src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_VdwLJ_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_ElecEw_VdwLJ_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEw_VdwLJ_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     int              nvdwtype;
  94     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  98     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  99     __m128i          ewitab;
 100     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 101     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 102     real             *ewtab;
 103     __m256d          dummy_mask,cutoff_mask;
 104     __m128           tmpmask0,tmpmask1;
 105     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 106     __m256d          one     = _mm256_set1_pd(1.0);
 107     __m256d          two     = _mm256_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm256_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 126     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 127     beta2            = _mm256_mul_pd(beta,beta);
 128     beta3            = _mm256_mul_pd(beta,beta2);
 129
 130     ewtab            = fr->ic->tabq_coul_FDV0;
 131     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 132     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 133
 134     /* Avoid stupid compiler warnings */
 135     jnrA = jnrB = jnrC = jnrD = 0;
 136     j_coord_offsetA = 0;
 137     j_coord_offsetB = 0;
 138     j_coord_offsetC = 0;
 139     j_coord_offsetD = 0;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     for(iidx=0;iidx<4*DIM;iidx++)
 145     {
 146         scratch[iidx] = 0.0;
 147     }
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 165
 166         fix0             = _mm256_setzero_pd();
 167         fiy0             = _mm256_setzero_pd();
 168         fiz0             = _mm256_setzero_pd();
 169
 170         /* Load parameters for i particles */
 171         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 172         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm256_setzero_pd();
 176         vvdwsum          = _mm256_setzero_pd();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             jnrC             = jjnr[jidx+2];
 186             jnrD             = jjnr[jidx+3];
 187             j_coord_offsetA  = DIM*jnrA;
 188             j_coord_offsetB  = DIM*jnrB;
 189             j_coord_offsetC  = DIM*jnrC;
 190             j_coord_offsetD  = DIM*jnrD;
 191
 192             /* load j atom coordinates */
 193             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 194                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 195                                                  &jx0,&jy0,&jz0);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _mm256_sub_pd(ix0,jx0);
 199             dy00             = _mm256_sub_pd(iy0,jy0);
 200             dz00             = _mm256_sub_pd(iz0,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 204
 205             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 206
 207             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 208
 209             /* Load parameters for j particles */
 210             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 211                                                                  charge+jnrC+0,charge+jnrD+0);
 212             vdwjidx0A        = 2*vdwtype[jnrA+0];
 213             vdwjidx0B        = 2*vdwtype[jnrB+0];
 214             vdwjidx0C        = 2*vdwtype[jnrC+0];
 215             vdwjidx0D        = 2*vdwtype[jnrD+0];
 216
 217             /**************************
 218              * CALCULATE INTERACTIONS *
 219              **************************/
 220
 221             r00              = _mm256_mul_pd(rsq00,rinv00);
 222
 223             /* Compute parameters for interactions between i and j atoms */
 224             qq00             = _mm256_mul_pd(iq0,jq0);
 225             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 226                                             vdwioffsetptr0+vdwjidx0B,
 227                                             vdwioffsetptr0+vdwjidx0C,
 228                                             vdwioffsetptr0+vdwjidx0D,
 229                                             &c6_00,&c12_00);
 230
 231             /* EWALD ELECTROSTATICS */
 232
 233             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 234             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 235             ewitab           = _mm256_cvttpd_epi32(ewrt);
 236             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 237             ewitab           = _mm_slli_epi32(ewitab,2);
 238             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 239             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 240             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 241             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 242             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 243             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 244             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 245             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 246             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 247
 248             /* LENNARD-JONES DISPERSION/REPULSION */
 249
 250             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 251             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 252             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 253             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 254             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 255
 256             /* Update potential sum for this i atom from the interaction with this j atom. */
 257             velecsum         = _mm256_add_pd(velecsum,velec);
 258             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 259
 260             fscal            = _mm256_add_pd(felec,fvdw);
 261
 262             /* Calculate temporary vectorial force */
 263             tx               = _mm256_mul_pd(fscal,dx00);
 264             ty               = _mm256_mul_pd(fscal,dy00);
 265             tz               = _mm256_mul_pd(fscal,dz00);
 266
 267             /* Update vectorial force */
 268             fix0             = _mm256_add_pd(fix0,tx);
 269             fiy0             = _mm256_add_pd(fiy0,ty);
 270             fiz0             = _mm256_add_pd(fiz0,tz);
 271
 272             fjptrA             = f+j_coord_offsetA;
 273             fjptrB             = f+j_coord_offsetB;
 274             fjptrC             = f+j_coord_offsetC;
 275             fjptrD             = f+j_coord_offsetD;
 276             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 277
 278             /* Inner loop uses 53 flops */
 279         }
 280
 281         if(jidx<j_index_end)
 282         {
 283
 284             /* Get j neighbor index, and coordinate index */
 285             jnrlistA         = jjnr[jidx];
 286             jnrlistB         = jjnr[jidx+1];
 287             jnrlistC         = jjnr[jidx+2];
 288             jnrlistD         = jjnr[jidx+3];
 289             /* Sign of each element will be negative for non-real atoms.
 290              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 291              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 292              */
 293             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 294
 295             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 296             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 297             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 298
 299             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 300             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 301             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 302             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 303             j_coord_offsetA  = DIM*jnrA;
 304             j_coord_offsetB  = DIM*jnrB;
 305             j_coord_offsetC  = DIM*jnrC;
 306             j_coord_offsetD  = DIM*jnrD;
 307
 308             /* load j atom coordinates */
 309             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 310                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 311                                                  &jx0,&jy0,&jz0);
 312
 313             /* Calculate displacement vector */
 314             dx00             = _mm256_sub_pd(ix0,jx0);
 315             dy00             = _mm256_sub_pd(iy0,jy0);
 316             dz00             = _mm256_sub_pd(iz0,jz0);
 317
 318             /* Calculate squared distance and things based on it */
 319             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 320
 321             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 322
 323             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 324
 325             /* Load parameters for j particles */
 326             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 327                                                                  charge+jnrC+0,charge+jnrD+0);
 328             vdwjidx0A        = 2*vdwtype[jnrA+0];
 329             vdwjidx0B        = 2*vdwtype[jnrB+0];
 330             vdwjidx0C        = 2*vdwtype[jnrC+0];
 331             vdwjidx0D        = 2*vdwtype[jnrD+0];
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             r00              = _mm256_mul_pd(rsq00,rinv00);
 338             r00              = _mm256_andnot_pd(dummy_mask,r00);
 339
 340             /* Compute parameters for interactions between i and j atoms */
 341             qq00             = _mm256_mul_pd(iq0,jq0);
 342             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 343                                             vdwioffsetptr0+vdwjidx0B,
 344                                             vdwioffsetptr0+vdwjidx0C,
 345                                             vdwioffsetptr0+vdwjidx0D,
 346                                             &c6_00,&c12_00);
 347
 348             /* EWALD ELECTROSTATICS */
 349
 350             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 351             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 352             ewitab           = _mm256_cvttpd_epi32(ewrt);
 353             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 354             ewitab           = _mm_slli_epi32(ewitab,2);
 355             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 356             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 357             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 358             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 359             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 360             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 361             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 362             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 363             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 364
 365             /* LENNARD-JONES DISPERSION/REPULSION */
 366
 367             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 368             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 369             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 370             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 371             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 372
 373             /* Update potential sum for this i atom from the interaction with this j atom. */
 374             velec            = _mm256_andnot_pd(dummy_mask,velec);
 375             velecsum         = _mm256_add_pd(velecsum,velec);
 376             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 377             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 378
 379             fscal            = _mm256_add_pd(felec,fvdw);
 380
 381             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 382
 383             /* Calculate temporary vectorial force */
 384             tx               = _mm256_mul_pd(fscal,dx00);
 385             ty               = _mm256_mul_pd(fscal,dy00);
 386             tz               = _mm256_mul_pd(fscal,dz00);
 387
 388             /* Update vectorial force */
 389             fix0             = _mm256_add_pd(fix0,tx);
 390             fiy0             = _mm256_add_pd(fiy0,ty);
 391             fiz0             = _mm256_add_pd(fiz0,tz);
 392
 393             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 394             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 395             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 396             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 397             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 398
 399             /* Inner loop uses 54 flops */
 400         }
 401
 402         /* End of innermost loop */
 403
 404         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 405                                                  f+i_coord_offset,fshift+i_shift_offset);
 406
 407         ggid                        = gid[iidx];
 408         /* Update potential energies */
 409         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 410         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 411
 412         /* Increment number of inner iterations */
 413         inneriter                  += j_index_end - j_index_start;
 414
 415         /* Outer loop uses 9 flops */
 416     }
 417
 418     /* Increment number of outer iterations */
 419     outeriter        += nri;
 420
 421     /* Update outer/inner flops */
 422
 423     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 424 }
 425 /*
 426  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 427  * Electrostatics interaction: Ewald
 428  * VdW interaction:            LennardJones
 429  * Geometry:                   Particle-Particle
 430  * Calculate force/pot:        Force
 431  */
 432 void
 433 nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 434                     (t_nblist                    * gmx_restrict       nlist,
 435                      rvec                        * gmx_restrict          xx,
 436                      rvec                        * gmx_restrict          ff,
 437                      t_forcerec                  * gmx_restrict          fr,
 438                      t_mdatoms                   * gmx_restrict     mdatoms,
 439                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 440                      t_nrnb                      * gmx_restrict        nrnb)
 441 {
 442     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 443      * just 0 for non-waters.
 444      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 445      * jnr indices corresponding to data put in the four positions in the SIMD register.
 446      */
 447     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 448     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 449     int              jnrA,jnrB,jnrC,jnrD;
 450     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 451     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 452     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 453     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 454     real             rcutoff_scalar;
 455     real             *shiftvec,*fshift,*x,*f;
 456     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 457     real             scratch[4*DIM];
 458     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 459     real *           vdwioffsetptr0;
 460     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 461     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 462     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 463     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 464     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 465     real             *charge;
 466     int              nvdwtype;
 467     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 468     int              *vdwtype;
 469     real             *vdwparam;
 470     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 471     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 472     __m128i          ewitab;
 473     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 474     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 475     real             *ewtab;
 476     __m256d          dummy_mask,cutoff_mask;
 477     __m128           tmpmask0,tmpmask1;
 478     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 479     __m256d          one     = _mm256_set1_pd(1.0);
 480     __m256d          two     = _mm256_set1_pd(2.0);
 481     x                = xx[0];
 482     f                = ff[0];
 483
 484     nri              = nlist->nri;
 485     iinr             = nlist->iinr;
 486     jindex           = nlist->jindex;
 487     jjnr             = nlist->jjnr;
 488     shiftidx         = nlist->shift;
 489     gid              = nlist->gid;
 490     shiftvec         = fr->shift_vec[0];
 491     fshift           = fr->fshift[0];
 492     facel            = _mm256_set1_pd(fr->epsfac);
 493     charge           = mdatoms->chargeA;
 494     nvdwtype         = fr->ntype;
 495     vdwparam         = fr->nbfp;
 496     vdwtype          = mdatoms->typeA;
 497
 498     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 499     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 500     beta2            = _mm256_mul_pd(beta,beta);
 501     beta3            = _mm256_mul_pd(beta,beta2);
 502
 503     ewtab            = fr->ic->tabq_coul_F;
 504     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 505     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 506
 507     /* Avoid stupid compiler warnings */
 508     jnrA = jnrB = jnrC = jnrD = 0;
 509     j_coord_offsetA = 0;
 510     j_coord_offsetB = 0;
 511     j_coord_offsetC = 0;
 512     j_coord_offsetD = 0;
 513
 514     outeriter        = 0;
 515     inneriter        = 0;
 516
 517     for(iidx=0;iidx<4*DIM;iidx++)
 518     {
 519         scratch[iidx] = 0.0;
 520     }
 521
 522     /* Start outer loop over neighborlists */
 523     for(iidx=0; iidx<nri; iidx++)
 524     {
 525         /* Load shift vector for this list */
 526         i_shift_offset   = DIM*shiftidx[iidx];
 527
 528         /* Load limits for loop over neighbors */
 529         j_index_start    = jindex[iidx];
 530         j_index_end      = jindex[iidx+1];
 531
 532         /* Get outer coordinate index */
 533         inr              = iinr[iidx];
 534         i_coord_offset   = DIM*inr;
 535
 536         /* Load i particle coords and add shift vector */
 537         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 538
 539         fix0             = _mm256_setzero_pd();
 540         fiy0             = _mm256_setzero_pd();
 541         fiz0             = _mm256_setzero_pd();
 542
 543         /* Load parameters for i particles */
 544         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 545         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 546
 547         /* Start inner kernel loop */
 548         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 549         {
 550
 551             /* Get j neighbor index, and coordinate index */
 552             jnrA             = jjnr[jidx];
 553             jnrB             = jjnr[jidx+1];
 554             jnrC             = jjnr[jidx+2];
 555             jnrD             = jjnr[jidx+3];
 556             j_coord_offsetA  = DIM*jnrA;
 557             j_coord_offsetB  = DIM*jnrB;
 558             j_coord_offsetC  = DIM*jnrC;
 559             j_coord_offsetD  = DIM*jnrD;
 560
 561             /* load j atom coordinates */
 562             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 563                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 564                                                  &jx0,&jy0,&jz0);
 565
 566             /* Calculate displacement vector */
 567             dx00             = _mm256_sub_pd(ix0,jx0);
 568             dy00             = _mm256_sub_pd(iy0,jy0);
 569             dz00             = _mm256_sub_pd(iz0,jz0);
 570
 571             /* Calculate squared distance and things based on it */
 572             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 573
 574             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 575
 576             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 577
 578             /* Load parameters for j particles */
 579             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 580                                                                  charge+jnrC+0,charge+jnrD+0);
 581             vdwjidx0A        = 2*vdwtype[jnrA+0];
 582             vdwjidx0B        = 2*vdwtype[jnrB+0];
 583             vdwjidx0C        = 2*vdwtype[jnrC+0];
 584             vdwjidx0D        = 2*vdwtype[jnrD+0];
 585
 586             /**************************
 587              * CALCULATE INTERACTIONS *
 588              **************************/
 589
 590             r00              = _mm256_mul_pd(rsq00,rinv00);
 591
 592             /* Compute parameters for interactions between i and j atoms */
 593             qq00             = _mm256_mul_pd(iq0,jq0);
 594             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 595                                             vdwioffsetptr0+vdwjidx0B,
 596                                             vdwioffsetptr0+vdwjidx0C,
 597                                             vdwioffsetptr0+vdwjidx0D,
 598                                             &c6_00,&c12_00);
 599
 600             /* EWALD ELECTROSTATICS */
 601
 602             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 603             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 604             ewitab           = _mm256_cvttpd_epi32(ewrt);
 605             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 606             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 607                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 608                                             &ewtabF,&ewtabFn);
 609             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 610             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 611
 612             /* LENNARD-JONES DISPERSION/REPULSION */
 613
 614             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 615             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 616
 617             fscal            = _mm256_add_pd(felec,fvdw);
 618
 619             /* Calculate temporary vectorial force */
 620             tx               = _mm256_mul_pd(fscal,dx00);
 621             ty               = _mm256_mul_pd(fscal,dy00);
 622             tz               = _mm256_mul_pd(fscal,dz00);
 623
 624             /* Update vectorial force */
 625             fix0             = _mm256_add_pd(fix0,tx);
 626             fiy0             = _mm256_add_pd(fiy0,ty);
 627             fiz0             = _mm256_add_pd(fiz0,tz);
 628
 629             fjptrA             = f+j_coord_offsetA;
 630             fjptrB             = f+j_coord_offsetB;
 631             fjptrC             = f+j_coord_offsetC;
 632             fjptrD             = f+j_coord_offsetD;
 633             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 634
 635             /* Inner loop uses 43 flops */
 636         }
 637
 638         if(jidx<j_index_end)
 639         {
 640
 641             /* Get j neighbor index, and coordinate index */
 642             jnrlistA         = jjnr[jidx];
 643             jnrlistB         = jjnr[jidx+1];
 644             jnrlistC         = jjnr[jidx+2];
 645             jnrlistD         = jjnr[jidx+3];
 646             /* Sign of each element will be negative for non-real atoms.
 647              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 648              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 649              */
 650             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 651
 652             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 653             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 654             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 655
 656             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 657             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 658             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 659             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 660             j_coord_offsetA  = DIM*jnrA;
 661             j_coord_offsetB  = DIM*jnrB;
 662             j_coord_offsetC  = DIM*jnrC;
 663             j_coord_offsetD  = DIM*jnrD;
 664
 665             /* load j atom coordinates */
 666             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 667                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 668                                                  &jx0,&jy0,&jz0);
 669
 670             /* Calculate displacement vector */
 671             dx00             = _mm256_sub_pd(ix0,jx0);
 672             dy00             = _mm256_sub_pd(iy0,jy0);
 673             dz00             = _mm256_sub_pd(iz0,jz0);
 674
 675             /* Calculate squared distance and things based on it */
 676             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 677
 678             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 679
 680             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 681
 682             /* Load parameters for j particles */
 683             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 684                                                                  charge+jnrC+0,charge+jnrD+0);
 685             vdwjidx0A        = 2*vdwtype[jnrA+0];
 686             vdwjidx0B        = 2*vdwtype[jnrB+0];
 687             vdwjidx0C        = 2*vdwtype[jnrC+0];
 688             vdwjidx0D        = 2*vdwtype[jnrD+0];
 689
 690             /**************************
 691              * CALCULATE INTERACTIONS *
 692              **************************/
 693
 694             r00              = _mm256_mul_pd(rsq00,rinv00);
 695             r00              = _mm256_andnot_pd(dummy_mask,r00);
 696
 697             /* Compute parameters for interactions between i and j atoms */
 698             qq00             = _mm256_mul_pd(iq0,jq0);
 699             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 700                                             vdwioffsetptr0+vdwjidx0B,
 701                                             vdwioffsetptr0+vdwjidx0C,
 702                                             vdwioffsetptr0+vdwjidx0D,
 703                                             &c6_00,&c12_00);
 704
 705             /* EWALD ELECTROSTATICS */
 706
 707             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 708             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 709             ewitab           = _mm256_cvttpd_epi32(ewrt);
 710             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 711             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 712                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 713                                             &ewtabF,&ewtabFn);
 714             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 715             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 716
 717             /* LENNARD-JONES DISPERSION/REPULSION */
 718
 719             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 720             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 721
 722             fscal            = _mm256_add_pd(felec,fvdw);
 723
 724             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 725
 726             /* Calculate temporary vectorial force */
 727             tx               = _mm256_mul_pd(fscal,dx00);
 728             ty               = _mm256_mul_pd(fscal,dy00);
 729             tz               = _mm256_mul_pd(fscal,dz00);
 730
 731             /* Update vectorial force */
 732             fix0             = _mm256_add_pd(fix0,tx);
 733             fiy0             = _mm256_add_pd(fiy0,ty);
 734             fiz0             = _mm256_add_pd(fiz0,tz);
 735
 736             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 737             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 738             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 739             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 740             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 741
 742             /* Inner loop uses 44 flops */
 743         }
 744
 745         /* End of innermost loop */
 746
 747         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 748                                                  f+i_coord_offset,fshift+i_shift_offset);
 749
 750         /* Increment number of inner iterations */
 751         inneriter                  += j_index_end - j_index_start;
 752
 753         /* Outer loop uses 7 flops */
 754     }
 755
 756     /* Increment number of outer iterations */
 757     outeriter        += nri;
 758
 759     /* Update outer/inner flops */
 760
 761     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*44);
 762 }