src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecGB_VdwLJ_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_VdwLJ_GeomP1P1_VF_avx_256_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwLJ_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     invsqrta         = fr->invsqrta;
 117     dvda             = fr->dvda;
 118     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 119     gbtab            = fr->gbtab.data;
 120     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128     j_coord_offsetE = 0;
 129     j_coord_offsetF = 0;
 130     j_coord_offsetG = 0;
 131     j_coord_offsetH = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     for(iidx=0;iidx<4*DIM;iidx++)
 137     {
 138         scratch[iidx] = 0.0;
 139     }
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 157
 158         fix0             = _mm256_setzero_ps();
 159         fiy0             = _mm256_setzero_ps();
 160         fiz0             = _mm256_setzero_ps();
 161
 162         /* Load parameters for i particles */
 163         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 164         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 165         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 166
 167         /* Reset potential sums */
 168         velecsum         = _mm256_setzero_ps();
 169         vgbsum           = _mm256_setzero_ps();
 170         vvdwsum          = _mm256_setzero_ps();
 171         dvdasum          = _mm256_setzero_ps();
 172
 173         /* Start inner kernel loop */
 174         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 175         {
 176
 177             /* Get j neighbor index, and coordinate index */
 178             jnrA             = jjnr[jidx];
 179             jnrB             = jjnr[jidx+1];
 180             jnrC             = jjnr[jidx+2];
 181             jnrD             = jjnr[jidx+3];
 182             jnrE             = jjnr[jidx+4];
 183             jnrF             = jjnr[jidx+5];
 184             jnrG             = jjnr[jidx+6];
 185             jnrH             = jjnr[jidx+7];
 186             j_coord_offsetA  = DIM*jnrA;
 187             j_coord_offsetB  = DIM*jnrB;
 188             j_coord_offsetC  = DIM*jnrC;
 189             j_coord_offsetD  = DIM*jnrD;
 190             j_coord_offsetE  = DIM*jnrE;
 191             j_coord_offsetF  = DIM*jnrF;
 192             j_coord_offsetG  = DIM*jnrG;
 193             j_coord_offsetH  = DIM*jnrH;
 194
 195             /* load j atom coordinates */
 196             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 197                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 198                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 199                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 200                                                  &jx0,&jy0,&jz0);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm256_sub_ps(ix0,jx0);
 204             dy00             = _mm256_sub_ps(iy0,jy0);
 205             dz00             = _mm256_sub_ps(iz0,jz0);
 206
 207             /* Calculate squared distance and things based on it */
 208             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 209
 210             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 211
 212             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 213
 214             /* Load parameters for j particles */
 215             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 216                                                                  charge+jnrC+0,charge+jnrD+0,
 217                                                                  charge+jnrE+0,charge+jnrF+0,
 218                                                                  charge+jnrG+0,charge+jnrH+0);
 219             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 220                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 221                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 222                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 223             vdwjidx0A        = 2*vdwtype[jnrA+0];
 224             vdwjidx0B        = 2*vdwtype[jnrB+0];
 225             vdwjidx0C        = 2*vdwtype[jnrC+0];
 226             vdwjidx0D        = 2*vdwtype[jnrD+0];
 227             vdwjidx0E        = 2*vdwtype[jnrE+0];
 228             vdwjidx0F        = 2*vdwtype[jnrF+0];
 229             vdwjidx0G        = 2*vdwtype[jnrG+0];
 230             vdwjidx0H        = 2*vdwtype[jnrH+0];
 231
 232             /**************************
 233              * CALCULATE INTERACTIONS *
 234              **************************/
 235
 236             r00              = _mm256_mul_ps(rsq00,rinv00);
 237
 238             /* Compute parameters for interactions between i and j atoms */
 239             qq00             = _mm256_mul_ps(iq0,jq0);
 240             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 241                                             vdwioffsetptr0+vdwjidx0B,
 242                                             vdwioffsetptr0+vdwjidx0C,
 243                                             vdwioffsetptr0+vdwjidx0D,
 244                                             vdwioffsetptr0+vdwjidx0E,
 245                                             vdwioffsetptr0+vdwjidx0F,
 246                                             vdwioffsetptr0+vdwjidx0G,
 247                                             vdwioffsetptr0+vdwjidx0H,
 248                                             &c6_00,&c12_00);
 249
 250             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 251             isaprod          = _mm256_mul_ps(isai0,isaj0);
 252             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 253             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 254
 255             /* Calculate generalized born table index - this is a separate table from the normal one,
 256              * but we use the same procedure by multiplying r with scale and truncating to integer.
 257              */
 258             rt               = _mm256_mul_ps(r00,gbscale);
 259             gbitab           = _mm256_cvttps_epi32(rt);
 260             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 261             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 262             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 263             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 264             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 265             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 266             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 267                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 268             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 269                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 270             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 271                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 272             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 273                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 274             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 275             Heps             = _mm256_mul_ps(gbeps,H);
 276             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 277             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 278             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 279
 280             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 281             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 282             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 283             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 284             fjptrA           = dvda+jnrA;
 285             fjptrB           = dvda+jnrB;
 286             fjptrC           = dvda+jnrC;
 287             fjptrD           = dvda+jnrD;
 288             fjptrE           = dvda+jnrE;
 289             fjptrF           = dvda+jnrF;
 290             fjptrG           = dvda+jnrG;
 291             fjptrH           = dvda+jnrH;
 292             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 293                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 294             velec            = _mm256_mul_ps(qq00,rinv00);
 295             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 296
 297             /* LENNARD-JONES DISPERSION/REPULSION */
 298
 299             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 300             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 301             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 302             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 303             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velecsum         = _mm256_add_ps(velecsum,velec);
 307             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 308             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 309
 310             fscal            = _mm256_add_ps(felec,fvdw);
 311
 312             /* Calculate temporary vectorial force */
 313             tx               = _mm256_mul_ps(fscal,dx00);
 314             ty               = _mm256_mul_ps(fscal,dy00);
 315             tz               = _mm256_mul_ps(fscal,dz00);
 316
 317             /* Update vectorial force */
 318             fix0             = _mm256_add_ps(fix0,tx);
 319             fiy0             = _mm256_add_ps(fiy0,ty);
 320             fiz0             = _mm256_add_ps(fiz0,tz);
 321
 322             fjptrA             = f+j_coord_offsetA;
 323             fjptrB             = f+j_coord_offsetB;
 324             fjptrC             = f+j_coord_offsetC;
 325             fjptrD             = f+j_coord_offsetD;
 326             fjptrE             = f+j_coord_offsetE;
 327             fjptrF             = f+j_coord_offsetF;
 328             fjptrG             = f+j_coord_offsetG;
 329             fjptrH             = f+j_coord_offsetH;
 330             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 331
 332             /* Inner loop uses 70 flops */
 333         }
 334
 335         if(jidx<j_index_end)
 336         {
 337
 338             /* Get j neighbor index, and coordinate index */
 339             jnrlistA         = jjnr[jidx];
 340             jnrlistB         = jjnr[jidx+1];
 341             jnrlistC         = jjnr[jidx+2];
 342             jnrlistD         = jjnr[jidx+3];
 343             jnrlistE         = jjnr[jidx+4];
 344             jnrlistF         = jjnr[jidx+5];
 345             jnrlistG         = jjnr[jidx+6];
 346             jnrlistH         = jjnr[jidx+7];
 347             /* Sign of each element will be negative for non-real atoms.
 348              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 349              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 350              */
 351             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 352                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 353
 354             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 355             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 356             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 357             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 358             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 359             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 360             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 361             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 362             j_coord_offsetA  = DIM*jnrA;
 363             j_coord_offsetB  = DIM*jnrB;
 364             j_coord_offsetC  = DIM*jnrC;
 365             j_coord_offsetD  = DIM*jnrD;
 366             j_coord_offsetE  = DIM*jnrE;
 367             j_coord_offsetF  = DIM*jnrF;
 368             j_coord_offsetG  = DIM*jnrG;
 369             j_coord_offsetH  = DIM*jnrH;
 370
 371             /* load j atom coordinates */
 372             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 373                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 374                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 375                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 376                                                  &jx0,&jy0,&jz0);
 377
 378             /* Calculate displacement vector */
 379             dx00             = _mm256_sub_ps(ix0,jx0);
 380             dy00             = _mm256_sub_ps(iy0,jy0);
 381             dz00             = _mm256_sub_ps(iz0,jz0);
 382
 383             /* Calculate squared distance and things based on it */
 384             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 385
 386             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 387
 388             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 389
 390             /* Load parameters for j particles */
 391             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 392                                                                  charge+jnrC+0,charge+jnrD+0,
 393                                                                  charge+jnrE+0,charge+jnrF+0,
 394                                                                  charge+jnrG+0,charge+jnrH+0);
 395             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 396                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 397                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 398                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 399             vdwjidx0A        = 2*vdwtype[jnrA+0];
 400             vdwjidx0B        = 2*vdwtype[jnrB+0];
 401             vdwjidx0C        = 2*vdwtype[jnrC+0];
 402             vdwjidx0D        = 2*vdwtype[jnrD+0];
 403             vdwjidx0E        = 2*vdwtype[jnrE+0];
 404             vdwjidx0F        = 2*vdwtype[jnrF+0];
 405             vdwjidx0G        = 2*vdwtype[jnrG+0];
 406             vdwjidx0H        = 2*vdwtype[jnrH+0];
 407
 408             /**************************
 409              * CALCULATE INTERACTIONS *
 410              **************************/
 411
 412             r00              = _mm256_mul_ps(rsq00,rinv00);
 413             r00              = _mm256_andnot_ps(dummy_mask,r00);
 414
 415             /* Compute parameters for interactions between i and j atoms */
 416             qq00             = _mm256_mul_ps(iq0,jq0);
 417             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 418                                             vdwioffsetptr0+vdwjidx0B,
 419                                             vdwioffsetptr0+vdwjidx0C,
 420                                             vdwioffsetptr0+vdwjidx0D,
 421                                             vdwioffsetptr0+vdwjidx0E,
 422                                             vdwioffsetptr0+vdwjidx0F,
 423                                             vdwioffsetptr0+vdwjidx0G,
 424                                             vdwioffsetptr0+vdwjidx0H,
 425                                             &c6_00,&c12_00);
 426
 427             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 428             isaprod          = _mm256_mul_ps(isai0,isaj0);
 429             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 430             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 431
 432             /* Calculate generalized born table index - this is a separate table from the normal one,
 433              * but we use the same procedure by multiplying r with scale and truncating to integer.
 434              */
 435             rt               = _mm256_mul_ps(r00,gbscale);
 436             gbitab           = _mm256_cvttps_epi32(rt);
 437             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 438             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 439             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 440             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 441             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 442             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 443             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 444                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 445             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 446                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 447             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 448                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 449             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 450                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 451             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 452             Heps             = _mm256_mul_ps(gbeps,H);
 453             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 454             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 455             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 456
 457             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 458             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 459             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 460             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 461             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 462             /* 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. */
 463             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 464             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 465             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 466             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 467             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 468             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 469             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 470             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 471             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 472                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 473             velec            = _mm256_mul_ps(qq00,rinv00);
 474             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 475
 476             /* LENNARD-JONES DISPERSION/REPULSION */
 477
 478             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 479             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 480             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 481             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 482             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velec            = _mm256_andnot_ps(dummy_mask,velec);
 486             velecsum         = _mm256_add_ps(velecsum,velec);
 487             vgb              = _mm256_andnot_ps(dummy_mask,vgb);
 488             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 489             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 490             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 491
 492             fscal            = _mm256_add_ps(felec,fvdw);
 493
 494             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 495
 496             /* Calculate temporary vectorial force */
 497             tx               = _mm256_mul_ps(fscal,dx00);
 498             ty               = _mm256_mul_ps(fscal,dy00);
 499             tz               = _mm256_mul_ps(fscal,dz00);
 500
 501             /* Update vectorial force */
 502             fix0             = _mm256_add_ps(fix0,tx);
 503             fiy0             = _mm256_add_ps(fiy0,ty);
 504             fiz0             = _mm256_add_ps(fiz0,tz);
 505
 506             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 507             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 508             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 509             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 510             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 511             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 512             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 513             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 514             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 515
 516             /* Inner loop uses 71 flops */
 517         }
 518
 519         /* End of innermost loop */
 520
 521         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 522                                                  f+i_coord_offset,fshift+i_shift_offset);
 523
 524         ggid                        = gid[iidx];
 525         /* Update potential energies */
 526         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 527         gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 528         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 529         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
 530         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
 531
 532         /* Increment number of inner iterations */
 533         inneriter                  += j_index_end - j_index_start;
 534
 535         /* Outer loop uses 10 flops */
 536     }
 537
 538     /* Increment number of outer iterations */
 539     outeriter        += nri;
 540
 541     /* Update outer/inner flops */
 542
 543     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
 544 }
 545 /*
 546  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 547  * Electrostatics interaction: GeneralizedBorn
 548  * VdW interaction:            LennardJones
 549  * Geometry:                   Particle-Particle
 550  * Calculate force/pot:        Force
 551  */
 552 void
 553 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 554                     (t_nblist * gmx_restrict                nlist,
 555                      rvec * gmx_restrict                    xx,
 556                      rvec * gmx_restrict                    ff,
 557                      t_forcerec * gmx_restrict              fr,
 558                      t_mdatoms * gmx_restrict               mdatoms,
 559                      nb_kernel_data_t * gmx_restrict        kernel_data,
 560                      t_nrnb * gmx_restrict                  nrnb)
 561 {
 562     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 563      * just 0 for non-waters.
 564      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 565      * jnr indices corresponding to data put in the four positions in the SIMD register.
 566      */
 567     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 568     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 569     int              jnrA,jnrB,jnrC,jnrD;
 570     int              jnrE,jnrF,jnrG,jnrH;
 571     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 572     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 573     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 574     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 575     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 576     real             rcutoff_scalar;
 577     real             *shiftvec,*fshift,*x,*f;
 578     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 579     real             scratch[4*DIM];
 580     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 581     real *           vdwioffsetptr0;
 582     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 583     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 584     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 585     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 586     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 587     real             *charge;
 588     __m256i          gbitab;
 589     __m128i          gbitab_lo,gbitab_hi;
 590     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 591     __m256           minushalf = _mm256_set1_ps(-0.5);
 592     real             *invsqrta,*dvda,*gbtab;
 593     int              nvdwtype;
 594     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 595     int              *vdwtype;
 596     real             *vdwparam;
 597     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 598     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 599     __m256i          vfitab;
 600     __m128i          vfitab_lo,vfitab_hi;
 601     __m128i          ifour       = _mm_set1_epi32(4);
 602     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 603     real             *vftab;
 604     __m256           dummy_mask,cutoff_mask;
 605     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 606     __m256           one     = _mm256_set1_ps(1.0);
 607     __m256           two     = _mm256_set1_ps(2.0);
 608     x                = xx[0];
 609     f                = ff[0];
 610
 611     nri              = nlist->nri;
 612     iinr             = nlist->iinr;
 613     jindex           = nlist->jindex;
 614     jjnr             = nlist->jjnr;
 615     shiftidx         = nlist->shift;
 616     gid              = nlist->gid;
 617     shiftvec         = fr->shift_vec[0];
 618     fshift           = fr->fshift[0];
 619     facel            = _mm256_set1_ps(fr->epsfac);
 620     charge           = mdatoms->chargeA;
 621     nvdwtype         = fr->ntype;
 622     vdwparam         = fr->nbfp;
 623     vdwtype          = mdatoms->typeA;
 624
 625     invsqrta         = fr->invsqrta;
 626     dvda             = fr->dvda;
 627     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 628     gbtab            = fr->gbtab.data;
 629     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 630
 631     /* Avoid stupid compiler warnings */
 632     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 633     j_coord_offsetA = 0;
 634     j_coord_offsetB = 0;
 635     j_coord_offsetC = 0;
 636     j_coord_offsetD = 0;
 637     j_coord_offsetE = 0;
 638     j_coord_offsetF = 0;
 639     j_coord_offsetG = 0;
 640     j_coord_offsetH = 0;
 641
 642     outeriter        = 0;
 643     inneriter        = 0;
 644
 645     for(iidx=0;iidx<4*DIM;iidx++)
 646     {
 647         scratch[iidx] = 0.0;
 648     }
 649
 650     /* Start outer loop over neighborlists */
 651     for(iidx=0; iidx<nri; iidx++)
 652     {
 653         /* Load shift vector for this list */
 654         i_shift_offset   = DIM*shiftidx[iidx];
 655
 656         /* Load limits for loop over neighbors */
 657         j_index_start    = jindex[iidx];
 658         j_index_end      = jindex[iidx+1];
 659
 660         /* Get outer coordinate index */
 661         inr              = iinr[iidx];
 662         i_coord_offset   = DIM*inr;
 663
 664         /* Load i particle coords and add shift vector */
 665         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 666
 667         fix0             = _mm256_setzero_ps();
 668         fiy0             = _mm256_setzero_ps();
 669         fiz0             = _mm256_setzero_ps();
 670
 671         /* Load parameters for i particles */
 672         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 673         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 674         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 675
 676         dvdasum          = _mm256_setzero_ps();
 677
 678         /* Start inner kernel loop */
 679         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 680         {
 681
 682             /* Get j neighbor index, and coordinate index */
 683             jnrA             = jjnr[jidx];
 684             jnrB             = jjnr[jidx+1];
 685             jnrC             = jjnr[jidx+2];
 686             jnrD             = jjnr[jidx+3];
 687             jnrE             = jjnr[jidx+4];
 688             jnrF             = jjnr[jidx+5];
 689             jnrG             = jjnr[jidx+6];
 690             jnrH             = jjnr[jidx+7];
 691             j_coord_offsetA  = DIM*jnrA;
 692             j_coord_offsetB  = DIM*jnrB;
 693             j_coord_offsetC  = DIM*jnrC;
 694             j_coord_offsetD  = DIM*jnrD;
 695             j_coord_offsetE  = DIM*jnrE;
 696             j_coord_offsetF  = DIM*jnrF;
 697             j_coord_offsetG  = DIM*jnrG;
 698             j_coord_offsetH  = DIM*jnrH;
 699
 700             /* load j atom coordinates */
 701             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 702                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 703                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 704                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 705                                                  &jx0,&jy0,&jz0);
 706
 707             /* Calculate displacement vector */
 708             dx00             = _mm256_sub_ps(ix0,jx0);
 709             dy00             = _mm256_sub_ps(iy0,jy0);
 710             dz00             = _mm256_sub_ps(iz0,jz0);
 711
 712             /* Calculate squared distance and things based on it */
 713             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 714
 715             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 716
 717             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 718
 719             /* Load parameters for j particles */
 720             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 721                                                                  charge+jnrC+0,charge+jnrD+0,
 722                                                                  charge+jnrE+0,charge+jnrF+0,
 723                                                                  charge+jnrG+0,charge+jnrH+0);
 724             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 725                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 726                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 727                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 728             vdwjidx0A        = 2*vdwtype[jnrA+0];
 729             vdwjidx0B        = 2*vdwtype[jnrB+0];
 730             vdwjidx0C        = 2*vdwtype[jnrC+0];
 731             vdwjidx0D        = 2*vdwtype[jnrD+0];
 732             vdwjidx0E        = 2*vdwtype[jnrE+0];
 733             vdwjidx0F        = 2*vdwtype[jnrF+0];
 734             vdwjidx0G        = 2*vdwtype[jnrG+0];
 735             vdwjidx0H        = 2*vdwtype[jnrH+0];
 736
 737             /**************************
 738              * CALCULATE INTERACTIONS *
 739              **************************/
 740
 741             r00              = _mm256_mul_ps(rsq00,rinv00);
 742
 743             /* Compute parameters for interactions between i and j atoms */
 744             qq00             = _mm256_mul_ps(iq0,jq0);
 745             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 746                                             vdwioffsetptr0+vdwjidx0B,
 747                                             vdwioffsetptr0+vdwjidx0C,
 748                                             vdwioffsetptr0+vdwjidx0D,
 749                                             vdwioffsetptr0+vdwjidx0E,
 750                                             vdwioffsetptr0+vdwjidx0F,
 751                                             vdwioffsetptr0+vdwjidx0G,
 752                                             vdwioffsetptr0+vdwjidx0H,
 753                                             &c6_00,&c12_00);
 754
 755             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 756             isaprod          = _mm256_mul_ps(isai0,isaj0);
 757             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 758             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 759
 760             /* Calculate generalized born table index - this is a separate table from the normal one,
 761              * but we use the same procedure by multiplying r with scale and truncating to integer.
 762              */
 763             rt               = _mm256_mul_ps(r00,gbscale);
 764             gbitab           = _mm256_cvttps_epi32(rt);
 765             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 766             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 767             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 768             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 769             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 770             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 771             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 772                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 773             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 774                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 775             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 776                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 777             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 778                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 779             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 780             Heps             = _mm256_mul_ps(gbeps,H);
 781             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 782             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 783             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 784
 785             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 786             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 787             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 788             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 789             fjptrA           = dvda+jnrA;
 790             fjptrB           = dvda+jnrB;
 791             fjptrC           = dvda+jnrC;
 792             fjptrD           = dvda+jnrD;
 793             fjptrE           = dvda+jnrE;
 794             fjptrF           = dvda+jnrF;
 795             fjptrG           = dvda+jnrG;
 796             fjptrH           = dvda+jnrH;
 797             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 798                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 799             velec            = _mm256_mul_ps(qq00,rinv00);
 800             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 801
 802             /* LENNARD-JONES DISPERSION/REPULSION */
 803
 804             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 805             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 806
 807             fscal            = _mm256_add_ps(felec,fvdw);
 808
 809             /* Calculate temporary vectorial force */
 810             tx               = _mm256_mul_ps(fscal,dx00);
 811             ty               = _mm256_mul_ps(fscal,dy00);
 812             tz               = _mm256_mul_ps(fscal,dz00);
 813
 814             /* Update vectorial force */
 815             fix0             = _mm256_add_ps(fix0,tx);
 816             fiy0             = _mm256_add_ps(fiy0,ty);
 817             fiz0             = _mm256_add_ps(fiz0,tz);
 818
 819             fjptrA             = f+j_coord_offsetA;
 820             fjptrB             = f+j_coord_offsetB;
 821             fjptrC             = f+j_coord_offsetC;
 822             fjptrD             = f+j_coord_offsetD;
 823             fjptrE             = f+j_coord_offsetE;
 824             fjptrF             = f+j_coord_offsetF;
 825             fjptrG             = f+j_coord_offsetG;
 826             fjptrH             = f+j_coord_offsetH;
 827             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 828
 829             /* Inner loop uses 63 flops */
 830         }
 831
 832         if(jidx<j_index_end)
 833         {
 834
 835             /* Get j neighbor index, and coordinate index */
 836             jnrlistA         = jjnr[jidx];
 837             jnrlistB         = jjnr[jidx+1];
 838             jnrlistC         = jjnr[jidx+2];
 839             jnrlistD         = jjnr[jidx+3];
 840             jnrlistE         = jjnr[jidx+4];
 841             jnrlistF         = jjnr[jidx+5];
 842             jnrlistG         = jjnr[jidx+6];
 843             jnrlistH         = jjnr[jidx+7];
 844             /* Sign of each element will be negative for non-real atoms.
 845              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 846              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 847              */
 848             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 849                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 850
 851             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 852             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 853             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 854             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 855             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 856             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 857             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 858             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 859             j_coord_offsetA  = DIM*jnrA;
 860             j_coord_offsetB  = DIM*jnrB;
 861             j_coord_offsetC  = DIM*jnrC;
 862             j_coord_offsetD  = DIM*jnrD;
 863             j_coord_offsetE  = DIM*jnrE;
 864             j_coord_offsetF  = DIM*jnrF;
 865             j_coord_offsetG  = DIM*jnrG;
 866             j_coord_offsetH  = DIM*jnrH;
 867
 868             /* load j atom coordinates */
 869             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 870                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 871                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 872                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 873                                                  &jx0,&jy0,&jz0);
 874
 875             /* Calculate displacement vector */
 876             dx00             = _mm256_sub_ps(ix0,jx0);
 877             dy00             = _mm256_sub_ps(iy0,jy0);
 878             dz00             = _mm256_sub_ps(iz0,jz0);
 879
 880             /* Calculate squared distance and things based on it */
 881             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 882
 883             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 884
 885             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 886
 887             /* Load parameters for j particles */
 888             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 889                                                                  charge+jnrC+0,charge+jnrD+0,
 890                                                                  charge+jnrE+0,charge+jnrF+0,
 891                                                                  charge+jnrG+0,charge+jnrH+0);
 892             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 893                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 894                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 895                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 896             vdwjidx0A        = 2*vdwtype[jnrA+0];
 897             vdwjidx0B        = 2*vdwtype[jnrB+0];
 898             vdwjidx0C        = 2*vdwtype[jnrC+0];
 899             vdwjidx0D        = 2*vdwtype[jnrD+0];
 900             vdwjidx0E        = 2*vdwtype[jnrE+0];
 901             vdwjidx0F        = 2*vdwtype[jnrF+0];
 902             vdwjidx0G        = 2*vdwtype[jnrG+0];
 903             vdwjidx0H        = 2*vdwtype[jnrH+0];
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             r00              = _mm256_mul_ps(rsq00,rinv00);
 910             r00              = _mm256_andnot_ps(dummy_mask,r00);
 911
 912             /* Compute parameters for interactions between i and j atoms */
 913             qq00             = _mm256_mul_ps(iq0,jq0);
 914             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 915                                             vdwioffsetptr0+vdwjidx0B,
 916                                             vdwioffsetptr0+vdwjidx0C,
 917                                             vdwioffsetptr0+vdwjidx0D,
 918                                             vdwioffsetptr0+vdwjidx0E,
 919                                             vdwioffsetptr0+vdwjidx0F,
 920                                             vdwioffsetptr0+vdwjidx0G,
 921                                             vdwioffsetptr0+vdwjidx0H,
 922                                             &c6_00,&c12_00);
 923
 924             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 925             isaprod          = _mm256_mul_ps(isai0,isaj0);
 926             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 927             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 928
 929             /* Calculate generalized born table index - this is a separate table from the normal one,
 930              * but we use the same procedure by multiplying r with scale and truncating to integer.
 931              */
 932             rt               = _mm256_mul_ps(r00,gbscale);
 933             gbitab           = _mm256_cvttps_epi32(rt);
 934             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 935             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 936             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 937             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 938             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 939             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 940             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 941                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 942             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 943                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 944             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 945                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 946             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 947                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 948             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 949             Heps             = _mm256_mul_ps(gbeps,H);
 950             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 951             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 952             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 953
 954             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 955             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 956             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 957             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 958             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 959             /* 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. */
 960             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 961             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 962             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 963             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 964             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 965             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 966             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 967             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 968             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 969                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 970             velec            = _mm256_mul_ps(qq00,rinv00);
 971             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 972
 973             /* LENNARD-JONES DISPERSION/REPULSION */
 974
 975             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 976             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 977
 978             fscal            = _mm256_add_ps(felec,fvdw);
 979
 980             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 981
 982             /* Calculate temporary vectorial force */
 983             tx               = _mm256_mul_ps(fscal,dx00);
 984             ty               = _mm256_mul_ps(fscal,dy00);
 985             tz               = _mm256_mul_ps(fscal,dz00);
 986
 987             /* Update vectorial force */
 988             fix0             = _mm256_add_ps(fix0,tx);
 989             fiy0             = _mm256_add_ps(fiy0,ty);
 990             fiz0             = _mm256_add_ps(fiz0,tz);
 991
 992             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 993             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 994             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 995             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 996             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 997             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 998             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 999             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
1000             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
1001
1002             /* Inner loop uses 64 flops */
1003         }
1004
1005         /* End of innermost loop */
1006
1007         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1008                                                  f+i_coord_offset,fshift+i_shift_offset);
1009
1010         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
1011         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
1012
1013         /* Increment number of inner iterations */
1014         inneriter                  += j_index_end - j_index_start;
1015
1016         /* Outer loop uses 7 flops */
1017     }
1018
1019     /* Increment number of outer iterations */
1020     outeriter        += nri;
1021
1022     /* Update outer/inner flops */
1023
1024     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);
1025 }