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, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: GeneralizedBorn
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  79     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  84     real             scratch[4*DIM];
  85     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     __m128i          gbitab;
  94     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  95     __m256d          minushalf = _mm256_set1_pd(-0.5);
  96     real             *invsqrta,*dvda,*gbtab;
  97     int              nvdwtype;
  98     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 102     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 103     __m128i          vfitab;
 104     __m128i          ifour       = _mm_set1_epi32(4);
 105     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 106     real             *vftab;
 107     __m256d          dummy_mask,cutoff_mask;
 108     __m128           tmpmask0,tmpmask1;
 109     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 110     __m256d          one     = _mm256_set1_pd(1.0);
 111     __m256d          two     = _mm256_set1_pd(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm256_set1_pd(fr->epsfac);
 124     charge           = mdatoms->chargeA;
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     vftab            = kernel_data->table_vdw->data;
 130     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 131
 132     invsqrta         = fr->invsqrta;
 133     dvda             = fr->dvda;
 134     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 135     gbtab            = fr->gbtab.data;
 136     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 137
 138     /* Avoid stupid compiler warnings */
 139     jnrA = jnrB = jnrC = jnrD = 0;
 140     j_coord_offsetA = 0;
 141     j_coord_offsetB = 0;
 142     j_coord_offsetC = 0;
 143     j_coord_offsetD = 0;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     for(iidx=0;iidx<4*DIM;iidx++)
 149     {
 150         scratch[iidx] = 0.0;
 151     }
 152
 153     /* Start outer loop over neighborlists */
 154     for(iidx=0; iidx<nri; iidx++)
 155     {
 156         /* Load shift vector for this list */
 157         i_shift_offset   = DIM*shiftidx[iidx];
 158
 159         /* Load limits for loop over neighbors */
 160         j_index_start    = jindex[iidx];
 161         j_index_end      = jindex[iidx+1];
 162
 163         /* Get outer coordinate index */
 164         inr              = iinr[iidx];
 165         i_coord_offset   = DIM*inr;
 166
 167         /* Load i particle coords and add shift vector */
 168         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 169
 170         fix0             = _mm256_setzero_pd();
 171         fiy0             = _mm256_setzero_pd();
 172         fiz0             = _mm256_setzero_pd();
 173
 174         /* Load parameters for i particles */
 175         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 176         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 177         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 178
 179         /* Reset potential sums */
 180         velecsum         = _mm256_setzero_pd();
 181         vgbsum           = _mm256_setzero_pd();
 182         vvdwsum          = _mm256_setzero_pd();
 183         dvdasum          = _mm256_setzero_pd();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             jnrC             = jjnr[jidx+2];
 193             jnrD             = jjnr[jidx+3];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196             j_coord_offsetC  = DIM*jnrC;
 197             j_coord_offsetD  = DIM*jnrD;
 198
 199             /* load j atom coordinates */
 200             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 201                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 202                                                  &jx0,&jy0,&jz0);
 203
 204             /* Calculate displacement vector */
 205             dx00             = _mm256_sub_pd(ix0,jx0);
 206             dy00             = _mm256_sub_pd(iy0,jy0);
 207             dz00             = _mm256_sub_pd(iz0,jz0);
 208
 209             /* Calculate squared distance and things based on it */
 210             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 211
 212             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 213
 214             /* Load parameters for j particles */
 215             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 216                                                                  charge+jnrC+0,charge+jnrD+0);
 217             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 218                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 219             vdwjidx0A        = 2*vdwtype[jnrA+0];
 220             vdwjidx0B        = 2*vdwtype[jnrB+0];
 221             vdwjidx0C        = 2*vdwtype[jnrC+0];
 222             vdwjidx0D        = 2*vdwtype[jnrD+0];
 223
 224             /**************************
 225              * CALCULATE INTERACTIONS *
 226              **************************/
 227
 228             r00              = _mm256_mul_pd(rsq00,rinv00);
 229
 230             /* Compute parameters for interactions between i and j atoms */
 231             qq00             = _mm256_mul_pd(iq0,jq0);
 232             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 233                                             vdwioffsetptr0+vdwjidx0B,
 234                                             vdwioffsetptr0+vdwjidx0C,
 235                                             vdwioffsetptr0+vdwjidx0D,
 236                                             &c6_00,&c12_00);
 237
 238             /* Calculate table index by multiplying r with table scale and truncate to integer */
 239             rt               = _mm256_mul_pd(r00,vftabscale);
 240             vfitab           = _mm256_cvttpd_epi32(rt);
 241             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 242             vfitab           = _mm_slli_epi32(vfitab,3);
 243
 244             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 245             isaprod          = _mm256_mul_pd(isai0,isaj0);
 246             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 247             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 248
 249             /* Calculate generalized born table index - this is a separate table from the normal one,
 250              * but we use the same procedure by multiplying r with scale and truncating to integer.
 251              */
 252             rt               = _mm256_mul_pd(r00,gbscale);
 253             gbitab           = _mm256_cvttpd_epi32(rt);
 254             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 255             gbitab           = _mm_slli_epi32(gbitab,2);
 256             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 257             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 258             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 259             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 260             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 261             Heps             = _mm256_mul_pd(gbeps,H);
 262             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 263             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 264             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 265
 266             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 267             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 268             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 269             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 270             fjptrA           = dvda+jnrA;
 271             fjptrB           = dvda+jnrB;
 272             fjptrC           = dvda+jnrC;
 273             fjptrD           = dvda+jnrD;
 274             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 275                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 276             velec            = _mm256_mul_pd(qq00,rinv00);
 277             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 278
 279             /* CUBIC SPLINE TABLE DISPERSION */
 280             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 281             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 282             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 283             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 284             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 285             Heps             = _mm256_mul_pd(vfeps,H);
 286             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 287             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 288             vvdw6            = _mm256_mul_pd(c6_00,VV);
 289             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 290             fvdw6            = _mm256_mul_pd(c6_00,FF);
 291
 292             /* CUBIC SPLINE TABLE REPULSION */
 293             vfitab           = _mm_add_epi32(vfitab,ifour);
 294             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 295             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 296             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 297             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 298             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 299             Heps             = _mm256_mul_pd(vfeps,H);
 300             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 301             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 302             vvdw12           = _mm256_mul_pd(c12_00,VV);
 303             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 304             fvdw12           = _mm256_mul_pd(c12_00,FF);
 305             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 306             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 307
 308             /* Update potential sum for this i atom from the interaction with this j atom. */
 309             velecsum         = _mm256_add_pd(velecsum,velec);
 310             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 311             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 312
 313             fscal            = _mm256_add_pd(felec,fvdw);
 314
 315             /* Calculate temporary vectorial force */
 316             tx               = _mm256_mul_pd(fscal,dx00);
 317             ty               = _mm256_mul_pd(fscal,dy00);
 318             tz               = _mm256_mul_pd(fscal,dz00);
 319
 320             /* Update vectorial force */
 321             fix0             = _mm256_add_pd(fix0,tx);
 322             fiy0             = _mm256_add_pd(fiy0,ty);
 323             fiz0             = _mm256_add_pd(fiz0,tz);
 324
 325             fjptrA             = f+j_coord_offsetA;
 326             fjptrB             = f+j_coord_offsetB;
 327             fjptrC             = f+j_coord_offsetC;
 328             fjptrD             = f+j_coord_offsetD;
 329             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 330
 331             /* Inner loop uses 91 flops */
 332         }
 333
 334         if(jidx<j_index_end)
 335         {
 336
 337             /* Get j neighbor index, and coordinate index */
 338             jnrlistA         = jjnr[jidx];
 339             jnrlistB         = jjnr[jidx+1];
 340             jnrlistC         = jjnr[jidx+2];
 341             jnrlistD         = jjnr[jidx+3];
 342             /* Sign of each element will be negative for non-real atoms.
 343              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 344              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 345              */
 346             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 347
 348             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 349             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 350             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 351
 352             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 353             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 354             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 355             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 356             j_coord_offsetA  = DIM*jnrA;
 357             j_coord_offsetB  = DIM*jnrB;
 358             j_coord_offsetC  = DIM*jnrC;
 359             j_coord_offsetD  = DIM*jnrD;
 360
 361             /* load j atom coordinates */
 362             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 363                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 364                                                  &jx0,&jy0,&jz0);
 365
 366             /* Calculate displacement vector */
 367             dx00             = _mm256_sub_pd(ix0,jx0);
 368             dy00             = _mm256_sub_pd(iy0,jy0);
 369             dz00             = _mm256_sub_pd(iz0,jz0);
 370
 371             /* Calculate squared distance and things based on it */
 372             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 373
 374             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 375
 376             /* Load parameters for j particles */
 377             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 378                                                                  charge+jnrC+0,charge+jnrD+0);
 379             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 380                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 381             vdwjidx0A        = 2*vdwtype[jnrA+0];
 382             vdwjidx0B        = 2*vdwtype[jnrB+0];
 383             vdwjidx0C        = 2*vdwtype[jnrC+0];
 384             vdwjidx0D        = 2*vdwtype[jnrD+0];
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             r00              = _mm256_mul_pd(rsq00,rinv00);
 391             r00              = _mm256_andnot_pd(dummy_mask,r00);
 392
 393             /* Compute parameters for interactions between i and j atoms */
 394             qq00             = _mm256_mul_pd(iq0,jq0);
 395             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 396                                             vdwioffsetptr0+vdwjidx0B,
 397                                             vdwioffsetptr0+vdwjidx0C,
 398                                             vdwioffsetptr0+vdwjidx0D,
 399                                             &c6_00,&c12_00);
 400
 401             /* Calculate table index by multiplying r with table scale and truncate to integer */
 402             rt               = _mm256_mul_pd(r00,vftabscale);
 403             vfitab           = _mm256_cvttpd_epi32(rt);
 404             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 405             vfitab           = _mm_slli_epi32(vfitab,3);
 406
 407             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 408             isaprod          = _mm256_mul_pd(isai0,isaj0);
 409             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 410             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 411
 412             /* Calculate generalized born table index - this is a separate table from the normal one,
 413              * but we use the same procedure by multiplying r with scale and truncating to integer.
 414              */
 415             rt               = _mm256_mul_pd(r00,gbscale);
 416             gbitab           = _mm256_cvttpd_epi32(rt);
 417             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 418             gbitab           = _mm_slli_epi32(gbitab,2);
 419             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 420             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 421             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 422             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 423             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 424             Heps             = _mm256_mul_pd(gbeps,H);
 425             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 426             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 427             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 428
 429             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 430             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 431             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 432             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 433             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 434             /* 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. */
 435             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 436             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 437             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 438             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 439             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 440                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 441             velec            = _mm256_mul_pd(qq00,rinv00);
 442             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 443
 444             /* CUBIC SPLINE TABLE DISPERSION */
 445             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 446             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 447             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 448             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 449             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 450             Heps             = _mm256_mul_pd(vfeps,H);
 451             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 452             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 453             vvdw6            = _mm256_mul_pd(c6_00,VV);
 454             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 455             fvdw6            = _mm256_mul_pd(c6_00,FF);
 456
 457             /* CUBIC SPLINE TABLE REPULSION */
 458             vfitab           = _mm_add_epi32(vfitab,ifour);
 459             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 460             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 461             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 462             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 463             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 464             Heps             = _mm256_mul_pd(vfeps,H);
 465             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 466             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 467             vvdw12           = _mm256_mul_pd(c12_00,VV);
 468             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 469             fvdw12           = _mm256_mul_pd(c12_00,FF);
 470             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 471             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 472
 473             /* Update potential sum for this i atom from the interaction with this j atom. */
 474             velec            = _mm256_andnot_pd(dummy_mask,velec);
 475             velecsum         = _mm256_add_pd(velecsum,velec);
 476             vgb              = _mm256_andnot_pd(dummy_mask,vgb);
 477             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 478             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 479             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 480
 481             fscal            = _mm256_add_pd(felec,fvdw);
 482
 483             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 484
 485             /* Calculate temporary vectorial force */
 486             tx               = _mm256_mul_pd(fscal,dx00);
 487             ty               = _mm256_mul_pd(fscal,dy00);
 488             tz               = _mm256_mul_pd(fscal,dz00);
 489
 490             /* Update vectorial force */
 491             fix0             = _mm256_add_pd(fix0,tx);
 492             fiy0             = _mm256_add_pd(fiy0,ty);
 493             fiz0             = _mm256_add_pd(fiz0,tz);
 494
 495             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 496             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 497             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 498             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 499             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 500
 501             /* Inner loop uses 92 flops */
 502         }
 503
 504         /* End of innermost loop */
 505
 506         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 507                                                  f+i_coord_offset,fshift+i_shift_offset);
 508
 509         ggid                        = gid[iidx];
 510         /* Update potential energies */
 511         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 512         gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 513         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 514         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 515         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 516
 517         /* Increment number of inner iterations */
 518         inneriter                  += j_index_end - j_index_start;
 519
 520         /* Outer loop uses 10 flops */
 521     }
 522
 523     /* Increment number of outer iterations */
 524     outeriter        += nri;
 525
 526     /* Update outer/inner flops */
 527
 528     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
 529 }
 530 /*
 531  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 532  * Electrostatics interaction: GeneralizedBorn
 533  * VdW interaction:            CubicSplineTable
 534  * Geometry:                   Particle-Particle
 535  * Calculate force/pot:        Force
 536  */
 537 void
 538 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 539                     (t_nblist                    * gmx_restrict       nlist,
 540                      rvec                        * gmx_restrict          xx,
 541                      rvec                        * gmx_restrict          ff,
 542                      t_forcerec                  * gmx_restrict          fr,
 543                      t_mdatoms                   * gmx_restrict     mdatoms,
 544                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 545                      t_nrnb                      * gmx_restrict        nrnb)
 546 {
 547     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 548      * just 0 for non-waters.
 549      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 550      * jnr indices corresponding to data put in the four positions in the SIMD register.
 551      */
 552     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 553     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 554     int              jnrA,jnrB,jnrC,jnrD;
 555     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 556     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 557     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 558     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 559     real             rcutoff_scalar;
 560     real             *shiftvec,*fshift,*x,*f;
 561     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 562     real             scratch[4*DIM];
 563     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 564     real *           vdwioffsetptr0;
 565     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 566     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 567     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 568     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 569     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 570     real             *charge;
 571     __m128i          gbitab;
 572     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 573     __m256d          minushalf = _mm256_set1_pd(-0.5);
 574     real             *invsqrta,*dvda,*gbtab;
 575     int              nvdwtype;
 576     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 577     int              *vdwtype;
 578     real             *vdwparam;
 579     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 580     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 581     __m128i          vfitab;
 582     __m128i          ifour       = _mm_set1_epi32(4);
 583     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 584     real             *vftab;
 585     __m256d          dummy_mask,cutoff_mask;
 586     __m128           tmpmask0,tmpmask1;
 587     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 588     __m256d          one     = _mm256_set1_pd(1.0);
 589     __m256d          two     = _mm256_set1_pd(2.0);
 590     x                = xx[0];
 591     f                = ff[0];
 592
 593     nri              = nlist->nri;
 594     iinr             = nlist->iinr;
 595     jindex           = nlist->jindex;
 596     jjnr             = nlist->jjnr;
 597     shiftidx         = nlist->shift;
 598     gid              = nlist->gid;
 599     shiftvec         = fr->shift_vec[0];
 600     fshift           = fr->fshift[0];
 601     facel            = _mm256_set1_pd(fr->epsfac);
 602     charge           = mdatoms->chargeA;
 603     nvdwtype         = fr->ntype;
 604     vdwparam         = fr->nbfp;
 605     vdwtype          = mdatoms->typeA;
 606
 607     vftab            = kernel_data->table_vdw->data;
 608     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 609
 610     invsqrta         = fr->invsqrta;
 611     dvda             = fr->dvda;
 612     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 613     gbtab            = fr->gbtab.data;
 614     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 615
 616     /* Avoid stupid compiler warnings */
 617     jnrA = jnrB = jnrC = jnrD = 0;
 618     j_coord_offsetA = 0;
 619     j_coord_offsetB = 0;
 620     j_coord_offsetC = 0;
 621     j_coord_offsetD = 0;
 622
 623     outeriter        = 0;
 624     inneriter        = 0;
 625
 626     for(iidx=0;iidx<4*DIM;iidx++)
 627     {
 628         scratch[iidx] = 0.0;
 629     }
 630
 631     /* Start outer loop over neighborlists */
 632     for(iidx=0; iidx<nri; iidx++)
 633     {
 634         /* Load shift vector for this list */
 635         i_shift_offset   = DIM*shiftidx[iidx];
 636
 637         /* Load limits for loop over neighbors */
 638         j_index_start    = jindex[iidx];
 639         j_index_end      = jindex[iidx+1];
 640
 641         /* Get outer coordinate index */
 642         inr              = iinr[iidx];
 643         i_coord_offset   = DIM*inr;
 644
 645         /* Load i particle coords and add shift vector */
 646         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 647
 648         fix0             = _mm256_setzero_pd();
 649         fiy0             = _mm256_setzero_pd();
 650         fiz0             = _mm256_setzero_pd();
 651
 652         /* Load parameters for i particles */
 653         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 654         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 655         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 656
 657         dvdasum          = _mm256_setzero_pd();
 658
 659         /* Start inner kernel loop */
 660         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 661         {
 662
 663             /* Get j neighbor index, and coordinate index */
 664             jnrA             = jjnr[jidx];
 665             jnrB             = jjnr[jidx+1];
 666             jnrC             = jjnr[jidx+2];
 667             jnrD             = jjnr[jidx+3];
 668             j_coord_offsetA  = DIM*jnrA;
 669             j_coord_offsetB  = DIM*jnrB;
 670             j_coord_offsetC  = DIM*jnrC;
 671             j_coord_offsetD  = DIM*jnrD;
 672
 673             /* load j atom coordinates */
 674             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 675                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 676                                                  &jx0,&jy0,&jz0);
 677
 678             /* Calculate displacement vector */
 679             dx00             = _mm256_sub_pd(ix0,jx0);
 680             dy00             = _mm256_sub_pd(iy0,jy0);
 681             dz00             = _mm256_sub_pd(iz0,jz0);
 682
 683             /* Calculate squared distance and things based on it */
 684             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 685
 686             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 687
 688             /* Load parameters for j particles */
 689             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 690                                                                  charge+jnrC+0,charge+jnrD+0);
 691             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 692                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 693             vdwjidx0A        = 2*vdwtype[jnrA+0];
 694             vdwjidx0B        = 2*vdwtype[jnrB+0];
 695             vdwjidx0C        = 2*vdwtype[jnrC+0];
 696             vdwjidx0D        = 2*vdwtype[jnrD+0];
 697
 698             /**************************
 699              * CALCULATE INTERACTIONS *
 700              **************************/
 701
 702             r00              = _mm256_mul_pd(rsq00,rinv00);
 703
 704             /* Compute parameters for interactions between i and j atoms */
 705             qq00             = _mm256_mul_pd(iq0,jq0);
 706             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 707                                             vdwioffsetptr0+vdwjidx0B,
 708                                             vdwioffsetptr0+vdwjidx0C,
 709                                             vdwioffsetptr0+vdwjidx0D,
 710                                             &c6_00,&c12_00);
 711
 712             /* Calculate table index by multiplying r with table scale and truncate to integer */
 713             rt               = _mm256_mul_pd(r00,vftabscale);
 714             vfitab           = _mm256_cvttpd_epi32(rt);
 715             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 716             vfitab           = _mm_slli_epi32(vfitab,3);
 717
 718             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 719             isaprod          = _mm256_mul_pd(isai0,isaj0);
 720             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 721             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 722
 723             /* Calculate generalized born table index - this is a separate table from the normal one,
 724              * but we use the same procedure by multiplying r with scale and truncating to integer.
 725              */
 726             rt               = _mm256_mul_pd(r00,gbscale);
 727             gbitab           = _mm256_cvttpd_epi32(rt);
 728             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 729             gbitab           = _mm_slli_epi32(gbitab,2);
 730             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 731             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 732             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 733             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 734             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 735             Heps             = _mm256_mul_pd(gbeps,H);
 736             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 737             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 738             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 739
 740             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 741             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 742             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 743             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 744             fjptrA           = dvda+jnrA;
 745             fjptrB           = dvda+jnrB;
 746             fjptrC           = dvda+jnrC;
 747             fjptrD           = dvda+jnrD;
 748             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 749                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 750             velec            = _mm256_mul_pd(qq00,rinv00);
 751             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 752
 753             /* CUBIC SPLINE TABLE DISPERSION */
 754             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 755             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 756             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 757             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 758             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 759             Heps             = _mm256_mul_pd(vfeps,H);
 760             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 761             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 762             fvdw6            = _mm256_mul_pd(c6_00,FF);
 763
 764             /* CUBIC SPLINE TABLE REPULSION */
 765             vfitab           = _mm_add_epi32(vfitab,ifour);
 766             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 767             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 768             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 769             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 770             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 771             Heps             = _mm256_mul_pd(vfeps,H);
 772             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 773             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 774             fvdw12           = _mm256_mul_pd(c12_00,FF);
 775             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 776
 777             fscal            = _mm256_add_pd(felec,fvdw);
 778
 779             /* Calculate temporary vectorial force */
 780             tx               = _mm256_mul_pd(fscal,dx00);
 781             ty               = _mm256_mul_pd(fscal,dy00);
 782             tz               = _mm256_mul_pd(fscal,dz00);
 783
 784             /* Update vectorial force */
 785             fix0             = _mm256_add_pd(fix0,tx);
 786             fiy0             = _mm256_add_pd(fiy0,ty);
 787             fiz0             = _mm256_add_pd(fiz0,tz);
 788
 789             fjptrA             = f+j_coord_offsetA;
 790             fjptrB             = f+j_coord_offsetB;
 791             fjptrC             = f+j_coord_offsetC;
 792             fjptrD             = f+j_coord_offsetD;
 793             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 794
 795             /* Inner loop uses 81 flops */
 796         }
 797
 798         if(jidx<j_index_end)
 799         {
 800
 801             /* Get j neighbor index, and coordinate index */
 802             jnrlistA         = jjnr[jidx];
 803             jnrlistB         = jjnr[jidx+1];
 804             jnrlistC         = jjnr[jidx+2];
 805             jnrlistD         = jjnr[jidx+3];
 806             /* Sign of each element will be negative for non-real atoms.
 807              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 808              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 809              */
 810             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 811
 812             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 813             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 814             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 815
 816             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 817             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 818             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 819             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 820             j_coord_offsetA  = DIM*jnrA;
 821             j_coord_offsetB  = DIM*jnrB;
 822             j_coord_offsetC  = DIM*jnrC;
 823             j_coord_offsetD  = DIM*jnrD;
 824
 825             /* load j atom coordinates */
 826             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 827                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 828                                                  &jx0,&jy0,&jz0);
 829
 830             /* Calculate displacement vector */
 831             dx00             = _mm256_sub_pd(ix0,jx0);
 832             dy00             = _mm256_sub_pd(iy0,jy0);
 833             dz00             = _mm256_sub_pd(iz0,jz0);
 834
 835             /* Calculate squared distance and things based on it */
 836             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 837
 838             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 839
 840             /* Load parameters for j particles */
 841             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 842                                                                  charge+jnrC+0,charge+jnrD+0);
 843             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 844                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 845             vdwjidx0A        = 2*vdwtype[jnrA+0];
 846             vdwjidx0B        = 2*vdwtype[jnrB+0];
 847             vdwjidx0C        = 2*vdwtype[jnrC+0];
 848             vdwjidx0D        = 2*vdwtype[jnrD+0];
 849
 850             /**************************
 851              * CALCULATE INTERACTIONS *
 852              **************************/
 853
 854             r00              = _mm256_mul_pd(rsq00,rinv00);
 855             r00              = _mm256_andnot_pd(dummy_mask,r00);
 856
 857             /* Compute parameters for interactions between i and j atoms */
 858             qq00             = _mm256_mul_pd(iq0,jq0);
 859             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 860                                             vdwioffsetptr0+vdwjidx0B,
 861                                             vdwioffsetptr0+vdwjidx0C,
 862                                             vdwioffsetptr0+vdwjidx0D,
 863                                             &c6_00,&c12_00);
 864
 865             /* Calculate table index by multiplying r with table scale and truncate to integer */
 866             rt               = _mm256_mul_pd(r00,vftabscale);
 867             vfitab           = _mm256_cvttpd_epi32(rt);
 868             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 869             vfitab           = _mm_slli_epi32(vfitab,3);
 870
 871             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 872             isaprod          = _mm256_mul_pd(isai0,isaj0);
 873             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 874             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 875
 876             /* Calculate generalized born table index - this is a separate table from the normal one,
 877              * but we use the same procedure by multiplying r with scale and truncating to integer.
 878              */
 879             rt               = _mm256_mul_pd(r00,gbscale);
 880             gbitab           = _mm256_cvttpd_epi32(rt);
 881             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 882             gbitab           = _mm_slli_epi32(gbitab,2);
 883             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 884             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 885             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 886             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 887             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 888             Heps             = _mm256_mul_pd(gbeps,H);
 889             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 890             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 891             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 892
 893             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 894             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 895             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 896             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 897             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 898             /* 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. */
 899             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 900             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 901             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 902             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 903             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 904                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 905             velec            = _mm256_mul_pd(qq00,rinv00);
 906             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 907
 908             /* CUBIC SPLINE TABLE DISPERSION */
 909             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 910             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 911             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 912             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 913             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 914             Heps             = _mm256_mul_pd(vfeps,H);
 915             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 916             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 917             fvdw6            = _mm256_mul_pd(c6_00,FF);
 918
 919             /* CUBIC SPLINE TABLE REPULSION */
 920             vfitab           = _mm_add_epi32(vfitab,ifour);
 921             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 922             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 923             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 924             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 925             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 926             Heps             = _mm256_mul_pd(vfeps,H);
 927             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 928             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 929             fvdw12           = _mm256_mul_pd(c12_00,FF);
 930             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 931
 932             fscal            = _mm256_add_pd(felec,fvdw);
 933
 934             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 935
 936             /* Calculate temporary vectorial force */
 937             tx               = _mm256_mul_pd(fscal,dx00);
 938             ty               = _mm256_mul_pd(fscal,dy00);
 939             tz               = _mm256_mul_pd(fscal,dz00);
 940
 941             /* Update vectorial force */
 942             fix0             = _mm256_add_pd(fix0,tx);
 943             fiy0             = _mm256_add_pd(fiy0,ty);
 944             fiz0             = _mm256_add_pd(fiz0,tz);
 945
 946             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 947             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 948             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 949             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 950             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 951
 952             /* Inner loop uses 82 flops */
 953         }
 954
 955         /* End of innermost loop */
 956
 957         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 958                                                  f+i_coord_offset,fshift+i_shift_offset);
 959
 960         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 961         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 962
 963         /* Increment number of inner iterations */
 964         inneriter                  += j_index_end - j_index_start;
 965
 966         /* Outer loop uses 7 flops */
 967     }
 968
 969     /* Increment number of outer iterations */
 970     outeriter        += nri;
 971
 972     /* Update outer/inner flops */
 973
 974     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);
 975 }