jnrD = 0;
 501     j_coord_offsetA = 0;
 502     j_coord_offsetB = 0;
 503     j_coord_offsetC = 0;
 504     j_coord_offsetD = 0;
 505
 506     outeriter        = 0;
 507     inneriter        = 0;
 508
 509     for(iidx=0;iidx<4*DIM;iidx++)
 510     {
 511         scratch[iidx] = 0.0;
 512     }
 513
 514     /* Start outer loop over neighborlists */
 515     for(iidx=0; iidx<nri; iidx++)
 516     {
 517         /* Load shift vector for this list */
 518         i_shift_offset   = DIM*shiftidx[iidx];
 519
 520         /* Load limits for loop over neighbors */
 521         j_index_start    = jindex[iidx];
 522         j_index_end      = jindex[iidx+1];
 523
 524         /* Get outer coordinate index */
 525         inr              = iinr[iidx];
 526         i_coord_offset   = DIM*inr;
 527
 528         /* Load i particle coords and add shift vector */
 529         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 530
 531         fix0             = _mm256_setzero_pd();
 532         fiy0             = _mm256_setzero_pd();
 533         fiz0             = _mm256_setzero_pd();
 534
 535         /* Load parameters for i particles */
 536         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 537         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 538
 539         /* Start inner kernel loop */
 540         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 541         {
 542
 543             /* Get j neighbor index, and coordinate index */
 544             jnrA             = jjnr[jidx];
 545             jnrB             = jjnr[jidx+1];
 546             jnrC             = jjnr[jidx+2];
 547             jnrD             = jjnr[jidx+3];
 548             j_coord_offsetA  = DIM*jnrA;
 549             j_coord_offsetB  = DIM*jnrB;
 550             j_coord_offsetC  = DIM*jnrC;
 551             j_coord_offsetD  = DIM*jnrD;
 552
 553             /* load j atom coordinates */
 554             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 555                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 556                                                  &jx0,&jy0,&jz0);
 557
 558             /* Calculate displacement vector */
 559             dx00             = _mm256_sub_pd(ix0,jx0);
 560             dy00             = _mm256_sub_pd(iy0,jy0);
 561             dz00             = _mm256_sub_pd(iz0,jz0);
 562
 563             /* Calculate squared distance and things based on it */
 564             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 565
 566             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 567
 568             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 569
 570             /* Load parameters for j particles */
 571             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 572                                                                  charge+jnrC+0,charge+jnrD+0);
 573             vdwjidx0A        = 2*vdwtype[jnrA+0];
 574             vdwjidx0B        = 2*vdwtype[jnrB+0];
 575             vdwjidx0C        = 2*vdwtype[jnrC+0];
 576             vdwjidx0D        = 2*vdwtype[jnrD+0];
 577
 578             /**************************
 579              * CALCULATE INTERACTIONS *
 580              **************************/
 581
 582             r00              = _mm256_mul_pd(rsq00,rinv00);
 583
 584             /* Compute parameters for interactions between i and j atoms */
 585             qq00             = _mm256_mul_pd(iq0,jq0);
 586             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 587                                             vdwioffsetptr0+vdwjidx0B,
 588                                             vdwioffsetptr0+vdwjidx0C,
 589                                             vdwioffsetptr0+vdwjidx0D,
 590                                             &c6_00,&c12_00);
 591
 592             /* Calculate table index by multiplying r with table scale and truncate to integer */
 593             rt               = _mm256_mul_pd(r00,vftabscale);
 594             vfitab           = _mm256_cvttpd_epi32(rt);
 595             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 596             vfitab           = _mm_slli_epi32(vfitab,2);
 597
 598             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 599             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 600             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 601             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 602             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 603             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 604             Heps             = _mm256_mul_pd(vfeps,H);
 605             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 606             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 607             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 608
 609             /* LENNARD-JONES DISPERSION/REPULSION */
 610
 611             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 612             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 613
 614             fscal            = _mm256_add_pd(felec,fvdw);
 615
 616             /* Calculate temporary vectorial force */
 617             tx               = _mm256_mul_pd(fscal,dx00);
 618             ty               = _mm256_mul_pd(fscal,dy00);
 619             tz               = _mm256_mul_pd(fscal,dz00);
 620
 621             /* Update vectorial force */
 622             fix0             = _mm256_add_pd(fix0,tx);
 623             fiy0             = _mm256_add_pd(fiy0,ty);
 624             fiz0             = _mm256_add_pd(fiz0,tz);
 625
 626             fjptrA             = f+j_coord_offsetA;
 627             fjptrB             = f+j_coord_offsetB;
 628             fjptrC             = f+j_coord_offsetC;
 629             fjptrD             = f+j_coord_offsetD;
 630             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 631
 632             /* Inner loop uses 47 flops */
 633         }
 634
 635         if(jidx<j_index_end)
 636         {
 637
 638             /* Get j neighbor index, and coordinate index */
 639             jnrlistA         = jjnr[jidx];
 640             jnrlistB         = jjnr[jidx+1];
 641             jnrlistC         = jjnr[jidx+2];
 642             jnrlistD         = jjnr[jidx+3];
 643             /* Sign of each element will be negative for non-real atoms.
 644              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 645              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 646              */
 647             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 648
 649             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 650             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 651             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 652
 653             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 654             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 655             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 656             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 657             j_coord_offsetA  = DIM*jnrA;
 658             j_coord_offsetB  = DIM*jnrB;
 659             j_coord_offsetC  = DIM*jnrC;
 660             j_coord_offsetD  = DIM*jnrD;
 661
 662             /* load j atom coordinates */
 663             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 664                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 665                                                  &jx0,&jy0,&jz0);
 666
 667             /* Calculate displacement vector */
 668             dx00             = _mm256_sub_pd(ix0,jx0);
 669             dy00             = _mm256_sub_pd(iy0,jy0);
 670             dz00             = _mm256_sub_pd(iz0,jz0);
 671
 672             /* Calculate squared distance and things based on it */
 673             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 674
 675             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 676
 677             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 678
 679             /* Load parameters for j particles */
 680             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 681                                                                  charge+jnrC+0,charge+jnrD+0);
 682             vdwjidx0A        = 2*vdwtype[jnrA+0];
 683             vdwjidx0B        = 2*vdwtype[jnrB+0];
 684             vdwjidx0C        = 2*vdwtype[jnrC+0];
 685             vdwjidx0D        = 2*vdwtype[jnrD+0];
 686
 687             /**************************
 688              * CALCULATE INTERACTIONS *
 689              **************************/
 690
 691             r00              = _mm256_mul_pd(rsq00,rinv00);
 692             r00              = _mm256_andnot_pd(dummy_mask,r00);
 693
 694             /* Compute parameters for interactions between i and j atoms */
 695             qq00             = _mm256_mul_pd(iq0,jq0);
 696             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 697                                             vdwioffsetptr0+vdwjidx0B,
 698                                             vdwioffsetptr0+vdwjidx0C,
 699                                             vdwioffsetptr0+vdwjidx0D,
 700                                             &c6_00,&c12_00);
 701
 702             /* Calculate table index by multiplying r with table scale and truncate to integer */
 703             rt               = _mm256_mul_pd(r00,vftabscale);
 704             vfitab           = _mm256_cvttpd_epi32(rt);
 705             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 706             vfitab           = _mm_slli_epi32(vfitab,2);
 707
 708             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 709             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 710             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 711             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 712             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 713             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 714             Heps             = _mm256_mul_pd(vfeps,H);
 715             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 716             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 717             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 718
 719             /* LENNARD-JONES DISPERSION/REPULSION */
 720
 721             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 722             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 723
 724             fscal            = _mm256_add_pd(felec,fvdw);
 725
 726             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 727
 728             /* Calculate temporary vectorial force */
 729             tx               = _mm256_mul_pd(fscal,dx00);
 730             ty               = _mm256_mul_pd(fscal,dy00);
 731             tz               = _mm256_mul_pd(fscal,dz00);
 732
 733             /* Update vectorial force */
 734             fix0             = _mm256_add_pd(fix0,tx);
 735             fiy0             = _mm256_add_pd(fiy0,ty);
 736             fiz0             = _mm256_add_pd(fiz0,tz);
 737
 738             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 739             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 740             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 741             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 742             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 743
 744             /* Inner loop uses 48 flops */
 745         }
 746
 747         /* End of innermost loop */
 748
 749         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 750                                                  f+i_coord_offset,fshift+i_shift_offset);
 751
 752         /* Increment number of inner iterations */
 753         inneriter                  += j_index_end - j_index_start;
 754
 755         /* Outer loop uses 7 flops */
 756     }
 757
 758     /* Increment number of outer iterations */
 759     outeriter        += nri;
 760
 761     /* Update outer/inner flops */
 762
 763     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*48);
 764 }