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             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 417             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 418             /* 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. */
 419             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 420             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 421             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 422             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 423             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 424                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 425             velec            = _mm256_mul_pd(qq00,rinv00);
 426             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 427
 428             /* CUBIC SPLINE TABLE DISPERSION */
 429             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 430             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 431             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 432             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 433             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 434             Heps             = _mm256_mul_pd(vfeps,H);
 435             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 436             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 437             vvdw6            = _mm256_mul_pd(c6_00,VV);
 438             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 439             fvdw6            = _mm256_mul_pd(c6_00,FF);
 440
 441             /* CUBIC SPLINE TABLE REPULSION */
 442             vfitab           = _mm_add_epi32(vfitab,ifour);
 443             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 444             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 445             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 446             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 447             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 448             Heps             = _mm256_mul_pd(vfeps,H);
 449             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 450             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 451             vvdw12           = _mm256_mul_pd(c12_00,VV);
 452             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 453             fvdw12           = _mm256_mul_pd(c12_00,FF);
 454             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 455             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velec            = _mm256_andnot_pd(dummy_mask,velec);
 459             velecsum         = _mm256_add_pd(velecsum,velec);
 460             vgb              = _mm256_andnot_pd(dummy_mask,vgb);
 461             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 462             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 463             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 464
 465             fscal            = _mm256_add_pd(felec,fvdw);
 466
 467             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 468
 469             /* Calculate temporary vectorial force */
 470             tx               = _mm256_mul_pd(fscal,dx00);
 471             ty               = _mm256_mul_pd(fscal,dy00);
 472             tz               = _mm256_mul_pd(fscal,dz00);
 473
 474             /* Update vectorial force */
 475             fix0             = _mm256_add_pd(fix0,tx);
 476             fiy0             = _mm256_add_pd(fiy0,ty);
 477             fiz0             = _mm256_add_pd(fiz0,tz);
 478
 479             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 480             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 481             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 482             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 483             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 484
 485             /* Inner loop uses 92 flops */
 486         }
 487
 488         /* End of innermost loop */
 489
 490         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 491                                                  f+i_coord_offset,fshift+i_shift_offset);
 492
 493         ggid                        = gid[iidx];
 494         /* Update potential energies */
 495         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 496         gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 497         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 498         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 499         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 500
 501         /* Increment number of inner iterations */
 502         inneriter                  += j_index_end - j_index_start;
 503
 504         /* Outer loop uses 10 flops */
 505     }
 506
 507     /* Increment number of outer iterations */
 508     outeriter        += nri;
 509
 510     /* Update outer/inner flops */
 511
 512     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
 513 }
 514 /*
 515  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 516  * Electrostatics interaction: GeneralizedBorn
 517  * VdW interaction:            CubicSplineTable
 518  * Geometry:                   Particle-Particle
 519  * Calculate force/pot:        Force
 520  */
 521 void
 522 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_double
 523                     (t_nblist * gmx_restrict                nlist,
 524                      rvec * gmx_restrict                    xx,
 525                      rvec * gmx_restrict                    ff,
 526                      t_forcerec * gmx_restrict              fr,
 527                      t_mdatoms * gmx_restrict               mdatoms,
 528                      nb_kernel_data_t * gmx_restrict        kernel_data,
 529                      t_nrnb * gmx_restrict                  nrnb)
 530 {
 531     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 532      * just 0 for non-waters.
 533      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 534      * jnr indices corresponding to data put in the four positions in the SIMD register.
 535      */
 536     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 537     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 538     int              jnrA,jnrB,jnrC,jnrD;
 539     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 540     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 541     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 542     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 543     real             rcutoff_scalar;
 544     real             *shiftvec,*fshift,*x,*f;
 545     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 546     real             scratch[4*DIM];
 547     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 548     real *           vdwioffsetptr0;
 549     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 550     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 551     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 552     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 553     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 554     real             *charge;
 555     __m128i          gbitab;
 556     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 557     __m256d          minushalf = _mm256_set1_pd(-0.5);
 558     real             *invsqrta,*dvda,*gbtab;
 559     int              nvdwtype;
 560     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 561     int              *vdwtype;
 562     real             *vdwparam;
 563     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 564     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 565     __m128i          vfitab;
 566     __m128i          ifour       = _mm_set1_epi32(4);
 567     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 568     real             *vftab;
 569     __m256d          dummy_mask,cutoff_mask;
 570     __m128           tmpmask0,tmpmask1;
 571     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 572     __m256d          one     = _mm256_set1_pd(1.0);
 573     __m256d          two     = _mm256_set1_pd(2.0);
 574     x                = xx[0];
 575     f                = ff[0];
 576
 577     nri              = nlist->nri;
 578     iinr             = nlist->iinr;
 579     jindex           = nlist->jindex;
 580     jjnr             = nlist->jjnr;
 581     shiftidx         = nlist->shift;
 582     gid              = nlist->gid;
 583     shiftvec         = fr->shift_vec[0];
 584     fshift           = fr->fshift[0];
 585     facel            = _mm256_set1_pd(fr->epsfac);
 586     charge           = mdatoms->chargeA;
 587     nvdwtype         = fr->ntype;
 588     vdwparam         = fr->nbfp;
 589     vdwtype          = mdatoms->typeA;
 590
 591     vftab            = kernel_data->table_vdw->data;
 592     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 593
 594     invsqrta         = fr->invsqrta;
 595     dvda             = fr->dvda;
 596     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 597     gbtab            = fr->gbtab.data;
 598     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 599
 600     /* Avoid stupid compiler warnings */
 601     jnrA = jnrB = jnrC = jnrD = 0;
 602     j_coord_offsetA = 0;
 603     j_coord_offsetB = 0;
 604     j_coord_offsetC = 0;
 605     j_coord_offsetD = 0;
 606
 607     outeriter        = 0;
 608     inneriter        = 0;
 609
 610     for(iidx=0;iidx<4*DIM;iidx++)
 611     {
 612         scratch[iidx] = 0.0;
 613     }
 614
 615     /* Start outer loop over neighborlists */
 616     for(iidx=0; iidx<nri; iidx++)
 617     {
 618         /* Load shift vector for this list */
 619         i_shift_offset   = DIM*shiftidx[iidx];
 620
 621         /* Load limits for loop over neighbors */
 622         j_index_start    = jindex[iidx];
 623         j_index_end      = jindex[iidx+1];
 624
 625         /* Get outer coordinate index */
 626         inr              = iinr[iidx];
 627         i_coord_offset   = DIM*inr;
 628
 629         /* Load i particle coords and add shift vector */
 630         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 631
 632         fix0             = _mm256_setzero_pd();
 633         fiy0             = _mm256_setzero_pd();
 634         fiz0             = _mm256_setzero_pd();
 635
 636         /* Load parameters for i particles */
 637         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 638         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 639         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 640
 641         dvdasum          = _mm256_setzero_pd();
 642
 643         /* Start inner kernel loop */
 644         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 645         {
 646
 647             /* Get j neighbor index, and coordinate index */
 648             jnrA             = jjnr[jidx];
 649             jnrB             = jjnr[jidx+1];
 650             jnrC             = jjnr[jidx+2];
 651             jnrD             = jjnr[jidx+3];
 652             j_coord_offsetA  = DIM*jnrA;
 653             j_coord_offsetB  = DIM*jnrB;
 654             j_coord_offsetC  = DIM*jnrC;
 655             j_coord_offsetD  = DIM*jnrD;
 656
 657             /* load j atom coordinates */
 658             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 659                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 660                                                  &jx0,&jy0,&jz0);
 661
 662             /* Calculate displacement vector */
 663             dx00             = _mm256_sub_pd(ix0,jx0);
 664             dy00             = _mm256_sub_pd(iy0,jy0);
 665             dz00             = _mm256_sub_pd(iz0,jz0);
 666
 667             /* Calculate squared distance and things based on it */
 668             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 669
 670             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 671
 672             /* Load parameters for j particles */
 673             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 674                                                                  charge+jnrC+0,charge+jnrD+0);
 675             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 676                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 677             vdwjidx0A        = 2*vdwtype[jnrA+0];
 678             vdwjidx0B        = 2*vdwtype[jnrB+0];
 679             vdwjidx0C        = 2*vdwtype[jnrC+0];
 680             vdwjidx0D        = 2*vdwtype[jnrD+0];
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             r00              = _mm256_mul_pd(rsq00,rinv00);
 687
 688             /* Compute parameters for interactions between i and j atoms */
 689             qq00             = _mm256_mul_pd(iq0,jq0);
 690             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 691                                             vdwioffsetptr0+vdwjidx0B,
 692                                             vdwioffsetptr0+vdwjidx0C,
 693                                             vdwioffsetptr0+vdwjidx0D,
 694                                             &c6_00,&c12_00);
 695
 696             /* Calculate table index by multiplying r with table scale and truncate to integer */
 697             rt               = _mm256_mul_pd(r00,vftabscale);
 698             vfitab           = _mm256_cvttpd_epi32(rt);
 699             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 700             vfitab           = _mm_slli_epi32(vfitab,3);
 701
 702             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 703             isaprod          = _mm256_mul_pd(isai0,isaj0);
 704             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 705             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 706
 707             /* Calculate generalized born table index - this is a separate table from the normal one,
 708              * but we use the same procedure by multiplying r with scale and truncating to integer.
 709              */
 710             rt               = _mm256_mul_pd(r00,gbscale);
 711             gbitab           = _mm256_cvttpd_epi32(rt);
 712             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 713             gbitab           = _mm_slli_epi32(gbitab,2);
 714             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 715             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 716             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 717             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 718             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 719             Heps             = _mm256_mul_pd(gbeps,H);
 720             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 721             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 722             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 723
 724             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 725             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 726             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 727             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 728             fjptrA           = dvda+jnrA;
 729             fjptrB           = dvda+jnrB;
 730             fjptrC           = dvda+jnrC;
 731             fjptrD           = dvda+jnrD;
 732             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 733                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 734             velec            = _mm256_mul_pd(qq00,rinv00);
 735             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 736
 737             /* CUBIC SPLINE TABLE DISPERSION */
 738             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 739             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 740             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 741             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 742             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 743             Heps             = _mm256_mul_pd(vfeps,H);
 744             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 745             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 746             fvdw6            = _mm256_mul_pd(c6_00,FF);
 747
 748             /* CUBIC SPLINE TABLE REPULSION */
 749             vfitab           = _mm_add_epi32(vfitab,ifour);
 750             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 751             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 752             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 753             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 754             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 755             Heps             = _mm256_mul_pd(vfeps,H);
 756             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 757             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 758             fvdw12           = _mm256_mul_pd(c12_00,FF);
 759             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 760
 761             fscal            = _mm256_add_pd(felec,fvdw);
 762
 763             /* Calculate temporary vectorial force */
 764             tx               = _mm256_mul_pd(fscal,dx00);
 765             ty               = _mm256_mul_pd(fscal,dy00);
 766             tz               = _mm256_mul_pd(fscal,dz00);
 767
 768             /* Update vectorial force */
 769             fix0             = _mm256_add_pd(fix0,tx);
 770             fiy0             = _mm256_add_pd(fiy0,ty);
 771             fiz0             = _mm256_add_pd(fiz0,tz);
 772
 773             fjptrA             = f+j_coord_offsetA;
 774             fjptrB             = f+j_coord_offsetB;
 775             fjptrC             = f+j_coord_offsetC;
 776             fjptrD             = f+j_coord_offsetD;
 777             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 778
 779             /* Inner loop uses 81 flops */
 780         }
 781
 782         if(jidx<j_index_end)
 783         {
 784
 785             /* Get j neighbor index, and coordinate index */
 786             jnrlistA         = jjnr[jidx];
 787             jnrlistB         = jjnr[jidx+1];
 788             jnrlistC         = jjnr[jidx+2];
 789             jnrlistD         = jjnr[jidx+3];
 790             /* Sign of each element will be negative for non-real atoms.
 791              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 792              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 793              */
 794             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 795
 796             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 797             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 798             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 799
 800             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 801             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 802             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 803             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 804             j_coord_offsetA  = DIM*jnrA;
 805             j_coord_offsetB  = DIM*jnrB;
 806             j_coord_offsetC  = DIM*jnrC;
 807             j_coord_offsetD  = DIM*jnrD;
 808
 809             /* load j atom coordinates */
 810             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 811                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 812                                                  &jx0,&jy0,&jz0);
 813
 814             /* Calculate displacement vector */
 815             dx00             = _mm256_sub_pd(ix0,jx0);
 816             dy00             = _mm256_sub_pd(iy0,jy0);
 817             dz00             = _mm256_sub_pd(iz0,jz0);
 818
 819             /* Calculate squared distance and things based on it */
 820             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 821
 822             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 823
 824             /* Load parameters for j particles */
 825             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 826                                                                  charge+jnrC+0,charge+jnrD+0);
 827             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 828                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 829             vdwjidx0A        = 2*vdwtype[jnrA+0];
 830             vdwjidx0B        = 2*vdwtype[jnrB+0];
 831             vdwjidx0C        = 2*vdwtype[jnrC+0];
 832             vdwjidx0D        = 2*vdwtype[jnrD+0];
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             r00              = _mm256_mul_pd(rsq00,rinv00);
 839             r00              = _mm256_andnot_pd(dummy_mask,r00);
 840
 841             /* Compute parameters for interactions between i and j atoms */
 842             qq00             = _mm256_mul_pd(iq0,jq0);
 843             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 844                                             vdwioffsetptr0+vdwjidx0B,
 845                                             vdwioffsetptr0+vdwjidx0C,
 846                                             vdwioffsetptr0+vdwjidx0D,
 847                                             &c6_00,&c12_00);
 848
 849             /* Calculate table index by multiplying r with table scale and truncate to integer */
 850             rt               = _mm256_mul_pd(r00,vftabscale);
 851             vfitab           = _mm256_cvttpd_epi32(rt);
 852             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 853             vfitab           = _mm_slli_epi32(vfitab,3);
 854
 855             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 856             isaprod          = _mm256_mul_pd(isai0,isaj0);
 857             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 858             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 859
 860             /* Calculate generalized born table index - this is a separate table from the normal one,
 861              * but we use the same procedure by multiplying r with scale and truncating to integer.
 862              */
 863             rt               = _mm256_mul_pd(r00,gbscale);
 864             gbitab           = _mm256_cvttpd_epi32(rt);
 865             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 866             gbitab           = _mm_slli_epi32(gbitab,2);
 867             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 868             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 869             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 870             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 871             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 872             Heps             = _mm256_mul_pd(gbeps,H);
 873             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 874             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 875             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 876
 877             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 878             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 879             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 880             dvdatmp          = _mm256_andnot_pd(dummy_mask,dvdatmp);
 881             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 882             /* 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. */
 883             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 884             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 885             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 886             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 887             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 888                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 889             velec            = _mm256_mul_pd(qq00,rinv00);
 890             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 891
 892             /* CUBIC SPLINE TABLE DISPERSION */
 893             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 894             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 895             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 896             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 897             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 898             Heps             = _mm256_mul_pd(vfeps,H);
 899             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 900             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 901             fvdw6            = _mm256_mul_pd(c6_00,FF);
 902
 903             /* CUBIC SPLINE TABLE REPULSION */
 904             vfitab           = _mm_add_epi32(vfitab,ifour);
 905             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 906             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 907             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 908             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 909             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 910             Heps             = _mm256_mul_pd(vfeps,H);
 911             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 912             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 913             fvdw12           = _mm256_mul_pd(c12_00,FF);
 914             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 915
 916             fscal            = _mm256_add_pd(felec,fvdw);
 917
 918             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 919
 920             /* Calculate temporary vectorial force */
 921             tx               = _mm256_mul_pd(fscal,dx00);
 922             ty               = _mm256_mul_pd(fscal,dy00);
 923             tz               = _mm256_mul_pd(fscal,dz00);
 924
 925             /* Update vectorial force */
 926             fix0             = _mm256_add_pd(fix0,tx);
 927             fiy0             = _mm256_add_pd(fiy0,ty);
 928             fiz0             = _mm256_add_pd(fiz0,tz);
 929
 930             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 931             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 932             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 933             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 934             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 935
 936             /* Inner loop uses 82 flops */
 937         }
 938
 939         /* End of innermost loop */
 940
 941         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 942                                                  f+i_coord_offset,fshift+i_shift_offset);
 943
 944         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 945         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 946
 947         /* Increment number of inner iterations */
 948         inneriter                  += j_index_end - j_index_start;
 949
 950         /* Outer loop uses 7 flops */
 951     }
 952
 953     /* Increment number of outer iterations */
 954     outeriter        += nri;
 955
 956     /* Update outer/inner flops */
 957
 958     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);
 959 }