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

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_256_double.h"
  34 #include "kernelutil_x86_avx_256_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  63     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  64     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  65     real             rcutoff_scalar;
  66     real             *shiftvec,*fshift,*x,*f;
  67     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  68     real             scratch[4*DIM];
  69     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  70     real *           vdwioffsetptr0;
  71     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  73     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  75     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  76     real             *charge;
  77     __m128i          gbitab;
  78     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  79     __m256d          minushalf = _mm256_set1_pd(-0.5);
  80     real             *invsqrta,*dvda,*gbtab;
  81     int              nvdwtype;
  82     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  83     int              *vdwtype;
  84     real             *vdwparam;
  85     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  86     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  87     __m128i          vfitab;
  88     __m128i          ifour       = _mm_set1_epi32(4);
  89     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  90     real             *vftab;
  91     __m256d          dummy_mask,cutoff_mask;
  92     __m128           tmpmask0,tmpmask1;
  93     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  94     __m256d          one     = _mm256_set1_pd(1.0);
  95     __m256d          two     = _mm256_set1_pd(2.0);
  96     x                = xx[0];
  97     f                = ff[0];
  98
  99     nri              = nlist->nri;
 100     iinr             = nlist->iinr;
 101     jindex           = nlist->jindex;
 102     jjnr             = nlist->jjnr;
 103     shiftidx         = nlist->shift;
 104     gid              = nlist->gid;
 105     shiftvec         = fr->shift_vec[0];
 106     fshift           = fr->fshift[0];
 107     facel            = _mm256_set1_pd(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109     nvdwtype         = fr->ntype;
 110     vdwparam         = fr->nbfp;
 111     vdwtype          = mdatoms->typeA;
 112
 113     vftab            = kernel_data->table_vdw->data;
 114     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 115
 116     invsqrta         = fr->invsqrta;
 117     dvda             = fr->dvda;
 118     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 119     gbtab            = fr->gbtab.data;
 120     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm256_setzero_pd();
 155         fiy0             = _mm256_setzero_pd();
 156         fiz0             = _mm256_setzero_pd();
 157
 158         /* Load parameters for i particles */
 159         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 160         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 161         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 162
 163         /* Reset potential sums */
 164         velecsum         = _mm256_setzero_pd();
 165         vgbsum           = _mm256_setzero_pd();
 166         vvdwsum          = _mm256_setzero_pd();
 167         dvdasum          = _mm256_setzero_pd();
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 171         {
 172
 173             /* Get j neighbor index, and coordinate index */
 174             jnrA             = jjnr[jidx];
 175             jnrB             = jjnr[jidx+1];
 176             jnrC             = jjnr[jidx+2];
 177             jnrD             = jjnr[jidx+3];
 178             j_coord_offsetA  = DIM*jnrA;
 179             j_coord_offsetB  = DIM*jnrB;
 180             j_coord_offsetC  = DIM*jnrC;
 181             j_coord_offsetD  = DIM*jnrD;
 182
 183             /* load j atom coordinates */
 184             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 185                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 186                                                  &jx0,&jy0,&jz0);
 187
 188             /* Calculate displacement vector */
 189             dx00             = _mm256_sub_pd(ix0,jx0);
 190             dy00             = _mm256_sub_pd(iy0,jy0);
 191             dz00             = _mm256_sub_pd(iz0,jz0);
 192
 193             /* Calculate squared distance and things based on it */
 194             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 195
 196             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 197
 198             /* Load parameters for j particles */
 199             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 200                                                                  charge+jnrC+0,charge+jnrD+0);
 201             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 202                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 203             vdwjidx0A        = 2*vdwtype[jnrA+0];
 204             vdwjidx0B        = 2*vdwtype[jnrB+0];
 205             vdwjidx0C        = 2*vdwtype[jnrC+0];
 206             vdwjidx0D        = 2*vdwtype[jnrD+0];
 207
 208             /**************************
 209              * CALCULATE INTERACTIONS *
 210              **************************/
 211
 212             r00              = _mm256_mul_pd(rsq00,rinv00);
 213
 214             /* Compute parameters for interactions between i and j atoms */
 215             qq00             = _mm256_mul_pd(iq0,jq0);
 216             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 217                                             vdwioffsetptr0+vdwjidx0B,
 218                                             vdwioffsetptr0+vdwjidx0C,
 219                                             vdwioffsetptr0+vdwjidx0D,
 220                                             &c6_00,&c12_00);
 221
 222             /* Calculate table index by multiplying r with table scale and truncate to integer */
 223             rt               = _mm256_mul_pd(r00,vftabscale);
 224             vfitab           = _mm256_cvttpd_epi32(rt);
 225             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 226             vfitab           = _mm_slli_epi32(vfitab,3);
 227
 228             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 229             isaprod          = _mm256_mul_pd(isai0,isaj0);
 230             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 231             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 232
 233             /* Calculate generalized born table index - this is a separate table from the normal one,
 234              * but we use the same procedure by multiplying r with scale and truncating to integer.
 235              */
 236             rt               = _mm256_mul_pd(r00,gbscale);
 237             gbitab           = _mm256_cvttpd_epi32(rt);
 238             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 239             gbitab           = _mm_slli_epi32(gbitab,2);
 240             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 241             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 242             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 243             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 244             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 245             Heps             = _mm256_mul_pd(gbeps,H);
 246             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 247             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 248             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 249
 250             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 251             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 252             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 253             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 254             fjptrA           = dvda+jnrA;
 255             fjptrB           = dvda+jnrB;
 256             fjptrC           = dvda+jnrC;
 257             fjptrD           = dvda+jnrD;
 258             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 259                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 260             velec            = _mm256_mul_pd(qq00,rinv00);
 261             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 262
 263             /* CUBIC SPLINE TABLE DISPERSION */
 264             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 265             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 266             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 267             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 268             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 269             Heps             = _mm256_mul_pd(vfeps,H);
 270             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 271             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 272             vvdw6            = _mm256_mul_pd(c6_00,VV);
 273             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 274             fvdw6            = _mm256_mul_pd(c6_00,FF);
 275
 276             /* CUBIC SPLINE TABLE REPULSION */
 277             vfitab           = _mm_add_epi32(vfitab,ifour);
 278             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 279             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 280             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 281             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 282             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 283             Heps             = _mm256_mul_pd(vfeps,H);
 284             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 285             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 286             vvdw12           = _mm256_mul_pd(c12_00,VV);
 287             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 288             fvdw12           = _mm256_mul_pd(c12_00,FF);
 289             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 290             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 291
 292             /* Update potential sum for this i atom from the interaction with this j atom. */
 293             velecsum         = _mm256_add_pd(velecsum,velec);
 294             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 295             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 296
 297             fscal            = _mm256_add_pd(felec,fvdw);
 298
 299             /* Calculate temporary vectorial force */
 300             tx               = _mm256_mul_pd(fscal,dx00);
 301             ty               = _mm256_mul_pd(fscal,dy00);
 302             tz               = _mm256_mul_pd(fscal,dz00);
 303
 304             /* Update vectorial force */
 305             fix0             = _mm256_add_pd(fix0,tx);
 306             fiy0             = _mm256_add_pd(fiy0,ty);
 307             fiz0             = _mm256_add_pd(fiz0,tz);
 308
 309             fjptrA             = f+j_coord_offsetA;
 310             fjptrB             = f+j_coord_offsetB;
 311             fjptrC             = f+j_coord_offsetC;
 312             fjptrD             = f+j_coord_offsetD;
 313             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 314
 315             /* Inner loop uses 91 flops */
 316         }
 317
 318         if(jidx<j_index_end)
 319         {
 320
 321             /* Get j neighbor index, and coordinate index */
 322             jnrlistA         = jjnr[jidx];
 323             jnrlistB         = jjnr[jidx+1];
 324             jnrlistC         = jjnr[jidx+2];
 325             jnrlistD         = jjnr[jidx+3];
 326             /* Sign of each element will be negative for non-real atoms.
 327              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 328              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 329              */
 330             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 331
 332             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 333             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 334             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 335
 336             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 337             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 338             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 339             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 340             j_coord_offsetA  = DIM*jnrA;
 341             j_coord_offsetB  = DIM*jnrB;
 342             j_coord_offsetC  = DIM*jnrC;
 343             j_coord_offsetD  = DIM*jnrD;
 344
 345             /* load j atom coordinates */
 346             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 347                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 348                                                  &jx0,&jy0,&jz0);
 349
 350             /* Calculate displacement vector */
 351             dx00             = _mm256_sub_pd(ix0,jx0);
 352             dy00             = _mm256_sub_pd(iy0,jy0);
 353             dz00             = _mm256_sub_pd(iz0,jz0);
 354
 355             /* Calculate squared distance and things based on it */
 356             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 357
 358             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 359
 360             /* Load parameters for j particles */
 361             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 362                                                                  charge+jnrC+0,charge+jnrD+0);
 363             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 364                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 365             vdwjidx0A        = 2*vdwtype[jnrA+0];
 366             vdwjidx0B        = 2*vdwtype[jnrB+0];
 367             vdwjidx0C        = 2*vdwtype[jnrC+0];
 368             vdwjidx0D        = 2*vdwtype[jnrD+0];
 369
 370             /**************************
 371              * CALCULATE INTERACTIONS *
 372              **************************/
 373
 374             r00              = _mm256_mul_pd(rsq00,rinv00);
 375             r00              = _mm256_andnot_pd(dummy_mask,r00);
 376
 377             /* Compute parameters for interactions between i and j atoms */
 378             qq00             = _mm256_mul_pd(iq0,jq0);
 379             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 380                                             vdwioffsetptr0+vdwjidx0B,
 381                                             vdwioffsetptr0+vdwjidx0C,
 382                                             vdwioffsetptr0+vdwjidx0D,
 383                                             &c6_00,&c12_00);
 384
 385             /* Calculate table index by multiplying r with table scale and truncate to integer */
 386             rt               = _mm256_mul_pd(r00,vftabscale);
 387             vfitab           = _mm256_cvttpd_epi32(rt);
 388             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 389             vfitab           = _mm_slli_epi32(vfitab,3);
 390
 391             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 392             isaprod          = _mm256_mul_pd(isai0,isaj0);
 393             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 394             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 395
 396             /* Calculate generalized born table index - this is a separate table from the normal one,
 397              * but we use the same procedure by multiplying r with scale and truncating to integer.
 398              */
 399             rt               = _mm256_mul_pd(r00,gbscale);
 400             gbitab           = _mm256_cvttpd_epi32(rt);
 401             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 402             gbitab           = _mm_slli_epi32(gbitab,2);
 403             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 404             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 405             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 406             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 407             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 408             Heps             = _mm256_mul_pd(gbeps,H);
 409             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 410             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 411             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 412
 413             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 414             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 415             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 416             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 417             /* 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. */
 418             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 419             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 420             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 421             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 422             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 423                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 424             velec            = _mm256_mul_pd(qq00,rinv00);
 425             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 426
 427             /* CUBIC SPLINE TABLE DISPERSION */
 428             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 429             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 430             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 431             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 432             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 433             Heps             = _mm256_mul_pd(vfeps,H);
 434             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 435             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 436             vvdw6            = _mm256_mul_pd(c6_00,VV);
 437             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 438             fvdw6            = _mm256_mul_pd(c6_00,FF);
 439
 440             /* CUBIC SPLINE TABLE REPULSION */
 441             vfitab           = _mm_add_epi32(vfitab,ifour);
 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             vvdw12           = _mm256_mul_pd(c12_00,VV);
 451             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 452             fvdw12           = _mm256_mul_pd(c12_00,FF);
 453             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 454             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             velec            = _mm256_andnot_pd(dummy_mask,velec);
 458             velecsum         = _mm256_add_pd(velecsum,velec);
 459             vgb              = _mm256_andnot_pd(dummy_mask,vgb);
 460             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 461             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 462             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 463
 464             fscal            = _mm256_add_pd(felec,fvdw);
 465
 466             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 467
 468             /* Calculate temporary vectorial force */
 469             tx               = _mm256_mul_pd(fscal,dx00);
 470             ty               = _mm256_mul_pd(fscal,dy00);
 471             tz               = _mm256_mul_pd(fscal,dz00);
 472
 473             /* Update vectorial force */
 474             fix0             = _mm256_add_pd(fix0,tx);
 475             fiy0             = _mm256_add_pd(fiy0,ty);
 476             fiz0             = _mm256_add_pd(fiz0,tz);
 477
 478             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 479             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 480             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 481             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 482             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 483
 484             /* Inner loop uses 92 flops */
 485         }
 486
 487         /* End of innermost loop */
 488
 489         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 490                                                  f+i_coord_offset,fshift+i_shift_offset);
 491
 492         ggid                        = gid[iidx];
 493         /* Update potential energies */
 494         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 495         gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 496         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 497         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 498         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 499
 500         /* Increment number of inner iterations */
 501         inneriter                  += j_index_end - j_index_start;
 502
 503         /* Outer loop uses 10 flops */
 504     }
 505
 506     /* Increment number of outer iterations */
 507     outeriter        += nri;
 508
 509     /* Update outer/inner flops */
 510
 511     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
 512 }
 513 /*
 514  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 515  * Electrostatics interaction: GeneralizedBorn
 516  * VdW interaction:            CubicSplineTable
 517  * Geometry:                   Particle-Particle
 518  * Calculate force/pot:        Force
 519  */
 520 void
 521 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 522                     (t_nblist * gmx_restrict                nlist,
 523                      rvec * gmx_restrict                    xx,
 524                      rvec * gmx_restrict                    ff,
 525                      t_forcerec * gmx_restrict              fr,
 526                      t_mdatoms * gmx_restrict               mdatoms,
 527                      nb_kernel_data_t * gmx_restrict        kernel_data,
 528                      t_nrnb * gmx_restrict                  nrnb)
 529 {
 530     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 531      * just 0 for non-waters.
 532      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 533      * jnr indices corresponding to data put in the four positions in the SIMD register.
 534      */
 535     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 536     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 537     int              jnrA,jnrB,jnrC,jnrD;
 538     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 539     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 540     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 541     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 542     real             rcutoff_scalar;
 543     real             *shiftvec,*fshift,*x,*f;
 544     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 545     real             scratch[4*DIM];
 546     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 547     real *           vdwioffsetptr0;
 548     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 549     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 550     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 551     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 552     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 553     real             *charge;
 554     __m128i          gbitab;
 555     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 556     __m256d          minushalf = _mm256_set1_pd(-0.5);
 557     real             *invsqrta,*dvda,*gbtab;
 558     int              nvdwtype;
 559     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 560     int              *vdwtype;
 561     real             *vdwparam;
 562     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 563     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 564     __m128i          vfitab;
 565     __m128i          ifour       = _mm_set1_epi32(4);
 566     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 567     real             *vftab;
 568     __m256d          dummy_mask,cutoff_mask;
 569     __m128           tmpmask0,tmpmask1;
 570     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 571     __m256d          one     = _mm256_set1_pd(1.0);
 572     __m256d          two     = _mm256_set1_pd(2.0);
 573     x                = xx[0];
 574     f                = ff[0];
 575
 576     nri              = nlist->nri;
 577     iinr             = nlist->iinr;
 578     jindex           = nlist->jindex;
 579     jjnr             = nlist->jjnr;
 580     shiftidx         = nlist->shift;
 581     gid              = nlist->gid;
 582     shiftvec         = fr->shift_vec[0];
 583     fshift           = fr->fshift[0];
 584     facel            = _mm256_set1_pd(fr->epsfac);
 585     charge           = mdatoms->chargeA;
 586     nvdwtype         = fr->ntype;
 587     vdwparam         = fr->nbfp;
 588     vdwtype          = mdatoms->typeA;
 589
 590     vftab            = kernel_data->table_vdw->data;
 591     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 592
 593     invsqrta         = fr->invsqrta;
 594     dvda             = fr->dvda;
 595     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 596     gbtab            = fr->gbtab.data;
 597     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 598
 599     /* Avoid stupid compiler warnings */
 600     jnrA = jnrB = jnrC = jnrD = 0;
 601     j_coord_offsetA = 0;
 602     j_coord_offsetB = 0;
 603     j_coord_offsetC = 0;
 604     j_coord_offsetD = 0;
 605
 606     outeriter        = 0;
 607     inneriter        = 0;
 608
 609     for(iidx=0;iidx<4*DIM;iidx++)
 610     {
 611         scratch[iidx] = 0.0;
 612     }
 613
 614     /* Start outer loop over neighborlists */
 615     for(iidx=0; iidx<nri; iidx++)
 616     {
 617         /* Load shift vector for this list */
 618         i_shift_offset   = DIM*shiftidx[iidx];
 619
 620         /* Load limits for loop over neighbors */
 621         j_index_start    = jindex[iidx];
 622         j_index_end      = jindex[iidx+1];
 623
 624         /* Get outer coordinate index */
 625         inr              = iinr[iidx];
 626         i_coord_offset   = DIM*inr;
 627
 628         /* Load i particle coords and add shift vector */
 629         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 630
 631         fix0             = _mm256_setzero_pd();
 632         fiy0             = _mm256_setzero_pd();
 633         fiz0             = _mm256_setzero_pd();
 634
 635         /* Load parameters for i particles */
 636         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 637         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 638         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 639
 640         dvdasum          = _mm256_setzero_pd();
 641
 642         /* Start inner kernel loop */
 643         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 644         {
 645
 646             /* Get j neighbor index, and coordinate index */
 647             jnrA             = jjnr[jidx];
 648             jnrB             = jjnr[jidx+1];
 649             jnrC             = jjnr[jidx+2];
 650             jnrD             = jjnr[jidx+3];
 651             j_coord_offsetA  = DIM*jnrA;
 652             j_coord_offsetB  = DIM*jnrB;
 653             j_coord_offsetC  = DIM*jnrC;
 654             j_coord_offsetD  = DIM*jnrD;
 655
 656             /* load j atom coordinates */
 657             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 658                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 659                                                  &jx0,&jy0,&jz0);
 660
 661             /* Calculate displacement vector */
 662             dx00             = _mm256_sub_pd(ix0,jx0);
 663             dy00             = _mm256_sub_pd(iy0,jy0);
 664             dz00             = _mm256_sub_pd(iz0,jz0);
 665
 666             /* Calculate squared distance and things based on it */
 667             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 668
 669             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 670
 671             /* Load parameters for j particles */
 672             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 673                                                                  charge+jnrC+0,charge+jnrD+0);
 674             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 675                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 676             vdwjidx0A        = 2*vdwtype[jnrA+0];
 677             vdwjidx0B        = 2*vdwtype[jnrB+0];
 678             vdwjidx0C        = 2*vdwtype[jnrC+0];
 679             vdwjidx0D        = 2*vdwtype[jnrD+0];
 680
 681             /**************************
 682              * CALCULATE INTERACTIONS *
 683              **************************/
 684
 685             r00              = _mm256_mul_pd(rsq00,rinv00);
 686
 687             /* Compute parameters for interactions between i and j atoms */
 688             qq00             = _mm256_mul_pd(iq0,jq0);
 689             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 690                                             vdwioffsetptr0+vdwjidx0B,
 691                                             vdwioffsetptr0+vdwjidx0C,
 692                                             vdwioffsetptr0+vdwjidx0D,
 693                                             &c6_00,&c12_00);
 694
 695             /* Calculate table index by multiplying r with table scale and truncate to integer */
 696             rt               = _mm256_mul_pd(r00,vftabscale);
 697             vfitab           = _mm256_cvttpd_epi32(rt);
 698             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 699             vfitab           = _mm_slli_epi32(vfitab,3);
 700
 701             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 702             isaprod          = _mm256_mul_pd(isai0,isaj0);
 703             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 704             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 705
 706             /* Calculate generalized born table index - this is a separate table from the normal one,
 707              * but we use the same procedure by multiplying r with scale and truncating to integer.
 708              */
 709             rt               = _mm256_mul_pd(r00,gbscale);
 710             gbitab           = _mm256_cvttpd_epi32(rt);
 711             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 712             gbitab           = _mm_slli_epi32(gbitab,2);
 713             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 714             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 715             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 716             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 717             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 718             Heps             = _mm256_mul_pd(gbeps,H);
 719             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 720             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 721             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 722
 723             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 724             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 725             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 726             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 727             fjptrA           = dvda+jnrA;
 728             fjptrB           = dvda+jnrB;
 729             fjptrC           = dvda+jnrC;
 730             fjptrD           = dvda+jnrD;
 731             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 732                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 733             velec            = _mm256_mul_pd(qq00,rinv00);
 734             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 735
 736             /* CUBIC SPLINE TABLE DISPERSION */
 737             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 738             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 739             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 740             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 741             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 742             Heps             = _mm256_mul_pd(vfeps,H);
 743             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 744             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 745             fvdw6            = _mm256_mul_pd(c6_00,FF);
 746
 747             /* CUBIC SPLINE TABLE REPULSION */
 748             vfitab           = _mm_add_epi32(vfitab,ifour);
 749             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 750             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 751             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 752             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 753             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 754             Heps             = _mm256_mul_pd(vfeps,H);
 755             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 756             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 757             fvdw12           = _mm256_mul_pd(c12_00,FF);
 758             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 759
 760             fscal            = _mm256_add_pd(felec,fvdw);
 761
 762             /* Calculate temporary vectorial force */
 763             tx               = _mm256_mul_pd(fscal,dx00);
 764             ty               = _mm256_mul_pd(fscal,dy00);
 765             tz               = _mm256_mul_pd(fscal,dz00);
 766
 767             /* Update vectorial force */
 768             fix0             = _mm256_add_pd(fix0,tx);
 769             fiy0             = _mm256_add_pd(fiy0,ty);
 770             fiz0             = _mm256_add_pd(fiz0,tz);
 771
 772             fjptrA             = f+j_coord_offsetA;
 773             fjptrB             = f+j_coord_offsetB;
 774             fjptrC             = f+j_coord_offsetC;
 775             fjptrD             = f+j_coord_offsetD;
 776             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 777
 778             /* Inner loop uses 81 flops */
 779         }
 780
 781         if(jidx<j_index_end)
 782         {
 783
 784             /* Get j neighbor index, and coordinate index */
 785             jnrlistA         = jjnr[jidx];
 786             jnrlistB         = jjnr[jidx+1];
 787             jnrlistC         = jjnr[jidx+2];
 788             jnrlistD         = jjnr[jidx+3];
 789             /* Sign of each element will be negative for non-real atoms.
 790              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 791              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 792              */
 793             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 794
 795             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 796             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 797             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 798
 799             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 800             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 801             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 802             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 803             j_coord_offsetA  = DIM*jnrA;
 804             j_coord_offsetB  = DIM*jnrB;
 805             j_coord_offsetC  = DIM*jnrC;
 806             j_coord_offsetD  = DIM*jnrD;
 807
 808             /* load j atom coordinates */
 809             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 810                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 811                                                  &jx0,&jy0,&jz0);
 812
 813             /* Calculate displacement vector */
 814             dx00             = _mm256_sub_pd(ix0,jx0);
 815             dy00             = _mm256_sub_pd(iy0,jy0);
 816             dz00             = _mm256_sub_pd(iz0,jz0);
 817
 818             /* Calculate squared distance and things based on it */
 819             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 820
 821             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 822
 823             /* Load parameters for j particles */
 824             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 825                                                                  charge+jnrC+0,charge+jnrD+0);
 826             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 827                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 828             vdwjidx0A        = 2*vdwtype[jnrA+0];
 829             vdwjidx0B        = 2*vdwtype[jnrB+0];
 830             vdwjidx0C        = 2*vdwtype[jnrC+0];
 831             vdwjidx0D        = 2*vdwtype[jnrD+0];
 832
 833             /**************************
 834              * CALCULATE INTERACTIONS *
 835              **************************/
 836
 837             r00              = _mm256_mul_pd(rsq00,rinv00);
 838             r00              = _mm256_andnot_pd(dummy_mask,r00);
 839
 840             /* Compute parameters for interactions between i and j atoms */
 841             qq00             = _mm256_mul_pd(iq0,jq0);
 842             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 843                                             vdwioffsetptr0+vdwjidx0B,
 844                                             vdwioffsetptr0+vdwjidx0C,
 845                                             vdwioffsetptr0+vdwjidx0D,
 846                                             &c6_00,&c12_00);
 847
 848             /* Calculate table index by multiplying r with table scale and truncate to integer */
 849             rt               = _mm256_mul_pd(r00,vftabscale);
 850             vfitab           = _mm256_cvttpd_epi32(rt);
 851             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 852             vfitab           = _mm_slli_epi32(vfitab,3);
 853
 854             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 855             isaprod          = _mm256_mul_pd(isai0,isaj0);
 856             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 857             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 858
 859             /* Calculate generalized born table index - this is a separate table from the normal one,
 860              * but we use the same procedure by multiplying r with scale and truncating to integer.
 861              */
 862             rt               = _mm256_mul_pd(r00,gbscale);
 863             gbitab           = _mm256_cvttpd_epi32(rt);
 864             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 865             gbitab           = _mm_slli_epi32(gbitab,2);
 866             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 867             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 868             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 869             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 870             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 871             Heps             = _mm256_mul_pd(gbeps,H);
 872             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 873             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 874             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 875
 876             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 877             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 878             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 879             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 880             /* 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. */
 881             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 882             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 883             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 884             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 885             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 886                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 887             velec            = _mm256_mul_pd(qq00,rinv00);
 888             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 889
 890             /* CUBIC SPLINE TABLE DISPERSION */
 891             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 892             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 893             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 894             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 895             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 896             Heps             = _mm256_mul_pd(vfeps,H);
 897             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 898             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 899             fvdw6            = _mm256_mul_pd(c6_00,FF);
 900
 901             /* CUBIC SPLINE TABLE REPULSION */
 902             vfitab           = _mm_add_epi32(vfitab,ifour);
 903             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 904             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 905             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 906             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 907             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 908             Heps             = _mm256_mul_pd(vfeps,H);
 909             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 910             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 911             fvdw12           = _mm256_mul_pd(c12_00,FF);
 912             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 913
 914             fscal            = _mm256_add_pd(felec,fvdw);
 915
 916             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 917
 918             /* Calculate temporary vectorial force */
 919             tx               = _mm256_mul_pd(fscal,dx00);
 920             ty               = _mm256_mul_pd(fscal,dy00);
 921             tz               = _mm256_mul_pd(fscal,dz00);
 922
 923             /* Update vectorial force */
 924             fix0             = _mm256_add_pd(fix0,tx);
 925             fiy0             = _mm256_add_pd(fiy0,ty);
 926             fiz0             = _mm256_add_pd(fiz0,tz);
 927
 928             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 929             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 930             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 931             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 932             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 933
 934             /* Inner loop uses 82 flops */
 935         }
 936
 937         /* End of innermost loop */
 938
 939         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 940                                                  f+i_coord_offset,fshift+i_shift_offset);
 941
 942         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 943         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 944
 945         /* Increment number of inner iterations */
 946         inneriter                  += j_index_end - j_index_start;
 947
 948         /* Outer loop uses 7 flops */
 949     }
 950
 951     /* Increment number of outer iterations */
 952     outeriter        += nri;
 953
 954     /* Update outer/inner flops */
 955
 956     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);
 957 }