src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCSTab_VdwNone_GeomP1P1_avx_256_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: CubicSplineTable
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     real *           vdwioffsetptr0;
  84     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     __m128i          vfitab;
  91     __m128i          ifour       = _mm_set1_epi32(4);
  92     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  93     real             *vftab;
  94     __m256d          dummy_mask,cutoff_mask;
  95     __m128           tmpmask0,tmpmask1;
  96     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  97     __m256d          one     = _mm256_set1_pd(1.0);
  98     __m256d          two     = _mm256_set1_pd(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm256_set1_pd(fr->ic->epsfac);
 111     charge           = mdatoms->chargeA;
 112
 113     vftab            = kernel_data->table_elec->data;
 114     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 115
 116     /* Avoid stupid compiler warnings */
 117     jnrA = jnrB = jnrC = jnrD = 0;
 118     j_coord_offsetA = 0;
 119     j_coord_offsetB = 0;
 120     j_coord_offsetC = 0;
 121     j_coord_offsetD = 0;
 122
 123     outeriter        = 0;
 124     inneriter        = 0;
 125
 126     for(iidx=0;iidx<4*DIM;iidx++)
 127     {
 128         scratch[iidx] = 0.0;
 129     }
 130
 131     /* Start outer loop over neighborlists */
 132     for(iidx=0; iidx<nri; iidx++)
 133     {
 134         /* Load shift vector for this list */
 135         i_shift_offset   = DIM*shiftidx[iidx];
 136
 137         /* Load limits for loop over neighbors */
 138         j_index_start    = jindex[iidx];
 139         j_index_end      = jindex[iidx+1];
 140
 141         /* Get outer coordinate index */
 142         inr              = iinr[iidx];
 143         i_coord_offset   = DIM*inr;
 144
 145         /* Load i particle coords and add shift vector */
 146         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 147
 148         fix0             = _mm256_setzero_pd();
 149         fiy0             = _mm256_setzero_pd();
 150         fiz0             = _mm256_setzero_pd();
 151
 152         /* Load parameters for i particles */
 153         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 154
 155         /* Reset potential sums */
 156         velecsum         = _mm256_setzero_pd();
 157
 158         /* Start inner kernel loop */
 159         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 160         {
 161
 162             /* Get j neighbor index, and coordinate index */
 163             jnrA             = jjnr[jidx];
 164             jnrB             = jjnr[jidx+1];
 165             jnrC             = jjnr[jidx+2];
 166             jnrD             = jjnr[jidx+3];
 167             j_coord_offsetA  = DIM*jnrA;
 168             j_coord_offsetB  = DIM*jnrB;
 169             j_coord_offsetC  = DIM*jnrC;
 170             j_coord_offsetD  = DIM*jnrD;
 171
 172             /* load j atom coordinates */
 173             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 174                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 175                                                  &jx0,&jy0,&jz0);
 176
 177             /* Calculate displacement vector */
 178             dx00             = _mm256_sub_pd(ix0,jx0);
 179             dy00             = _mm256_sub_pd(iy0,jy0);
 180             dz00             = _mm256_sub_pd(iz0,jz0);
 181
 182             /* Calculate squared distance and things based on it */
 183             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 184
 185             rinv00           = avx256_invsqrt_d(rsq00);
 186
 187             /* Load parameters for j particles */
 188             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 189                                                                  charge+jnrC+0,charge+jnrD+0);
 190
 191             /**************************
 192              * CALCULATE INTERACTIONS *
 193              **************************/
 194
 195             r00              = _mm256_mul_pd(rsq00,rinv00);
 196
 197             /* Compute parameters for interactions between i and j atoms */
 198             qq00             = _mm256_mul_pd(iq0,jq0);
 199
 200             /* Calculate table index by multiplying r with table scale and truncate to integer */
 201             rt               = _mm256_mul_pd(r00,vftabscale);
 202             vfitab           = _mm256_cvttpd_epi32(rt);
 203             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 204             vfitab           = _mm_slli_epi32(vfitab,2);
 205
 206             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 207             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 208             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 209             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 210             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 211             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 212             Heps             = _mm256_mul_pd(vfeps,H);
 213             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 214             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 215             velec            = _mm256_mul_pd(qq00,VV);
 216             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 217             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 218
 219             /* Update potential sum for this i atom from the interaction with this j atom. */
 220             velecsum         = _mm256_add_pd(velecsum,velec);
 221
 222             fscal            = felec;
 223
 224             /* Calculate temporary vectorial force */
 225             tx               = _mm256_mul_pd(fscal,dx00);
 226             ty               = _mm256_mul_pd(fscal,dy00);
 227             tz               = _mm256_mul_pd(fscal,dz00);
 228
 229             /* Update vectorial force */
 230             fix0             = _mm256_add_pd(fix0,tx);
 231             fiy0             = _mm256_add_pd(fiy0,ty);
 232             fiz0             = _mm256_add_pd(fiz0,tz);
 233
 234             fjptrA             = f+j_coord_offsetA;
 235             fjptrB             = f+j_coord_offsetB;
 236             fjptrC             = f+j_coord_offsetC;
 237             fjptrD             = f+j_coord_offsetD;
 238             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 239
 240             /* Inner loop uses 43 flops */
 241         }
 242
 243         if(jidx<j_index_end)
 244         {
 245
 246             /* Get j neighbor index, and coordinate index */
 247             jnrlistA         = jjnr[jidx];
 248             jnrlistB         = jjnr[jidx+1];
 249             jnrlistC         = jjnr[jidx+2];
 250             jnrlistD         = jjnr[jidx+3];
 251             /* Sign of each element will be negative for non-real atoms.
 252              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 253              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 254              */
 255             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 256
 257             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 258             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 259             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 260
 261             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 262             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 263             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 264             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 265             j_coord_offsetA  = DIM*jnrA;
 266             j_coord_offsetB  = DIM*jnrB;
 267             j_coord_offsetC  = DIM*jnrC;
 268             j_coord_offsetD  = DIM*jnrD;
 269
 270             /* load j atom coordinates */
 271             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 272                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 273                                                  &jx0,&jy0,&jz0);
 274
 275             /* Calculate displacement vector */
 276             dx00             = _mm256_sub_pd(ix0,jx0);
 277             dy00             = _mm256_sub_pd(iy0,jy0);
 278             dz00             = _mm256_sub_pd(iz0,jz0);
 279
 280             /* Calculate squared distance and things based on it */
 281             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 282
 283             rinv00           = avx256_invsqrt_d(rsq00);
 284
 285             /* Load parameters for j particles */
 286             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 287                                                                  charge+jnrC+0,charge+jnrD+0);
 288
 289             /**************************
 290              * CALCULATE INTERACTIONS *
 291              **************************/
 292
 293             r00              = _mm256_mul_pd(rsq00,rinv00);
 294             r00              = _mm256_andnot_pd(dummy_mask,r00);
 295
 296             /* Compute parameters for interactions between i and j atoms */
 297             qq00             = _mm256_mul_pd(iq0,jq0);
 298
 299             /* Calculate table index by multiplying r with table scale and truncate to integer */
 300             rt               = _mm256_mul_pd(r00,vftabscale);
 301             vfitab           = _mm256_cvttpd_epi32(rt);
 302             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 303             vfitab           = _mm_slli_epi32(vfitab,2);
 304
 305             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 306             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 307             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 308             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 309             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 310             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 311             Heps             = _mm256_mul_pd(vfeps,H);
 312             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 313             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 314             velec            = _mm256_mul_pd(qq00,VV);
 315             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 316             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velec            = _mm256_andnot_pd(dummy_mask,velec);
 320             velecsum         = _mm256_add_pd(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 325
 326             /* Calculate temporary vectorial force */
 327             tx               = _mm256_mul_pd(fscal,dx00);
 328             ty               = _mm256_mul_pd(fscal,dy00);
 329             tz               = _mm256_mul_pd(fscal,dz00);
 330
 331             /* Update vectorial force */
 332             fix0             = _mm256_add_pd(fix0,tx);
 333             fiy0             = _mm256_add_pd(fiy0,ty);
 334             fiz0             = _mm256_add_pd(fiz0,tz);
 335
 336             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 337             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 338             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 339             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 340             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 341
 342             /* Inner loop uses 44 flops */
 343         }
 344
 345         /* End of innermost loop */
 346
 347         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 348                                                  f+i_coord_offset,fshift+i_shift_offset);
 349
 350         ggid                        = gid[iidx];
 351         /* Update potential energies */
 352         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 353
 354         /* Increment number of inner iterations */
 355         inneriter                  += j_index_end - j_index_start;
 356
 357         /* Outer loop uses 8 flops */
 358     }
 359
 360     /* Increment number of outer iterations */
 361     outeriter        += nri;
 362
 363     /* Update outer/inner flops */
 364
 365     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*44);
 366 }
 367 /*
 368  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_double
 369  * Electrostatics interaction: CubicSplineTable
 370  * VdW interaction:            None
 371  * Geometry:                   Particle-Particle
 372  * Calculate force/pot:        Force
 373  */
 374 void
 375 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_double
 376                     (t_nblist                    * gmx_restrict       nlist,
 377                      rvec                        * gmx_restrict          xx,
 378                      rvec                        * gmx_restrict          ff,
 379                      struct t_forcerec           * gmx_restrict          fr,
 380                      t_mdatoms                   * gmx_restrict     mdatoms,
 381                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 382                      t_nrnb                      * gmx_restrict        nrnb)
 383 {
 384     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 385      * just 0 for non-waters.
 386      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 387      * jnr indices corresponding to data put in the four positions in the SIMD register.
 388      */
 389     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 390     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 391     int              jnrA,jnrB,jnrC,jnrD;
 392     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 393     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 394     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 395     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 396     real             rcutoff_scalar;
 397     real             *shiftvec,*fshift,*x,*f;
 398     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 399     real             scratch[4*DIM];
 400     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 401     real *           vdwioffsetptr0;
 402     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 403     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 404     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 405     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 406     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 407     real             *charge;
 408     __m128i          vfitab;
 409     __m128i          ifour       = _mm_set1_epi32(4);
 410     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 411     real             *vftab;
 412     __m256d          dummy_mask,cutoff_mask;
 413     __m128           tmpmask0,tmpmask1;
 414     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 415     __m256d          one     = _mm256_set1_pd(1.0);
 416     __m256d          two     = _mm256_set1_pd(2.0);
 417     x                = xx[0];
 418     f                = ff[0];
 419
 420     nri              = nlist->nri;
 421     iinr             = nlist->iinr;
 422     jindex           = nlist->jindex;
 423     jjnr             = nlist->jjnr;
 424     shiftidx         = nlist->shift;
 425     gid              = nlist->gid;
 426     shiftvec         = fr->shift_vec[0];
 427     fshift           = fr->fshift[0];
 428     facel            = _mm256_set1_pd(fr->ic->epsfac);
 429     charge           = mdatoms->chargeA;
 430
 431     vftab            = kernel_data->table_elec->data;
 432     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 433
 434     /* Avoid stupid compiler warnings */
 435     jnrA = jnrB = jnrC = jnrD = 0;
 436     j_coord_offsetA = 0;
 437     j_coord_offsetB = 0;
 438     j_coord_offsetC = 0;
 439     j_coord_offsetD = 0;
 440
 441     outeriter        = 0;
 442     inneriter        = 0;
 443
 444     for(iidx=0;iidx<4*DIM;iidx++)
 445     {
 446         scratch[iidx] = 0.0;
 447     }
 448
 449     /* Start outer loop over neighborlists */
 450     for(iidx=0; iidx<nri; iidx++)
 451     {
 452         /* Load shift vector for this list */
 453         i_shift_offset   = DIM*shiftidx[iidx];
 454
 455         /* Load limits for loop over neighbors */
 456         j_index_start    = jindex[iidx];
 457         j_index_end      = jindex[iidx+1];
 458
 459         /* Get outer coordinate index */
 460         inr              = iinr[iidx];
 461         i_coord_offset   = DIM*inr;
 462
 463         /* Load i particle coords and add shift vector */
 464         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 465
 466         fix0             = _mm256_setzero_pd();
 467         fiy0             = _mm256_setzero_pd();
 468         fiz0             = _mm256_setzero_pd();
 469
 470         /* Load parameters for i particles */
 471         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 472
 473         /* Start inner kernel loop */
 474         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 475         {
 476
 477             /* Get j neighbor index, and coordinate index */
 478             jnrA             = jjnr[jidx];
 479             jnrB             = jjnr[jidx+1];
 480             jnrC             = jjnr[jidx+2];
 481             jnrD             = jjnr[jidx+3];
 482             j_coord_offsetA  = DIM*jnrA;
 483             j_coord_offsetB  = DIM*jnrB;
 484             j_coord_offsetC  = DIM*jnrC;
 485             j_coord_offsetD  = DIM*jnrD;
 486
 487             /* load j atom coordinates */
 488             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 489                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 490                                                  &jx0,&jy0,&jz0);
 491
 492             /* Calculate displacement vector */
 493             dx00             = _mm256_sub_pd(ix0,jx0);
 494             dy00             = _mm256_sub_pd(iy0,jy0);
 495             dz00             = _mm256_sub_pd(iz0,jz0);
 496
 497             /* Calculate squared distance and things based on it */
 498             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 499
 500             rinv00           = avx256_invsqrt_d(rsq00);
 501
 502             /* Load parameters for j particles */
 503             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 504                                                                  charge+jnrC+0,charge+jnrD+0);
 505
 506             /**************************
 507              * CALCULATE INTERACTIONS *
 508              **************************/
 509
 510             r00              = _mm256_mul_pd(rsq00,rinv00);
 511
 512             /* Compute parameters for interactions between i and j atoms */
 513             qq00             = _mm256_mul_pd(iq0,jq0);
 514
 515             /* Calculate table index by multiplying r with table scale and truncate to integer */
 516             rt               = _mm256_mul_pd(r00,vftabscale);
 517             vfitab           = _mm256_cvttpd_epi32(rt);
 518             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 519             vfitab           = _mm_slli_epi32(vfitab,2);
 520
 521             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 522             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 523             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 524             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 525             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 526             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 527             Heps             = _mm256_mul_pd(vfeps,H);
 528             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 529             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 530             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 531
 532             fscal            = felec;
 533
 534             /* Calculate temporary vectorial force */
 535             tx               = _mm256_mul_pd(fscal,dx00);
 536             ty               = _mm256_mul_pd(fscal,dy00);
 537             tz               = _mm256_mul_pd(fscal,dz00);
 538
 539             /* Update vectorial force */
 540             fix0             = _mm256_add_pd(fix0,tx);
 541             fiy0             = _mm256_add_pd(fiy0,ty);
 542             fiz0             = _mm256_add_pd(fiz0,tz);
 543
 544             fjptrA             = f+j_coord_offsetA;
 545             fjptrB             = f+j_coord_offsetB;
 546             fjptrC             = f+j_coord_offsetC;
 547             fjptrD             = f+j_coord_offsetD;
 548             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 549
 550             /* Inner loop uses 39 flops */
 551         }
 552
 553         if(jidx<j_index_end)
 554         {
 555
 556             /* Get j neighbor index, and coordinate index */
 557             jnrlistA         = jjnr[jidx];
 558             jnrlistB         = jjnr[jidx+1];
 559             jnrlistC         = jjnr[jidx+2];
 560             jnrlistD         = jjnr[jidx+3];
 561             /* Sign of each element will be negative for non-real atoms.
 562              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 563              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 564              */
 565             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 566
 567             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 568             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 569             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 570
 571             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 572             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 573             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 574             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 575             j_coord_offsetA  = DIM*jnrA;
 576             j_coord_offsetB  = DIM*jnrB;
 577             j_coord_offsetC  = DIM*jnrC;
 578             j_coord_offsetD  = DIM*jnrD;
 579
 580             /* load j atom coordinates */
 581             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 582                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 583                                                  &jx0,&jy0,&jz0);
 584
 585             /* Calculate displacement vector */
 586             dx00             = _mm256_sub_pd(ix0,jx0);
 587             dy00             = _mm256_sub_pd(iy0,jy0);
 588             dz00             = _mm256_sub_pd(iz0,jz0);
 589
 590             /* Calculate squared distance and things based on it */
 591             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 592
 593             rinv00           = avx256_invsqrt_d(rsq00);
 594
 595             /* Load parameters for j particles */
 596             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 597                                                                  charge+jnrC+0,charge+jnrD+0);
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             r00              = _mm256_mul_pd(rsq00,rinv00);
 604             r00              = _mm256_andnot_pd(dummy_mask,r00);
 605
 606             /* Compute parameters for interactions between i and j atoms */
 607             qq00             = _mm256_mul_pd(iq0,jq0);
 608
 609             /* Calculate table index by multiplying r with table scale and truncate to integer */
 610             rt               = _mm256_mul_pd(r00,vftabscale);
 611             vfitab           = _mm256_cvttpd_epi32(rt);
 612             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 613             vfitab           = _mm_slli_epi32(vfitab,2);
 614
 615             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 616             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 617             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 618             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 619             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 620             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 621             Heps             = _mm256_mul_pd(vfeps,H);
 622             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 623             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 624             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 625
 626             fscal            = felec;
 627
 628             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 629
 630             /* Calculate temporary vectorial force */
 631             tx               = _mm256_mul_pd(fscal,dx00);
 632             ty               = _mm256_mul_pd(fscal,dy00);
 633             tz               = _mm256_mul_pd(fscal,dz00);
 634
 635             /* Update vectorial force */
 636             fix0             = _mm256_add_pd(fix0,tx);
 637             fiy0             = _mm256_add_pd(fiy0,ty);
 638             fiz0             = _mm256_add_pd(fiz0,tz);
 639
 640             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 641             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 642             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 643             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 644             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 645
 646             /* Inner loop uses 40 flops */
 647         }
 648
 649         /* End of innermost loop */
 650
 651         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 652                                                  f+i_coord_offset,fshift+i_shift_offset);
 653
 654         /* Increment number of inner iterations */
 655         inneriter                  += j_index_end - j_index_start;
 656
 657         /* Outer loop uses 7 flops */
 658     }
 659
 660     /* Increment number of outer iterations */
 661     outeriter        += nri;
 662
 663     /* Update outer/inner flops */
 664
 665     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 666 }