src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_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_ElecEw_VdwNone_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_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          ewitab;
  91     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  92     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  93     real             *ewtab;
  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     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 114     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 115     beta2            = _mm256_mul_pd(beta,beta);
 116     beta3            = _mm256_mul_pd(beta,beta2);
 117
 118     ewtab            = fr->ic->tabq_coul_FDV0;
 119     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 120     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm256_setzero_pd();
 155         fiy0             = _mm256_setzero_pd();
 156         fiz0             = _mm256_setzero_pd();
 157
 158         /* Load parameters for i particles */
 159         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 160
 161         /* Reset potential sums */
 162         velecsum         = _mm256_setzero_pd();
 163
 164         /* Start inner kernel loop */
 165         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 166         {
 167
 168             /* Get j neighbor index, and coordinate index */
 169             jnrA             = jjnr[jidx];
 170             jnrB             = jjnr[jidx+1];
 171             jnrC             = jjnr[jidx+2];
 172             jnrD             = jjnr[jidx+3];
 173             j_coord_offsetA  = DIM*jnrA;
 174             j_coord_offsetB  = DIM*jnrB;
 175             j_coord_offsetC  = DIM*jnrC;
 176             j_coord_offsetD  = DIM*jnrD;
 177
 178             /* load j atom coordinates */
 179             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 180                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 181                                                  &jx0,&jy0,&jz0);
 182
 183             /* Calculate displacement vector */
 184             dx00             = _mm256_sub_pd(ix0,jx0);
 185             dy00             = _mm256_sub_pd(iy0,jy0);
 186             dz00             = _mm256_sub_pd(iz0,jz0);
 187
 188             /* Calculate squared distance and things based on it */
 189             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 190
 191             rinv00           = avx256_invsqrt_d(rsq00);
 192
 193             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 194
 195             /* Load parameters for j particles */
 196             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 197                                                                  charge+jnrC+0,charge+jnrD+0);
 198
 199             /**************************
 200              * CALCULATE INTERACTIONS *
 201              **************************/
 202
 203             r00              = _mm256_mul_pd(rsq00,rinv00);
 204
 205             /* Compute parameters for interactions between i and j atoms */
 206             qq00             = _mm256_mul_pd(iq0,jq0);
 207
 208             /* EWALD ELECTROSTATICS */
 209
 210             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 211             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 212             ewitab           = _mm256_cvttpd_epi32(ewrt);
 213             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 214             ewitab           = _mm_slli_epi32(ewitab,2);
 215             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 216             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 217             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 218             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 219             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 220             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 221             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 222             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 223             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 224
 225             /* Update potential sum for this i atom from the interaction with this j atom. */
 226             velecsum         = _mm256_add_pd(velecsum,velec);
 227
 228             fscal            = felec;
 229
 230             /* Calculate temporary vectorial force */
 231             tx               = _mm256_mul_pd(fscal,dx00);
 232             ty               = _mm256_mul_pd(fscal,dy00);
 233             tz               = _mm256_mul_pd(fscal,dz00);
 234
 235             /* Update vectorial force */
 236             fix0             = _mm256_add_pd(fix0,tx);
 237             fiy0             = _mm256_add_pd(fiy0,ty);
 238             fiz0             = _mm256_add_pd(fiz0,tz);
 239
 240             fjptrA             = f+j_coord_offsetA;
 241             fjptrB             = f+j_coord_offsetB;
 242             fjptrC             = f+j_coord_offsetC;
 243             fjptrD             = f+j_coord_offsetD;
 244             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 245
 246             /* Inner loop uses 41 flops */
 247         }
 248
 249         if(jidx<j_index_end)
 250         {
 251
 252             /* Get j neighbor index, and coordinate index */
 253             jnrlistA         = jjnr[jidx];
 254             jnrlistB         = jjnr[jidx+1];
 255             jnrlistC         = jjnr[jidx+2];
 256             jnrlistD         = jjnr[jidx+3];
 257             /* Sign of each element will be negative for non-real atoms.
 258              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 259              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 260              */
 261             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 262
 263             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 264             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 265             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 266
 267             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 268             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 269             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 270             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 271             j_coord_offsetA  = DIM*jnrA;
 272             j_coord_offsetB  = DIM*jnrB;
 273             j_coord_offsetC  = DIM*jnrC;
 274             j_coord_offsetD  = DIM*jnrD;
 275
 276             /* load j atom coordinates */
 277             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 278                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 279                                                  &jx0,&jy0,&jz0);
 280
 281             /* Calculate displacement vector */
 282             dx00             = _mm256_sub_pd(ix0,jx0);
 283             dy00             = _mm256_sub_pd(iy0,jy0);
 284             dz00             = _mm256_sub_pd(iz0,jz0);
 285
 286             /* Calculate squared distance and things based on it */
 287             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 288
 289             rinv00           = avx256_invsqrt_d(rsq00);
 290
 291             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 292
 293             /* Load parameters for j particles */
 294             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 295                                                                  charge+jnrC+0,charge+jnrD+0);
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             r00              = _mm256_mul_pd(rsq00,rinv00);
 302             r00              = _mm256_andnot_pd(dummy_mask,r00);
 303
 304             /* Compute parameters for interactions between i and j atoms */
 305             qq00             = _mm256_mul_pd(iq0,jq0);
 306
 307             /* EWALD ELECTROSTATICS */
 308
 309             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 310             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 311             ewitab           = _mm256_cvttpd_epi32(ewrt);
 312             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 313             ewitab           = _mm_slli_epi32(ewitab,2);
 314             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 315             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 316             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 317             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 318             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 319             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 320             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 321             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 322             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 323
 324             /* Update potential sum for this i atom from the interaction with this j atom. */
 325             velec            = _mm256_andnot_pd(dummy_mask,velec);
 326             velecsum         = _mm256_add_pd(velecsum,velec);
 327
 328             fscal            = felec;
 329
 330             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 331
 332             /* Calculate temporary vectorial force */
 333             tx               = _mm256_mul_pd(fscal,dx00);
 334             ty               = _mm256_mul_pd(fscal,dy00);
 335             tz               = _mm256_mul_pd(fscal,dz00);
 336
 337             /* Update vectorial force */
 338             fix0             = _mm256_add_pd(fix0,tx);
 339             fiy0             = _mm256_add_pd(fiy0,ty);
 340             fiz0             = _mm256_add_pd(fiz0,tz);
 341
 342             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 343             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 344             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 345             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 346             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 347
 348             /* Inner loop uses 42 flops */
 349         }
 350
 351         /* End of innermost loop */
 352
 353         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 354                                                  f+i_coord_offset,fshift+i_shift_offset);
 355
 356         ggid                        = gid[iidx];
 357         /* Update potential energies */
 358         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 359
 360         /* Increment number of inner iterations */
 361         inneriter                  += j_index_end - j_index_start;
 362
 363         /* Outer loop uses 8 flops */
 364     }
 365
 366     /* Increment number of outer iterations */
 367     outeriter        += nri;
 368
 369     /* Update outer/inner flops */
 370
 371     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*42);
 372 }
 373 /*
 374  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomP1P1_F_avx_256_double
 375  * Electrostatics interaction: Ewald
 376  * VdW interaction:            None
 377  * Geometry:                   Particle-Particle
 378  * Calculate force/pot:        Force
 379  */
 380 void
 381 nb_kernel_ElecEw_VdwNone_GeomP1P1_F_avx_256_double
 382                     (t_nblist                    * gmx_restrict       nlist,
 383                      rvec                        * gmx_restrict          xx,
 384                      rvec                        * gmx_restrict          ff,
 385                      struct t_forcerec           * gmx_restrict          fr,
 386                      t_mdatoms                   * gmx_restrict     mdatoms,
 387                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 388                      t_nrnb                      * gmx_restrict        nrnb)
 389 {
 390     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 391      * just 0 for non-waters.
 392      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 393      * jnr indices corresponding to data put in the four positions in the SIMD register.
 394      */
 395     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 396     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 397     int              jnrA,jnrB,jnrC,jnrD;
 398     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 399     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 400     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 401     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 402     real             rcutoff_scalar;
 403     real             *shiftvec,*fshift,*x,*f;
 404     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 405     real             scratch[4*DIM];
 406     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 407     real *           vdwioffsetptr0;
 408     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 409     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 410     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 411     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 412     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 413     real             *charge;
 414     __m128i          ewitab;
 415     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 416     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 417     real             *ewtab;
 418     __m256d          dummy_mask,cutoff_mask;
 419     __m128           tmpmask0,tmpmask1;
 420     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 421     __m256d          one     = _mm256_set1_pd(1.0);
 422     __m256d          two     = _mm256_set1_pd(2.0);
 423     x                = xx[0];
 424     f                = ff[0];
 425
 426     nri              = nlist->nri;
 427     iinr             = nlist->iinr;
 428     jindex           = nlist->jindex;
 429     jjnr             = nlist->jjnr;
 430     shiftidx         = nlist->shift;
 431     gid              = nlist->gid;
 432     shiftvec         = fr->shift_vec[0];
 433     fshift           = fr->fshift[0];
 434     facel            = _mm256_set1_pd(fr->ic->epsfac);
 435     charge           = mdatoms->chargeA;
 436
 437     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 438     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 439     beta2            = _mm256_mul_pd(beta,beta);
 440     beta3            = _mm256_mul_pd(beta,beta2);
 441
 442     ewtab            = fr->ic->tabq_coul_F;
 443     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 444     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 445
 446     /* Avoid stupid compiler warnings */
 447     jnrA = jnrB = jnrC = jnrD = 0;
 448     j_coord_offsetA = 0;
 449     j_coord_offsetB = 0;
 450     j_coord_offsetC = 0;
 451     j_coord_offsetD = 0;
 452
 453     outeriter        = 0;
 454     inneriter        = 0;
 455
 456     for(iidx=0;iidx<4*DIM;iidx++)
 457     {
 458         scratch[iidx] = 0.0;
 459     }
 460
 461     /* Start outer loop over neighborlists */
 462     for(iidx=0; iidx<nri; iidx++)
 463     {
 464         /* Load shift vector for this list */
 465         i_shift_offset   = DIM*shiftidx[iidx];
 466
 467         /* Load limits for loop over neighbors */
 468         j_index_start    = jindex[iidx];
 469         j_index_end      = jindex[iidx+1];
 470
 471         /* Get outer coordinate index */
 472         inr              = iinr[iidx];
 473         i_coord_offset   = DIM*inr;
 474
 475         /* Load i particle coords and add shift vector */
 476         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 477
 478         fix0             = _mm256_setzero_pd();
 479         fiy0             = _mm256_setzero_pd();
 480         fiz0             = _mm256_setzero_pd();
 481
 482         /* Load parameters for i particles */
 483         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 484
 485         /* Start inner kernel loop */
 486         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 487         {
 488
 489             /* Get j neighbor index, and coordinate index */
 490             jnrA             = jjnr[jidx];
 491             jnrB             = jjnr[jidx+1];
 492             jnrC             = jjnr[jidx+2];
 493             jnrD             = jjnr[jidx+3];
 494             j_coord_offsetA  = DIM*jnrA;
 495             j_coord_offsetB  = DIM*jnrB;
 496             j_coord_offsetC  = DIM*jnrC;
 497             j_coord_offsetD  = DIM*jnrD;
 498
 499             /* load j atom coordinates */
 500             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 501                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 502                                                  &jx0,&jy0,&jz0);
 503
 504             /* Calculate displacement vector */
 505             dx00             = _mm256_sub_pd(ix0,jx0);
 506             dy00             = _mm256_sub_pd(iy0,jy0);
 507             dz00             = _mm256_sub_pd(iz0,jz0);
 508
 509             /* Calculate squared distance and things based on it */
 510             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 511
 512             rinv00           = avx256_invsqrt_d(rsq00);
 513
 514             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 515
 516             /* Load parameters for j particles */
 517             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 518                                                                  charge+jnrC+0,charge+jnrD+0);
 519
 520             /**************************
 521              * CALCULATE INTERACTIONS *
 522              **************************/
 523
 524             r00              = _mm256_mul_pd(rsq00,rinv00);
 525
 526             /* Compute parameters for interactions between i and j atoms */
 527             qq00             = _mm256_mul_pd(iq0,jq0);
 528
 529             /* EWALD ELECTROSTATICS */
 530
 531             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 532             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 533             ewitab           = _mm256_cvttpd_epi32(ewrt);
 534             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 535             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 536                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 537                                             &ewtabF,&ewtabFn);
 538             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 539             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 540
 541             fscal            = felec;
 542
 543             /* Calculate temporary vectorial force */
 544             tx               = _mm256_mul_pd(fscal,dx00);
 545             ty               = _mm256_mul_pd(fscal,dy00);
 546             tz               = _mm256_mul_pd(fscal,dz00);
 547
 548             /* Update vectorial force */
 549             fix0             = _mm256_add_pd(fix0,tx);
 550             fiy0             = _mm256_add_pd(fiy0,ty);
 551             fiz0             = _mm256_add_pd(fiz0,tz);
 552
 553             fjptrA             = f+j_coord_offsetA;
 554             fjptrB             = f+j_coord_offsetB;
 555             fjptrC             = f+j_coord_offsetC;
 556             fjptrD             = f+j_coord_offsetD;
 557             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 558
 559             /* Inner loop uses 36 flops */
 560         }
 561
 562         if(jidx<j_index_end)
 563         {
 564
 565             /* Get j neighbor index, and coordinate index */
 566             jnrlistA         = jjnr[jidx];
 567             jnrlistB         = jjnr[jidx+1];
 568             jnrlistC         = jjnr[jidx+2];
 569             jnrlistD         = jjnr[jidx+3];
 570             /* Sign of each element will be negative for non-real atoms.
 571              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 572              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 573              */
 574             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 575
 576             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 577             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 578             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 579
 580             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 581             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 582             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 583             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 584             j_coord_offsetA  = DIM*jnrA;
 585             j_coord_offsetB  = DIM*jnrB;
 586             j_coord_offsetC  = DIM*jnrC;
 587             j_coord_offsetD  = DIM*jnrD;
 588
 589             /* load j atom coordinates */
 590             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 591                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 592                                                  &jx0,&jy0,&jz0);
 593
 594             /* Calculate displacement vector */
 595             dx00             = _mm256_sub_pd(ix0,jx0);
 596             dy00             = _mm256_sub_pd(iy0,jy0);
 597             dz00             = _mm256_sub_pd(iz0,jz0);
 598
 599             /* Calculate squared distance and things based on it */
 600             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 601
 602             rinv00           = avx256_invsqrt_d(rsq00);
 603
 604             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 605
 606             /* Load parameters for j particles */
 607             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 608                                                                  charge+jnrC+0,charge+jnrD+0);
 609
 610             /**************************
 611              * CALCULATE INTERACTIONS *
 612              **************************/
 613
 614             r00              = _mm256_mul_pd(rsq00,rinv00);
 615             r00              = _mm256_andnot_pd(dummy_mask,r00);
 616
 617             /* Compute parameters for interactions between i and j atoms */
 618             qq00             = _mm256_mul_pd(iq0,jq0);
 619
 620             /* EWALD ELECTROSTATICS */
 621
 622             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 623             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 624             ewitab           = _mm256_cvttpd_epi32(ewrt);
 625             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 626             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 627                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 628                                             &ewtabF,&ewtabFn);
 629             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 630             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 631
 632             fscal            = felec;
 633
 634             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 635
 636             /* Calculate temporary vectorial force */
 637             tx               = _mm256_mul_pd(fscal,dx00);
 638             ty               = _mm256_mul_pd(fscal,dy00);
 639             tz               = _mm256_mul_pd(fscal,dz00);
 640
 641             /* Update vectorial force */
 642             fix0             = _mm256_add_pd(fix0,tx);
 643             fiy0             = _mm256_add_pd(fiy0,ty);
 644             fiz0             = _mm256_add_pd(fiz0,tz);
 645
 646             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 647             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 648             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 649             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 650             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 651
 652             /* Inner loop uses 37 flops */
 653         }
 654
 655         /* End of innermost loop */
 656
 657         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 658                                                  f+i_coord_offset,fshift+i_shift_offset);
 659
 660         /* Increment number of inner iterations */
 661         inneriter                  += j_index_end - j_index_start;
 662
 663         /* Outer loop uses 7 flops */
 664     }
 665
 666     /* Increment number of outer iterations */
 667     outeriter        += nri;
 668
 669     /* Update outer/inner flops */
 670
 671     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*37);
 672 }