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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/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_ElecEwSh_VdwNone_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSh_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          ewitab;
  94     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  95     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  96     real             *ewtab;
  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     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 117     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 118     beta2            = _mm256_mul_pd(beta,beta);
 119     beta3            = _mm256_mul_pd(beta,beta2);
 120
 121     ewtab            = fr->ic->tabq_coul_FDV0;
 122     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 123     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 124
 125     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 126     rcutoff_scalar   = fr->rcoulomb;
 127     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 128     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 129
 130     /* Avoid stupid compiler warnings */
 131     jnrA = jnrB = jnrC = jnrD = 0;
 132     j_coord_offsetA = 0;
 133     j_coord_offsetB = 0;
 134     j_coord_offsetC = 0;
 135     j_coord_offsetD = 0;
 136
 137     outeriter        = 0;
 138     inneriter        = 0;
 139
 140     for(iidx=0;iidx<4*DIM;iidx++)
 141     {
 142         scratch[iidx] = 0.0;
 143     }
 144
 145     /* Start outer loop over neighborlists */
 146     for(iidx=0; iidx<nri; iidx++)
 147     {
 148         /* Load shift vector for this list */
 149         i_shift_offset   = DIM*shiftidx[iidx];
 150
 151         /* Load limits for loop over neighbors */
 152         j_index_start    = jindex[iidx];
 153         j_index_end      = jindex[iidx+1];
 154
 155         /* Get outer coordinate index */
 156         inr              = iinr[iidx];
 157         i_coord_offset   = DIM*inr;
 158
 159         /* Load i particle coords and add shift vector */
 160         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 161
 162         fix0             = _mm256_setzero_pd();
 163         fiy0             = _mm256_setzero_pd();
 164         fiz0             = _mm256_setzero_pd();
 165
 166         /* Load parameters for i particles */
 167         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 168
 169         /* Reset potential sums */
 170         velecsum         = _mm256_setzero_pd();
 171
 172         /* Start inner kernel loop */
 173         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 174         {
 175
 176             /* Get j neighbor index, and coordinate index */
 177             jnrA             = jjnr[jidx];
 178             jnrB             = jjnr[jidx+1];
 179             jnrC             = jjnr[jidx+2];
 180             jnrD             = jjnr[jidx+3];
 181             j_coord_offsetA  = DIM*jnrA;
 182             j_coord_offsetB  = DIM*jnrB;
 183             j_coord_offsetC  = DIM*jnrC;
 184             j_coord_offsetD  = DIM*jnrD;
 185
 186             /* load j atom coordinates */
 187             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 188                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 189                                                  &jx0,&jy0,&jz0);
 190
 191             /* Calculate displacement vector */
 192             dx00             = _mm256_sub_pd(ix0,jx0);
 193             dy00             = _mm256_sub_pd(iy0,jy0);
 194             dz00             = _mm256_sub_pd(iz0,jz0);
 195
 196             /* Calculate squared distance and things based on it */
 197             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 198
 199             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 200
 201             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 202
 203             /* Load parameters for j particles */
 204             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 205                                                                  charge+jnrC+0,charge+jnrD+0);
 206
 207             /**************************
 208              * CALCULATE INTERACTIONS *
 209              **************************/
 210
 211             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 212             {
 213
 214             r00              = _mm256_mul_pd(rsq00,rinv00);
 215
 216             /* Compute parameters for interactions between i and j atoms */
 217             qq00             = _mm256_mul_pd(iq0,jq0);
 218
 219             /* EWALD ELECTROSTATICS */
 220
 221             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 222             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 223             ewitab           = _mm256_cvttpd_epi32(ewrt);
 224             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 225             ewitab           = _mm_slli_epi32(ewitab,2);
 226             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 227             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 228             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 229             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 230             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 231             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 232             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 233             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 234             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 235
 236             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 237
 238             /* Update potential sum for this i atom from the interaction with this j atom. */
 239             velec            = _mm256_and_pd(velec,cutoff_mask);
 240             velecsum         = _mm256_add_pd(velecsum,velec);
 241
 242             fscal            = felec;
 243
 244             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 245
 246             /* Calculate temporary vectorial force */
 247             tx               = _mm256_mul_pd(fscal,dx00);
 248             ty               = _mm256_mul_pd(fscal,dy00);
 249             tz               = _mm256_mul_pd(fscal,dz00);
 250
 251             /* Update vectorial force */
 252             fix0             = _mm256_add_pd(fix0,tx);
 253             fiy0             = _mm256_add_pd(fiy0,ty);
 254             fiz0             = _mm256_add_pd(fiz0,tz);
 255
 256             fjptrA             = f+j_coord_offsetA;
 257             fjptrB             = f+j_coord_offsetB;
 258             fjptrC             = f+j_coord_offsetC;
 259             fjptrD             = f+j_coord_offsetD;
 260             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 261
 262             }
 263
 264             /* Inner loop uses 46 flops */
 265         }
 266
 267         if(jidx<j_index_end)
 268         {
 269
 270             /* Get j neighbor index, and coordinate index */
 271             jnrlistA         = jjnr[jidx];
 272             jnrlistB         = jjnr[jidx+1];
 273             jnrlistC         = jjnr[jidx+2];
 274             jnrlistD         = jjnr[jidx+3];
 275             /* Sign of each element will be negative for non-real atoms.
 276              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 277              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 278              */
 279             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 280
 281             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 282             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 283             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 284
 285             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 286             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 287             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 288             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 289             j_coord_offsetA  = DIM*jnrA;
 290             j_coord_offsetB  = DIM*jnrB;
 291             j_coord_offsetC  = DIM*jnrC;
 292             j_coord_offsetD  = DIM*jnrD;
 293
 294             /* load j atom coordinates */
 295             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 296                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 297                                                  &jx0,&jy0,&jz0);
 298
 299             /* Calculate displacement vector */
 300             dx00             = _mm256_sub_pd(ix0,jx0);
 301             dy00             = _mm256_sub_pd(iy0,jy0);
 302             dz00             = _mm256_sub_pd(iz0,jz0);
 303
 304             /* Calculate squared distance and things based on it */
 305             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 306
 307             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 308
 309             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 310
 311             /* Load parameters for j particles */
 312             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 313                                                                  charge+jnrC+0,charge+jnrD+0);
 314
 315             /**************************
 316              * CALCULATE INTERACTIONS *
 317              **************************/
 318
 319             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 320             {
 321
 322             r00              = _mm256_mul_pd(rsq00,rinv00);
 323             r00              = _mm256_andnot_pd(dummy_mask,r00);
 324
 325             /* Compute parameters for interactions between i and j atoms */
 326             qq00             = _mm256_mul_pd(iq0,jq0);
 327
 328             /* EWALD ELECTROSTATICS */
 329
 330             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 331             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 332             ewitab           = _mm256_cvttpd_epi32(ewrt);
 333             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 334             ewitab           = _mm_slli_epi32(ewitab,2);
 335             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 336             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 337             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 338             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 339             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 340             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 341             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 342             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 343             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 344
 345             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velec            = _mm256_and_pd(velec,cutoff_mask);
 349             velec            = _mm256_andnot_pd(dummy_mask,velec);
 350             velecsum         = _mm256_add_pd(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 355
 356             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 357
 358             /* Calculate temporary vectorial force */
 359             tx               = _mm256_mul_pd(fscal,dx00);
 360             ty               = _mm256_mul_pd(fscal,dy00);
 361             tz               = _mm256_mul_pd(fscal,dz00);
 362
 363             /* Update vectorial force */
 364             fix0             = _mm256_add_pd(fix0,tx);
 365             fiy0             = _mm256_add_pd(fiy0,ty);
 366             fiz0             = _mm256_add_pd(fiz0,tz);
 367
 368             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 369             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 370             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 371             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 372             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 373
 374             }
 375
 376             /* Inner loop uses 47 flops */
 377         }
 378
 379         /* End of innermost loop */
 380
 381         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 382                                                  f+i_coord_offset,fshift+i_shift_offset);
 383
 384         ggid                        = gid[iidx];
 385         /* Update potential energies */
 386         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 387
 388         /* Increment number of inner iterations */
 389         inneriter                  += j_index_end - j_index_start;
 390
 391         /* Outer loop uses 8 flops */
 392     }
 393
 394     /* Increment number of outer iterations */
 395     outeriter        += nri;
 396
 397     /* Update outer/inner flops */
 398
 399     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*47);
 400 }
 401 /*
 402  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_256_double
 403  * Electrostatics interaction: Ewald
 404  * VdW interaction:            None
 405  * Geometry:                   Particle-Particle
 406  * Calculate force/pot:        Force
 407  */
 408 void
 409 nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_256_double
 410                     (t_nblist                    * gmx_restrict       nlist,
 411                      rvec                        * gmx_restrict          xx,
 412                      rvec                        * gmx_restrict          ff,
 413                      t_forcerec                  * gmx_restrict          fr,
 414                      t_mdatoms                   * gmx_restrict     mdatoms,
 415                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 416                      t_nrnb                      * gmx_restrict        nrnb)
 417 {
 418     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 419      * just 0 for non-waters.
 420      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 421      * jnr indices corresponding to data put in the four positions in the SIMD register.
 422      */
 423     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 424     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 425     int              jnrA,jnrB,jnrC,jnrD;
 426     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 427     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 428     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 429     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 430     real             rcutoff_scalar;
 431     real             *shiftvec,*fshift,*x,*f;
 432     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 433     real             scratch[4*DIM];
 434     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 435     real *           vdwioffsetptr0;
 436     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 437     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 438     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 439     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 440     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 441     real             *charge;
 442     __m128i          ewitab;
 443     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 444     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 445     real             *ewtab;
 446     __m256d          dummy_mask,cutoff_mask;
 447     __m128           tmpmask0,tmpmask1;
 448     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 449     __m256d          one     = _mm256_set1_pd(1.0);
 450     __m256d          two     = _mm256_set1_pd(2.0);
 451     x                = xx[0];
 452     f                = ff[0];
 453
 454     nri              = nlist->nri;
 455     iinr             = nlist->iinr;
 456     jindex           = nlist->jindex;
 457     jjnr             = nlist->jjnr;
 458     shiftidx         = nlist->shift;
 459     gid              = nlist->gid;
 460     shiftvec         = fr->shift_vec[0];
 461     fshift           = fr->fshift[0];
 462     facel            = _mm256_set1_pd(fr->epsfac);
 463     charge           = mdatoms->chargeA;
 464
 465     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 466     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 467     beta2            = _mm256_mul_pd(beta,beta);
 468     beta3            = _mm256_mul_pd(beta,beta2);
 469
 470     ewtab            = fr->ic->tabq_coul_F;
 471     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 472     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 473
 474     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 475     rcutoff_scalar   = fr->rcoulomb;
 476     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 477     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 478
 479     /* Avoid stupid compiler warnings */
 480     jnrA = jnrB = jnrC = jnrD = 0;
 481     j_coord_offsetA = 0;
 482     j_coord_offsetB = 0;
 483     j_coord_offsetC = 0;
 484     j_coord_offsetD = 0;
 485
 486     outeriter        = 0;
 487     inneriter        = 0;
 488
 489     for(iidx=0;iidx<4*DIM;iidx++)
 490     {
 491         scratch[iidx] = 0.0;
 492     }
 493
 494     /* Start outer loop over neighborlists */
 495     for(iidx=0; iidx<nri; iidx++)
 496     {
 497         /* Load shift vector for this list */
 498         i_shift_offset   = DIM*shiftidx[iidx];
 499
 500         /* Load limits for loop over neighbors */
 501         j_index_start    = jindex[iidx];
 502         j_index_end      = jindex[iidx+1];
 503
 504         /* Get outer coordinate index */
 505         inr              = iinr[iidx];
 506         i_coord_offset   = DIM*inr;
 507
 508         /* Load i particle coords and add shift vector */
 509         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 510
 511         fix0             = _mm256_setzero_pd();
 512         fiy0             = _mm256_setzero_pd();
 513         fiz0             = _mm256_setzero_pd();
 514
 515         /* Load parameters for i particles */
 516         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 517
 518         /* Start inner kernel loop */
 519         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 520         {
 521
 522             /* Get j neighbor index, and coordinate index */
 523             jnrA             = jjnr[jidx];
 524             jnrB             = jjnr[jidx+1];
 525             jnrC             = jjnr[jidx+2];
 526             jnrD             = jjnr[jidx+3];
 527             j_coord_offsetA  = DIM*jnrA;
 528             j_coord_offsetB  = DIM*jnrB;
 529             j_coord_offsetC  = DIM*jnrC;
 530             j_coord_offsetD  = DIM*jnrD;
 531
 532             /* load j atom coordinates */
 533             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 534                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 535                                                  &jx0,&jy0,&jz0);
 536
 537             /* Calculate displacement vector */
 538             dx00             = _mm256_sub_pd(ix0,jx0);
 539             dy00             = _mm256_sub_pd(iy0,jy0);
 540             dz00             = _mm256_sub_pd(iz0,jz0);
 541
 542             /* Calculate squared distance and things based on it */
 543             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 544
 545             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 546
 547             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 548
 549             /* Load parameters for j particles */
 550             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 551                                                                  charge+jnrC+0,charge+jnrD+0);
 552
 553             /**************************
 554              * CALCULATE INTERACTIONS *
 555              **************************/
 556
 557             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 558             {
 559
 560             r00              = _mm256_mul_pd(rsq00,rinv00);
 561
 562             /* Compute parameters for interactions between i and j atoms */
 563             qq00             = _mm256_mul_pd(iq0,jq0);
 564
 565             /* EWALD ELECTROSTATICS */
 566
 567             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 568             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 569             ewitab           = _mm256_cvttpd_epi32(ewrt);
 570             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 571             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 572                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 573                                             &ewtabF,&ewtabFn);
 574             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 575             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 576
 577             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 578
 579             fscal            = felec;
 580
 581             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 582
 583             /* Calculate temporary vectorial force */
 584             tx               = _mm256_mul_pd(fscal,dx00);
 585             ty               = _mm256_mul_pd(fscal,dy00);
 586             tz               = _mm256_mul_pd(fscal,dz00);
 587
 588             /* Update vectorial force */
 589             fix0             = _mm256_add_pd(fix0,tx);
 590             fiy0             = _mm256_add_pd(fiy0,ty);
 591             fiz0             = _mm256_add_pd(fiz0,tz);
 592
 593             fjptrA             = f+j_coord_offsetA;
 594             fjptrB             = f+j_coord_offsetB;
 595             fjptrC             = f+j_coord_offsetC;
 596             fjptrD             = f+j_coord_offsetD;
 597             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 598
 599             }
 600
 601             /* Inner loop uses 39 flops */
 602         }
 603
 604         if(jidx<j_index_end)
 605         {
 606
 607             /* Get j neighbor index, and coordinate index */
 608             jnrlistA         = jjnr[jidx];
 609             jnrlistB         = jjnr[jidx+1];
 610             jnrlistC         = jjnr[jidx+2];
 611             jnrlistD         = jjnr[jidx+3];
 612             /* Sign of each element will be negative for non-real atoms.
 613              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 614              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 615              */
 616             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 617
 618             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 619             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 620             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 621
 622             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 623             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 624             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 625             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 626             j_coord_offsetA  = DIM*jnrA;
 627             j_coord_offsetB  = DIM*jnrB;
 628             j_coord_offsetC  = DIM*jnrC;
 629             j_coord_offsetD  = DIM*jnrD;
 630
 631             /* load j atom coordinates */
 632             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 633                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 634                                                  &jx0,&jy0,&jz0);
 635
 636             /* Calculate displacement vector */
 637             dx00             = _mm256_sub_pd(ix0,jx0);
 638             dy00             = _mm256_sub_pd(iy0,jy0);
 639             dz00             = _mm256_sub_pd(iz0,jz0);
 640
 641             /* Calculate squared distance and things based on it */
 642             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 643
 644             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 645
 646             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 647
 648             /* Load parameters for j particles */
 649             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 650                                                                  charge+jnrC+0,charge+jnrD+0);
 651
 652             /**************************
 653              * CALCULATE INTERACTIONS *
 654              **************************/
 655
 656             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 657             {
 658
 659             r00              = _mm256_mul_pd(rsq00,rinv00);
 660             r00              = _mm256_andnot_pd(dummy_mask,r00);
 661
 662             /* Compute parameters for interactions between i and j atoms */
 663             qq00             = _mm256_mul_pd(iq0,jq0);
 664
 665             /* EWALD ELECTROSTATICS */
 666
 667             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 668             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 669             ewitab           = _mm256_cvttpd_epi32(ewrt);
 670             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 671             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 672                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 673                                             &ewtabF,&ewtabFn);
 674             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 675             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 676
 677             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 678
 679             fscal            = felec;
 680
 681             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 682
 683             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 684
 685             /* Calculate temporary vectorial force */
 686             tx               = _mm256_mul_pd(fscal,dx00);
 687             ty               = _mm256_mul_pd(fscal,dy00);
 688             tz               = _mm256_mul_pd(fscal,dz00);
 689
 690             /* Update vectorial force */
 691             fix0             = _mm256_add_pd(fix0,tx);
 692             fiy0             = _mm256_add_pd(fiy0,ty);
 693             fiz0             = _mm256_add_pd(fiz0,tz);
 694
 695             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 696             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 697             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 698             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 699             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 700
 701             }
 702
 703             /* Inner loop uses 40 flops */
 704         }
 705
 706         /* End of innermost loop */
 707
 708         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 709                                                  f+i_coord_offset,fshift+i_shift_offset);
 710
 711         /* Increment number of inner iterations */
 712         inneriter                  += j_index_end - j_index_start;
 713
 714         /* Outer loop uses 7 flops */
 715     }
 716
 717     /* Increment number of outer iterations */
 718     outeriter        += nri;
 719
 720     /* Update outer/inner flops */
 721
 722     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 723 }