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) 2012,2013,2014,2015,2017, 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_VdwCSTab_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: GeneralizedBorn
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecGB_VdwCSTab_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     int              nvdwtype;
  95     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  99     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 100     __m128i          vfitab;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 103     real             *vftab;
 104     __m256d          dummy_mask,cutoff_mask;
 105     __m128           tmpmask0,tmpmask1;
 106     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 107     __m256d          one     = _mm256_set1_pd(1.0);
 108     __m256d          two     = _mm256_set1_pd(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm256_set1_pd(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     vftab            = kernel_data->table_vdw->data;
 127     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 128
 129     invsqrta         = fr->invsqrta;
 130     dvda             = fr->dvda;
 131     gbtabscale       = _mm256_set1_pd(fr->gbtab->scale);
 132     gbtab            = fr->gbtab->data;
 133     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 134
 135     /* Avoid stupid compiler warnings */
 136     jnrA = jnrB = jnrC = jnrD = 0;
 137     j_coord_offsetA = 0;
 138     j_coord_offsetB = 0;
 139     j_coord_offsetC = 0;
 140     j_coord_offsetD = 0;
 141
 142     outeriter        = 0;
 143     inneriter        = 0;
 144
 145     for(iidx=0;iidx<4*DIM;iidx++)
 146     {
 147         scratch[iidx] = 0.0;
 148     }
 149
 150     /* Start outer loop over neighborlists */
 151     for(iidx=0; iidx<nri; iidx++)
 152     {
 153         /* Load shift vector for this list */
 154         i_shift_offset   = DIM*shiftidx[iidx];
 155
 156         /* Load limits for loop over neighbors */
 157         j_index_start    = jindex[iidx];
 158         j_index_end      = jindex[iidx+1];
 159
 160         /* Get outer coordinate index */
 161         inr              = iinr[iidx];
 162         i_coord_offset   = DIM*inr;
 163
 164         /* Load i particle coords and add shift vector */
 165         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 166
 167         fix0             = _mm256_setzero_pd();
 168         fiy0             = _mm256_setzero_pd();
 169         fiz0             = _mm256_setzero_pd();
 170
 171         /* Load parameters for i particles */
 172         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 173         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 174         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 175
 176         /* Reset potential sums */
 177         velecsum         = _mm256_setzero_pd();
 178         vgbsum           = _mm256_setzero_pd();
 179         vvdwsum          = _mm256_setzero_pd();
 180         dvdasum          = _mm256_setzero_pd();
 181
 182         /* Start inner kernel loop */
 183         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 184         {
 185
 186             /* Get j neighbor index, and coordinate index */
 187             jnrA             = jjnr[jidx];
 188             jnrB             = jjnr[jidx+1];
 189             jnrC             = jjnr[jidx+2];
 190             jnrD             = jjnr[jidx+3];
 191             j_coord_offsetA  = DIM*jnrA;
 192             j_coord_offsetB  = DIM*jnrB;
 193             j_coord_offsetC  = DIM*jnrC;
 194             j_coord_offsetD  = DIM*jnrD;
 195
 196             /* load j atom coordinates */
 197             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 198                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 199                                                  &jx0,&jy0,&jz0);
 200
 201             /* Calculate displacement vector */
 202             dx00             = _mm256_sub_pd(ix0,jx0);
 203             dy00             = _mm256_sub_pd(iy0,jy0);
 204             dz00             = _mm256_sub_pd(iz0,jz0);
 205
 206             /* Calculate squared distance and things based on it */
 207             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 208
 209             rinv00           = avx256_invsqrt_d(rsq00);
 210
 211             /* Load parameters for j particles */
 212             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 213                                                                  charge+jnrC+0,charge+jnrD+0);
 214             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 215                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 216             vdwjidx0A        = 2*vdwtype[jnrA+0];
 217             vdwjidx0B        = 2*vdwtype[jnrB+0];
 218             vdwjidx0C        = 2*vdwtype[jnrC+0];
 219             vdwjidx0D        = 2*vdwtype[jnrD+0];
 220
 221             /**************************
 222              * CALCULATE INTERACTIONS *
 223              **************************/
 224
 225             r00              = _mm256_mul_pd(rsq00,rinv00);
 226
 227             /* Compute parameters for interactions between i and j atoms */
 228             qq00             = _mm256_mul_pd(iq0,jq0);
 229             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 230                                             vdwioffsetptr0+vdwjidx0B,
 231                                             vdwioffsetptr0+vdwjidx0C,
 232                                             vdwioffsetptr0+vdwjidx0D,
 233                                             &c6_00,&c12_00);
 234
 235             /* Calculate table index by multiplying r with table scale and truncate to integer */
 236             rt               = _mm256_mul_pd(r00,vftabscale);
 237             vfitab           = _mm256_cvttpd_epi32(rt);
 238             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 239             vfitab           = _mm_slli_epi32(vfitab,3);
 240
 241             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 242             isaprod          = _mm256_mul_pd(isai0,isaj0);
 243             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 244             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 245
 246             /* Calculate generalized born table index - this is a separate table from the normal one,
 247              * but we use the same procedure by multiplying r with scale and truncating to integer.
 248              */
 249             rt               = _mm256_mul_pd(r00,gbscale);
 250             gbitab           = _mm256_cvttpd_epi32(rt);
 251             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 252             gbitab           = _mm_slli_epi32(gbitab,2);
 253             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 254             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 255             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 256             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 257             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 258             Heps             = _mm256_mul_pd(gbeps,H);
 259             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 260             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 261             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 262
 263             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 264             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 265             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 266             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 267             fjptrA           = dvda+jnrA;
 268             fjptrB           = dvda+jnrB;
 269             fjptrC           = dvda+jnrC;
 270             fjptrD           = dvda+jnrD;
 271             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 272                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 273             velec            = _mm256_mul_pd(qq00,rinv00);
 274             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 275
 276             /* CUBIC SPLINE TABLE DISPERSION */
 277             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 278             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 279             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 280             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 281             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 282             Heps             = _mm256_mul_pd(vfeps,H);
 283             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 284             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 285             vvdw6            = _mm256_mul_pd(c6_00,VV);
 286             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 287             fvdw6            = _mm256_mul_pd(c6_00,FF);
 288
 289             /* CUBIC SPLINE TABLE REPULSION */
 290             vfitab           = _mm_add_epi32(vfitab,ifour);
 291             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 292             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 293             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 294             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 295             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 296             Heps             = _mm256_mul_pd(vfeps,H);
 297             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 298             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 299             vvdw12           = _mm256_mul_pd(c12_00,VV);
 300             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 301             fvdw12           = _mm256_mul_pd(c12_00,FF);
 302             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 303             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velecsum         = _mm256_add_pd(velecsum,velec);
 307             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 308             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 309
 310             fscal            = _mm256_add_pd(felec,fvdw);
 311
 312             /* Calculate temporary vectorial force */
 313             tx               = _mm256_mul_pd(fscal,dx00);
 314             ty               = _mm256_mul_pd(fscal,dy00);
 315             tz               = _mm256_mul_pd(fscal,dz00);
 316
 317             /* Update vectorial force */
 318             fix0             = _mm256_add_pd(fix0,tx);
 319             fiy0             = _mm256_add_pd(fiy0,ty);
 320             fiz0             = _mm256_add_pd(fiz0,tz);
 321
 322             fjptrA             = f+j_coord_offsetA;
 323             fjptrB             = f+j_coord_offsetB;
 324             fjptrC             = f+j_coord_offsetC;
 325             fjptrD             = f+j_coord_offsetD;
 326             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 327
 328             /* Inner loop uses 91 flops */
 329         }
 330
 331         if(jidx<j_index_end)
 332         {
 333
 334             /* Get j neighbor index, and coordinate index */
 335             jnrlistA         = jjnr[jidx];
 336             jnrlistB         = jjnr[jidx+1];
 337             jnrlistC         = jjnr[jidx+2];
 338             jnrlistD         = jjnr[jidx+3];
 339             /* Sign of each element will be negative for non-real atoms.
 340              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 341              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 342              */
 343             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 344
 345             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 346             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 347             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 348
 349             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 350             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 351             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 352             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 353             j_coord_offsetA  = DIM*jnrA;
 354             j_coord_offsetB  = DIM*jnrB;
 355             j_coord_offsetC  = DIM*jnrC;
 356             j_coord_offsetD  = DIM*jnrD;
 357
 358             /* load j atom coordinates */
 359             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 360                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 361                                                  &jx0,&jy0,&jz0);
 362
 363             /* Calculate displacement vector */
 364             dx00             = _mm256_sub_pd(ix0,jx0);
 365             dy00             = _mm256_sub_pd(iy0,jy0);
 366             dz00             = _mm256_sub_pd(iz0,jz0);
 367
 368             /* Calculate squared distance and things based on it */
 369             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 370
 371             rinv00           = avx256_invsqrt_d(rsq00);
 372
 373             /* Load parameters for j particles */
 374             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 375                                                                  charge+jnrC+0,charge+jnrD+0);
 376             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 377                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 378             vdwjidx0A        = 2*vdwtype[jnrA+0];
 379             vdwjidx0B        = 2*vdwtype[jnrB+0];
 380             vdwjidx0C        = 2*vdwtype[jnrC+0];
 381             vdwjidx0D        = 2*vdwtype[jnrD+0];
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             r00              = _mm256_mul_pd(rsq00,rinv00);
 388             r00              = _mm256_andnot_pd(dummy_mask,r00);
 389
 390             /* Compute parameters for interactions between i and j atoms */
 391             qq00             = _mm256_mul_pd(iq0,jq0);
 392             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 393                                             vdwioffsetptr0+vdwjidx0B,
 394                                             vdwioffsetptr0+vdwjidx0C,
 395                                             vdwioffsetptr0+vdwjidx0D,
 396                                             &c6_00,&c12_00);
 397
 398             /* Calculate table index by multiplying r with table scale and truncate to integer */
 399             rt               = _mm256_mul_pd(r00,vftabscale);
 400             vfitab           = _mm256_cvttpd_epi32(rt);
 401             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 402             vfitab           = _mm_slli_epi32(vfitab,3);
 403
 404             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 405             isaprod          = _mm256_mul_pd(isai0,isaj0);
 406             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 407             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 408
 409             /* Calculate generalized born table index - this is a separate table from the normal one,
 410              * but we use the same procedure by multiplying r with scale and truncating to integer.
 411              */
 412             rt               = _mm256_mul_pd(r00,gbscale);
 413             gbitab           = _mm256_cvttpd_epi32(rt);
 414             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 415             gbitab           = _mm_slli_epi32(gbitab,2);
 416             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 417             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 418             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 419             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 420             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 421             Heps             = _mm256_mul_pd(gbeps,H);
 422             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 423             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 424             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 425
 426             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 427             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 428             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 429             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 430             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 431             /* 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. */
 432             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 433             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 434             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 435             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 436             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 437                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 438             velec            = _mm256_mul_pd(qq00,rinv00);
 439             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 440
 441             /* CUBIC SPLINE TABLE DISPERSION */
 442             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 443             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 444             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 445             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 446             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 447             Heps             = _mm256_mul_pd(vfeps,H);
 448             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 449             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 450             vvdw6            = _mm256_mul_pd(c6_00,VV);
 451             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 452             fvdw6            = _mm256_mul_pd(c6_00,FF);
 453
 454             /* CUBIC SPLINE TABLE REPULSION */
 455             vfitab           = _mm_add_epi32(vfitab,ifour);
 456             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 457             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 458             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 459             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 460             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 461             Heps             = _mm256_mul_pd(vfeps,H);
 462             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 463             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 464             vvdw12           = _mm256_mul_pd(c12_00,VV);
 465             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 466             fvdw12           = _mm256_mul_pd(c12_00,FF);
 467             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 468             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velec            = _mm256_andnot_pd(dummy_mask,velec);
 472             velecsum         = _mm256_add_pd(velecsum,velec);
 473             vgb              = _mm256_andnot_pd(dummy_mask,vgb);
 474             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 475             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 476             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 477
 478             fscal            = _mm256_add_pd(felec,fvdw);
 479
 480             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 481
 482             /* Calculate temporary vectorial force */
 483             tx               = _mm256_mul_pd(fscal,dx00);
 484             ty               = _mm256_mul_pd(fscal,dy00);
 485             tz               = _mm256_mul_pd(fscal,dz00);
 486
 487             /* Update vectorial force */
 488             fix0             = _mm256_add_pd(fix0,tx);
 489             fiy0             = _mm256_add_pd(fiy0,ty);
 490             fiz0             = _mm256_add_pd(fiz0,tz);
 491
 492             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 493             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 494             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 495             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 496             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 497
 498             /* Inner loop uses 92 flops */
 499         }
 500
 501         /* End of innermost loop */
 502
 503         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 504                                                  f+i_coord_offset,fshift+i_shift_offset);
 505
 506         ggid                        = gid[iidx];
 507         /* Update potential energies */
 508         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 509         gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 510         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 511         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 512         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 513
 514         /* Increment number of inner iterations */
 515         inneriter                  += j_index_end - j_index_start;
 516
 517         /* Outer loop uses 10 flops */
 518     }
 519
 520     /* Increment number of outer iterations */
 521     outeriter        += nri;
 522
 523     /* Update outer/inner flops */
 524
 525     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
 526 }
 527 /*
 528  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 529  * Electrostatics interaction: GeneralizedBorn
 530  * VdW interaction:            CubicSplineTable
 531  * Geometry:                   Particle-Particle
 532  * Calculate force/pot:        Force
 533  */
 534 void
 535 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 536                     (t_nblist                    * gmx_restrict       nlist,
 537                      rvec                        * gmx_restrict          xx,
 538                      rvec                        * gmx_restrict          ff,
 539                      struct t_forcerec           * gmx_restrict          fr,
 540                      t_mdatoms                   * gmx_restrict     mdatoms,
 541                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 542                      t_nrnb                      * gmx_restrict        nrnb)
 543 {
 544     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 545      * just 0 for non-waters.
 546      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 547      * jnr indices corresponding to data put in the four positions in the SIMD register.
 548      */
 549     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 550     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 551     int              jnrA,jnrB,jnrC,jnrD;
 552     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 553     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 554     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 555     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 556     real             rcutoff_scalar;
 557     real             *shiftvec,*fshift,*x,*f;
 558     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 559     real             scratch[4*DIM];
 560     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 561     real *           vdwioffsetptr0;
 562     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 563     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 564     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 565     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 566     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 567     real             *charge;
 568     __m128i          gbitab;
 569     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 570     __m256d          minushalf = _mm256_set1_pd(-0.5);
 571     real             *invsqrta,*dvda,*gbtab;
 572     int              nvdwtype;
 573     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 574     int              *vdwtype;
 575     real             *vdwparam;
 576     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 577     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 578     __m128i          vfitab;
 579     __m128i          ifour       = _mm_set1_epi32(4);
 580     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 581     real             *vftab;
 582     __m256d          dummy_mask,cutoff_mask;
 583     __m128           tmpmask0,tmpmask1;
 584     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 585     __m256d          one     = _mm256_set1_pd(1.0);
 586     __m256d          two     = _mm256_set1_pd(2.0);
 587     x                = xx[0];
 588     f                = ff[0];
 589
 590     nri              = nlist->nri;
 591     iinr             = nlist->iinr;
 592     jindex           = nlist->jindex;
 593     jjnr             = nlist->jjnr;
 594     shiftidx         = nlist->shift;
 595     gid              = nlist->gid;
 596     shiftvec         = fr->shift_vec[0];
 597     fshift           = fr->fshift[0];
 598     facel            = _mm256_set1_pd(fr->ic->epsfac);
 599     charge           = mdatoms->chargeA;
 600     nvdwtype         = fr->ntype;
 601     vdwparam         = fr->nbfp;
 602     vdwtype          = mdatoms->typeA;
 603
 604     vftab            = kernel_data->table_vdw->data;
 605     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 606
 607     invsqrta         = fr->invsqrta;
 608     dvda             = fr->dvda;
 609     gbtabscale       = _mm256_set1_pd(fr->gbtab->scale);
 610     gbtab            = fr->gbtab->data;
 611     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 612
 613     /* Avoid stupid compiler warnings */
 614     jnrA = jnrB = jnrC = jnrD = 0;
 615     j_coord_offsetA = 0;
 616     j_coord_offsetB = 0;
 617     j_coord_offsetC = 0;
 618     j_coord_offsetD = 0;
 619
 620     outeriter        = 0;
 621     inneriter        = 0;
 622
 623     for(iidx=0;iidx<4*DIM;iidx++)
 624     {
 625         scratch[iidx] = 0.0;
 626     }
 627
 628     /* Start outer loop over neighborlists */
 629     for(iidx=0; iidx<nri; iidx++)
 630     {
 631         /* Load shift vector for this list */
 632         i_shift_offset   = DIM*shiftidx[iidx];
 633
 634         /* Load limits for loop over neighbors */
 635         j_index_start    = jindex[iidx];
 636         j_index_end      = jindex[iidx+1];
 637
 638         /* Get outer coordinate index */
 639         inr              = iinr[iidx];
 640         i_coord_offset   = DIM*inr;
 641
 642         /* Load i particle coords and add shift vector */
 643         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 644
 645         fix0             = _mm256_setzero_pd();
 646         fiy0             = _mm256_setzero_pd();
 647         fiz0             = _mm256_setzero_pd();
 648
 649         /* Load parameters for i particles */
 650         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 651         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 652         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 653
 654         dvdasum          = _mm256_setzero_pd();
 655
 656         /* Start inner kernel loop */
 657         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 658         {
 659
 660             /* Get j neighbor index, and coordinate index */
 661             jnrA             = jjnr[jidx];
 662             jnrB             = jjnr[jidx+1];
 663             jnrC             = jjnr[jidx+2];
 664             jnrD             = jjnr[jidx+3];
 665             j_coord_offsetA  = DIM*jnrA;
 666             j_coord_offsetB  = DIM*jnrB;
 667             j_coord_offsetC  = DIM*jnrC;
 668             j_coord_offsetD  = DIM*jnrD;
 669
 670             /* load j atom coordinates */
 671             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 672                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 673                                                  &jx0,&jy0,&jz0);
 674
 675             /* Calculate displacement vector */
 676             dx00             = _mm256_sub_pd(ix0,jx0);
 677             dy00             = _mm256_sub_pd(iy0,jy0);
 678             dz00             = _mm256_sub_pd(iz0,jz0);
 679
 680             /* Calculate squared distance and things based on it */
 681             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 682
 683             rinv00           = avx256_invsqrt_d(rsq00);
 684
 685             /* Load parameters for j particles */
 686             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 687                                                                  charge+jnrC+0,charge+jnrD+0);
 688             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 689                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 690             vdwjidx0A        = 2*vdwtype[jnrA+0];
 691             vdwjidx0B        = 2*vdwtype[jnrB+0];
 692             vdwjidx0C        = 2*vdwtype[jnrC+0];
 693             vdwjidx0D        = 2*vdwtype[jnrD+0];
 694
 695             /**************************
 696              * CALCULATE INTERACTIONS *
 697              **************************/
 698
 699             r00              = _mm256_mul_pd(rsq00,rinv00);
 700
 701             /* Compute parameters for interactions between i and j atoms */
 702             qq00             = _mm256_mul_pd(iq0,jq0);
 703             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 704                                             vdwioffsetptr0+vdwjidx0B,
 705                                             vdwioffsetptr0+vdwjidx0C,
 706                                             vdwioffsetptr0+vdwjidx0D,
 707                                             &c6_00,&c12_00);
 708
 709             /* Calculate table index by multiplying r with table scale and truncate to integer */
 710             rt               = _mm256_mul_pd(r00,vftabscale);
 711             vfitab           = _mm256_cvttpd_epi32(rt);
 712             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 713             vfitab           = _mm_slli_epi32(vfitab,3);
 714
 715             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 716             isaprod          = _mm256_mul_pd(isai0,isaj0);
 717             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 718             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 719
 720             /* Calculate generalized born table index - this is a separate table from the normal one,
 721              * but we use the same procedure by multiplying r with scale and truncating to integer.
 722              */
 723             rt               = _mm256_mul_pd(r00,gbscale);
 724             gbitab           = _mm256_cvttpd_epi32(rt);
 725             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 726             gbitab           = _mm_slli_epi32(gbitab,2);
 727             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 728             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 729             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 730             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 731             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 732             Heps             = _mm256_mul_pd(gbeps,H);
 733             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 734             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 735             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 736
 737             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 738             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 739             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 740             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 741             fjptrA           = dvda+jnrA;
 742             fjptrB           = dvda+jnrB;
 743             fjptrC           = dvda+jnrC;
 744             fjptrD           = dvda+jnrD;
 745             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 746                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 747             velec            = _mm256_mul_pd(qq00,rinv00);
 748             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 749
 750             /* CUBIC SPLINE TABLE DISPERSION */
 751             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 752             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 753             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 754             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 755             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 756             Heps             = _mm256_mul_pd(vfeps,H);
 757             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 758             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 759             fvdw6            = _mm256_mul_pd(c6_00,FF);
 760
 761             /* CUBIC SPLINE TABLE REPULSION */
 762             vfitab           = _mm_add_epi32(vfitab,ifour);
 763             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 764             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 765             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 766             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 767             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 768             Heps             = _mm256_mul_pd(vfeps,H);
 769             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 770             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 771             fvdw12           = _mm256_mul_pd(c12_00,FF);
 772             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 773
 774             fscal            = _mm256_add_pd(felec,fvdw);
 775
 776             /* Calculate temporary vectorial force */
 777             tx               = _mm256_mul_pd(fscal,dx00);
 778             ty               = _mm256_mul_pd(fscal,dy00);
 779             tz               = _mm256_mul_pd(fscal,dz00);
 780
 781             /* Update vectorial force */
 782             fix0             = _mm256_add_pd(fix0,tx);
 783             fiy0             = _mm256_add_pd(fiy0,ty);
 784             fiz0             = _mm256_add_pd(fiz0,tz);
 785
 786             fjptrA             = f+j_coord_offsetA;
 787             fjptrB             = f+j_coord_offsetB;
 788             fjptrC             = f+j_coord_offsetC;
 789             fjptrD             = f+j_coord_offsetD;
 790             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 791
 792             /* Inner loop uses 81 flops */
 793         }
 794
 795         if(jidx<j_index_end)
 796         {
 797
 798             /* Get j neighbor index, and coordinate index */
 799             jnrlistA         = jjnr[jidx];
 800             jnrlistB         = jjnr[jidx+1];
 801             jnrlistC         = jjnr[jidx+2];
 802             jnrlistD         = jjnr[jidx+3];
 803             /* Sign of each element will be negative for non-real atoms.
 804              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 805              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 806              */
 807             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 808
 809             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 810             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 811             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 812
 813             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 814             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 815             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 816             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 817             j_coord_offsetA  = DIM*jnrA;
 818             j_coord_offsetB  = DIM*jnrB;
 819             j_coord_offsetC  = DIM*jnrC;
 820             j_coord_offsetD  = DIM*jnrD;
 821
 822             /* load j atom coordinates */
 823             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 824                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 825                                                  &jx0,&jy0,&jz0);
 826
 827             /* Calculate displacement vector */
 828             dx00             = _mm256_sub_pd(ix0,jx0);
 829             dy00             = _mm256_sub_pd(iy0,jy0);
 830             dz00             = _mm256_sub_pd(iz0,jz0);
 831
 832             /* Calculate squared distance and things based on it */
 833             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 834
 835             rinv00           = avx256_invsqrt_d(rsq00);
 836
 837             /* Load parameters for j particles */
 838             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 839                                                                  charge+jnrC+0,charge+jnrD+0);
 840             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 841                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 842             vdwjidx0A        = 2*vdwtype[jnrA+0];
 843             vdwjidx0B        = 2*vdwtype[jnrB+0];
 844             vdwjidx0C        = 2*vdwtype[jnrC+0];
 845             vdwjidx0D        = 2*vdwtype[jnrD+0];
 846
 847             /**************************
 848              * CALCULATE INTERACTIONS *
 849              **************************/
 850
 851             r00              = _mm256_mul_pd(rsq00,rinv00);
 852             r00              = _mm256_andnot_pd(dummy_mask,r00);
 853
 854             /* Compute parameters for interactions between i and j atoms */
 855             qq00             = _mm256_mul_pd(iq0,jq0);
 856             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 857                                             vdwioffsetptr0+vdwjidx0B,
 858                                             vdwioffsetptr0+vdwjidx0C,
 859                                             vdwioffsetptr0+vdwjidx0D,
 860                                             &c6_00,&c12_00);
 861
 862             /* Calculate table index by multiplying r with table scale and truncate to integer */
 863             rt               = _mm256_mul_pd(r00,vftabscale);
 864             vfitab           = _mm256_cvttpd_epi32(rt);
 865             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 866             vfitab           = _mm_slli_epi32(vfitab,3);
 867
 868             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 869             isaprod          = _mm256_mul_pd(isai0,isaj0);
 870             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 871             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 872
 873             /* Calculate generalized born table index - this is a separate table from the normal one,
 874              * but we use the same procedure by multiplying r with scale and truncating to integer.
 875              */
 876             rt               = _mm256_mul_pd(r00,gbscale);
 877             gbitab           = _mm256_cvttpd_epi32(rt);
 878             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 879             gbitab           = _mm_slli_epi32(gbitab,2);
 880             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 881             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 882             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 883             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 884             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 885             Heps             = _mm256_mul_pd(gbeps,H);
 886             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 887             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 888             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 889
 890             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 891             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 892             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 893             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 894             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 895             /* 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. */
 896             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 897             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 898             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 899             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 900             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 901                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 902             velec            = _mm256_mul_pd(qq00,rinv00);
 903             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 904
 905             /* CUBIC SPLINE TABLE DISPERSION */
 906             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 907             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 908             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 909             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 910             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 911             Heps             = _mm256_mul_pd(vfeps,H);
 912             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 913             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 914             fvdw6            = _mm256_mul_pd(c6_00,FF);
 915
 916             /* CUBIC SPLINE TABLE REPULSION */
 917             vfitab           = _mm_add_epi32(vfitab,ifour);
 918             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 919             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 920             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 921             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 922             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 923             Heps             = _mm256_mul_pd(vfeps,H);
 924             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 925             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 926             fvdw12           = _mm256_mul_pd(c12_00,FF);
 927             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 928
 929             fscal            = _mm256_add_pd(felec,fvdw);
 930
 931             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 932
 933             /* Calculate temporary vectorial force */
 934             tx               = _mm256_mul_pd(fscal,dx00);
 935             ty               = _mm256_mul_pd(fscal,dy00);
 936             tz               = _mm256_mul_pd(fscal,dz00);
 937
 938             /* Update vectorial force */
 939             fix0             = _mm256_add_pd(fix0,tx);
 940             fiy0             = _mm256_add_pd(fiy0,ty);
 941             fiz0             = _mm256_add_pd(fiz0,tz);
 942
 943             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 944             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 945             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 946             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 947             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 948
 949             /* Inner loop uses 82 flops */
 950         }
 951
 952         /* End of innermost loop */
 953
 954         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 955                                                  f+i_coord_offset,fshift+i_shift_offset);
 956
 957         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 958         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 959
 960         /* Increment number of inner iterations */
 961         inneriter                  += j_index_end - j_index_start;
 962
 963         /* Outer loop uses 7 flops */
 964     }
 965
 966     /* Increment number of outer iterations */
 967     outeriter        += nri;
 968
 969     /* Update outer/inner flops */
 970
 971     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);
 972 }