src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEwSh_VdwNone_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            None
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSh_VdwNone_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  90     real             *charge;
  91     __m128i          ewitab;
  92     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  93     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  94     real             *ewtab;
  95     __m256d          dummy_mask,cutoff_mask;
  96     __m128           tmpmask0,tmpmask1;
  97     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  98     __m256d          one     = _mm256_set1_pd(1.0);
  99     __m256d          two     = _mm256_set1_pd(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = _mm256_set1_pd(fr->epsfac);
 112     charge           = mdatoms->chargeA;
 113
 114     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 115     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 116     beta2            = _mm256_mul_pd(beta,beta);
 117     beta3            = _mm256_mul_pd(beta,beta2);
 118
 119     ewtab            = fr->ic->tabq_coul_FDV0;
 120     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 121     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 122
 123     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 124     rcutoff_scalar   = fr->rcoulomb;
 125     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 126     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 127
 128     /* Avoid stupid compiler warnings */
 129     jnrA = jnrB = jnrC = jnrD = 0;
 130     j_coord_offsetA = 0;
 131     j_coord_offsetB = 0;
 132     j_coord_offsetC = 0;
 133     j_coord_offsetD = 0;
 134
 135     outeriter        = 0;
 136     inneriter        = 0;
 137
 138     for(iidx=0;iidx<4*DIM;iidx++)
 139     {
 140         scratch[iidx] = 0.0;
 141     }
 142
 143     /* Start outer loop over neighborlists */
 144     for(iidx=0; iidx<nri; iidx++)
 145     {
 146         /* Load shift vector for this list */
 147         i_shift_offset   = DIM*shiftidx[iidx];
 148
 149         /* Load limits for loop over neighbors */
 150         j_index_start    = jindex[iidx];
 151         j_index_end      = jindex[iidx+1];
 152
 153         /* Get outer coordinate index */
 154         inr              = iinr[iidx];
 155         i_coord_offset   = DIM*inr;
 156
 157         /* Load i particle coords and add shift vector */
 158         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 159
 160         fix0             = _mm256_setzero_pd();
 161         fiy0             = _mm256_setzero_pd();
 162         fiz0             = _mm256_setzero_pd();
 163
 164         /* Load parameters for i particles */
 165         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 166
 167         /* Reset potential sums */
 168         velecsum         = _mm256_setzero_pd();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             jnrC             = jjnr[jidx+2];
 178             jnrD             = jjnr[jidx+3];
 179             j_coord_offsetA  = DIM*jnrA;
 180             j_coord_offsetB  = DIM*jnrB;
 181             j_coord_offsetC  = DIM*jnrC;
 182             j_coord_offsetD  = DIM*jnrD;
 183
 184             /* load j atom coordinates */
 185             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 186                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 187                                                  &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _mm256_sub_pd(ix0,jx0);
 191             dy00             = _mm256_sub_pd(iy0,jy0);
 192             dz00             = _mm256_sub_pd(iz0,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 196
 197             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 198
 199             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 200
 201             /* Load parameters for j particles */
 202             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 203                                                                  charge+jnrC+0,charge+jnrD+0);
 204
 205             /**************************
 206              * CALCULATE INTERACTIONS *
 207              **************************/
 208
 209             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 210             {
 211
 212             r00              = _mm256_mul_pd(rsq00,rinv00);
 213
 214             /* Compute parameters for interactions between i and j atoms */
 215             qq00             = _mm256_mul_pd(iq0,jq0);
 216
 217             /* EWALD ELECTROSTATICS */
 218
 219             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 220             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 221             ewitab           = _mm256_cvttpd_epi32(ewrt);
 222             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 223             ewitab           = _mm_slli_epi32(ewitab,2);
 224             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 225             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 226             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 227             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 228             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 229             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 230             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 231             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 232             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 233
 234             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 235
 236             /* Update potential sum for this i atom from the interaction with this j atom. */
 237             velec            = _mm256_and_pd(velec,cutoff_mask);
 238             velecsum         = _mm256_add_pd(velecsum,velec);
 239
 240             fscal            = felec;
 241
 242             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 243
 244             /* Calculate temporary vectorial force */
 245             tx               = _mm256_mul_pd(fscal,dx00);
 246             ty               = _mm256_mul_pd(fscal,dy00);
 247             tz               = _mm256_mul_pd(fscal,dz00);
 248
 249             /* Update vectorial force */
 250             fix0             = _mm256_add_pd(fix0,tx);
 251             fiy0             = _mm256_add_pd(fiy0,ty);
 252             fiz0             = _mm256_add_pd(fiz0,tz);
 253
 254             fjptrA             = f+j_coord_offsetA;
 255             fjptrB             = f+j_coord_offsetB;
 256             fjptrC             = f+j_coord_offsetC;
 257             fjptrD             = f+j_coord_offsetD;
 258             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 259
 260             }
 261
 262             /* Inner loop uses 46 flops */
 263         }
 264
 265         if(jidx<j_index_end)
 266         {
 267
 268             /* Get j neighbor index, and coordinate index */
 269             jnrlistA         = jjnr[jidx];
 270             jnrlistB         = jjnr[jidx+1];
 271             jnrlistC         = jjnr[jidx+2];
 272             jnrlistD         = jjnr[jidx+3];
 273             /* Sign of each element will be negative for non-real atoms.
 274              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 275              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 276              */
 277             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 278
 279             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 280             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 281             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 282
 283             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 284             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 285             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 286             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 287             j_coord_offsetA  = DIM*jnrA;
 288             j_coord_offsetB  = DIM*jnrB;
 289             j_coord_offsetC  = DIM*jnrC;
 290             j_coord_offsetD  = DIM*jnrD;
 291
 292             /* load j atom coordinates */
 293             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 294                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 295                                                  &jx0,&jy0,&jz0);
 296
 297             /* Calculate displacement vector */
 298             dx00             = _mm256_sub_pd(ix0,jx0);
 299             dy00             = _mm256_sub_pd(iy0,jy0);
 300             dz00             = _mm256_sub_pd(iz0,jz0);
 301
 302             /* Calculate squared distance and things based on it */
 303             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 304
 305             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 306
 307             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 308
 309             /* Load parameters for j particles */
 310             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 311                                                                  charge+jnrC+0,charge+jnrD+0);
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 318             {
 319
 320             r00              = _mm256_mul_pd(rsq00,rinv00);
 321             r00              = _mm256_andnot_pd(dummy_mask,r00);
 322
 323             /* Compute parameters for interactions between i and j atoms */
 324             qq00             = _mm256_mul_pd(iq0,jq0);
 325
 326             /* EWALD ELECTROSTATICS */
 327
 328             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 329             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 330             ewitab           = _mm256_cvttpd_epi32(ewrt);
 331             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 332             ewitab           = _mm_slli_epi32(ewitab,2);
 333             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 334             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 335             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 336             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 337             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 338             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 339             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 340             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 341             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 342
 343             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 344
 345             /* Update potential sum for this i atom from the interaction with this j atom. */
 346             velec            = _mm256_and_pd(velec,cutoff_mask);
 347             velec            = _mm256_andnot_pd(dummy_mask,velec);
 348             velecsum         = _mm256_add_pd(velecsum,velec);
 349
 350             fscal            = felec;
 351
 352             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 353
 354             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 355
 356             /* Calculate temporary vectorial force */
 357             tx               = _mm256_mul_pd(fscal,dx00);
 358             ty               = _mm256_mul_pd(fscal,dy00);
 359             tz               = _mm256_mul_pd(fscal,dz00);
 360
 361             /* Update vectorial force */
 362             fix0             = _mm256_add_pd(fix0,tx);
 363             fiy0             = _mm256_add_pd(fiy0,ty);
 364             fiz0             = _mm256_add_pd(fiz0,tz);
 365
 366             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 367             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 368             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 369             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 370             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 371
 372             }
 373
 374             /* Inner loop uses 47 flops */
 375         }
 376
 377         /* End of innermost loop */
 378
 379         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 380                                                  f+i_coord_offset,fshift+i_shift_offset);
 381
 382         ggid                        = gid[iidx];
 383         /* Update potential energies */
 384         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 385
 386         /* Increment number of inner iterations */
 387         inneriter                  += j_index_end - j_index_start;
 388
 389         /* Outer loop uses 8 flops */
 390     }
 391
 392     /* Increment number of outer iterations */
 393     outeriter        += nri;
 394
 395     /* Update outer/inner flops */
 396
 397     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*47);
 398 }
 399 /*
 400  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_256_double
 401  * Electrostatics interaction: Ewald
 402  * VdW interaction:            None
 403  * Geometry:                   Particle-Particle
 404  * Calculate force/pot:        Force
 405  */
 406 void
 407 nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_256_double
 408                     (t_nblist                    * gmx_restrict       nlist,
 409                      rvec                        * gmx_restrict          xx,
 410                      rvec                        * gmx_restrict          ff,
 411                      t_forcerec                  * gmx_restrict          fr,
 412                      t_mdatoms                   * gmx_restrict     mdatoms,
 413                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 414                      t_nrnb                      * gmx_restrict        nrnb)
 415 {
 416     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 417      * just 0 for non-waters.
 418      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 419      * jnr indices corresponding to data put in the four positions in the SIMD register.
 420      */
 421     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 422     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 423     int              jnrA,jnrB,jnrC,jnrD;
 424     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 425     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 426     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 427     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 428     real             rcutoff_scalar;
 429     real             *shiftvec,*fshift,*x,*f;
 430     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 431     real             scratch[4*DIM];
 432     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 433     real *           vdwioffsetptr0;
 434     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 435     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 436     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 437     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 438     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 439     real             *charge;
 440     __m128i          ewitab;
 441     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 442     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 443     real             *ewtab;
 444     __m256d          dummy_mask,cutoff_mask;
 445     __m128           tmpmask0,tmpmask1;
 446     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 447     __m256d          one     = _mm256_set1_pd(1.0);
 448     __m256d          two     = _mm256_set1_pd(2.0);
 449     x                = xx[0];
 450     f                = ff[0];
 451
 452     nri              = nlist->nri;
 453     iinr             = nlist->iinr;
 454     jindex           = nlist->jindex;
 455     jjnr             = nlist->jjnr;
 456     shiftidx         = nlist->shift;
 457     gid              = nlist->gid;
 458     shiftvec         = fr->shift_vec[0];
 459     fshift           = fr->fshift[0];
 460     facel            = _mm256_set1_pd(fr->epsfac);
 461     charge           = mdatoms->chargeA;
 462
 463     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 464     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 465     beta2            = _mm256_mul_pd(beta,beta);
 466     beta3            = _mm256_mul_pd(beta,beta2);
 467
 468     ewtab            = fr->ic->tabq_coul_F;
 469     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 470     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 471
 472     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 473     rcutoff_scalar   = fr->rcoulomb;
 474     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 475     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 476
 477     /* Avoid stupid compiler warnings */
 478     jnrA = jnrB = jnrC = jnrD = 0;
 479     j_coord_offsetA = 0;
 480     j_coord_offsetB = 0;
 481     j_coord_offsetC = 0;
 482     j_coord_offsetD = 0;
 483
 484     outeriter        = 0;
 485     inneriter        = 0;
 486
 487     for(iidx=0;iidx<4*DIM;iidx++)
 488     {
 489         scratch[iidx] = 0.0;
 490     }
 491
 492     /* Start outer loop over neighborlists */
 493     for(iidx=0; iidx<nri; iidx++)
 494     {
 495         /* Load shift vector for this list */
 496         i_shift_offset   = DIM*shiftidx[iidx];
 497
 498         /* Load limits for loop over neighbors */
 499         j_index_start    = jindex[iidx];
 500         j_index_end      = jindex[iidx+1];
 501
 502         /* Get outer coordinate index */
 503         inr              = iinr[iidx];
 504         i_coord_offset   = DIM*inr;
 505
 506         /* Load i particle coords and add shift vector */
 507         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 508
 509         fix0             = _mm256_setzero_pd();
 510         fiy0             = _mm256_setzero_pd();
 511         fiz0             = _mm256_setzero_pd();
 512
 513         /* Load parameters for i particles */
 514         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 515
 516         /* Start inner kernel loop */
 517         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 518         {
 519
 520             /* Get j neighbor index, and coordinate index */
 521             jnrA             = jjnr[jidx];
 522             jnrB             = jjnr[jidx+1];
 523             jnrC             = jjnr[jidx+2];
 524             jnrD             = jjnr[jidx+3];
 525             j_coord_offsetA  = DIM*jnrA;
 526             j_coord_offsetB  = DIM*jnrB;
 527             j_coord_offsetC  = DIM*jnrC;
 528             j_coord_offsetD  = DIM*jnrD;
 529
 530             /* load j atom coordinates */
 531             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 532                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 533                                                  &jx0,&jy0,&jz0);
 534
 535             /* Calculate displacement vector */
 536             dx00             = _mm256_sub_pd(ix0,jx0);
 537             dy00             = _mm256_sub_pd(iy0,jy0);
 538             dz00             = _mm256_sub_pd(iz0,jz0);
 539
 540             /* Calculate squared distance and things based on it */
 541             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 542
 543             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 544
 545             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 546
 547             /* Load parameters for j particles */
 548             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 549                                                                  charge+jnrC+0,charge+jnrD+0);
 550
 551             /**************************
 552              * CALCULATE INTERACTIONS *
 553              **************************/
 554
 555             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 556             {
 557
 558             r00              = _mm256_mul_pd(rsq00,rinv00);
 559
 560             /* Compute parameters for interactions between i and j atoms */
 561             qq00             = _mm256_mul_pd(iq0,jq0);
 562
 563             /* EWALD ELECTROSTATICS */
 564
 565             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 566             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 567             ewitab           = _mm256_cvttpd_epi32(ewrt);
 568             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 569             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 570                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 571                                             &ewtabF,&ewtabFn);
 572             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 573             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 574
 575             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 576
 577             fscal            = felec;
 578
 579             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 580
 581             /* Calculate temporary vectorial force */
 582             tx               = _mm256_mul_pd(fscal,dx00);
 583             ty               = _mm256_mul_pd(fscal,dy00);
 584             tz               = _mm256_mul_pd(fscal,dz00);
 585
 586             /* Update vectorial force */
 587             fix0             = _mm256_add_pd(fix0,tx);
 588             fiy0             = _mm256_add_pd(fiy0,ty);
 589             fiz0             = _mm256_add_pd(fiz0,tz);
 590
 591             fjptrA             = f+j_coord_offsetA;
 592             fjptrB             = f+j_coord_offsetB;
 593             fjptrC             = f+j_coord_offsetC;
 594             fjptrD             = f+j_coord_offsetD;
 595             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 596
 597             }
 598
 599             /* Inner loop uses 39 flops */
 600         }
 601
 602         if(jidx<j_index_end)
 603         {
 604
 605             /* Get j neighbor index, and coordinate index */
 606             jnrlistA         = jjnr[jidx];
 607             jnrlistB         = jjnr[jidx+1];
 608             jnrlistC         = jjnr[jidx+2];
 609             jnrlistD         = jjnr[jidx+3];
 610             /* Sign of each element will be negative for non-real atoms.
 611              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 612              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 613              */
 614             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 615
 616             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 617             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 618             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 619
 620             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 621             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 622             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 623             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 624             j_coord_offsetA  = DIM*jnrA;
 625             j_coord_offsetB  = DIM*jnrB;
 626             j_coord_offsetC  = DIM*jnrC;
 627             j_coord_offsetD  = DIM*jnrD;
 628
 629             /* load j atom coordinates */
 630             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 631                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 632                                                  &jx0,&jy0,&jz0);
 633
 634             /* Calculate displacement vector */
 635             dx00             = _mm256_sub_pd(ix0,jx0);
 636             dy00             = _mm256_sub_pd(iy0,jy0);
 637             dz00             = _mm256_sub_pd(iz0,jz0);
 638
 639             /* Calculate squared distance and things based on it */
 640             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 641
 642             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 643
 644             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 645
 646             /* Load parameters for j particles */
 647             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 648                                                                  charge+jnrC+0,charge+jnrD+0);
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 655             {
 656
 657             r00              = _mm256_mul_pd(rsq00,rinv00);
 658             r00              = _mm256_andnot_pd(dummy_mask,r00);
 659
 660             /* Compute parameters for interactions between i and j atoms */
 661             qq00             = _mm256_mul_pd(iq0,jq0);
 662
 663             /* EWALD ELECTROSTATICS */
 664
 665             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 666             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 667             ewitab           = _mm256_cvttpd_epi32(ewrt);
 668             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 669             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 670                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 671                                             &ewtabF,&ewtabFn);
 672             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 673             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 674
 675             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 676
 677             fscal            = felec;
 678
 679             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 680
 681             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 682
 683             /* Calculate temporary vectorial force */
 684             tx               = _mm256_mul_pd(fscal,dx00);
 685             ty               = _mm256_mul_pd(fscal,dy00);
 686             tz               = _mm256_mul_pd(fscal,dz00);
 687
 688             /* Update vectorial force */
 689             fix0             = _mm256_add_pd(fix0,tx);
 690             fiy0             = _mm256_add_pd(fiy0,ty);
 691             fiz0             = _mm256_add_pd(fiz0,tz);
 692
 693             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 694             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 695             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 696             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 697             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 698
 699             }
 700
 701             /* Inner loop uses 40 flops */
 702         }
 703
 704         /* End of innermost loop */
 705
 706         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 707                                                  f+i_coord_offset,fshift+i_shift_offset);
 708
 709         /* Increment number of inner iterations */
 710         inneriter                  += j_index_end - j_index_start;
 711
 712         /* Outer loop uses 7 flops */
 713     }
 714
 715     /* Increment number of outer iterations */
 716     outeriter        += nri;
 717
 718     /* Update outer/inner flops */
 719
 720     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 721 }