src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCSTab_VdwNone_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 "gromacs/simd/math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: CubicSplineTable
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCSTab_VdwNone_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     __m128i          vfitab;
  94     __m128i          ifour       = _mm_set1_epi32(4);
  95     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  96     real             *vftab;
  97     __m256d          dummy_mask,cutoff_mask;
  98     __m128           tmpmask0,tmpmask1;
  99     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 100     __m256d          one     = _mm256_set1_pd(1.0);
 101     __m256d          two     = _mm256_set1_pd(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm256_set1_pd(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115
 116     vftab            = kernel_data->table_elec->data;
 117     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 118
 119     /* Avoid stupid compiler warnings */
 120     jnrA = jnrB = jnrC = jnrD = 0;
 121     j_coord_offsetA = 0;
 122     j_coord_offsetB = 0;
 123     j_coord_offsetC = 0;
 124     j_coord_offsetD = 0;
 125
 126     outeriter        = 0;
 127     inneriter        = 0;
 128
 129     for(iidx=0;iidx<4*DIM;iidx++)
 130     {
 131         scratch[iidx] = 0.0;
 132     }
 133
 134     /* Start outer loop over neighborlists */
 135     for(iidx=0; iidx<nri; iidx++)
 136     {
 137         /* Load shift vector for this list */
 138         i_shift_offset   = DIM*shiftidx[iidx];
 139
 140         /* Load limits for loop over neighbors */
 141         j_index_start    = jindex[iidx];
 142         j_index_end      = jindex[iidx+1];
 143
 144         /* Get outer coordinate index */
 145         inr              = iinr[iidx];
 146         i_coord_offset   = DIM*inr;
 147
 148         /* Load i particle coords and add shift vector */
 149         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 150
 151         fix0             = _mm256_setzero_pd();
 152         fiy0             = _mm256_setzero_pd();
 153         fiz0             = _mm256_setzero_pd();
 154
 155         /* Load parameters for i particles */
 156         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 157
 158         /* Reset potential sums */
 159         velecsum         = _mm256_setzero_pd();
 160
 161         /* Start inner kernel loop */
 162         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 163         {
 164
 165             /* Get j neighbor index, and coordinate index */
 166             jnrA             = jjnr[jidx];
 167             jnrB             = jjnr[jidx+1];
 168             jnrC             = jjnr[jidx+2];
 169             jnrD             = jjnr[jidx+3];
 170             j_coord_offsetA  = DIM*jnrA;
 171             j_coord_offsetB  = DIM*jnrB;
 172             j_coord_offsetC  = DIM*jnrC;
 173             j_coord_offsetD  = DIM*jnrD;
 174
 175             /* load j atom coordinates */
 176             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 177                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 178                                                  &jx0,&jy0,&jz0);
 179
 180             /* Calculate displacement vector */
 181             dx00             = _mm256_sub_pd(ix0,jx0);
 182             dy00             = _mm256_sub_pd(iy0,jy0);
 183             dz00             = _mm256_sub_pd(iz0,jz0);
 184
 185             /* Calculate squared distance and things based on it */
 186             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 187
 188             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 189
 190             /* Load parameters for j particles */
 191             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 192                                                                  charge+jnrC+0,charge+jnrD+0);
 193
 194             /**************************
 195              * CALCULATE INTERACTIONS *
 196              **************************/
 197
 198             r00              = _mm256_mul_pd(rsq00,rinv00);
 199
 200             /* Compute parameters for interactions between i and j atoms */
 201             qq00             = _mm256_mul_pd(iq0,jq0);
 202
 203             /* Calculate table index by multiplying r with table scale and truncate to integer */
 204             rt               = _mm256_mul_pd(r00,vftabscale);
 205             vfitab           = _mm256_cvttpd_epi32(rt);
 206             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 207             vfitab           = _mm_slli_epi32(vfitab,2);
 208
 209             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 210             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 211             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 212             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 213             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 214             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 215             Heps             = _mm256_mul_pd(vfeps,H);
 216             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 217             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 218             velec            = _mm256_mul_pd(qq00,VV);
 219             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 220             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 221
 222             /* Update potential sum for this i atom from the interaction with this j atom. */
 223             velecsum         = _mm256_add_pd(velecsum,velec);
 224
 225             fscal            = felec;
 226
 227             /* Calculate temporary vectorial force */
 228             tx               = _mm256_mul_pd(fscal,dx00);
 229             ty               = _mm256_mul_pd(fscal,dy00);
 230             tz               = _mm256_mul_pd(fscal,dz00);
 231
 232             /* Update vectorial force */
 233             fix0             = _mm256_add_pd(fix0,tx);
 234             fiy0             = _mm256_add_pd(fiy0,ty);
 235             fiz0             = _mm256_add_pd(fiz0,tz);
 236
 237             fjptrA             = f+j_coord_offsetA;
 238             fjptrB             = f+j_coord_offsetB;
 239             fjptrC             = f+j_coord_offsetC;
 240             fjptrD             = f+j_coord_offsetD;
 241             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 242
 243             /* Inner loop uses 43 flops */
 244         }
 245
 246         if(jidx<j_index_end)
 247         {
 248
 249             /* Get j neighbor index, and coordinate index */
 250             jnrlistA         = jjnr[jidx];
 251             jnrlistB         = jjnr[jidx+1];
 252             jnrlistC         = jjnr[jidx+2];
 253             jnrlistD         = jjnr[jidx+3];
 254             /* Sign of each element will be negative for non-real atoms.
 255              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 256              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 257              */
 258             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 259
 260             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 261             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 262             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 263
 264             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 265             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 266             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 267             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 268             j_coord_offsetA  = DIM*jnrA;
 269             j_coord_offsetB  = DIM*jnrB;
 270             j_coord_offsetC  = DIM*jnrC;
 271             j_coord_offsetD  = DIM*jnrD;
 272
 273             /* load j atom coordinates */
 274             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 275                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 276                                                  &jx0,&jy0,&jz0);
 277
 278             /* Calculate displacement vector */
 279             dx00             = _mm256_sub_pd(ix0,jx0);
 280             dy00             = _mm256_sub_pd(iy0,jy0);
 281             dz00             = _mm256_sub_pd(iz0,jz0);
 282
 283             /* Calculate squared distance and things based on it */
 284             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 285
 286             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 287
 288             /* Load parameters for j particles */
 289             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 290                                                                  charge+jnrC+0,charge+jnrD+0);
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             r00              = _mm256_mul_pd(rsq00,rinv00);
 297             r00              = _mm256_andnot_pd(dummy_mask,r00);
 298
 299             /* Compute parameters for interactions between i and j atoms */
 300             qq00             = _mm256_mul_pd(iq0,jq0);
 301
 302             /* Calculate table index by multiplying r with table scale and truncate to integer */
 303             rt               = _mm256_mul_pd(r00,vftabscale);
 304             vfitab           = _mm256_cvttpd_epi32(rt);
 305             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 306             vfitab           = _mm_slli_epi32(vfitab,2);
 307
 308             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 309             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 310             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 311             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 312             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 313             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 314             Heps             = _mm256_mul_pd(vfeps,H);
 315             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 316             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 317             velec            = _mm256_mul_pd(qq00,VV);
 318             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 319             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 320
 321             /* Update potential sum for this i atom from the interaction with this j atom. */
 322             velec            = _mm256_andnot_pd(dummy_mask,velec);
 323             velecsum         = _mm256_add_pd(velecsum,velec);
 324
 325             fscal            = felec;
 326
 327             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 328
 329             /* Calculate temporary vectorial force */
 330             tx               = _mm256_mul_pd(fscal,dx00);
 331             ty               = _mm256_mul_pd(fscal,dy00);
 332             tz               = _mm256_mul_pd(fscal,dz00);
 333
 334             /* Update vectorial force */
 335             fix0             = _mm256_add_pd(fix0,tx);
 336             fiy0             = _mm256_add_pd(fiy0,ty);
 337             fiz0             = _mm256_add_pd(fiz0,tz);
 338
 339             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 340             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 341             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 342             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 343             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 344
 345             /* Inner loop uses 44 flops */
 346         }
 347
 348         /* End of innermost loop */
 349
 350         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 351                                                  f+i_coord_offset,fshift+i_shift_offset);
 352
 353         ggid                        = gid[iidx];
 354         /* Update potential energies */
 355         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 356
 357         /* Increment number of inner iterations */
 358         inneriter                  += j_index_end - j_index_start;
 359
 360         /* Outer loop uses 8 flops */
 361     }
 362
 363     /* Increment number of outer iterations */
 364     outeriter        += nri;
 365
 366     /* Update outer/inner flops */
 367
 368     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*44);
 369 }
 370 /*
 371  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_double
 372  * Electrostatics interaction: CubicSplineTable
 373  * VdW interaction:            None
 374  * Geometry:                   Particle-Particle
 375  * Calculate force/pot:        Force
 376  */
 377 void
 378 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_double
 379                     (t_nblist                    * gmx_restrict       nlist,
 380                      rvec                        * gmx_restrict          xx,
 381                      rvec                        * gmx_restrict          ff,
 382                      t_forcerec                  * gmx_restrict          fr,
 383                      t_mdatoms                   * gmx_restrict     mdatoms,
 384                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 385                      t_nrnb                      * gmx_restrict        nrnb)
 386 {
 387     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 388      * just 0 for non-waters.
 389      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 390      * jnr indices corresponding to data put in the four positions in the SIMD register.
 391      */
 392     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 393     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 394     int              jnrA,jnrB,jnrC,jnrD;
 395     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 396     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 397     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 398     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 399     real             rcutoff_scalar;
 400     real             *shiftvec,*fshift,*x,*f;
 401     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 402     real             scratch[4*DIM];
 403     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 404     real *           vdwioffsetptr0;
 405     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 406     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 407     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 408     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 409     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 410     real             *charge;
 411     __m128i          vfitab;
 412     __m128i          ifour       = _mm_set1_epi32(4);
 413     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 414     real             *vftab;
 415     __m256d          dummy_mask,cutoff_mask;
 416     __m128           tmpmask0,tmpmask1;
 417     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 418     __m256d          one     = _mm256_set1_pd(1.0);
 419     __m256d          two     = _mm256_set1_pd(2.0);
 420     x                = xx[0];
 421     f                = ff[0];
 422
 423     nri              = nlist->nri;
 424     iinr             = nlist->iinr;
 425     jindex           = nlist->jindex;
 426     jjnr             = nlist->jjnr;
 427     shiftidx         = nlist->shift;
 428     gid              = nlist->gid;
 429     shiftvec         = fr->shift_vec[0];
 430     fshift           = fr->fshift[0];
 431     facel            = _mm256_set1_pd(fr->epsfac);
 432     charge           = mdatoms->chargeA;
 433
 434     vftab            = kernel_data->table_elec->data;
 435     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 436
 437     /* Avoid stupid compiler warnings */
 438     jnrA = jnrB = jnrC = jnrD = 0;
 439     j_coord_offsetA = 0;
 440     j_coord_offsetB = 0;
 441     j_coord_offsetC = 0;
 442     j_coord_offsetD = 0;
 443
 444     outeriter        = 0;
 445     inneriter        = 0;
 446
 447     for(iidx=0;iidx<4*DIM;iidx++)
 448     {
 449         scratch[iidx] = 0.0;
 450     }
 451
 452     /* Start outer loop over neighborlists */
 453     for(iidx=0; iidx<nri; iidx++)
 454     {
 455         /* Load shift vector for this list */
 456         i_shift_offset   = DIM*shiftidx[iidx];
 457
 458         /* Load limits for loop over neighbors */
 459         j_index_start    = jindex[iidx];
 460         j_index_end      = jindex[iidx+1];
 461
 462         /* Get outer coordinate index */
 463         inr              = iinr[iidx];
 464         i_coord_offset   = DIM*inr;
 465
 466         /* Load i particle coords and add shift vector */
 467         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 468
 469         fix0             = _mm256_setzero_pd();
 470         fiy0             = _mm256_setzero_pd();
 471         fiz0             = _mm256_setzero_pd();
 472
 473         /* Load parameters for i particles */
 474         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 475
 476         /* Start inner kernel loop */
 477         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 478         {
 479
 480             /* Get j neighbor index, and coordinate index */
 481             jnrA             = jjnr[jidx];
 482             jnrB             = jjnr[jidx+1];
 483             jnrC             = jjnr[jidx+2];
 484             jnrD             = jjnr[jidx+3];
 485             j_coord_offsetA  = DIM*jnrA;
 486             j_coord_offsetB  = DIM*jnrB;
 487             j_coord_offsetC  = DIM*jnrC;
 488             j_coord_offsetD  = DIM*jnrD;
 489
 490             /* load j atom coordinates */
 491             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 492                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 493                                                  &jx0,&jy0,&jz0);
 494
 495             /* Calculate displacement vector */
 496             dx00             = _mm256_sub_pd(ix0,jx0);
 497             dy00             = _mm256_sub_pd(iy0,jy0);
 498             dz00             = _mm256_sub_pd(iz0,jz0);
 499
 500             /* Calculate squared distance and things based on it */
 501             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 502
 503             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 504
 505             /* Load parameters for j particles */
 506             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 507                                                                  charge+jnrC+0,charge+jnrD+0);
 508
 509             /**************************
 510              * CALCULATE INTERACTIONS *
 511              **************************/
 512
 513             r00              = _mm256_mul_pd(rsq00,rinv00);
 514
 515             /* Compute parameters for interactions between i and j atoms */
 516             qq00             = _mm256_mul_pd(iq0,jq0);
 517
 518             /* Calculate table index by multiplying r with table scale and truncate to integer */
 519             rt               = _mm256_mul_pd(r00,vftabscale);
 520             vfitab           = _mm256_cvttpd_epi32(rt);
 521             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 522             vfitab           = _mm_slli_epi32(vfitab,2);
 523
 524             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 525             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 526             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 527             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 528             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 529             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 530             Heps             = _mm256_mul_pd(vfeps,H);
 531             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 532             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 533             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 534
 535             fscal            = felec;
 536
 537             /* Calculate temporary vectorial force */
 538             tx               = _mm256_mul_pd(fscal,dx00);
 539             ty               = _mm256_mul_pd(fscal,dy00);
 540             tz               = _mm256_mul_pd(fscal,dz00);
 541
 542             /* Update vectorial force */
 543             fix0             = _mm256_add_pd(fix0,tx);
 544             fiy0             = _mm256_add_pd(fiy0,ty);
 545             fiz0             = _mm256_add_pd(fiz0,tz);
 546
 547             fjptrA             = f+j_coord_offsetA;
 548             fjptrB             = f+j_coord_offsetB;
 549             fjptrC             = f+j_coord_offsetC;
 550             fjptrD             = f+j_coord_offsetD;
 551             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 552
 553             /* Inner loop uses 39 flops */
 554         }
 555
 556         if(jidx<j_index_end)
 557         {
 558
 559             /* Get j neighbor index, and coordinate index */
 560             jnrlistA         = jjnr[jidx];
 561             jnrlistB         = jjnr[jidx+1];
 562             jnrlistC         = jjnr[jidx+2];
 563             jnrlistD         = jjnr[jidx+3];
 564             /* Sign of each element will be negative for non-real atoms.
 565              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 566              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 567              */
 568             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 569
 570             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 571             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 572             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 573
 574             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 575             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 576             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 577             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 578             j_coord_offsetA  = DIM*jnrA;
 579             j_coord_offsetB  = DIM*jnrB;
 580             j_coord_offsetC  = DIM*jnrC;
 581             j_coord_offsetD  = DIM*jnrD;
 582
 583             /* load j atom coordinates */
 584             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 585                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 586                                                  &jx0,&jy0,&jz0);
 587
 588             /* Calculate displacement vector */
 589             dx00             = _mm256_sub_pd(ix0,jx0);
 590             dy00             = _mm256_sub_pd(iy0,jy0);
 591             dz00             = _mm256_sub_pd(iz0,jz0);
 592
 593             /* Calculate squared distance and things based on it */
 594             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 595
 596             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 597
 598             /* Load parameters for j particles */
 599             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 600                                                                  charge+jnrC+0,charge+jnrD+0);
 601
 602             /**************************
 603              * CALCULATE INTERACTIONS *
 604              **************************/
 605
 606             r00              = _mm256_mul_pd(rsq00,rinv00);
 607             r00              = _mm256_andnot_pd(dummy_mask,r00);
 608
 609             /* Compute parameters for interactions between i and j atoms */
 610             qq00             = _mm256_mul_pd(iq0,jq0);
 611
 612             /* Calculate table index by multiplying r with table scale and truncate to integer */
 613             rt               = _mm256_mul_pd(r00,vftabscale);
 614             vfitab           = _mm256_cvttpd_epi32(rt);
 615             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 616             vfitab           = _mm_slli_epi32(vfitab,2);
 617
 618             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 619             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 620             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 621             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 622             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 623             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 624             Heps             = _mm256_mul_pd(vfeps,H);
 625             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 626             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 627             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 628
 629             fscal            = felec;
 630
 631             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 632
 633             /* Calculate temporary vectorial force */
 634             tx               = _mm256_mul_pd(fscal,dx00);
 635             ty               = _mm256_mul_pd(fscal,dy00);
 636             tz               = _mm256_mul_pd(fscal,dz00);
 637
 638             /* Update vectorial force */
 639             fix0             = _mm256_add_pd(fix0,tx);
 640             fiy0             = _mm256_add_pd(fiy0,ty);
 641             fiz0             = _mm256_add_pd(fiz0,tz);
 642
 643             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 644             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 645             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 646             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 647             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 648
 649             /* Inner loop uses 40 flops */
 650         }
 651
 652         /* End of innermost loop */
 653
 654         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 655                                                  f+i_coord_offset,fshift+i_shift_offset);
 656
 657         /* Increment number of inner iterations */
 658         inneriter                  += j_index_end - j_index_start;
 659
 660         /* Outer loop uses 7 flops */
 661     }
 662
 663     /* Increment number of outer iterations */
 664     outeriter        += nri;
 665
 666     /* Update outer/inner flops */
 667
 668     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 669 }