src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_256_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_single 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_single.h"
  34 #include "kernelutil_x86_avx_256_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_single
  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_single
  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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
  62     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  63     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  64     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  65     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  66     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  67     real             rcutoff_scalar;
  68     real             *shiftvec,*fshift,*x,*f;
  69     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  70     real             scratch[4*DIM];
  71     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  72     real *           vdwioffsetptr0;
  73     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  75     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  77     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     __m256i          gbitab;
  80     __m128i          gbitab_lo,gbitab_hi;
  81     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  82     __m256           minushalf = _mm256_set1_ps(-0.5);
  83     real             *invsqrta,*dvda,*gbtab;
  84     int              nvdwtype;
  85     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  86     int              *vdwtype;
  87     real             *vdwparam;
  88     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  89     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  90     __m256i          vfitab;
  91     __m128i          vfitab_lo,vfitab_hi;
  92     __m128i          ifour       = _mm_set1_epi32(4);
  93     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  94     real             *vftab;
  95     __m256           dummy_mask,cutoff_mask;
  96     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  97     __m256           one     = _mm256_set1_ps(1.0);
  98     __m256           two     = _mm256_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm256_set1_ps(fr->epsfac);
 111     charge           = mdatoms->chargeA;
 112     nvdwtype         = fr->ntype;
 113     vdwparam         = fr->nbfp;
 114     vdwtype          = mdatoms->typeA;
 115
 116     vftab            = kernel_data->table_vdw->data;
 117     vftabscale       = _mm256_set1_ps(kernel_data->table_vdw->scale);
 118
 119     invsqrta         = fr->invsqrta;
 120     dvda             = fr->dvda;
 121     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 122     gbtab            = fr->gbtab.data;
 123     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 124
 125     /* Avoid stupid compiler warnings */
 126     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 127     j_coord_offsetA = 0;
 128     j_coord_offsetB = 0;
 129     j_coord_offsetC = 0;
 130     j_coord_offsetD = 0;
 131     j_coord_offsetE = 0;
 132     j_coord_offsetF = 0;
 133     j_coord_offsetG = 0;
 134     j_coord_offsetH = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     for(iidx=0;iidx<4*DIM;iidx++)
 140     {
 141         scratch[iidx] = 0.0;
 142     }
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 160
 161         fix0             = _mm256_setzero_ps();
 162         fiy0             = _mm256_setzero_ps();
 163         fiz0             = _mm256_setzero_ps();
 164
 165         /* Load parameters for i particles */
 166         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 167         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 168         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 169
 170         /* Reset potential sums */
 171         velecsum         = _mm256_setzero_ps();
 172         vgbsum           = _mm256_setzero_ps();
 173         vvdwsum          = _mm256_setzero_ps();
 174         dvdasum          = _mm256_setzero_ps();
 175
 176         /* Start inner kernel loop */
 177         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 178         {
 179
 180             /* Get j neighbor index, and coordinate index */
 181             jnrA             = jjnr[jidx];
 182             jnrB             = jjnr[jidx+1];
 183             jnrC             = jjnr[jidx+2];
 184             jnrD             = jjnr[jidx+3];
 185             jnrE             = jjnr[jidx+4];
 186             jnrF             = jjnr[jidx+5];
 187             jnrG             = jjnr[jidx+6];
 188             jnrH             = jjnr[jidx+7];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191             j_coord_offsetC  = DIM*jnrC;
 192             j_coord_offsetD  = DIM*jnrD;
 193             j_coord_offsetE  = DIM*jnrE;
 194             j_coord_offsetF  = DIM*jnrF;
 195             j_coord_offsetG  = DIM*jnrG;
 196             j_coord_offsetH  = DIM*jnrH;
 197
 198             /* load j atom coordinates */
 199             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 200                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 201                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 202                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 203                                                  &jx0,&jy0,&jz0);
 204
 205             /* Calculate displacement vector */
 206             dx00             = _mm256_sub_ps(ix0,jx0);
 207             dy00             = _mm256_sub_ps(iy0,jy0);
 208             dz00             = _mm256_sub_ps(iz0,jz0);
 209
 210             /* Calculate squared distance and things based on it */
 211             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 212
 213             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 214
 215             /* Load parameters for j particles */
 216             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 217                                                                  charge+jnrC+0,charge+jnrD+0,
 218                                                                  charge+jnrE+0,charge+jnrF+0,
 219                                                                  charge+jnrG+0,charge+jnrH+0);
 220             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 221                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 222                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 223                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 224             vdwjidx0A        = 2*vdwtype[jnrA+0];
 225             vdwjidx0B        = 2*vdwtype[jnrB+0];
 226             vdwjidx0C        = 2*vdwtype[jnrC+0];
 227             vdwjidx0D        = 2*vdwtype[jnrD+0];
 228             vdwjidx0E        = 2*vdwtype[jnrE+0];
 229             vdwjidx0F        = 2*vdwtype[jnrF+0];
 230             vdwjidx0G        = 2*vdwtype[jnrG+0];
 231             vdwjidx0H        = 2*vdwtype[jnrH+0];
 232
 233             /**************************
 234              * CALCULATE INTERACTIONS *
 235              **************************/
 236
 237             r00              = _mm256_mul_ps(rsq00,rinv00);
 238
 239             /* Compute parameters for interactions between i and j atoms */
 240             qq00             = _mm256_mul_ps(iq0,jq0);
 241             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 242                                             vdwioffsetptr0+vdwjidx0B,
 243                                             vdwioffsetptr0+vdwjidx0C,
 244                                             vdwioffsetptr0+vdwjidx0D,
 245                                             vdwioffsetptr0+vdwjidx0E,
 246                                             vdwioffsetptr0+vdwjidx0F,
 247                                             vdwioffsetptr0+vdwjidx0G,
 248                                             vdwioffsetptr0+vdwjidx0H,
 249                                             &c6_00,&c12_00);
 250
 251             /* Calculate table index by multiplying r with table scale and truncate to integer */
 252             rt               = _mm256_mul_ps(r00,vftabscale);
 253             vfitab           = _mm256_cvttps_epi32(rt);
 254             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 255             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 256             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 257             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 258             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 259             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 260
 261             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 262             isaprod          = _mm256_mul_ps(isai0,isaj0);
 263             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 264             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 265
 266             /* Calculate generalized born table index - this is a separate table from the normal one,
 267              * but we use the same procedure by multiplying r with scale and truncating to integer.
 268              */
 269             rt               = _mm256_mul_ps(r00,gbscale);
 270             gbitab           = _mm256_cvttps_epi32(rt);
 271             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 272             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 273             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 274             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 275             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 276             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 277             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 278                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 279             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 280                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 281             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 282                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 283             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 284                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 285             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 286             Heps             = _mm256_mul_ps(gbeps,H);
 287             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 288             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 289             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 290
 291             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 292             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 293             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 294             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 295             fjptrA           = dvda+jnrA;
 296             fjptrB           = dvda+jnrB;
 297             fjptrC           = dvda+jnrC;
 298             fjptrD           = dvda+jnrD;
 299             fjptrE           = dvda+jnrE;
 300             fjptrF           = dvda+jnrF;
 301             fjptrG           = dvda+jnrG;
 302             fjptrH           = dvda+jnrH;
 303             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 304                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 305             velec            = _mm256_mul_ps(qq00,rinv00);
 306             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 307
 308             /* CUBIC SPLINE TABLE DISPERSION */
 309             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 310                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 311             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 312                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 313             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 314                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 315             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 316                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 317             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 318             Heps             = _mm256_mul_ps(vfeps,H);
 319             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 320             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 321             vvdw6            = _mm256_mul_ps(c6_00,VV);
 322             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 323             fvdw6            = _mm256_mul_ps(c6_00,FF);
 324
 325             /* CUBIC SPLINE TABLE REPULSION */
 326             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 327             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 328             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 329                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 330             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 331                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 332             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 333                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 334             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 335                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 336             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 337             Heps             = _mm256_mul_ps(vfeps,H);
 338             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 339             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 340             vvdw12           = _mm256_mul_ps(c12_00,VV);
 341             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 342             fvdw12           = _mm256_mul_ps(c12_00,FF);
 343             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 344             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velecsum         = _mm256_add_ps(velecsum,velec);
 348             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 349             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 350
 351             fscal            = _mm256_add_ps(felec,fvdw);
 352
 353             /* Calculate temporary vectorial force */
 354             tx               = _mm256_mul_ps(fscal,dx00);
 355             ty               = _mm256_mul_ps(fscal,dy00);
 356             tz               = _mm256_mul_ps(fscal,dz00);
 357
 358             /* Update vectorial force */
 359             fix0             = _mm256_add_ps(fix0,tx);
 360             fiy0             = _mm256_add_ps(fiy0,ty);
 361             fiz0             = _mm256_add_ps(fiz0,tz);
 362
 363             fjptrA             = f+j_coord_offsetA;
 364             fjptrB             = f+j_coord_offsetB;
 365             fjptrC             = f+j_coord_offsetC;
 366             fjptrD             = f+j_coord_offsetD;
 367             fjptrE             = f+j_coord_offsetE;
 368             fjptrF             = f+j_coord_offsetF;
 369             fjptrG             = f+j_coord_offsetG;
 370             fjptrH             = f+j_coord_offsetH;
 371             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 372
 373             /* Inner loop uses 91 flops */
 374         }
 375
 376         if(jidx<j_index_end)
 377         {
 378
 379             /* Get j neighbor index, and coordinate index */
 380             jnrlistA         = jjnr[jidx];
 381             jnrlistB         = jjnr[jidx+1];
 382             jnrlistC         = jjnr[jidx+2];
 383             jnrlistD         = jjnr[jidx+3];
 384             jnrlistE         = jjnr[jidx+4];
 385             jnrlistF         = jjnr[jidx+5];
 386             jnrlistG         = jjnr[jidx+6];
 387             jnrlistH         = jjnr[jidx+7];
 388             /* Sign of each element will be negative for non-real atoms.
 389              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 390              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 391              */
 392             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 393                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 394
 395             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 396             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 397             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 398             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 399             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 400             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 401             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 402             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 403             j_coord_offsetA  = DIM*jnrA;
 404             j_coord_offsetB  = DIM*jnrB;
 405             j_coord_offsetC  = DIM*jnrC;
 406             j_coord_offsetD  = DIM*jnrD;
 407             j_coord_offsetE  = DIM*jnrE;
 408             j_coord_offsetF  = DIM*jnrF;
 409             j_coord_offsetG  = DIM*jnrG;
 410             j_coord_offsetH  = DIM*jnrH;
 411
 412             /* load j atom coordinates */
 413             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 414                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 415                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 416                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 417                                                  &jx0,&jy0,&jz0);
 418
 419             /* Calculate displacement vector */
 420             dx00             = _mm256_sub_ps(ix0,jx0);
 421             dy00             = _mm256_sub_ps(iy0,jy0);
 422             dz00             = _mm256_sub_ps(iz0,jz0);
 423
 424             /* Calculate squared distance and things based on it */
 425             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 426
 427             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 428
 429             /* Load parameters for j particles */
 430             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 431                                                                  charge+jnrC+0,charge+jnrD+0,
 432                                                                  charge+jnrE+0,charge+jnrF+0,
 433                                                                  charge+jnrG+0,charge+jnrH+0);
 434             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 435                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 436                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 437                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 438             vdwjidx0A        = 2*vdwtype[jnrA+0];
 439             vdwjidx0B        = 2*vdwtype[jnrB+0];
 440             vdwjidx0C        = 2*vdwtype[jnrC+0];
 441             vdwjidx0D        = 2*vdwtype[jnrD+0];
 442             vdwjidx0E        = 2*vdwtype[jnrE+0];
 443             vdwjidx0F        = 2*vdwtype[jnrF+0];
 444             vdwjidx0G        = 2*vdwtype[jnrG+0];
 445             vdwjidx0H        = 2*vdwtype[jnrH+0];
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             r00              = _mm256_mul_ps(rsq00,rinv00);
 452             r00              = _mm256_andnot_ps(dummy_mask,r00);
 453
 454             /* Compute parameters for interactions between i and j atoms */
 455             qq00             = _mm256_mul_ps(iq0,jq0);
 456             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 457                                             vdwioffsetptr0+vdwjidx0B,
 458                                             vdwioffsetptr0+vdwjidx0C,
 459                                             vdwioffsetptr0+vdwjidx0D,
 460                                             vdwioffsetptr0+vdwjidx0E,
 461                                             vdwioffsetptr0+vdwjidx0F,
 462                                             vdwioffsetptr0+vdwjidx0G,
 463                                             vdwioffsetptr0+vdwjidx0H,
 464                                             &c6_00,&c12_00);
 465
 466             /* Calculate table index by multiplying r with table scale and truncate to integer */
 467             rt               = _mm256_mul_ps(r00,vftabscale);
 468             vfitab           = _mm256_cvttps_epi32(rt);
 469             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 470             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 471             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 472             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 473             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 474             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 475
 476             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 477             isaprod          = _mm256_mul_ps(isai0,isaj0);
 478             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 479             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 480
 481             /* Calculate generalized born table index - this is a separate table from the normal one,
 482              * but we use the same procedure by multiplying r with scale and truncating to integer.
 483              */
 484             rt               = _mm256_mul_ps(r00,gbscale);
 485             gbitab           = _mm256_cvttps_epi32(rt);
 486             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 487             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 488             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 489             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 490             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 491             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 492             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 493                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 494             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 495                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 496             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 497                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 498             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 499                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 500             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 501             Heps             = _mm256_mul_ps(gbeps,H);
 502             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 503             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 504             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 505
 506             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 507             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 508             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 509             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 510             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 511             /* 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. */
 512             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 513             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 514             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 515             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 516             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 517             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 518             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 519             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 520             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 521                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 522             velec            = _mm256_mul_ps(qq00,rinv00);
 523             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 524
 525             /* CUBIC SPLINE TABLE DISPERSION */
 526             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 527                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 528             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 529                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 530             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 531                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 532             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 533                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 534             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 535             Heps             = _mm256_mul_ps(vfeps,H);
 536             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 537             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 538             vvdw6            = _mm256_mul_ps(c6_00,VV);
 539             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 540             fvdw6            = _mm256_mul_ps(c6_00,FF);
 541
 542             /* CUBIC SPLINE TABLE REPULSION */
 543             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 544             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 545             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 546                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 547             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 548                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 549             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 550                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 551             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 552                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 553             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 554             Heps             = _mm256_mul_ps(vfeps,H);
 555             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 556             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 557             vvdw12           = _mm256_mul_ps(c12_00,VV);
 558             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 559             fvdw12           = _mm256_mul_ps(c12_00,FF);
 560             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 561             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 562
 563             /* Update potential sum for this i atom from the interaction with this j atom. */
 564             velec            = _mm256_andnot_ps(dummy_mask,velec);
 565             velecsum         = _mm256_add_ps(velecsum,velec);
 566             vgb              = _mm256_andnot_ps(dummy_mask,vgb);
 567             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 568             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 569             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 570
 571             fscal            = _mm256_add_ps(felec,fvdw);
 572
 573             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 574
 575             /* Calculate temporary vectorial force */
 576             tx               = _mm256_mul_ps(fscal,dx00);
 577             ty               = _mm256_mul_ps(fscal,dy00);
 578             tz               = _mm256_mul_ps(fscal,dz00);
 579
 580             /* Update vectorial force */
 581             fix0             = _mm256_add_ps(fix0,tx);
 582             fiy0             = _mm256_add_ps(fiy0,ty);
 583             fiz0             = _mm256_add_ps(fiz0,tz);
 584
 585             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 586             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 587             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 588             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 589             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 590             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 591             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 592             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 593             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 594
 595             /* Inner loop uses 92 flops */
 596         }
 597
 598         /* End of innermost loop */
 599
 600         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 601                                                  f+i_coord_offset,fshift+i_shift_offset);
 602
 603         ggid                        = gid[iidx];
 604         /* Update potential energies */
 605         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 606         gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 607         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 608         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
 609         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
 610
 611         /* Increment number of inner iterations */
 612         inneriter                  += j_index_end - j_index_start;
 613
 614         /* Outer loop uses 10 flops */
 615     }
 616
 617     /* Increment number of outer iterations */
 618     outeriter        += nri;
 619
 620     /* Update outer/inner flops */
 621
 622     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
 623 }
 624 /*
 625  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_single
 626  * Electrostatics interaction: GeneralizedBorn
 627  * VdW interaction:            CubicSplineTable
 628  * Geometry:                   Particle-Particle
 629  * Calculate force/pot:        Force
 630  */
 631 void
 632 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_single
 633                     (t_nblist * gmx_restrict                nlist,
 634                      rvec * gmx_restrict                    xx,
 635                      rvec * gmx_restrict                    ff,
 636                      t_forcerec * gmx_restrict              fr,
 637                      t_mdatoms * gmx_restrict               mdatoms,
 638                      nb_kernel_data_t * gmx_restrict        kernel_data,
 639                      t_nrnb * gmx_restrict                  nrnb)
 640 {
 641     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 642      * just 0 for non-waters.
 643      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 644      * jnr indices corresponding to data put in the four positions in the SIMD register.
 645      */
 646     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 647     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 648     int              jnrA,jnrB,jnrC,jnrD;
 649     int              jnrE,jnrF,jnrG,jnrH;
 650     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 651     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 652     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 653     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 654     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 655     real             rcutoff_scalar;
 656     real             *shiftvec,*fshift,*x,*f;
 657     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 658     real             scratch[4*DIM];
 659     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 660     real *           vdwioffsetptr0;
 661     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 662     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 663     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 664     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 665     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 666     real             *charge;
 667     __m256i          gbitab;
 668     __m128i          gbitab_lo,gbitab_hi;
 669     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 670     __m256           minushalf = _mm256_set1_ps(-0.5);
 671     real             *invsqrta,*dvda,*gbtab;
 672     int              nvdwtype;
 673     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 674     int              *vdwtype;
 675     real             *vdwparam;
 676     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 677     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 678     __m256i          vfitab;
 679     __m128i          vfitab_lo,vfitab_hi;
 680     __m128i          ifour       = _mm_set1_epi32(4);
 681     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 682     real             *vftab;
 683     __m256           dummy_mask,cutoff_mask;
 684     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 685     __m256           one     = _mm256_set1_ps(1.0);
 686     __m256           two     = _mm256_set1_ps(2.0);
 687     x                = xx[0];
 688     f                = ff[0];
 689
 690     nri              = nlist->nri;
 691     iinr             = nlist->iinr;
 692     jindex           = nlist->jindex;
 693     jjnr             = nlist->jjnr;
 694     shiftidx         = nlist->shift;
 695     gid              = nlist->gid;
 696     shiftvec         = fr->shift_vec[0];
 697     fshift           = fr->fshift[0];
 698     facel            = _mm256_set1_ps(fr->epsfac);
 699     charge           = mdatoms->chargeA;
 700     nvdwtype         = fr->ntype;
 701     vdwparam         = fr->nbfp;
 702     vdwtype          = mdatoms->typeA;
 703
 704     vftab            = kernel_data->table_vdw->data;
 705     vftabscale       = _mm256_set1_ps(kernel_data->table_vdw->scale);
 706
 707     invsqrta         = fr->invsqrta;
 708     dvda             = fr->dvda;
 709     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 710     gbtab            = fr->gbtab.data;
 711     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 712
 713     /* Avoid stupid compiler warnings */
 714     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 715     j_coord_offsetA = 0;
 716     j_coord_offsetB = 0;
 717     j_coord_offsetC = 0;
 718     j_coord_offsetD = 0;
 719     j_coord_offsetE = 0;
 720     j_coord_offsetF = 0;
 721     j_coord_offsetG = 0;
 722     j_coord_offsetH = 0;
 723
 724     outeriter        = 0;
 725     inneriter        = 0;
 726
 727     for(iidx=0;iidx<4*DIM;iidx++)
 728     {
 729         scratch[iidx] = 0.0;
 730     }
 731
 732     /* Start outer loop over neighborlists */
 733     for(iidx=0; iidx<nri; iidx++)
 734     {
 735         /* Load shift vector for this list */
 736         i_shift_offset   = DIM*shiftidx[iidx];
 737
 738         /* Load limits for loop over neighbors */
 739         j_index_start    = jindex[iidx];
 740         j_index_end      = jindex[iidx+1];
 741
 742         /* Get outer coordinate index */
 743         inr              = iinr[iidx];
 744         i_coord_offset   = DIM*inr;
 745
 746         /* Load i particle coords and add shift vector */
 747         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 748
 749         fix0             = _mm256_setzero_ps();
 750         fiy0             = _mm256_setzero_ps();
 751         fiz0             = _mm256_setzero_ps();
 752
 753         /* Load parameters for i particles */
 754         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 755         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 756         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 757
 758         dvdasum          = _mm256_setzero_ps();
 759
 760         /* Start inner kernel loop */
 761         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 762         {
 763
 764             /* Get j neighbor index, and coordinate index */
 765             jnrA             = jjnr[jidx];
 766             jnrB             = jjnr[jidx+1];
 767             jnrC             = jjnr[jidx+2];
 768             jnrD             = jjnr[jidx+3];
 769             jnrE             = jjnr[jidx+4];
 770             jnrF             = jjnr[jidx+5];
 771             jnrG             = jjnr[jidx+6];
 772             jnrH             = jjnr[jidx+7];
 773             j_coord_offsetA  = DIM*jnrA;
 774             j_coord_offsetB  = DIM*jnrB;
 775             j_coord_offsetC  = DIM*jnrC;
 776             j_coord_offsetD  = DIM*jnrD;
 777             j_coord_offsetE  = DIM*jnrE;
 778             j_coord_offsetF  = DIM*jnrF;
 779             j_coord_offsetG  = DIM*jnrG;
 780             j_coord_offsetH  = DIM*jnrH;
 781
 782             /* load j atom coordinates */
 783             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 784                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 785                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 786                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 787                                                  &jx0,&jy0,&jz0);
 788
 789             /* Calculate displacement vector */
 790             dx00             = _mm256_sub_ps(ix0,jx0);
 791             dy00             = _mm256_sub_ps(iy0,jy0);
 792             dz00             = _mm256_sub_ps(iz0,jz0);
 793
 794             /* Calculate squared distance and things based on it */
 795             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 796
 797             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 798
 799             /* Load parameters for j particles */
 800             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 801                                                                  charge+jnrC+0,charge+jnrD+0,
 802                                                                  charge+jnrE+0,charge+jnrF+0,
 803                                                                  charge+jnrG+0,charge+jnrH+0);
 804             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 805                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 806                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 807                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 808             vdwjidx0A        = 2*vdwtype[jnrA+0];
 809             vdwjidx0B        = 2*vdwtype[jnrB+0];
 810             vdwjidx0C        = 2*vdwtype[jnrC+0];
 811             vdwjidx0D        = 2*vdwtype[jnrD+0];
 812             vdwjidx0E        = 2*vdwtype[jnrE+0];
 813             vdwjidx0F        = 2*vdwtype[jnrF+0];
 814             vdwjidx0G        = 2*vdwtype[jnrG+0];
 815             vdwjidx0H        = 2*vdwtype[jnrH+0];
 816
 817             /**************************
 818              * CALCULATE INTERACTIONS *
 819              **************************/
 820
 821             r00              = _mm256_mul_ps(rsq00,rinv00);
 822
 823             /* Compute parameters for interactions between i and j atoms */
 824             qq00             = _mm256_mul_ps(iq0,jq0);
 825             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 826                                             vdwioffsetptr0+vdwjidx0B,
 827                                             vdwioffsetptr0+vdwjidx0C,
 828                                             vdwioffsetptr0+vdwjidx0D,
 829                                             vdwioffsetptr0+vdwjidx0E,
 830                                             vdwioffsetptr0+vdwjidx0F,
 831                                             vdwioffsetptr0+vdwjidx0G,
 832                                             vdwioffsetptr0+vdwjidx0H,
 833                                             &c6_00,&c12_00);
 834
 835             /* Calculate table index by multiplying r with table scale and truncate to integer */
 836             rt               = _mm256_mul_ps(r00,vftabscale);
 837             vfitab           = _mm256_cvttps_epi32(rt);
 838             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 839             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 840             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 841             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 842             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 843             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 844
 845             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 846             isaprod          = _mm256_mul_ps(isai0,isaj0);
 847             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 848             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 849
 850             /* Calculate generalized born table index - this is a separate table from the normal one,
 851              * but we use the same procedure by multiplying r with scale and truncating to integer.
 852              */
 853             rt               = _mm256_mul_ps(r00,gbscale);
 854             gbitab           = _mm256_cvttps_epi32(rt);
 855             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 856             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 857             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 858             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 859             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 860             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 861             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 862                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 863             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 864                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 865             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 866                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 867             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 868                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 869             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 870             Heps             = _mm256_mul_ps(gbeps,H);
 871             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 872             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 873             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 874
 875             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 876             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 877             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 878             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 879             fjptrA           = dvda+jnrA;
 880             fjptrB           = dvda+jnrB;
 881             fjptrC           = dvda+jnrC;
 882             fjptrD           = dvda+jnrD;
 883             fjptrE           = dvda+jnrE;
 884             fjptrF           = dvda+jnrF;
 885             fjptrG           = dvda+jnrG;
 886             fjptrH           = dvda+jnrH;
 887             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 888                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 889             velec            = _mm256_mul_ps(qq00,rinv00);
 890             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 891
 892             /* CUBIC SPLINE TABLE DISPERSION */
 893             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 894                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 895             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 896                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 897             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 898                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 899             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 900                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 901             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 902             Heps             = _mm256_mul_ps(vfeps,H);
 903             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 904             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 905             fvdw6            = _mm256_mul_ps(c6_00,FF);
 906
 907             /* CUBIC SPLINE TABLE REPULSION */
 908             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 909             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 910             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 911                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 912             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 913                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 914             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 915                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 916             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 917                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 918             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 919             Heps             = _mm256_mul_ps(vfeps,H);
 920             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 921             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 922             fvdw12           = _mm256_mul_ps(c12_00,FF);
 923             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 924
 925             fscal            = _mm256_add_ps(felec,fvdw);
 926
 927             /* Calculate temporary vectorial force */
 928             tx               = _mm256_mul_ps(fscal,dx00);
 929             ty               = _mm256_mul_ps(fscal,dy00);
 930             tz               = _mm256_mul_ps(fscal,dz00);
 931
 932             /* Update vectorial force */
 933             fix0             = _mm256_add_ps(fix0,tx);
 934             fiy0             = _mm256_add_ps(fiy0,ty);
 935             fiz0             = _mm256_add_ps(fiz0,tz);
 936
 937             fjptrA             = f+j_coord_offsetA;
 938             fjptrB             = f+j_coord_offsetB;
 939             fjptrC             = f+j_coord_offsetC;
 940             fjptrD             = f+j_coord_offsetD;
 941             fjptrE             = f+j_coord_offsetE;
 942             fjptrF             = f+j_coord_offsetF;
 943             fjptrG             = f+j_coord_offsetG;
 944             fjptrH             = f+j_coord_offsetH;
 945             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 946
 947             /* Inner loop uses 81 flops */
 948         }
 949
 950         if(jidx<j_index_end)
 951         {
 952
 953             /* Get j neighbor index, and coordinate index */
 954             jnrlistA         = jjnr[jidx];
 955             jnrlistB         = jjnr[jidx+1];
 956             jnrlistC         = jjnr[jidx+2];
 957             jnrlistD         = jjnr[jidx+3];
 958             jnrlistE         = jjnr[jidx+4];
 959             jnrlistF         = jjnr[jidx+5];
 960             jnrlistG         = jjnr[jidx+6];
 961             jnrlistH         = jjnr[jidx+7];
 962             /* Sign of each element will be negative for non-real atoms.
 963              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 964              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 965              */
 966             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 967                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 968
 969             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 970             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 971             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 972             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 973             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 974             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 975             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 976             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 977             j_coord_offsetA  = DIM*jnrA;
 978             j_coord_offsetB  = DIM*jnrB;
 979             j_coord_offsetC  = DIM*jnrC;
 980             j_coord_offsetD  = DIM*jnrD;
 981             j_coord_offsetE  = DIM*jnrE;
 982             j_coord_offsetF  = DIM*jnrF;
 983             j_coord_offsetG  = DIM*jnrG;
 984             j_coord_offsetH  = DIM*jnrH;
 985
 986             /* load j atom coordinates */
 987             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 988                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 989                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 990                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 991                                                  &jx0,&jy0,&jz0);
 992
 993             /* Calculate displacement vector */
 994             dx00             = _mm256_sub_ps(ix0,jx0);
 995             dy00             = _mm256_sub_ps(iy0,jy0);
 996             dz00             = _mm256_sub_ps(iz0,jz0);
 997
 998             /* Calculate squared distance and things based on it */
 999             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
1000
1001             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
1002
1003             /* Load parameters for j particles */
1004             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1005                                                                  charge+jnrC+0,charge+jnrD+0,
1006                                                                  charge+jnrE+0,charge+jnrF+0,
1007                                                                  charge+jnrG+0,charge+jnrH+0);
1008             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
1009                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
1010                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
1011                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
1012             vdwjidx0A        = 2*vdwtype[jnrA+0];
1013             vdwjidx0B        = 2*vdwtype[jnrB+0];
1014             vdwjidx0C        = 2*vdwtype[jnrC+0];
1015             vdwjidx0D        = 2*vdwtype[jnrD+0];
1016             vdwjidx0E        = 2*vdwtype[jnrE+0];
1017             vdwjidx0F        = 2*vdwtype[jnrF+0];
1018             vdwjidx0G        = 2*vdwtype[jnrG+0];
1019             vdwjidx0H        = 2*vdwtype[jnrH+0];
1020
1021             /**************************
1022              * CALCULATE INTERACTIONS *
1023              **************************/
1024
1025             r00              = _mm256_mul_ps(rsq00,rinv00);
1026             r00              = _mm256_andnot_ps(dummy_mask,r00);
1027
1028             /* Compute parameters for interactions between i and j atoms */
1029             qq00             = _mm256_mul_ps(iq0,jq0);
1030             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
1031                                             vdwioffsetptr0+vdwjidx0B,
1032                                             vdwioffsetptr0+vdwjidx0C,
1033                                             vdwioffsetptr0+vdwjidx0D,
1034                                             vdwioffsetptr0+vdwjidx0E,
1035                                             vdwioffsetptr0+vdwjidx0F,
1036                                             vdwioffsetptr0+vdwjidx0G,
1037                                             vdwioffsetptr0+vdwjidx0H,
1038                                             &c6_00,&c12_00);
1039
1040             /* Calculate table index by multiplying r with table scale and truncate to integer */
1041             rt               = _mm256_mul_ps(r00,vftabscale);
1042             vfitab           = _mm256_cvttps_epi32(rt);
1043             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1044             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1045             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
1046             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
1047             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
1048             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
1049
1050             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
1051             isaprod          = _mm256_mul_ps(isai0,isaj0);
1052             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
1053             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
1054
1055             /* Calculate generalized born table index - this is a separate table from the normal one,
1056              * but we use the same procedure by multiplying r with scale and truncating to integer.
1057              */
1058             rt               = _mm256_mul_ps(r00,gbscale);
1059             gbitab           = _mm256_cvttps_epi32(rt);
1060             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1061             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1062             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
1063             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
1064             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
1065             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
1066             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
1067                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
1068             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
1069                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
1070             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
1071                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
1072             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
1073                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
1074             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1075             Heps             = _mm256_mul_ps(gbeps,H);
1076             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
1077             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
1078             vgb              = _mm256_mul_ps(gbqqfactor,VV);
1079
1080             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1081             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
1082             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
1083             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
1084             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
1085             /* 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. */
1086             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
1087             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
1088             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
1089             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
1090             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
1091             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
1092             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
1093             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
1094             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
1095                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
1096             velec            = _mm256_mul_ps(qq00,rinv00);
1097             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
1098
1099             /* CUBIC SPLINE TABLE DISPERSION */
1100             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1101                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1102             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1103                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1104             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1105                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1106             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1107                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1108             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1109             Heps             = _mm256_mul_ps(vfeps,H);
1110             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1111             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1112             fvdw6            = _mm256_mul_ps(c6_00,FF);
1113
1114             /* CUBIC SPLINE TABLE REPULSION */
1115             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
1116             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
1117             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1118                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1119             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1120                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1121             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1122                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1123             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1124                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1125             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1126             Heps             = _mm256_mul_ps(vfeps,H);
1127             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1128             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1129             fvdw12           = _mm256_mul_ps(c12_00,FF);
1130             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
1131
1132             fscal            = _mm256_add_ps(felec,fvdw);
1133
1134             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
1135
1136             /* Calculate temporary vectorial force */
1137             tx               = _mm256_mul_ps(fscal,dx00);
1138             ty               = _mm256_mul_ps(fscal,dy00);
1139             tz               = _mm256_mul_ps(fscal,dz00);
1140
1141             /* Update vectorial force */
1142             fix0             = _mm256_add_ps(fix0,tx);
1143             fiy0             = _mm256_add_ps(fiy0,ty);
1144             fiz0             = _mm256_add_ps(fiz0,tz);
1145
1146             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1147             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1148             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1149             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1150             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
1151             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
1152             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
1153             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
1154             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
1155
1156             /* Inner loop uses 82 flops */
1157         }
1158
1159         /* End of innermost loop */
1160
1161         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1162                                                  f+i_coord_offset,fshift+i_shift_offset);
1163
1164         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
1165         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
1166
1167         /* Increment number of inner iterations */
1168         inneriter                  += j_index_end - j_index_start;
1169
1170         /* Outer loop uses 7 flops */
1171     }
1172
1173     /* Increment number of outer iterations */
1174     outeriter        += nri;
1175
1176     /* Update outer/inner flops */
1177
1178     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);
1179 }