src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecGB_VdwNone_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: GeneralizedBorn
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_256_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     real *           vdwioffsetptr0;
  84     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     __m128i          gbitab;
  91     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  92     __m256d          minushalf = _mm256_set1_pd(-0.5);
  93     real             *invsqrta,*dvda,*gbtab;
  94     __m128i          vfitab;
  95     __m128i          ifour       = _mm_set1_epi32(4);
  96     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  97     real             *vftab;
  98     __m256d          dummy_mask,cutoff_mask;
  99     __m128           tmpmask0,tmpmask1;
 100     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 101     __m256d          one     = _mm256_set1_pd(1.0);
 102     __m256d          two     = _mm256_set1_pd(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm256_set1_pd(fr->ic->epsfac);
 115     charge           = mdatoms->chargeA;
 116
 117     invsqrta         = fr->invsqrta;
 118     dvda             = fr->dvda;
 119     gbtabscale       = _mm256_set1_pd(fr->gbtab->scale);
 120     gbtab            = fr->gbtab->data;
 121     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 122
 123     /* Avoid stupid compiler warnings */
 124     jnrA = jnrB = jnrC = jnrD = 0;
 125     j_coord_offsetA = 0;
 126     j_coord_offsetB = 0;
 127     j_coord_offsetC = 0;
 128     j_coord_offsetD = 0;
 129
 130     outeriter        = 0;
 131     inneriter        = 0;
 132
 133     for(iidx=0;iidx<4*DIM;iidx++)
 134     {
 135         scratch[iidx] = 0.0;
 136     }
 137
 138     /* Start outer loop over neighborlists */
 139     for(iidx=0; iidx<nri; iidx++)
 140     {
 141         /* Load shift vector for this list */
 142         i_shift_offset   = DIM*shiftidx[iidx];
 143
 144         /* Load limits for loop over neighbors */
 145         j_index_start    = jindex[iidx];
 146         j_index_end      = jindex[iidx+1];
 147
 148         /* Get outer coordinate index */
 149         inr              = iinr[iidx];
 150         i_coord_offset   = DIM*inr;
 151
 152         /* Load i particle coords and add shift vector */
 153         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 154
 155         fix0             = _mm256_setzero_pd();
 156         fiy0             = _mm256_setzero_pd();
 157         fiz0             = _mm256_setzero_pd();
 158
 159         /* Load parameters for i particles */
 160         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 161         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 162
 163         /* Reset potential sums */
 164         velecsum         = _mm256_setzero_pd();
 165         vgbsum           = _mm256_setzero_pd();
 166         dvdasum          = _mm256_setzero_pd();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             jnrC             = jjnr[jidx+2];
 176             jnrD             = jjnr[jidx+3];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179             j_coord_offsetC  = DIM*jnrC;
 180             j_coord_offsetD  = DIM*jnrD;
 181
 182             /* load j atom coordinates */
 183             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 184                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 185                                                  &jx0,&jy0,&jz0);
 186
 187             /* Calculate displacement vector */
 188             dx00             = _mm256_sub_pd(ix0,jx0);
 189             dy00             = _mm256_sub_pd(iy0,jy0);
 190             dz00             = _mm256_sub_pd(iz0,jz0);
 191
 192             /* Calculate squared distance and things based on it */
 193             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 194
 195             rinv00           = avx256_invsqrt_d(rsq00);
 196
 197             /* Load parameters for j particles */
 198             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 199                                                                  charge+jnrC+0,charge+jnrD+0);
 200             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 201                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 202
 203             /**************************
 204              * CALCULATE INTERACTIONS *
 205              **************************/
 206
 207             r00              = _mm256_mul_pd(rsq00,rinv00);
 208
 209             /* Compute parameters for interactions between i and j atoms */
 210             qq00             = _mm256_mul_pd(iq0,jq0);
 211
 212             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 213             isaprod          = _mm256_mul_pd(isai0,isaj0);
 214             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 215             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 216
 217             /* Calculate generalized born table index - this is a separate table from the normal one,
 218              * but we use the same procedure by multiplying r with scale and truncating to integer.
 219              */
 220             rt               = _mm256_mul_pd(r00,gbscale);
 221             gbitab           = _mm256_cvttpd_epi32(rt);
 222             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 223             gbitab           = _mm_slli_epi32(gbitab,2);
 224             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 225             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 226             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 227             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 228             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 229             Heps             = _mm256_mul_pd(gbeps,H);
 230             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 231             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 232             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 233
 234             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 235             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 236             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 237             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 238             fjptrA           = dvda+jnrA;
 239             fjptrB           = dvda+jnrB;
 240             fjptrC           = dvda+jnrC;
 241             fjptrD           = dvda+jnrD;
 242             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 243                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 244             velec            = _mm256_mul_pd(qq00,rinv00);
 245             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 246
 247             /* Update potential sum for this i atom from the interaction with this j atom. */
 248             velecsum         = _mm256_add_pd(velecsum,velec);
 249             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 250
 251             fscal            = felec;
 252
 253             /* Calculate temporary vectorial force */
 254             tx               = _mm256_mul_pd(fscal,dx00);
 255             ty               = _mm256_mul_pd(fscal,dy00);
 256             tz               = _mm256_mul_pd(fscal,dz00);
 257
 258             /* Update vectorial force */
 259             fix0             = _mm256_add_pd(fix0,tx);
 260             fiy0             = _mm256_add_pd(fiy0,ty);
 261             fiz0             = _mm256_add_pd(fiz0,tz);
 262
 263             fjptrA             = f+j_coord_offsetA;
 264             fjptrB             = f+j_coord_offsetB;
 265             fjptrC             = f+j_coord_offsetC;
 266             fjptrD             = f+j_coord_offsetD;
 267             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 268
 269             /* Inner loop uses 57 flops */
 270         }
 271
 272         if(jidx<j_index_end)
 273         {
 274
 275             /* Get j neighbor index, and coordinate index */
 276             jnrlistA         = jjnr[jidx];
 277             jnrlistB         = jjnr[jidx+1];
 278             jnrlistC         = jjnr[jidx+2];
 279             jnrlistD         = jjnr[jidx+3];
 280             /* Sign of each element will be negative for non-real atoms.
 281              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 282              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 283              */
 284             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 285
 286             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 287             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 288             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 289
 290             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 291             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 292             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 293             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 294             j_coord_offsetA  = DIM*jnrA;
 295             j_coord_offsetB  = DIM*jnrB;
 296             j_coord_offsetC  = DIM*jnrC;
 297             j_coord_offsetD  = DIM*jnrD;
 298
 299             /* load j atom coordinates */
 300             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 301                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 302                                                  &jx0,&jy0,&jz0);
 303
 304             /* Calculate displacement vector */
 305             dx00             = _mm256_sub_pd(ix0,jx0);
 306             dy00             = _mm256_sub_pd(iy0,jy0);
 307             dz00             = _mm256_sub_pd(iz0,jz0);
 308
 309             /* Calculate squared distance and things based on it */
 310             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 311
 312             rinv00           = avx256_invsqrt_d(rsq00);
 313
 314             /* Load parameters for j particles */
 315             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 316                                                                  charge+jnrC+0,charge+jnrD+0);
 317             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 318                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 319
 320             /**************************
 321              * CALCULATE INTERACTIONS *
 322              **************************/
 323
 324             r00              = _mm256_mul_pd(rsq00,rinv00);
 325             r00              = _mm256_andnot_pd(dummy_mask,r00);
 326
 327             /* Compute parameters for interactions between i and j atoms */
 328             qq00             = _mm256_mul_pd(iq0,jq0);
 329
 330             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 331             isaprod          = _mm256_mul_pd(isai0,isaj0);
 332             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 333             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 334
 335             /* Calculate generalized born table index - this is a separate table from the normal one,
 336              * but we use the same procedure by multiplying r with scale and truncating to integer.
 337              */
 338             rt               = _mm256_mul_pd(r00,gbscale);
 339             gbitab           = _mm256_cvttpd_epi32(rt);
 340             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 341             gbitab           = _mm_slli_epi32(gbitab,2);
 342             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 343             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 344             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 345             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 346             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 347             Heps             = _mm256_mul_pd(gbeps,H);
 348             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 349             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 350             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 351
 352             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 353             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 354             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 355             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 356             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 357             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 358             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 359             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 360             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 361             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 362             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 363                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 364             velec            = _mm256_mul_pd(qq00,rinv00);
 365             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velec            = _mm256_andnot_pd(dummy_mask,velec);
 369             velecsum         = _mm256_add_pd(velecsum,velec);
 370             vgb              = _mm256_andnot_pd(dummy_mask,vgb);
 371             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 372
 373             fscal            = felec;
 374
 375             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 376
 377             /* Calculate temporary vectorial force */
 378             tx               = _mm256_mul_pd(fscal,dx00);
 379             ty               = _mm256_mul_pd(fscal,dy00);
 380             tz               = _mm256_mul_pd(fscal,dz00);
 381
 382             /* Update vectorial force */
 383             fix0             = _mm256_add_pd(fix0,tx);
 384             fiy0             = _mm256_add_pd(fiy0,ty);
 385             fiz0             = _mm256_add_pd(fiz0,tz);
 386
 387             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 388             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 389             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 390             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 391             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 392
 393             /* Inner loop uses 58 flops */
 394         }
 395
 396         /* End of innermost loop */
 397
 398         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 399                                                  f+i_coord_offset,fshift+i_shift_offset);
 400
 401         ggid                        = gid[iidx];
 402         /* Update potential energies */
 403         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 404         gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 405         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 406         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 407
 408         /* Increment number of inner iterations */
 409         inneriter                  += j_index_end - j_index_start;
 410
 411         /* Outer loop uses 9 flops */
 412     }
 413
 414     /* Increment number of outer iterations */
 415     outeriter        += nri;
 416
 417     /* Update outer/inner flops */
 418
 419     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
 420 }
 421 /*
 422  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
 423  * Electrostatics interaction: GeneralizedBorn
 424  * VdW interaction:            None
 425  * Geometry:                   Particle-Particle
 426  * Calculate force/pot:        Force
 427  */
 428 void
 429 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
 430                     (t_nblist                    * gmx_restrict       nlist,
 431                      rvec                        * gmx_restrict          xx,
 432                      rvec                        * gmx_restrict          ff,
 433                      struct t_forcerec           * gmx_restrict          fr,
 434                      t_mdatoms                   * gmx_restrict     mdatoms,
 435                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 436                      t_nrnb                      * gmx_restrict        nrnb)
 437 {
 438     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 439      * just 0 for non-waters.
 440      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 441      * jnr indices corresponding to data put in the four positions in the SIMD register.
 442      */
 443     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 444     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 445     int              jnrA,jnrB,jnrC,jnrD;
 446     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 447     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 448     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 449     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 450     real             rcutoff_scalar;
 451     real             *shiftvec,*fshift,*x,*f;
 452     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 453     real             scratch[4*DIM];
 454     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 455     real *           vdwioffsetptr0;
 456     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 457     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 458     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 459     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 460     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 461     real             *charge;
 462     __m128i          gbitab;
 463     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 464     __m256d          minushalf = _mm256_set1_pd(-0.5);
 465     real             *invsqrta,*dvda,*gbtab;
 466     __m128i          vfitab;
 467     __m128i          ifour       = _mm_set1_epi32(4);
 468     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 469     real             *vftab;
 470     __m256d          dummy_mask,cutoff_mask;
 471     __m128           tmpmask0,tmpmask1;
 472     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 473     __m256d          one     = _mm256_set1_pd(1.0);
 474     __m256d          two     = _mm256_set1_pd(2.0);
 475     x                = xx[0];
 476     f                = ff[0];
 477
 478     nri              = nlist->nri;
 479     iinr             = nlist->iinr;
 480     jindex           = nlist->jindex;
 481     jjnr             = nlist->jjnr;
 482     shiftidx         = nlist->shift;
 483     gid              = nlist->gid;
 484     shiftvec         = fr->shift_vec[0];
 485     fshift           = fr->fshift[0];
 486     facel            = _mm256_set1_pd(fr->ic->epsfac);
 487     charge           = mdatoms->chargeA;
 488
 489     invsqrta         = fr->invsqrta;
 490     dvda             = fr->dvda;
 491     gbtabscale       = _mm256_set1_pd(fr->gbtab->scale);
 492     gbtab            = fr->gbtab->data;
 493     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 494
 495     /* Avoid stupid compiler warnings */
 496     jnrA = jnrB = jnrC = jnrD = 0;
 497     j_coord_offsetA = 0;
 498     j_coord_offsetB = 0;
 499     j_coord_offsetC = 0;
 500     j_coord_offsetD = 0;
 501
 502     outeriter        = 0;
 503     inneriter        = 0;
 504
 505     for(iidx=0;iidx<4*DIM;iidx++)
 506     {
 507         scratch[iidx] = 0.0;
 508     }
 509
 510     /* Start outer loop over neighborlists */
 511     for(iidx=0; iidx<nri; iidx++)
 512     {
 513         /* Load shift vector for this list */
 514         i_shift_offset   = DIM*shiftidx[iidx];
 515
 516         /* Load limits for loop over neighbors */
 517         j_index_start    = jindex[iidx];
 518         j_index_end      = jindex[iidx+1];
 519
 520         /* Get outer coordinate index */
 521         inr              = iinr[iidx];
 522         i_coord_offset   = DIM*inr;
 523
 524         /* Load i particle coords and add shift vector */
 525         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 526
 527         fix0             = _mm256_setzero_pd();
 528         fiy0             = _mm256_setzero_pd();
 529         fiz0             = _mm256_setzero_pd();
 530
 531         /* Load parameters for i particles */
 532         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 533         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 534
 535         dvdasum          = _mm256_setzero_pd();
 536
 537         /* Start inner kernel loop */
 538         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 539         {
 540
 541             /* Get j neighbor index, and coordinate index */
 542             jnrA             = jjnr[jidx];
 543             jnrB             = jjnr[jidx+1];
 544             jnrC             = jjnr[jidx+2];
 545             jnrD             = jjnr[jidx+3];
 546             j_coord_offsetA  = DIM*jnrA;
 547             j_coord_offsetB  = DIM*jnrB;
 548             j_coord_offsetC  = DIM*jnrC;
 549             j_coord_offsetD  = DIM*jnrD;
 550
 551             /* load j atom coordinates */
 552             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 553                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 554                                                  &jx0,&jy0,&jz0);
 555
 556             /* Calculate displacement vector */
 557             dx00             = _mm256_sub_pd(ix0,jx0);
 558             dy00             = _mm256_sub_pd(iy0,jy0);
 559             dz00             = _mm256_sub_pd(iz0,jz0);
 560
 561             /* Calculate squared distance and things based on it */
 562             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 563
 564             rinv00           = avx256_invsqrt_d(rsq00);
 565
 566             /* Load parameters for j particles */
 567             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 568                                                                  charge+jnrC+0,charge+jnrD+0);
 569             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 570                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 571
 572             /**************************
 573              * CALCULATE INTERACTIONS *
 574              **************************/
 575
 576             r00              = _mm256_mul_pd(rsq00,rinv00);
 577
 578             /* Compute parameters for interactions between i and j atoms */
 579             qq00             = _mm256_mul_pd(iq0,jq0);
 580
 581             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 582             isaprod          = _mm256_mul_pd(isai0,isaj0);
 583             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 584             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 585
 586             /* Calculate generalized born table index - this is a separate table from the normal one,
 587              * but we use the same procedure by multiplying r with scale and truncating to integer.
 588              */
 589             rt               = _mm256_mul_pd(r00,gbscale);
 590             gbitab           = _mm256_cvttpd_epi32(rt);
 591             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 592             gbitab           = _mm_slli_epi32(gbitab,2);
 593             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 594             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 595             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 596             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 597             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 598             Heps             = _mm256_mul_pd(gbeps,H);
 599             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 600             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 601             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 602
 603             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 604             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 605             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 606             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 607             fjptrA           = dvda+jnrA;
 608             fjptrB           = dvda+jnrB;
 609             fjptrC           = dvda+jnrC;
 610             fjptrD           = dvda+jnrD;
 611             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 612                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 613             velec            = _mm256_mul_pd(qq00,rinv00);
 614             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 615
 616             fscal            = felec;
 617
 618             /* Calculate temporary vectorial force */
 619             tx               = _mm256_mul_pd(fscal,dx00);
 620             ty               = _mm256_mul_pd(fscal,dy00);
 621             tz               = _mm256_mul_pd(fscal,dz00);
 622
 623             /* Update vectorial force */
 624             fix0             = _mm256_add_pd(fix0,tx);
 625             fiy0             = _mm256_add_pd(fiy0,ty);
 626             fiz0             = _mm256_add_pd(fiz0,tz);
 627
 628             fjptrA             = f+j_coord_offsetA;
 629             fjptrB             = f+j_coord_offsetB;
 630             fjptrC             = f+j_coord_offsetC;
 631             fjptrD             = f+j_coord_offsetD;
 632             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 633
 634             /* Inner loop uses 55 flops */
 635         }
 636
 637         if(jidx<j_index_end)
 638         {
 639
 640             /* Get j neighbor index, and coordinate index */
 641             jnrlistA         = jjnr[jidx];
 642             jnrlistB         = jjnr[jidx+1];
 643             jnrlistC         = jjnr[jidx+2];
 644             jnrlistD         = jjnr[jidx+3];
 645             /* Sign of each element will be negative for non-real atoms.
 646              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 647              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 648              */
 649             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 650
 651             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 652             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 653             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 654
 655             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 656             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 657             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 658             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 659             j_coord_offsetA  = DIM*jnrA;
 660             j_coord_offsetB  = DIM*jnrB;
 661             j_coord_offsetC  = DIM*jnrC;
 662             j_coord_offsetD  = DIM*jnrD;
 663
 664             /* load j atom coordinates */
 665             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 666                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 667                                                  &jx0,&jy0,&jz0);
 668
 669             /* Calculate displacement vector */
 670             dx00             = _mm256_sub_pd(ix0,jx0);
 671             dy00             = _mm256_sub_pd(iy0,jy0);
 672             dz00             = _mm256_sub_pd(iz0,jz0);
 673
 674             /* Calculate squared distance and things based on it */
 675             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 676
 677             rinv00           = avx256_invsqrt_d(rsq00);
 678
 679             /* Load parameters for j particles */
 680             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 681                                                                  charge+jnrC+0,charge+jnrD+0);
 682             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 683                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 684
 685             /**************************
 686              * CALCULATE INTERACTIONS *
 687              **************************/
 688
 689             r00              = _mm256_mul_pd(rsq00,rinv00);
 690             r00              = _mm256_andnot_pd(dummy_mask,r00);
 691
 692             /* Compute parameters for interactions between i and j atoms */
 693             qq00             = _mm256_mul_pd(iq0,jq0);
 694
 695             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 696             isaprod          = _mm256_mul_pd(isai0,isaj0);
 697             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 698             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 699
 700             /* Calculate generalized born table index - this is a separate table from the normal one,
 701              * but we use the same procedure by multiplying r with scale and truncating to integer.
 702              */
 703             rt               = _mm256_mul_pd(r00,gbscale);
 704             gbitab           = _mm256_cvttpd_epi32(rt);
 705             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 706             gbitab           = _mm_slli_epi32(gbitab,2);
 707             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 708             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 709             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 710             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 711             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 712             Heps             = _mm256_mul_pd(gbeps,H);
 713             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 714             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 715             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 716
 717             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 718             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 719             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 720             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 721             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 722             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 723             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 724             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 725             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 726             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 727             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 728                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 729             velec            = _mm256_mul_pd(qq00,rinv00);
 730             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 731
 732             fscal            = felec;
 733
 734             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 735
 736             /* Calculate temporary vectorial force */
 737             tx               = _mm256_mul_pd(fscal,dx00);
 738             ty               = _mm256_mul_pd(fscal,dy00);
 739             tz               = _mm256_mul_pd(fscal,dz00);
 740
 741             /* Update vectorial force */
 742             fix0             = _mm256_add_pd(fix0,tx);
 743             fiy0             = _mm256_add_pd(fiy0,ty);
 744             fiz0             = _mm256_add_pd(fiz0,tz);
 745
 746             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 747             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 748             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 749             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 750             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 751
 752             /* Inner loop uses 56 flops */
 753         }
 754
 755         /* End of innermost loop */
 756
 757         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 758                                                  f+i_coord_offset,fshift+i_shift_offset);
 759
 760         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 761         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 762
 763         /* Increment number of inner iterations */
 764         inneriter                  += j_index_end - j_index_start;
 765
 766         /* Outer loop uses 7 flops */
 767     }
 768
 769     /* Increment number of outer iterations */
 770     outeriter        += nri;
 771
 772     /* Update outer/inner flops */
 773
 774     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);
 775 }