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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_single
  51  * Electrostatics interaction: GeneralizedBorn
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_256_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrE,jnrF,jnrG,jnrH;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  83     real             scratch[4*DIM];
  84     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     real *           vdwioffsetptr0;
  86     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  88     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m256i          gbitab;
  93     __m128i          gbitab_lo,gbitab_hi;
  94     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  95     __m256           minushalf = _mm256_set1_ps(-0.5);
  96     real             *invsqrta,*dvda,*gbtab;
  97     int              nvdwtype;
  98     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 102     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 103     __m256i          vfitab;
 104     __m128i          vfitab_lo,vfitab_hi;
 105     __m128i          ifour       = _mm_set1_epi32(4);
 106     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 107     real             *vftab;
 108     __m256           dummy_mask,cutoff_mask;
 109     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 110     __m256           one     = _mm256_set1_ps(1.0);
 111     __m256           two     = _mm256_set1_ps(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm256_set1_ps(fr->ic->epsfac);
 124     charge           = mdatoms->chargeA;
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     vftab            = kernel_data->table_vdw->data;
 130     vftabscale       = _mm256_set1_ps(kernel_data->table_vdw->scale);
 131
 132     invsqrta         = fr->invsqrta;
 133     dvda             = fr->dvda;
 134     gbtabscale       = _mm256_set1_ps(fr->gbtab->scale);
 135     gbtab            = fr->gbtab->data;
 136     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 137
 138     /* Avoid stupid compiler warnings */
 139     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 140     j_coord_offsetA = 0;
 141     j_coord_offsetB = 0;
 142     j_coord_offsetC = 0;
 143     j_coord_offsetD = 0;
 144     j_coord_offsetE = 0;
 145     j_coord_offsetF = 0;
 146     j_coord_offsetG = 0;
 147     j_coord_offsetH = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     for(iidx=0;iidx<4*DIM;iidx++)
 153     {
 154         scratch[iidx] = 0.0;
 155     }
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 173
 174         fix0             = _mm256_setzero_ps();
 175         fiy0             = _mm256_setzero_ps();
 176         fiz0             = _mm256_setzero_ps();
 177
 178         /* Load parameters for i particles */
 179         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 180         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 181         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 182
 183         /* Reset potential sums */
 184         velecsum         = _mm256_setzero_ps();
 185         vgbsum           = _mm256_setzero_ps();
 186         vvdwsum          = _mm256_setzero_ps();
 187         dvdasum          = _mm256_setzero_ps();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             jnrC             = jjnr[jidx+2];
 197             jnrD             = jjnr[jidx+3];
 198             jnrE             = jjnr[jidx+4];
 199             jnrF             = jjnr[jidx+5];
 200             jnrG             = jjnr[jidx+6];
 201             jnrH             = jjnr[jidx+7];
 202             j_coord_offsetA  = DIM*jnrA;
 203             j_coord_offsetB  = DIM*jnrB;
 204             j_coord_offsetC  = DIM*jnrC;
 205             j_coord_offsetD  = DIM*jnrD;
 206             j_coord_offsetE  = DIM*jnrE;
 207             j_coord_offsetF  = DIM*jnrF;
 208             j_coord_offsetG  = DIM*jnrG;
 209             j_coord_offsetH  = DIM*jnrH;
 210
 211             /* load j atom coordinates */
 212             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 213                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 214                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 215                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 216                                                  &jx0,&jy0,&jz0);
 217
 218             /* Calculate displacement vector */
 219             dx00             = _mm256_sub_ps(ix0,jx0);
 220             dy00             = _mm256_sub_ps(iy0,jy0);
 221             dz00             = _mm256_sub_ps(iz0,jz0);
 222
 223             /* Calculate squared distance and things based on it */
 224             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 225
 226             rinv00           = avx256_invsqrt_f(rsq00);
 227
 228             /* Load parameters for j particles */
 229             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 230                                                                  charge+jnrC+0,charge+jnrD+0,
 231                                                                  charge+jnrE+0,charge+jnrF+0,
 232                                                                  charge+jnrG+0,charge+jnrH+0);
 233             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 234                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 235                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 236                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 237             vdwjidx0A        = 2*vdwtype[jnrA+0];
 238             vdwjidx0B        = 2*vdwtype[jnrB+0];
 239             vdwjidx0C        = 2*vdwtype[jnrC+0];
 240             vdwjidx0D        = 2*vdwtype[jnrD+0];
 241             vdwjidx0E        = 2*vdwtype[jnrE+0];
 242             vdwjidx0F        = 2*vdwtype[jnrF+0];
 243             vdwjidx0G        = 2*vdwtype[jnrG+0];
 244             vdwjidx0H        = 2*vdwtype[jnrH+0];
 245
 246             /**************************
 247              * CALCULATE INTERACTIONS *
 248              **************************/
 249
 250             r00              = _mm256_mul_ps(rsq00,rinv00);
 251
 252             /* Compute parameters for interactions between i and j atoms */
 253             qq00             = _mm256_mul_ps(iq0,jq0);
 254             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 255                                             vdwioffsetptr0+vdwjidx0B,
 256                                             vdwioffsetptr0+vdwjidx0C,
 257                                             vdwioffsetptr0+vdwjidx0D,
 258                                             vdwioffsetptr0+vdwjidx0E,
 259                                             vdwioffsetptr0+vdwjidx0F,
 260                                             vdwioffsetptr0+vdwjidx0G,
 261                                             vdwioffsetptr0+vdwjidx0H,
 262                                             &c6_00,&c12_00);
 263
 264             /* Calculate table index by multiplying r with table scale and truncate to integer */
 265             rt               = _mm256_mul_ps(r00,vftabscale);
 266             vfitab           = _mm256_cvttps_epi32(rt);
 267             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 268             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 269             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 270             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 271             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 272             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 273
 274             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 275             isaprod          = _mm256_mul_ps(isai0,isaj0);
 276             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 277             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 278
 279             /* Calculate generalized born table index - this is a separate table from the normal one,
 280              * but we use the same procedure by multiplying r with scale and truncating to integer.
 281              */
 282             rt               = _mm256_mul_ps(r00,gbscale);
 283             gbitab           = _mm256_cvttps_epi32(rt);
 284             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 285             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 286             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 287             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 288             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 289             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 290             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 291                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 292             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 293                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 294             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 295                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 296             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 297                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 298             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 299             Heps             = _mm256_mul_ps(gbeps,H);
 300             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 301             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 302             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 303
 304             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 305             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 306             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 307             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 308             fjptrA           = dvda+jnrA;
 309             fjptrB           = dvda+jnrB;
 310             fjptrC           = dvda+jnrC;
 311             fjptrD           = dvda+jnrD;
 312             fjptrE           = dvda+jnrE;
 313             fjptrF           = dvda+jnrF;
 314             fjptrG           = dvda+jnrG;
 315             fjptrH           = dvda+jnrH;
 316             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 317                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 318             velec            = _mm256_mul_ps(qq00,rinv00);
 319             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 320
 321             /* CUBIC SPLINE TABLE DISPERSION */
 322             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 323                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 324             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 325                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 326             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 327                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 328             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 329                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 330             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 331             Heps             = _mm256_mul_ps(vfeps,H);
 332             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 333             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 334             vvdw6            = _mm256_mul_ps(c6_00,VV);
 335             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 336             fvdw6            = _mm256_mul_ps(c6_00,FF);
 337
 338             /* CUBIC SPLINE TABLE REPULSION */
 339             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 340             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 341             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 342                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 343             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 344                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 345             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 346                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 347             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 348                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 349             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 350             Heps             = _mm256_mul_ps(vfeps,H);
 351             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 352             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 353             vvdw12           = _mm256_mul_ps(c12_00,VV);
 354             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 355             fvdw12           = _mm256_mul_ps(c12_00,FF);
 356             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 357             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velecsum         = _mm256_add_ps(velecsum,velec);
 361             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 362             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 363
 364             fscal            = _mm256_add_ps(felec,fvdw);
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = _mm256_mul_ps(fscal,dx00);
 368             ty               = _mm256_mul_ps(fscal,dy00);
 369             tz               = _mm256_mul_ps(fscal,dz00);
 370
 371             /* Update vectorial force */
 372             fix0             = _mm256_add_ps(fix0,tx);
 373             fiy0             = _mm256_add_ps(fiy0,ty);
 374             fiz0             = _mm256_add_ps(fiz0,tz);
 375
 376             fjptrA             = f+j_coord_offsetA;
 377             fjptrB             = f+j_coord_offsetB;
 378             fjptrC             = f+j_coord_offsetC;
 379             fjptrD             = f+j_coord_offsetD;
 380             fjptrE             = f+j_coord_offsetE;
 381             fjptrF             = f+j_coord_offsetF;
 382             fjptrG             = f+j_coord_offsetG;
 383             fjptrH             = f+j_coord_offsetH;
 384             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 385
 386             /* Inner loop uses 91 flops */
 387         }
 388
 389         if(jidx<j_index_end)
 390         {
 391
 392             /* Get j neighbor index, and coordinate index */
 393             jnrlistA         = jjnr[jidx];
 394             jnrlistB         = jjnr[jidx+1];
 395             jnrlistC         = jjnr[jidx+2];
 396             jnrlistD         = jjnr[jidx+3];
 397             jnrlistE         = jjnr[jidx+4];
 398             jnrlistF         = jjnr[jidx+5];
 399             jnrlistG         = jjnr[jidx+6];
 400             jnrlistH         = jjnr[jidx+7];
 401             /* Sign of each element will be negative for non-real atoms.
 402              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 403              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 404              */
 405             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 406                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 407
 408             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 409             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 410             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 411             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 412             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 413             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 414             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 415             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 416             j_coord_offsetA  = DIM*jnrA;
 417             j_coord_offsetB  = DIM*jnrB;
 418             j_coord_offsetC  = DIM*jnrC;
 419             j_coord_offsetD  = DIM*jnrD;
 420             j_coord_offsetE  = DIM*jnrE;
 421             j_coord_offsetF  = DIM*jnrF;
 422             j_coord_offsetG  = DIM*jnrG;
 423             j_coord_offsetH  = DIM*jnrH;
 424
 425             /* load j atom coordinates */
 426             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 427                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 428                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 429                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 430                                                  &jx0,&jy0,&jz0);
 431
 432             /* Calculate displacement vector */
 433             dx00             = _mm256_sub_ps(ix0,jx0);
 434             dy00             = _mm256_sub_ps(iy0,jy0);
 435             dz00             = _mm256_sub_ps(iz0,jz0);
 436
 437             /* Calculate squared distance and things based on it */
 438             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 439
 440             rinv00           = avx256_invsqrt_f(rsq00);
 441
 442             /* Load parameters for j particles */
 443             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 444                                                                  charge+jnrC+0,charge+jnrD+0,
 445                                                                  charge+jnrE+0,charge+jnrF+0,
 446                                                                  charge+jnrG+0,charge+jnrH+0);
 447             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 448                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 449                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 450                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 451             vdwjidx0A        = 2*vdwtype[jnrA+0];
 452             vdwjidx0B        = 2*vdwtype[jnrB+0];
 453             vdwjidx0C        = 2*vdwtype[jnrC+0];
 454             vdwjidx0D        = 2*vdwtype[jnrD+0];
 455             vdwjidx0E        = 2*vdwtype[jnrE+0];
 456             vdwjidx0F        = 2*vdwtype[jnrF+0];
 457             vdwjidx0G        = 2*vdwtype[jnrG+0];
 458             vdwjidx0H        = 2*vdwtype[jnrH+0];
 459
 460             /**************************
 461              * CALCULATE INTERACTIONS *
 462              **************************/
 463
 464             r00              = _mm256_mul_ps(rsq00,rinv00);
 465             r00              = _mm256_andnot_ps(dummy_mask,r00);
 466
 467             /* Compute parameters for interactions between i and j atoms */
 468             qq00             = _mm256_mul_ps(iq0,jq0);
 469             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 470                                             vdwioffsetptr0+vdwjidx0B,
 471                                             vdwioffsetptr0+vdwjidx0C,
 472                                             vdwioffsetptr0+vdwjidx0D,
 473                                             vdwioffsetptr0+vdwjidx0E,
 474                                             vdwioffsetptr0+vdwjidx0F,
 475                                             vdwioffsetptr0+vdwjidx0G,
 476                                             vdwioffsetptr0+vdwjidx0H,
 477                                             &c6_00,&c12_00);
 478
 479             /* Calculate table index by multiplying r with table scale and truncate to integer */
 480             rt               = _mm256_mul_ps(r00,vftabscale);
 481             vfitab           = _mm256_cvttps_epi32(rt);
 482             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 483             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 484             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 485             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 486             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 487             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 488
 489             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 490             isaprod          = _mm256_mul_ps(isai0,isaj0);
 491             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 492             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 493
 494             /* Calculate generalized born table index - this is a separate table from the normal one,
 495              * but we use the same procedure by multiplying r with scale and truncating to integer.
 496              */
 497             rt               = _mm256_mul_ps(r00,gbscale);
 498             gbitab           = _mm256_cvttps_epi32(rt);
 499             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 500             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 501             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 502             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 503             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 504             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 505             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 506                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 507             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 508                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 509             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 510                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 511             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 512                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 513             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 514             Heps             = _mm256_mul_ps(gbeps,H);
 515             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 516             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 517             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 518
 519             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 520             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 521             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 522             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 523             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 524             /* 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. */
 525             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 526             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 527             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 528             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 529             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 530             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 531             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 532             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 533             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 534                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 535             velec            = _mm256_mul_ps(qq00,rinv00);
 536             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 537
 538             /* CUBIC SPLINE TABLE DISPERSION */
 539             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 540                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 541             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 542                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 543             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 544                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 545             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 546                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 547             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 548             Heps             = _mm256_mul_ps(vfeps,H);
 549             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 550             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 551             vvdw6            = _mm256_mul_ps(c6_00,VV);
 552             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 553             fvdw6            = _mm256_mul_ps(c6_00,FF);
 554
 555             /* CUBIC SPLINE TABLE REPULSION */
 556             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 557             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 558             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 559                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 560             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 561                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 562             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 563                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 564             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 565                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 566             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 567             Heps             = _mm256_mul_ps(vfeps,H);
 568             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 569             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 570             vvdw12           = _mm256_mul_ps(c12_00,VV);
 571             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 572             fvdw12           = _mm256_mul_ps(c12_00,FF);
 573             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 574             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 575
 576             /* Update potential sum for this i atom from the interaction with this j atom. */
 577             velec            = _mm256_andnot_ps(dummy_mask,velec);
 578             velecsum         = _mm256_add_ps(velecsum,velec);
 579             vgb              = _mm256_andnot_ps(dummy_mask,vgb);
 580             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 581             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 582             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 583
 584             fscal            = _mm256_add_ps(felec,fvdw);
 585
 586             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 587
 588             /* Calculate temporary vectorial force */
 589             tx               = _mm256_mul_ps(fscal,dx00);
 590             ty               = _mm256_mul_ps(fscal,dy00);
 591             tz               = _mm256_mul_ps(fscal,dz00);
 592
 593             /* Update vectorial force */
 594             fix0             = _mm256_add_ps(fix0,tx);
 595             fiy0             = _mm256_add_ps(fiy0,ty);
 596             fiz0             = _mm256_add_ps(fiz0,tz);
 597
 598             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 599             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 600             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 601             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 602             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 603             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 604             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 605             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 606             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 607
 608             /* Inner loop uses 92 flops */
 609         }
 610
 611         /* End of innermost loop */
 612
 613         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 614                                                  f+i_coord_offset,fshift+i_shift_offset);
 615
 616         ggid                        = gid[iidx];
 617         /* Update potential energies */
 618         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 619         gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 620         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 621         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
 622         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
 623
 624         /* Increment number of inner iterations */
 625         inneriter                  += j_index_end - j_index_start;
 626
 627         /* Outer loop uses 10 flops */
 628     }
 629
 630     /* Increment number of outer iterations */
 631     outeriter        += nri;
 632
 633     /* Update outer/inner flops */
 634
 635     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
 636 }
 637 /*
 638  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_single
 639  * Electrostatics interaction: GeneralizedBorn
 640  * VdW interaction:            CubicSplineTable
 641  * Geometry:                   Particle-Particle
 642  * Calculate force/pot:        Force
 643  */
 644 void
 645 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_256_single
 646                     (t_nblist                    * gmx_restrict       nlist,
 647                      rvec                        * gmx_restrict          xx,
 648                      rvec                        * gmx_restrict          ff,
 649                      struct t_forcerec           * gmx_restrict          fr,
 650                      t_mdatoms                   * gmx_restrict     mdatoms,
 651                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 652                      t_nrnb                      * gmx_restrict        nrnb)
 653 {
 654     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 655      * just 0 for non-waters.
 656      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 657      * jnr indices corresponding to data put in the four positions in the SIMD register.
 658      */
 659     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 660     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 661     int              jnrA,jnrB,jnrC,jnrD;
 662     int              jnrE,jnrF,jnrG,jnrH;
 663     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 664     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 665     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 666     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 667     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 668     real             rcutoff_scalar;
 669     real             *shiftvec,*fshift,*x,*f;
 670     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 671     real             scratch[4*DIM];
 672     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 673     real *           vdwioffsetptr0;
 674     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 675     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 676     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 677     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 678     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 679     real             *charge;
 680     __m256i          gbitab;
 681     __m128i          gbitab_lo,gbitab_hi;
 682     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 683     __m256           minushalf = _mm256_set1_ps(-0.5);
 684     real             *invsqrta,*dvda,*gbtab;
 685     int              nvdwtype;
 686     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 687     int              *vdwtype;
 688     real             *vdwparam;
 689     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 690     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 691     __m256i          vfitab;
 692     __m128i          vfitab_lo,vfitab_hi;
 693     __m128i          ifour       = _mm_set1_epi32(4);
 694     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 695     real             *vftab;
 696     __m256           dummy_mask,cutoff_mask;
 697     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 698     __m256           one     = _mm256_set1_ps(1.0);
 699     __m256           two     = _mm256_set1_ps(2.0);
 700     x                = xx[0];
 701     f                = ff[0];
 702
 703     nri              = nlist->nri;
 704     iinr             = nlist->iinr;
 705     jindex           = nlist->jindex;
 706     jjnr             = nlist->jjnr;
 707     shiftidx         = nlist->shift;
 708     gid              = nlist->gid;
 709     shiftvec         = fr->shift_vec[0];
 710     fshift           = fr->fshift[0];
 711     facel            = _mm256_set1_ps(fr->ic->epsfac);
 712     charge           = mdatoms->chargeA;
 713     nvdwtype         = fr->ntype;
 714     vdwparam         = fr->nbfp;
 715     vdwtype          = mdatoms->typeA;
 716
 717     vftab            = kernel_data->table_vdw->data;
 718     vftabscale       = _mm256_set1_ps(kernel_data->table_vdw->scale);
 719
 720     invsqrta         = fr->invsqrta;
 721     dvda             = fr->dvda;
 722     gbtabscale       = _mm256_set1_ps(fr->gbtab->scale);
 723     gbtab            = fr->gbtab->data;
 724     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 725
 726     /* Avoid stupid compiler warnings */
 727     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 728     j_coord_offsetA = 0;
 729     j_coord_offsetB = 0;
 730     j_coord_offsetC = 0;
 731     j_coord_offsetD = 0;
 732     j_coord_offsetE = 0;
 733     j_coord_offsetF = 0;
 734     j_coord_offsetG = 0;
 735     j_coord_offsetH = 0;
 736
 737     outeriter        = 0;
 738     inneriter        = 0;
 739
 740     for(iidx=0;iidx<4*DIM;iidx++)
 741     {
 742         scratch[iidx] = 0.0;
 743     }
 744
 745     /* Start outer loop over neighborlists */
 746     for(iidx=0; iidx<nri; iidx++)
 747     {
 748         /* Load shift vector for this list */
 749         i_shift_offset   = DIM*shiftidx[iidx];
 750
 751         /* Load limits for loop over neighbors */
 752         j_index_start    = jindex[iidx];
 753         j_index_end      = jindex[iidx+1];
 754
 755         /* Get outer coordinate index */
 756         inr              = iinr[iidx];
 757         i_coord_offset   = DIM*inr;
 758
 759         /* Load i particle coords and add shift vector */
 760         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 761
 762         fix0             = _mm256_setzero_ps();
 763         fiy0             = _mm256_setzero_ps();
 764         fiz0             = _mm256_setzero_ps();
 765
 766         /* Load parameters for i particles */
 767         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 768         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 769         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 770
 771         dvdasum          = _mm256_setzero_ps();
 772
 773         /* Start inner kernel loop */
 774         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 775         {
 776
 777             /* Get j neighbor index, and coordinate index */
 778             jnrA             = jjnr[jidx];
 779             jnrB             = jjnr[jidx+1];
 780             jnrC             = jjnr[jidx+2];
 781             jnrD             = jjnr[jidx+3];
 782             jnrE             = jjnr[jidx+4];
 783             jnrF             = jjnr[jidx+5];
 784             jnrG             = jjnr[jidx+6];
 785             jnrH             = jjnr[jidx+7];
 786             j_coord_offsetA  = DIM*jnrA;
 787             j_coord_offsetB  = DIM*jnrB;
 788             j_coord_offsetC  = DIM*jnrC;
 789             j_coord_offsetD  = DIM*jnrD;
 790             j_coord_offsetE  = DIM*jnrE;
 791             j_coord_offsetF  = DIM*jnrF;
 792             j_coord_offsetG  = DIM*jnrG;
 793             j_coord_offsetH  = DIM*jnrH;
 794
 795             /* load j atom coordinates */
 796             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 797                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 798                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 799                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 800                                                  &jx0,&jy0,&jz0);
 801
 802             /* Calculate displacement vector */
 803             dx00             = _mm256_sub_ps(ix0,jx0);
 804             dy00             = _mm256_sub_ps(iy0,jy0);
 805             dz00             = _mm256_sub_ps(iz0,jz0);
 806
 807             /* Calculate squared distance and things based on it */
 808             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 809
 810             rinv00           = avx256_invsqrt_f(rsq00);
 811
 812             /* Load parameters for j particles */
 813             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 814                                                                  charge+jnrC+0,charge+jnrD+0,
 815                                                                  charge+jnrE+0,charge+jnrF+0,
 816                                                                  charge+jnrG+0,charge+jnrH+0);
 817             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 818                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 819                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 820                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 821             vdwjidx0A        = 2*vdwtype[jnrA+0];
 822             vdwjidx0B        = 2*vdwtype[jnrB+0];
 823             vdwjidx0C        = 2*vdwtype[jnrC+0];
 824             vdwjidx0D        = 2*vdwtype[jnrD+0];
 825             vdwjidx0E        = 2*vdwtype[jnrE+0];
 826             vdwjidx0F        = 2*vdwtype[jnrF+0];
 827             vdwjidx0G        = 2*vdwtype[jnrG+0];
 828             vdwjidx0H        = 2*vdwtype[jnrH+0];
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             r00              = _mm256_mul_ps(rsq00,rinv00);
 835
 836             /* Compute parameters for interactions between i and j atoms */
 837             qq00             = _mm256_mul_ps(iq0,jq0);
 838             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 839                                             vdwioffsetptr0+vdwjidx0B,
 840                                             vdwioffsetptr0+vdwjidx0C,
 841                                             vdwioffsetptr0+vdwjidx0D,
 842                                             vdwioffsetptr0+vdwjidx0E,
 843                                             vdwioffsetptr0+vdwjidx0F,
 844                                             vdwioffsetptr0+vdwjidx0G,
 845                                             vdwioffsetptr0+vdwjidx0H,
 846                                             &c6_00,&c12_00);
 847
 848             /* Calculate table index by multiplying r with table scale and truncate to integer */
 849             rt               = _mm256_mul_ps(r00,vftabscale);
 850             vfitab           = _mm256_cvttps_epi32(rt);
 851             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 852             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 853             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 854             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 855             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 856             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 857
 858             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 859             isaprod          = _mm256_mul_ps(isai0,isaj0);
 860             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 861             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 862
 863             /* Calculate generalized born table index - this is a separate table from the normal one,
 864              * but we use the same procedure by multiplying r with scale and truncating to integer.
 865              */
 866             rt               = _mm256_mul_ps(r00,gbscale);
 867             gbitab           = _mm256_cvttps_epi32(rt);
 868             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 869             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 870             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 871             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 872             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 873             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 874             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 875                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 876             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 877                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 878             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 879                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 880             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 881                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 882             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 883             Heps             = _mm256_mul_ps(gbeps,H);
 884             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 885             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 886             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 887
 888             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 889             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 890             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 891             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 892             fjptrA           = dvda+jnrA;
 893             fjptrB           = dvda+jnrB;
 894             fjptrC           = dvda+jnrC;
 895             fjptrD           = dvda+jnrD;
 896             fjptrE           = dvda+jnrE;
 897             fjptrF           = dvda+jnrF;
 898             fjptrG           = dvda+jnrG;
 899             fjptrH           = dvda+jnrH;
 900             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 901                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 902             velec            = _mm256_mul_ps(qq00,rinv00);
 903             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 904
 905             /* CUBIC SPLINE TABLE DISPERSION */
 906             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 907                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 908             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 909                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 910             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 911                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 912             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 913                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 914             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 915             Heps             = _mm256_mul_ps(vfeps,H);
 916             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 917             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 918             fvdw6            = _mm256_mul_ps(c6_00,FF);
 919
 920             /* CUBIC SPLINE TABLE REPULSION */
 921             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 922             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 923             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 924                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 925             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 926                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 927             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 928                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 929             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 930                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 931             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 932             Heps             = _mm256_mul_ps(vfeps,H);
 933             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 934             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 935             fvdw12           = _mm256_mul_ps(c12_00,FF);
 936             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 937
 938             fscal            = _mm256_add_ps(felec,fvdw);
 939
 940             /* Calculate temporary vectorial force */
 941             tx               = _mm256_mul_ps(fscal,dx00);
 942             ty               = _mm256_mul_ps(fscal,dy00);
 943             tz               = _mm256_mul_ps(fscal,dz00);
 944
 945             /* Update vectorial force */
 946             fix0             = _mm256_add_ps(fix0,tx);
 947             fiy0             = _mm256_add_ps(fiy0,ty);
 948             fiz0             = _mm256_add_ps(fiz0,tz);
 949
 950             fjptrA             = f+j_coord_offsetA;
 951             fjptrB             = f+j_coord_offsetB;
 952             fjptrC             = f+j_coord_offsetC;
 953             fjptrD             = f+j_coord_offsetD;
 954             fjptrE             = f+j_coord_offsetE;
 955             fjptrF             = f+j_coord_offsetF;
 956             fjptrG             = f+j_coord_offsetG;
 957             fjptrH             = f+j_coord_offsetH;
 958             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 959
 960             /* Inner loop uses 81 flops */
 961         }
 962
 963         if(jidx<j_index_end)
 964         {
 965
 966             /* Get j neighbor index, and coordinate index */
 967             jnrlistA         = jjnr[jidx];
 968             jnrlistB         = jjnr[jidx+1];
 969             jnrlistC         = jjnr[jidx+2];
 970             jnrlistD         = jjnr[jidx+3];
 971             jnrlistE         = jjnr[jidx+4];
 972             jnrlistF         = jjnr[jidx+5];
 973             jnrlistG         = jjnr[jidx+6];
 974             jnrlistH         = jjnr[jidx+7];
 975             /* Sign of each element will be negative for non-real atoms.
 976              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 977              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 978              */
 979             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 980                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 981
 982             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 983             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 984             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 985             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 986             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 987             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 988             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 989             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 990             j_coord_offsetA  = DIM*jnrA;
 991             j_coord_offsetB  = DIM*jnrB;
 992             j_coord_offsetC  = DIM*jnrC;
 993             j_coord_offsetD  = DIM*jnrD;
 994             j_coord_offsetE  = DIM*jnrE;
 995             j_coord_offsetF  = DIM*jnrF;
 996             j_coord_offsetG  = DIM*jnrG;
 997             j_coord_offsetH  = DIM*jnrH;
 998
 999             /* load j atom coordinates */
1000             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1001                                                  x+j_coord_offsetC,x+j_coord_offsetD,
1002                                                  x+j_coord_offsetE,x+j_coord_offsetF,
1003                                                  x+j_coord_offsetG,x+j_coord_offsetH,
1004                                                  &jx0,&jy0,&jz0);
1005
1006             /* Calculate displacement vector */
1007             dx00             = _mm256_sub_ps(ix0,jx0);
1008             dy00             = _mm256_sub_ps(iy0,jy0);
1009             dz00             = _mm256_sub_ps(iz0,jz0);
1010
1011             /* Calculate squared distance and things based on it */
1012             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
1013
1014             rinv00           = avx256_invsqrt_f(rsq00);
1015
1016             /* Load parameters for j particles */
1017             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1018                                                                  charge+jnrC+0,charge+jnrD+0,
1019                                                                  charge+jnrE+0,charge+jnrF+0,
1020                                                                  charge+jnrG+0,charge+jnrH+0);
1021             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
1022                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
1023                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
1024                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
1025             vdwjidx0A        = 2*vdwtype[jnrA+0];
1026             vdwjidx0B        = 2*vdwtype[jnrB+0];
1027             vdwjidx0C        = 2*vdwtype[jnrC+0];
1028             vdwjidx0D        = 2*vdwtype[jnrD+0];
1029             vdwjidx0E        = 2*vdwtype[jnrE+0];
1030             vdwjidx0F        = 2*vdwtype[jnrF+0];
1031             vdwjidx0G        = 2*vdwtype[jnrG+0];
1032             vdwjidx0H        = 2*vdwtype[jnrH+0];
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             r00              = _mm256_mul_ps(rsq00,rinv00);
1039             r00              = _mm256_andnot_ps(dummy_mask,r00);
1040
1041             /* Compute parameters for interactions between i and j atoms */
1042             qq00             = _mm256_mul_ps(iq0,jq0);
1043             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
1044                                             vdwioffsetptr0+vdwjidx0B,
1045                                             vdwioffsetptr0+vdwjidx0C,
1046                                             vdwioffsetptr0+vdwjidx0D,
1047                                             vdwioffsetptr0+vdwjidx0E,
1048                                             vdwioffsetptr0+vdwjidx0F,
1049                                             vdwioffsetptr0+vdwjidx0G,
1050                                             vdwioffsetptr0+vdwjidx0H,
1051                                             &c6_00,&c12_00);
1052
1053             /* Calculate table index by multiplying r with table scale and truncate to integer */
1054             rt               = _mm256_mul_ps(r00,vftabscale);
1055             vfitab           = _mm256_cvttps_epi32(rt);
1056             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1057             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1058             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
1059             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
1060             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
1061             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
1062
1063             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
1064             isaprod          = _mm256_mul_ps(isai0,isaj0);
1065             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
1066             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
1067
1068             /* Calculate generalized born table index - this is a separate table from the normal one,
1069              * but we use the same procedure by multiplying r with scale and truncating to integer.
1070              */
1071             rt               = _mm256_mul_ps(r00,gbscale);
1072             gbitab           = _mm256_cvttps_epi32(rt);
1073             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1074             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1075             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
1076             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
1077             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
1078             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
1079             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
1080                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
1081             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
1082                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
1083             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
1084                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
1085             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
1086                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
1087             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1088             Heps             = _mm256_mul_ps(gbeps,H);
1089             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
1090             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
1091             vgb              = _mm256_mul_ps(gbqqfactor,VV);
1092
1093             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1094             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
1095             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
1096             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
1097             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
1098             /* 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. */
1099             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
1100             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
1101             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
1102             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
1103             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
1104             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
1105             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
1106             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
1107             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
1108                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
1109             velec            = _mm256_mul_ps(qq00,rinv00);
1110             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
1111
1112             /* CUBIC SPLINE TABLE DISPERSION */
1113             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1114                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1115             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1116                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1117             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1118                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1119             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1120                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1121             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1122             Heps             = _mm256_mul_ps(vfeps,H);
1123             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1124             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1125             fvdw6            = _mm256_mul_ps(c6_00,FF);
1126
1127             /* CUBIC SPLINE TABLE REPULSION */
1128             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
1129             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
1130             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1131                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1132             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1133                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1134             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1135                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1136             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1137                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1138             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1139             Heps             = _mm256_mul_ps(vfeps,H);
1140             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1141             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1142             fvdw12           = _mm256_mul_ps(c12_00,FF);
1143             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
1144
1145             fscal            = _mm256_add_ps(felec,fvdw);
1146
1147             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
1148
1149             /* Calculate temporary vectorial force */
1150             tx               = _mm256_mul_ps(fscal,dx00);
1151             ty               = _mm256_mul_ps(fscal,dy00);
1152             tz               = _mm256_mul_ps(fscal,dz00);
1153
1154             /* Update vectorial force */
1155             fix0             = _mm256_add_ps(fix0,tx);
1156             fiy0             = _mm256_add_ps(fiy0,ty);
1157             fiz0             = _mm256_add_ps(fiz0,tz);
1158
1159             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1160             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1161             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1162             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1163             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
1164             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
1165             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
1166             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
1167             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
1168
1169             /* Inner loop uses 82 flops */
1170         }
1171
1172         /* End of innermost loop */
1173
1174         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1175                                                  f+i_coord_offset,fshift+i_shift_offset);
1176
1177         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
1178         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
1179
1180         /* Increment number of inner iterations */
1181         inneriter                  += j_index_end - j_index_start;
1182
1183         /* Outer loop uses 7 flops */
1184     }
1185
1186     /* Increment number of outer iterations */
1187     outeriter        += nri;
1188
1189     /* Update outer/inner flops */
1190
1191     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);
1192 }