src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: GeneralizedBorn
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  90     real             *charge;
  91     __m128i          gbitab;
  92     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  93     __m256d          minushalf = _mm256_set1_pd(-0.5);
  94     real             *invsqrta,*dvda,*gbtab;
  95     int              nvdwtype;
  96     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 100     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 101     __m128i          vfitab;
 102     __m128i          ifour       = _mm_set1_epi32(4);
 103     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 104     real             *vftab;
 105     __m256d          dummy_mask,cutoff_mask;
 106     __m128           tmpmask0,tmpmask1;
 107     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 108     __m256d          one     = _mm256_set1_pd(1.0);
 109     __m256d          two     = _mm256_set1_pd(2.0);
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = _mm256_set1_pd(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     vftab            = kernel_data->table_vdw->data;
 128     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 129
 130     invsqrta         = fr->invsqrta;
 131     dvda             = fr->dvda;
 132     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 133     gbtab            = fr->gbtab.data;
 134     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 135
 136     /* Avoid stupid compiler warnings */
 137     jnrA = jnrB = jnrC = jnrD = 0;
 138     j_coord_offsetA = 0;
 139     j_coord_offsetB = 0;
 140     j_coord_offsetC = 0;
 141     j_coord_offsetD = 0;
 142
 143     outeriter        = 0;
 144     inneriter        = 0;
 145
 146     for(iidx=0;iidx<4*DIM;iidx++)
 147     {
 148         scratch[iidx] = 0.0;
 149     }
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 167
 168         fix0             = _mm256_setzero_pd();
 169         fiy0             = _mm256_setzero_pd();
 170         fiz0             = _mm256_setzero_pd();
 171
 172         /* Load parameters for i particles */
 173         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 174         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 175         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 176
 177         /* Reset potential sums */
 178         velecsum         = _mm256_setzero_pd();
 179         vgbsum           = _mm256_setzero_pd();
 180         vvdwsum          = _mm256_setzero_pd();
 181         dvdasum          = _mm256_setzero_pd();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             jnrC             = jjnr[jidx+2];
 191             jnrD             = jjnr[jidx+3];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194             j_coord_offsetC  = DIM*jnrC;
 195             j_coord_offsetD  = DIM*jnrD;
 196
 197             /* load j atom coordinates */
 198             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 200                                                  &jx0,&jy0,&jz0);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm256_sub_pd(ix0,jx0);
 204             dy00             = _mm256_sub_pd(iy0,jy0);
 205             dz00             = _mm256_sub_pd(iz0,jz0);
 206
 207             /* Calculate squared distance and things based on it */
 208             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 209
 210             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 211
 212             /* Load parameters for j particles */
 213             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 214                                                                  charge+jnrC+0,charge+jnrD+0);
 215             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 216                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 217             vdwjidx0A        = 2*vdwtype[jnrA+0];
 218             vdwjidx0B        = 2*vdwtype[jnrB+0];
 219             vdwjidx0C        = 2*vdwtype[jnrC+0];
 220             vdwjidx0D        = 2*vdwtype[jnrD+0];
 221
 222             /**************************
 223              * CALCULATE INTERACTIONS *
 224              **************************/
 225
 226             r00              = _mm256_mul_pd(rsq00,rinv00);
 227
 228             /* Compute parameters for interactions between i and j atoms */
 229             qq00             = _mm256_mul_pd(iq0,jq0);
 230             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 231                                             vdwioffsetptr0+vdwjidx0B,
 232                                             vdwioffsetptr0+vdwjidx0C,
 233                                             vdwioffsetptr0+vdwjidx0D,
 234                                             &c6_00,&c12_00);
 235
 236             /* Calculate table index by multiplying r with table scale and truncate to integer */
 237             rt               = _mm256_mul_pd(r00,vftabscale);
 238             vfitab           = _mm256_cvttpd_epi32(rt);
 239             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 240             vfitab           = _mm_slli_epi32(vfitab,3);
 241
 242             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 243             isaprod          = _mm256_mul_pd(isai0,isaj0);
 244             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 245             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 246
 247             /* Calculate generalized born table index - this is a separate table from the normal one,
 248              * but we use the same procedure by multiplying r with scale and truncating to integer.
 249              */
 250             rt               = _mm256_mul_pd(r00,gbscale);
 251             gbitab           = _mm256_cvttpd_epi32(rt);
 252             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 253             gbitab           = _mm_slli_epi32(gbitab,2);
 254             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 255             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 256             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 257             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 258             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 259             Heps             = _mm256_mul_pd(gbeps,H);
 260             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 261             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 262             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 263
 264             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 265             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 266             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 267             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 268             fjptrA           = dvda+jnrA;
 269             fjptrB           = dvda+jnrB;
 270             fjptrC           = dvda+jnrC;
 271             fjptrD           = dvda+jnrD;
 272             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 273                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 274             velec            = _mm256_mul_pd(qq00,rinv00);
 275             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 276
 277             /* CUBIC SPLINE TABLE DISPERSION */
 278             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 279             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 280             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 281             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 282             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 283             Heps             = _mm256_mul_pd(vfeps,H);
 284             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 285             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 286             vvdw6            = _mm256_mul_pd(c6_00,VV);
 287             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 288             fvdw6            = _mm256_mul_pd(c6_00,FF);
 289
 290             /* CUBIC SPLINE TABLE REPULSION */
 291             vfitab           = _mm_add_epi32(vfitab,ifour);
 292             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 293             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 294             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 295             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 296             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 297             Heps             = _mm256_mul_pd(vfeps,H);
 298             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 299             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 300             vvdw12           = _mm256_mul_pd(c12_00,VV);
 301             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 302             fvdw12           = _mm256_mul_pd(c12_00,FF);
 303             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 304             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 305
 306             /* Update potential sum for this i atom from the interaction with this j atom. */
 307             velecsum         = _mm256_add_pd(velecsum,velec);
 308             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 309             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 310
 311             fscal            = _mm256_add_pd(felec,fvdw);
 312
 313             /* Calculate temporary vectorial force */
 314             tx               = _mm256_mul_pd(fscal,dx00);
 315             ty               = _mm256_mul_pd(fscal,dy00);
 316             tz               = _mm256_mul_pd(fscal,dz00);
 317
 318             /* Update vectorial force */
 319             fix0             = _mm256_add_pd(fix0,tx);
 320             fiy0             = _mm256_add_pd(fiy0,ty);
 321             fiz0             = _mm256_add_pd(fiz0,tz);
 322
 323             fjptrA             = f+j_coord_offsetA;
 324             fjptrB             = f+j_coord_offsetB;
 325             fjptrC             = f+j_coord_offsetC;
 326             fjptrD             = f+j_coord_offsetD;
 327             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 328
 329             /* Inner loop uses 91 flops */
 330         }
 331
 332         if(jidx<j_index_end)
 333         {
 334
 335             /* Get j neighbor index, and coordinate index */
 336             jnrlistA         = jjnr[jidx];
 337             jnrlistB         = jjnr[jidx+1];
 338             jnrlistC         = jjnr[jidx+2];
 339             jnrlistD         = jjnr[jidx+3];
 340             /* Sign of each element will be negative for non-real atoms.
 341              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 342              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 343              */
 344             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 345
 346             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 347             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 348             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 349
 350             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 351             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 352             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 353             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 354             j_coord_offsetA  = DIM*jnrA;
 355             j_coord_offsetB  = DIM*jnrB;
 356             j_coord_offsetC  = DIM*jnrC;
 357             j_coord_offsetD  = DIM*jnrD;
 358
 359             /* load j atom coordinates */
 360             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 361                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 362                                                  &jx0,&jy0,&jz0);
 363
 364             /* Calculate displacement vector */
 365             dx00             = _mm256_sub_pd(ix0,jx0);
 366             dy00             = _mm256_sub_pd(iy0,jy0);
 367             dz00             = _mm256_sub_pd(iz0,jz0);
 368
 369             /* Calculate squared distance and things based on it */
 370             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 371
 372             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 373
 374             /* Load parameters for j particles */
 375             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 376                                                                  charge+jnrC+0,charge+jnrD+0);
 377             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 378                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 379             vdwjidx0A        = 2*vdwtype[jnrA+0];
 380             vdwjidx0B        = 2*vdwtype[jnrB+0];
 381             vdwjidx0C        = 2*vdwtype[jnrC+0];
 382             vdwjidx0D        = 2*vdwtype[jnrD+0];
 383
 384             /**************************
 385              * CALCULATE INTERACTIONS *
 386              **************************/
 387
 388             r00              = _mm256_mul_pd(rsq00,rinv00);
 389             r00              = _mm256_andnot_pd(dummy_mask,r00);
 390
 391             /* Compute parameters for interactions between i and j atoms */
 392             qq00             = _mm256_mul_pd(iq0,jq0);
 393             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 394                                             vdwioffsetptr0+vdwjidx0B,
 395                                             vdwioffsetptr0+vdwjidx0C,
 396                                             vdwioffsetptr0+vdwjidx0D,
 397                                             &c6_00,&c12_00);
 398
 399             /* Calculate table index by multiplying r with table scale and truncate to integer */
 400             rt               = _mm256_mul_pd(r00,vftabscale);
 401             vfitab           = _mm256_cvttpd_epi32(rt);
 402             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 403             vfitab           = _mm_slli_epi32(vfitab,3);
 404
 405             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 406             isaprod          = _mm256_mul_pd(isai0,isaj0);
 407             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 408             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 409
 410             /* Calculate generalized born table index - this is a separate table from the normal one,
 411              * but we use the same procedure by multiplying r with scale and truncating to integer.
 412              */
 413             rt               = _mm256_mul_pd(r00,gbscale);
 414             gbitab           = _mm256_cvttpd_epi32(rt);
 415             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 416             gbitab           = _mm_slli_epi32(gbitab,2);
 417             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 418             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 419             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 420             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 421             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 422             Heps             = _mm256_mul_pd(gbeps,H);
 423             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 424             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 425             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 426
 427             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 428             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 429             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 430             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 431             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 432             /* 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. */
 433             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 434             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 435             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 436             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 437             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 438                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 439             velec            = _mm256_mul_pd(qq00,rinv00);
 440             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 441
 442             /* CUBIC SPLINE TABLE DISPERSION */
 443             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 444             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 445             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 446             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 447             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 448             Heps             = _mm256_mul_pd(vfeps,H);
 449             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 450             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 451             vvdw6            = _mm256_mul_pd(c6_00,VV);
 452             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 453             fvdw6            = _mm256_mul_pd(c6_00,FF);
 454
 455             /* CUBIC SPLINE TABLE REPULSION */
 456             vfitab           = _mm_add_epi32(vfitab,ifour);
 457             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 458             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 459             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 460             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 461             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 462             Heps             = _mm256_mul_pd(vfeps,H);
 463             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 464             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 465             vvdw12           = _mm256_mul_pd(c12_00,VV);
 466             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 467             fvdw12           = _mm256_mul_pd(c12_00,FF);
 468             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 469             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 470
 471             /* Update potential sum for this i atom from the interaction with this j atom. */
 472             velec            = _mm256_andnot_pd(dummy_mask,velec);
 473             velecsum         = _mm256_add_pd(velecsum,velec);
 474             vgb              = _mm256_andnot_pd(dummy_mask,vgb);
 475             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 476             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 477             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 478
 479             fscal            = _mm256_add_pd(felec,fvdw);
 480
 481             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 482
 483             /* Calculate temporary vectorial force */
 484             tx               = _mm256_mul_pd(fscal,dx00);
 485             ty               = _mm256_mul_pd(fscal,dy00);
 486             tz               = _mm256_mul_pd(fscal,dz00);
 487
 488             /* Update vectorial force */
 489             fix0             = _mm256_add_pd(fix0,tx);
 490             fiy0             = _mm256_add_pd(fiy0,ty);
 491             fiz0             = _mm256_add_pd(fiz0,tz);
 492
 493             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 494             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 495             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 496             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 497             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 498
 499             /* Inner loop uses 92 flops */
 500         }
 501
 502         /* End of innermost loop */
 503
 504         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 505                                                  f+i_coord_offset,fshift+i_shift_offset);
 506
 507         ggid                        = gid[iidx];
 508         /* Update potential energies */
 509         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 510         gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 511         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 512         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 513         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 514
 515         /* Increment number of inner iterations */
 516         inneriter                  += j_index_end - j_index_start;
 517
 518         /* Outer loop uses 10 flops */
 519     }
 520
 521     /* Increment number of outer iterations */
 522     outeriter        += nri;
 523
 524     /* Update outer/inner flops */
 525
 526     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
 527 }
 528 /*
 529  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 530  * Electrostatics interaction: GeneralizedBorn
 531  * VdW interaction:            CubicSplineTable
 532  * Geometry:                   Particle-Particle
 533  * Calculate force/pot:        Force
 534  */
 535 void
 536 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 537                     (t_nblist                    * gmx_restrict       nlist,
 538                      rvec                        * gmx_restrict          xx,
 539                      rvec                        * gmx_restrict          ff,
 540                      t_forcerec                  * gmx_restrict          fr,
 541                      t_mdatoms                   * gmx_restrict     mdatoms,
 542                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 543                      t_nrnb                      * gmx_restrict        nrnb)
 544 {
 545     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 546      * just 0 for non-waters.
 547      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 548      * jnr indices corresponding to data put in the four positions in the SIMD register.
 549      */
 550     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 551     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 552     int              jnrA,jnrB,jnrC,jnrD;
 553     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 554     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 555     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 556     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 557     real             rcutoff_scalar;
 558     real             *shiftvec,*fshift,*x,*f;
 559     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 560     real             scratch[4*DIM];
 561     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 562     real *           vdwioffsetptr0;
 563     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 564     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 565     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 566     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 567     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 568     real             *charge;
 569     __m128i          gbitab;
 570     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 571     __m256d          minushalf = _mm256_set1_pd(-0.5);
 572     real             *invsqrta,*dvda,*gbtab;
 573     int              nvdwtype;
 574     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 575     int              *vdwtype;
 576     real             *vdwparam;
 577     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 578     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 579     __m128i          vfitab;
 580     __m128i          ifour       = _mm_set1_epi32(4);
 581     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 582     real             *vftab;
 583     __m256d          dummy_mask,cutoff_mask;
 584     __m128           tmpmask0,tmpmask1;
 585     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 586     __m256d          one     = _mm256_set1_pd(1.0);
 587     __m256d          two     = _mm256_set1_pd(2.0);
 588     x                = xx[0];
 589     f                = ff[0];
 590
 591     nri              = nlist->nri;
 592     iinr             = nlist->iinr;
 593     jindex           = nlist->jindex;
 594     jjnr             = nlist->jjnr;
 595     shiftidx         = nlist->shift;
 596     gid              = nlist->gid;
 597     shiftvec         = fr->shift_vec[0];
 598     fshift           = fr->fshift[0];
 599     facel            = _mm256_set1_pd(fr->epsfac);
 600     charge           = mdatoms->chargeA;
 601     nvdwtype         = fr->ntype;
 602     vdwparam         = fr->nbfp;
 603     vdwtype          = mdatoms->typeA;
 604
 605     vftab            = kernel_data->table_vdw->data;
 606     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 607
 608     invsqrta         = fr->invsqrta;
 609     dvda             = fr->dvda;
 610     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 611     gbtab            = fr->gbtab.data;
 612     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 613
 614     /* Avoid stupid compiler warnings */
 615     jnrA = jnrB = jnrC = jnrD = 0;
 616     j_coord_offsetA = 0;
 617     j_coord_offsetB = 0;
 618     j_coord_offsetC = 0;
 619     j_coord_offsetD = 0;
 620
 621     outeriter        = 0;
 622     inneriter        = 0;
 623
 624     for(iidx=0;iidx<4*DIM;iidx++)
 625     {
 626         scratch[iidx] = 0.0;
 627     }
 628
 629     /* Start outer loop over neighborlists */
 630     for(iidx=0; iidx<nri; iidx++)
 631     {
 632         /* Load shift vector for this list */
 633         i_shift_offset   = DIM*shiftidx[iidx];
 634
 635         /* Load limits for loop over neighbors */
 636         j_index_start    = jindex[iidx];
 637         j_index_end      = jindex[iidx+1];
 638
 639         /* Get outer coordinate index */
 640         inr              = iinr[iidx];
 641         i_coord_offset   = DIM*inr;
 642
 643         /* Load i particle coords and add shift vector */
 644         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 645
 646         fix0             = _mm256_setzero_pd();
 647         fiy0             = _mm256_setzero_pd();
 648         fiz0             = _mm256_setzero_pd();
 649
 650         /* Load parameters for i particles */
 651         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 652         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 653         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 654
 655         dvdasum          = _mm256_setzero_pd();
 656
 657         /* Start inner kernel loop */
 658         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 659         {
 660
 661             /* Get j neighbor index, and coordinate index */
 662             jnrA             = jjnr[jidx];
 663             jnrB             = jjnr[jidx+1];
 664             jnrC             = jjnr[jidx+2];
 665             jnrD             = jjnr[jidx+3];
 666             j_coord_offsetA  = DIM*jnrA;
 667             j_coord_offsetB  = DIM*jnrB;
 668             j_coord_offsetC  = DIM*jnrC;
 669             j_coord_offsetD  = DIM*jnrD;
 670
 671             /* load j atom coordinates */
 672             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 673                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 674                                                  &jx0,&jy0,&jz0);
 675
 676             /* Calculate displacement vector */
 677             dx00             = _mm256_sub_pd(ix0,jx0);
 678             dy00             = _mm256_sub_pd(iy0,jy0);
 679             dz00             = _mm256_sub_pd(iz0,jz0);
 680
 681             /* Calculate squared distance and things based on it */
 682             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 683
 684             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 685
 686             /* Load parameters for j particles */
 687             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 688                                                                  charge+jnrC+0,charge+jnrD+0);
 689             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 690                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 691             vdwjidx0A        = 2*vdwtype[jnrA+0];
 692             vdwjidx0B        = 2*vdwtype[jnrB+0];
 693             vdwjidx0C        = 2*vdwtype[jnrC+0];
 694             vdwjidx0D        = 2*vdwtype[jnrD+0];
 695
 696             /**************************
 697              * CALCULATE INTERACTIONS *
 698              **************************/
 699
 700             r00              = _mm256_mul_pd(rsq00,rinv00);
 701
 702             /* Compute parameters for interactions between i and j atoms */
 703             qq00             = _mm256_mul_pd(iq0,jq0);
 704             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 705                                             vdwioffsetptr0+vdwjidx0B,
 706                                             vdwioffsetptr0+vdwjidx0C,
 707                                             vdwioffsetptr0+vdwjidx0D,
 708                                             &c6_00,&c12_00);
 709
 710             /* Calculate table index by multiplying r with table scale and truncate to integer */
 711             rt               = _mm256_mul_pd(r00,vftabscale);
 712             vfitab           = _mm256_cvttpd_epi32(rt);
 713             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 714             vfitab           = _mm_slli_epi32(vfitab,3);
 715
 716             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 717             isaprod          = _mm256_mul_pd(isai0,isaj0);
 718             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 719             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 720
 721             /* Calculate generalized born table index - this is a separate table from the normal one,
 722              * but we use the same procedure by multiplying r with scale and truncating to integer.
 723              */
 724             rt               = _mm256_mul_pd(r00,gbscale);
 725             gbitab           = _mm256_cvttpd_epi32(rt);
 726             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 727             gbitab           = _mm_slli_epi32(gbitab,2);
 728             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 729             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 730             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 731             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 732             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 733             Heps             = _mm256_mul_pd(gbeps,H);
 734             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 735             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 736             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 737
 738             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 739             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 740             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 741             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 742             fjptrA           = dvda+jnrA;
 743             fjptrB           = dvda+jnrB;
 744             fjptrC           = dvda+jnrC;
 745             fjptrD           = dvda+jnrD;
 746             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 747                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 748             velec            = _mm256_mul_pd(qq00,rinv00);
 749             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 750
 751             /* CUBIC SPLINE TABLE DISPERSION */
 752             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 753             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 754             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 755             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 756             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 757             Heps             = _mm256_mul_pd(vfeps,H);
 758             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 759             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 760             fvdw6            = _mm256_mul_pd(c6_00,FF);
 761
 762             /* CUBIC SPLINE TABLE REPULSION */
 763             vfitab           = _mm_add_epi32(vfitab,ifour);
 764             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 765             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 766             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 767             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 768             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 769             Heps             = _mm256_mul_pd(vfeps,H);
 770             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 771             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 772             fvdw12           = _mm256_mul_pd(c12_00,FF);
 773             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 774
 775             fscal            = _mm256_add_pd(felec,fvdw);
 776
 777             /* Calculate temporary vectorial force */
 778             tx               = _mm256_mul_pd(fscal,dx00);
 779             ty               = _mm256_mul_pd(fscal,dy00);
 780             tz               = _mm256_mul_pd(fscal,dz00);
 781
 782             /* Update vectorial force */
 783             fix0             = _mm256_add_pd(fix0,tx);
 784             fiy0             = _mm256_add_pd(fiy0,ty);
 785             fiz0             = _mm256_add_pd(fiz0,tz);
 786
 787             fjptrA             = f+j_coord_offsetA;
 788             fjptrB             = f+j_coord_offsetB;
 789             fjptrC             = f+j_coord_offsetC;
 790             fjptrD             = f+j_coord_offsetD;
 791             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 792
 793             /* Inner loop uses 81 flops */
 794         }
 795
 796         if(jidx<j_index_end)
 797         {
 798
 799             /* Get j neighbor index, and coordinate index */
 800             jnrlistA         = jjnr[jidx];
 801             jnrlistB         = jjnr[jidx+1];
 802             jnrlistC         = jjnr[jidx+2];
 803             jnrlistD         = jjnr[jidx+3];
 804             /* Sign of each element will be negative for non-real atoms.
 805              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 806              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 807              */
 808             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 809
 810             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 811             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 812             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 813
 814             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 815             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 816             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 817             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 818             j_coord_offsetA  = DIM*jnrA;
 819             j_coord_offsetB  = DIM*jnrB;
 820             j_coord_offsetC  = DIM*jnrC;
 821             j_coord_offsetD  = DIM*jnrD;
 822
 823             /* load j atom coordinates */
 824             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 825                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 826                                                  &jx0,&jy0,&jz0);
 827
 828             /* Calculate displacement vector */
 829             dx00             = _mm256_sub_pd(ix0,jx0);
 830             dy00             = _mm256_sub_pd(iy0,jy0);
 831             dz00             = _mm256_sub_pd(iz0,jz0);
 832
 833             /* Calculate squared distance and things based on it */
 834             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 835
 836             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 837
 838             /* Load parameters for j particles */
 839             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 840                                                                  charge+jnrC+0,charge+jnrD+0);
 841             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 842                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 843             vdwjidx0A        = 2*vdwtype[jnrA+0];
 844             vdwjidx0B        = 2*vdwtype[jnrB+0];
 845             vdwjidx0C        = 2*vdwtype[jnrC+0];
 846             vdwjidx0D        = 2*vdwtype[jnrD+0];
 847
 848             /**************************
 849              * CALCULATE INTERACTIONS *
 850              **************************/
 851
 852             r00              = _mm256_mul_pd(rsq00,rinv00);
 853             r00              = _mm256_andnot_pd(dummy_mask,r00);
 854
 855             /* Compute parameters for interactions between i and j atoms */
 856             qq00             = _mm256_mul_pd(iq0,jq0);
 857             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 858                                             vdwioffsetptr0+vdwjidx0B,
 859                                             vdwioffsetptr0+vdwjidx0C,
 860                                             vdwioffsetptr0+vdwjidx0D,
 861                                             &c6_00,&c12_00);
 862
 863             /* Calculate table index by multiplying r with table scale and truncate to integer */
 864             rt               = _mm256_mul_pd(r00,vftabscale);
 865             vfitab           = _mm256_cvttpd_epi32(rt);
 866             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 867             vfitab           = _mm_slli_epi32(vfitab,3);
 868
 869             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 870             isaprod          = _mm256_mul_pd(isai0,isaj0);
 871             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 872             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 873
 874             /* Calculate generalized born table index - this is a separate table from the normal one,
 875              * but we use the same procedure by multiplying r with scale and truncating to integer.
 876              */
 877             rt               = _mm256_mul_pd(r00,gbscale);
 878             gbitab           = _mm256_cvttpd_epi32(rt);
 879             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 880             gbitab           = _mm_slli_epi32(gbitab,2);
 881             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 882             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 883             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 884             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 885             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 886             Heps             = _mm256_mul_pd(gbeps,H);
 887             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 888             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 889             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 890
 891             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 892             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 893             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 894             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 895             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 896             /* 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. */
 897             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 898             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 899             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 900             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 901             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 902                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 903             velec            = _mm256_mul_pd(qq00,rinv00);
 904             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 905
 906             /* CUBIC SPLINE TABLE DISPERSION */
 907             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 908             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 909             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 910             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 911             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 912             Heps             = _mm256_mul_pd(vfeps,H);
 913             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 914             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 915             fvdw6            = _mm256_mul_pd(c6_00,FF);
 916
 917             /* CUBIC SPLINE TABLE REPULSION */
 918             vfitab           = _mm_add_epi32(vfitab,ifour);
 919             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 920             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 921             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 922             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 923             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 924             Heps             = _mm256_mul_pd(vfeps,H);
 925             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 926             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 927             fvdw12           = _mm256_mul_pd(c12_00,FF);
 928             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 929
 930             fscal            = _mm256_add_pd(felec,fvdw);
 931
 932             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 933
 934             /* Calculate temporary vectorial force */
 935             tx               = _mm256_mul_pd(fscal,dx00);
 936             ty               = _mm256_mul_pd(fscal,dy00);
 937             tz               = _mm256_mul_pd(fscal,dz00);
 938
 939             /* Update vectorial force */
 940             fix0             = _mm256_add_pd(fix0,tx);
 941             fiy0             = _mm256_add_pd(fiy0,ty);
 942             fiz0             = _mm256_add_pd(fiz0,tz);
 943
 944             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 945             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 946             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 947             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 948             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 949
 950             /* Inner loop uses 82 flops */
 951         }
 952
 953         /* End of innermost loop */
 954
 955         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 956                                                  f+i_coord_offset,fshift+i_shift_offset);
 957
 958         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 959         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 960
 961         /* Increment number of inner iterations */
 962         inneriter                  += j_index_end - j_index_start;
 963
 964         /* Outer loop uses 7 flops */
 965     }
 966
 967     /* Increment number of outer iterations */
 968     outeriter        += nri;
 969
 970     /* Update outer/inner flops */
 971
 972     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);
 973 }