src/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecEw_VdwCSTab_GeomP1P1_avx_256_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_256_single.h"
  34 #include "kernelutil_x86_avx_256_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_avx_256_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_avx_256_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrE,jnrF,jnrG,jnrH;
  62     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  63     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  64     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  65     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  66     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  67     real             rcutoff_scalar;
  68     real             *shiftvec,*fshift,*x,*f;
  69     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  70     real             scratch[4*DIM];
  71     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  72     real *           vdwioffsetptr0;
  73     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  75     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  77     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     int              nvdwtype;
  80     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  81     int              *vdwtype;
  82     real             *vdwparam;
  83     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  84     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  85     __m256i          vfitab;
  86     __m128i          vfitab_lo,vfitab_hi;
  87     __m128i          ifour       = _mm_set1_epi32(4);
  88     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  89     real             *vftab;
  90     __m256i          ewitab;
  91     __m128i          ewitab_lo,ewitab_hi;
  92     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  93     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  94     real             *ewtab;
  95     __m256           dummy_mask,cutoff_mask;
  96     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  97     __m256           one     = _mm256_set1_ps(1.0);
  98     __m256           two     = _mm256_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm256_set1_ps(fr->epsfac);
 111     charge           = mdatoms->chargeA;
 112     nvdwtype         = fr->ntype;
 113     vdwparam         = fr->nbfp;
 114     vdwtype          = mdatoms->typeA;
 115
 116     vftab            = kernel_data->table_vdw->data;
 117     vftabscale       = _mm256_set1_ps(kernel_data->table_vdw->scale);
 118
 119     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 120     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff);
 121     beta2            = _mm256_mul_ps(beta,beta);
 122     beta3            = _mm256_mul_ps(beta,beta2);
 123
 124     ewtab            = fr->ic->tabq_coul_FDV0;
 125     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 126     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 127
 128     /* Avoid stupid compiler warnings */
 129     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 130     j_coord_offsetA = 0;
 131     j_coord_offsetB = 0;
 132     j_coord_offsetC = 0;
 133     j_coord_offsetD = 0;
 134     j_coord_offsetE = 0;
 135     j_coord_offsetF = 0;
 136     j_coord_offsetG = 0;
 137     j_coord_offsetH = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     for(iidx=0;iidx<4*DIM;iidx++)
 143     {
 144         scratch[iidx] = 0.0;
 145     }
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 163
 164         fix0             = _mm256_setzero_ps();
 165         fiy0             = _mm256_setzero_ps();
 166         fiz0             = _mm256_setzero_ps();
 167
 168         /* Load parameters for i particles */
 169         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 170         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 171
 172         /* Reset potential sums */
 173         velecsum         = _mm256_setzero_ps();
 174         vvdwsum          = _mm256_setzero_ps();
 175
 176         /* Start inner kernel loop */
 177         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 178         {
 179
 180             /* Get j neighbor index, and coordinate index */
 181             jnrA             = jjnr[jidx];
 182             jnrB             = jjnr[jidx+1];
 183             jnrC             = jjnr[jidx+2];
 184             jnrD             = jjnr[jidx+3];
 185             jnrE             = jjnr[jidx+4];
 186             jnrF             = jjnr[jidx+5];
 187             jnrG             = jjnr[jidx+6];
 188             jnrH             = jjnr[jidx+7];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191             j_coord_offsetC  = DIM*jnrC;
 192             j_coord_offsetD  = DIM*jnrD;
 193             j_coord_offsetE  = DIM*jnrE;
 194             j_coord_offsetF  = DIM*jnrF;
 195             j_coord_offsetG  = DIM*jnrG;
 196             j_coord_offsetH  = DIM*jnrH;
 197
 198             /* load j atom coordinates */
 199             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 200                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 201                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 202                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 203                                                  &jx0,&jy0,&jz0);
 204
 205             /* Calculate displacement vector */
 206             dx00             = _mm256_sub_ps(ix0,jx0);
 207             dy00             = _mm256_sub_ps(iy0,jy0);
 208             dz00             = _mm256_sub_ps(iz0,jz0);
 209
 210             /* Calculate squared distance and things based on it */
 211             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 212
 213             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 214
 215             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 216
 217             /* Load parameters for j particles */
 218             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 219                                                                  charge+jnrC+0,charge+jnrD+0,
 220                                                                  charge+jnrE+0,charge+jnrF+0,
 221                                                                  charge+jnrG+0,charge+jnrH+0);
 222             vdwjidx0A        = 2*vdwtype[jnrA+0];
 223             vdwjidx0B        = 2*vdwtype[jnrB+0];
 224             vdwjidx0C        = 2*vdwtype[jnrC+0];
 225             vdwjidx0D        = 2*vdwtype[jnrD+0];
 226             vdwjidx0E        = 2*vdwtype[jnrE+0];
 227             vdwjidx0F        = 2*vdwtype[jnrF+0];
 228             vdwjidx0G        = 2*vdwtype[jnrG+0];
 229             vdwjidx0H        = 2*vdwtype[jnrH+0];
 230
 231             /**************************
 232              * CALCULATE INTERACTIONS *
 233              **************************/
 234
 235             r00              = _mm256_mul_ps(rsq00,rinv00);
 236
 237             /* Compute parameters for interactions between i and j atoms */
 238             qq00             = _mm256_mul_ps(iq0,jq0);
 239             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 240                                             vdwioffsetptr0+vdwjidx0B,
 241                                             vdwioffsetptr0+vdwjidx0C,
 242                                             vdwioffsetptr0+vdwjidx0D,
 243                                             vdwioffsetptr0+vdwjidx0E,
 244                                             vdwioffsetptr0+vdwjidx0F,
 245                                             vdwioffsetptr0+vdwjidx0G,
 246                                             vdwioffsetptr0+vdwjidx0H,
 247                                             &c6_00,&c12_00);
 248
 249             /* Calculate table index by multiplying r with table scale and truncate to integer */
 250             rt               = _mm256_mul_ps(r00,vftabscale);
 251             vfitab           = _mm256_cvttps_epi32(rt);
 252             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 253             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 254             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 255             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 256             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 257             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 258
 259             /* EWALD ELECTROSTATICS */
 260
 261             /* Analytical PME correction */
 262             zeta2            = _mm256_mul_ps(beta2,rsq00);
 263             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 264             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 265             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 266             felec            = _mm256_mul_ps(qq00,felec);
 267             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 268             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 269             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 270             velec            = _mm256_mul_ps(qq00,velec);
 271
 272             /* CUBIC SPLINE TABLE DISPERSION */
 273             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 274                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 275             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 276                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 277             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 278                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 279             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 280                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 281             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 282             Heps             = _mm256_mul_ps(vfeps,H);
 283             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 284             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 285             vvdw6            = _mm256_mul_ps(c6_00,VV);
 286             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 287             fvdw6            = _mm256_mul_ps(c6_00,FF);
 288
 289             /* CUBIC SPLINE TABLE REPULSION */
 290             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 291             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 292             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 293                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 294             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 295                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 296             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 297                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 298             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 299                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 300             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 301             Heps             = _mm256_mul_ps(vfeps,H);
 302             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 303             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 304             vvdw12           = _mm256_mul_ps(c12_00,VV);
 305             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 306             fvdw12           = _mm256_mul_ps(c12_00,FF);
 307             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 308             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm256_add_ps(velecsum,velec);
 312             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 313
 314             fscal            = _mm256_add_ps(felec,fvdw);
 315
 316             /* Calculate temporary vectorial force */
 317             tx               = _mm256_mul_ps(fscal,dx00);
 318             ty               = _mm256_mul_ps(fscal,dy00);
 319             tz               = _mm256_mul_ps(fscal,dz00);
 320
 321             /* Update vectorial force */
 322             fix0             = _mm256_add_ps(fix0,tx);
 323             fiy0             = _mm256_add_ps(fiy0,ty);
 324             fiz0             = _mm256_add_ps(fiz0,tz);
 325
 326             fjptrA             = f+j_coord_offsetA;
 327             fjptrB             = f+j_coord_offsetB;
 328             fjptrC             = f+j_coord_offsetC;
 329             fjptrD             = f+j_coord_offsetD;
 330             fjptrE             = f+j_coord_offsetE;
 331             fjptrF             = f+j_coord_offsetF;
 332             fjptrG             = f+j_coord_offsetG;
 333             fjptrH             = f+j_coord_offsetH;
 334             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 335
 336             /* Inner loop uses 118 flops */
 337         }
 338
 339         if(jidx<j_index_end)
 340         {
 341
 342             /* Get j neighbor index, and coordinate index */
 343             jnrlistA         = jjnr[jidx];
 344             jnrlistB         = jjnr[jidx+1];
 345             jnrlistC         = jjnr[jidx+2];
 346             jnrlistD         = jjnr[jidx+3];
 347             jnrlistE         = jjnr[jidx+4];
 348             jnrlistF         = jjnr[jidx+5];
 349             jnrlistG         = jjnr[jidx+6];
 350             jnrlistH         = jjnr[jidx+7];
 351             /* Sign of each element will be negative for non-real atoms.
 352              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 353              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 354              */
 355             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 356                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 357
 358             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 359             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 360             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 361             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 362             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 363             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 364             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 365             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 366             j_coord_offsetA  = DIM*jnrA;
 367             j_coord_offsetB  = DIM*jnrB;
 368             j_coord_offsetC  = DIM*jnrC;
 369             j_coord_offsetD  = DIM*jnrD;
 370             j_coord_offsetE  = DIM*jnrE;
 371             j_coord_offsetF  = DIM*jnrF;
 372             j_coord_offsetG  = DIM*jnrG;
 373             j_coord_offsetH  = DIM*jnrH;
 374
 375             /* load j atom coordinates */
 376             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 377                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 378                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 379                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 380                                                  &jx0,&jy0,&jz0);
 381
 382             /* Calculate displacement vector */
 383             dx00             = _mm256_sub_ps(ix0,jx0);
 384             dy00             = _mm256_sub_ps(iy0,jy0);
 385             dz00             = _mm256_sub_ps(iz0,jz0);
 386
 387             /* Calculate squared distance and things based on it */
 388             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 389
 390             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 391
 392             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 393
 394             /* Load parameters for j particles */
 395             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 396                                                                  charge+jnrC+0,charge+jnrD+0,
 397                                                                  charge+jnrE+0,charge+jnrF+0,
 398                                                                  charge+jnrG+0,charge+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             /* Calculate table index by multiplying r with table scale and truncate to integer */
 428             rt               = _mm256_mul_ps(r00,vftabscale);
 429             vfitab           = _mm256_cvttps_epi32(rt);
 430             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 431             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 432             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 433             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 434             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 435             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 436
 437             /* EWALD ELECTROSTATICS */
 438
 439             /* Analytical PME correction */
 440             zeta2            = _mm256_mul_ps(beta2,rsq00);
 441             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 442             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 443             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 444             felec            = _mm256_mul_ps(qq00,felec);
 445             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 446             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 447             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 448             velec            = _mm256_mul_ps(qq00,velec);
 449
 450             /* CUBIC SPLINE TABLE DISPERSION */
 451             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 452                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 453             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 454                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 455             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 456                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 457             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 458                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 459             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 460             Heps             = _mm256_mul_ps(vfeps,H);
 461             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 462             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 463             vvdw6            = _mm256_mul_ps(c6_00,VV);
 464             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 465             fvdw6            = _mm256_mul_ps(c6_00,FF);
 466
 467             /* CUBIC SPLINE TABLE REPULSION */
 468             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 469             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 470             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 471                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 472             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 473                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 474             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 475                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 476             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 477                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 478             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 479             Heps             = _mm256_mul_ps(vfeps,H);
 480             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 481             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 482             vvdw12           = _mm256_mul_ps(c12_00,VV);
 483             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 484             fvdw12           = _mm256_mul_ps(c12_00,FF);
 485             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 486             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 487
 488             /* Update potential sum for this i atom from the interaction with this j atom. */
 489             velec            = _mm256_andnot_ps(dummy_mask,velec);
 490             velecsum         = _mm256_add_ps(velecsum,velec);
 491             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 492             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 493
 494             fscal            = _mm256_add_ps(felec,fvdw);
 495
 496             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 497
 498             /* Calculate temporary vectorial force */
 499             tx               = _mm256_mul_ps(fscal,dx00);
 500             ty               = _mm256_mul_ps(fscal,dy00);
 501             tz               = _mm256_mul_ps(fscal,dz00);
 502
 503             /* Update vectorial force */
 504             fix0             = _mm256_add_ps(fix0,tx);
 505             fiy0             = _mm256_add_ps(fiy0,ty);
 506             fiz0             = _mm256_add_ps(fiz0,tz);
 507
 508             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 509             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 510             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 511             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 512             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 513             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 514             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 515             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 516             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 517
 518             /* Inner loop uses 119 flops */
 519         }
 520
 521         /* End of innermost loop */
 522
 523         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 524                                                  f+i_coord_offset,fshift+i_shift_offset);
 525
 526         ggid                        = gid[iidx];
 527         /* Update potential energies */
 528         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 529         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 530
 531         /* Increment number of inner iterations */
 532         inneriter                  += j_index_end - j_index_start;
 533
 534         /* Outer loop uses 9 flops */
 535     }
 536
 537     /* Increment number of outer iterations */
 538     outeriter        += nri;
 539
 540     /* Update outer/inner flops */
 541
 542     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*119);
 543 }
 544 /*
 545  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_single
 546  * Electrostatics interaction: Ewald
 547  * VdW interaction:            CubicSplineTable
 548  * Geometry:                   Particle-Particle
 549  * Calculate force/pot:        Force
 550  */
 551 void
 552 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_single
 553                     (t_nblist * gmx_restrict                nlist,
 554                      rvec * gmx_restrict                    xx,
 555                      rvec * gmx_restrict                    ff,
 556                      t_forcerec * gmx_restrict              fr,
 557                      t_mdatoms * gmx_restrict               mdatoms,
 558                      nb_kernel_data_t * gmx_restrict        kernel_data,
 559                      t_nrnb * gmx_restrict                  nrnb)
 560 {
 561     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 562      * just 0 for non-waters.
 563      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 564      * jnr indices corresponding to data put in the four positions in the SIMD register.
 565      */
 566     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 567     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 568     int              jnrA,jnrB,jnrC,jnrD;
 569     int              jnrE,jnrF,jnrG,jnrH;
 570     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 571     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 572     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 573     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 574     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 575     real             rcutoff_scalar;
 576     real             *shiftvec,*fshift,*x,*f;
 577     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 578     real             scratch[4*DIM];
 579     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 580     real *           vdwioffsetptr0;
 581     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 582     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 583     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 584     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 585     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 586     real             *charge;
 587     int              nvdwtype;
 588     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 589     int              *vdwtype;
 590     real             *vdwparam;
 591     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 592     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 593     __m256i          vfitab;
 594     __m128i          vfitab_lo,vfitab_hi;
 595     __m128i          ifour       = _mm_set1_epi32(4);
 596     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 597     real             *vftab;
 598     __m256i          ewitab;
 599     __m128i          ewitab_lo,ewitab_hi;
 600     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 601     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 602     real             *ewtab;
 603     __m256           dummy_mask,cutoff_mask;
 604     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 605     __m256           one     = _mm256_set1_ps(1.0);
 606     __m256           two     = _mm256_set1_ps(2.0);
 607     x                = xx[0];
 608     f                = ff[0];
 609
 610     nri              = nlist->nri;
 611     iinr             = nlist->iinr;
 612     jindex           = nlist->jindex;
 613     jjnr             = nlist->jjnr;
 614     shiftidx         = nlist->shift;
 615     gid              = nlist->gid;
 616     shiftvec         = fr->shift_vec[0];
 617     fshift           = fr->fshift[0];
 618     facel            = _mm256_set1_ps(fr->epsfac);
 619     charge           = mdatoms->chargeA;
 620     nvdwtype         = fr->ntype;
 621     vdwparam         = fr->nbfp;
 622     vdwtype          = mdatoms->typeA;
 623
 624     vftab            = kernel_data->table_vdw->data;
 625     vftabscale       = _mm256_set1_ps(kernel_data->table_vdw->scale);
 626
 627     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 628     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff);
 629     beta2            = _mm256_mul_ps(beta,beta);
 630     beta3            = _mm256_mul_ps(beta,beta2);
 631
 632     ewtab            = fr->ic->tabq_coul_F;
 633     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 634     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 635
 636     /* Avoid stupid compiler warnings */
 637     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 638     j_coord_offsetA = 0;
 639     j_coord_offsetB = 0;
 640     j_coord_offsetC = 0;
 641     j_coord_offsetD = 0;
 642     j_coord_offsetE = 0;
 643     j_coord_offsetF = 0;
 644     j_coord_offsetG = 0;
 645     j_coord_offsetH = 0;
 646
 647     outeriter        = 0;
 648     inneriter        = 0;
 649
 650     for(iidx=0;iidx<4*DIM;iidx++)
 651     {
 652         scratch[iidx] = 0.0;
 653     }
 654
 655     /* Start outer loop over neighborlists */
 656     for(iidx=0; iidx<nri; iidx++)
 657     {
 658         /* Load shift vector for this list */
 659         i_shift_offset   = DIM*shiftidx[iidx];
 660
 661         /* Load limits for loop over neighbors */
 662         j_index_start    = jindex[iidx];
 663         j_index_end      = jindex[iidx+1];
 664
 665         /* Get outer coordinate index */
 666         inr              = iinr[iidx];
 667         i_coord_offset   = DIM*inr;
 668
 669         /* Load i particle coords and add shift vector */
 670         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 671
 672         fix0             = _mm256_setzero_ps();
 673         fiy0             = _mm256_setzero_ps();
 674         fiz0             = _mm256_setzero_ps();
 675
 676         /* Load parameters for i particles */
 677         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 678         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 679
 680         /* Start inner kernel loop */
 681         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 682         {
 683
 684             /* Get j neighbor index, and coordinate index */
 685             jnrA             = jjnr[jidx];
 686             jnrB             = jjnr[jidx+1];
 687             jnrC             = jjnr[jidx+2];
 688             jnrD             = jjnr[jidx+3];
 689             jnrE             = jjnr[jidx+4];
 690             jnrF             = jjnr[jidx+5];
 691             jnrG             = jjnr[jidx+6];
 692             jnrH             = jjnr[jidx+7];
 693             j_coord_offsetA  = DIM*jnrA;
 694             j_coord_offsetB  = DIM*jnrB;
 695             j_coord_offsetC  = DIM*jnrC;
 696             j_coord_offsetD  = DIM*jnrD;
 697             j_coord_offsetE  = DIM*jnrE;
 698             j_coord_offsetF  = DIM*jnrF;
 699             j_coord_offsetG  = DIM*jnrG;
 700             j_coord_offsetH  = DIM*jnrH;
 701
 702             /* load j atom coordinates */
 703             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 704                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 705                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 706                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 707                                                  &jx0,&jy0,&jz0);
 708
 709             /* Calculate displacement vector */
 710             dx00             = _mm256_sub_ps(ix0,jx0);
 711             dy00             = _mm256_sub_ps(iy0,jy0);
 712             dz00             = _mm256_sub_ps(iz0,jz0);
 713
 714             /* Calculate squared distance and things based on it */
 715             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 716
 717             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 718
 719             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 720
 721             /* Load parameters for j particles */
 722             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 723                                                                  charge+jnrC+0,charge+jnrD+0,
 724                                                                  charge+jnrE+0,charge+jnrF+0,
 725                                                                  charge+jnrG+0,charge+jnrH+0);
 726             vdwjidx0A        = 2*vdwtype[jnrA+0];
 727             vdwjidx0B        = 2*vdwtype[jnrB+0];
 728             vdwjidx0C        = 2*vdwtype[jnrC+0];
 729             vdwjidx0D        = 2*vdwtype[jnrD+0];
 730             vdwjidx0E        = 2*vdwtype[jnrE+0];
 731             vdwjidx0F        = 2*vdwtype[jnrF+0];
 732             vdwjidx0G        = 2*vdwtype[jnrG+0];
 733             vdwjidx0H        = 2*vdwtype[jnrH+0];
 734
 735             /**************************
 736              * CALCULATE INTERACTIONS *
 737              **************************/
 738
 739             r00              = _mm256_mul_ps(rsq00,rinv00);
 740
 741             /* Compute parameters for interactions between i and j atoms */
 742             qq00             = _mm256_mul_ps(iq0,jq0);
 743             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 744                                             vdwioffsetptr0+vdwjidx0B,
 745                                             vdwioffsetptr0+vdwjidx0C,
 746                                             vdwioffsetptr0+vdwjidx0D,
 747                                             vdwioffsetptr0+vdwjidx0E,
 748                                             vdwioffsetptr0+vdwjidx0F,
 749                                             vdwioffsetptr0+vdwjidx0G,
 750                                             vdwioffsetptr0+vdwjidx0H,
 751                                             &c6_00,&c12_00);
 752
 753             /* Calculate table index by multiplying r with table scale and truncate to integer */
 754             rt               = _mm256_mul_ps(r00,vftabscale);
 755             vfitab           = _mm256_cvttps_epi32(rt);
 756             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 757             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 758             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 759             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 760             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 761             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 762
 763             /* EWALD ELECTROSTATICS */
 764
 765             /* Analytical PME correction */
 766             zeta2            = _mm256_mul_ps(beta2,rsq00);
 767             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 768             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 769             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 770             felec            = _mm256_mul_ps(qq00,felec);
 771
 772             /* CUBIC SPLINE TABLE DISPERSION */
 773             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 774                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 775             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 776                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 777             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 778                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 779             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 780                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 781             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 782             Heps             = _mm256_mul_ps(vfeps,H);
 783             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 784             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 785             fvdw6            = _mm256_mul_ps(c6_00,FF);
 786
 787             /* CUBIC SPLINE TABLE REPULSION */
 788             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 789             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 790             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 791                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 792             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 793                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 794             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 795                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 796             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 797                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 798             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 799             Heps             = _mm256_mul_ps(vfeps,H);
 800             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 801             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 802             fvdw12           = _mm256_mul_ps(c12_00,FF);
 803             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 804
 805             fscal            = _mm256_add_ps(felec,fvdw);
 806
 807             /* Calculate temporary vectorial force */
 808             tx               = _mm256_mul_ps(fscal,dx00);
 809             ty               = _mm256_mul_ps(fscal,dy00);
 810             tz               = _mm256_mul_ps(fscal,dz00);
 811
 812             /* Update vectorial force */
 813             fix0             = _mm256_add_ps(fix0,tx);
 814             fiy0             = _mm256_add_ps(fiy0,ty);
 815             fiz0             = _mm256_add_ps(fiz0,tz);
 816
 817             fjptrA             = f+j_coord_offsetA;
 818             fjptrB             = f+j_coord_offsetB;
 819             fjptrC             = f+j_coord_offsetC;
 820             fjptrD             = f+j_coord_offsetD;
 821             fjptrE             = f+j_coord_offsetE;
 822             fjptrF             = f+j_coord_offsetF;
 823             fjptrG             = f+j_coord_offsetG;
 824             fjptrH             = f+j_coord_offsetH;
 825             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 826
 827             /* Inner loop uses 82 flops */
 828         }
 829
 830         if(jidx<j_index_end)
 831         {
 832
 833             /* Get j neighbor index, and coordinate index */
 834             jnrlistA         = jjnr[jidx];
 835             jnrlistB         = jjnr[jidx+1];
 836             jnrlistC         = jjnr[jidx+2];
 837             jnrlistD         = jjnr[jidx+3];
 838             jnrlistE         = jjnr[jidx+4];
 839             jnrlistF         = jjnr[jidx+5];
 840             jnrlistG         = jjnr[jidx+6];
 841             jnrlistH         = jjnr[jidx+7];
 842             /* Sign of each element will be negative for non-real atoms.
 843              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 844              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 845              */
 846             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 847                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 848
 849             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 850             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 851             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 852             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 853             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 854             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 855             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 856             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 857             j_coord_offsetA  = DIM*jnrA;
 858             j_coord_offsetB  = DIM*jnrB;
 859             j_coord_offsetC  = DIM*jnrC;
 860             j_coord_offsetD  = DIM*jnrD;
 861             j_coord_offsetE  = DIM*jnrE;
 862             j_coord_offsetF  = DIM*jnrF;
 863             j_coord_offsetG  = DIM*jnrG;
 864             j_coord_offsetH  = DIM*jnrH;
 865
 866             /* load j atom coordinates */
 867             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 868                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 869                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 870                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 871                                                  &jx0,&jy0,&jz0);
 872
 873             /* Calculate displacement vector */
 874             dx00             = _mm256_sub_ps(ix0,jx0);
 875             dy00             = _mm256_sub_ps(iy0,jy0);
 876             dz00             = _mm256_sub_ps(iz0,jz0);
 877
 878             /* Calculate squared distance and things based on it */
 879             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 880
 881             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 882
 883             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 884
 885             /* Load parameters for j particles */
 886             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 887                                                                  charge+jnrC+0,charge+jnrD+0,
 888                                                                  charge+jnrE+0,charge+jnrF+0,
 889                                                                  charge+jnrG+0,charge+jnrH+0);
 890             vdwjidx0A        = 2*vdwtype[jnrA+0];
 891             vdwjidx0B        = 2*vdwtype[jnrB+0];
 892             vdwjidx0C        = 2*vdwtype[jnrC+0];
 893             vdwjidx0D        = 2*vdwtype[jnrD+0];
 894             vdwjidx0E        = 2*vdwtype[jnrE+0];
 895             vdwjidx0F        = 2*vdwtype[jnrF+0];
 896             vdwjidx0G        = 2*vdwtype[jnrG+0];
 897             vdwjidx0H        = 2*vdwtype[jnrH+0];
 898
 899             /**************************
 900              * CALCULATE INTERACTIONS *
 901              **************************/
 902
 903             r00              = _mm256_mul_ps(rsq00,rinv00);
 904             r00              = _mm256_andnot_ps(dummy_mask,r00);
 905
 906             /* Compute parameters for interactions between i and j atoms */
 907             qq00             = _mm256_mul_ps(iq0,jq0);
 908             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 909                                             vdwioffsetptr0+vdwjidx0B,
 910                                             vdwioffsetptr0+vdwjidx0C,
 911                                             vdwioffsetptr0+vdwjidx0D,
 912                                             vdwioffsetptr0+vdwjidx0E,
 913                                             vdwioffsetptr0+vdwjidx0F,
 914                                             vdwioffsetptr0+vdwjidx0G,
 915                                             vdwioffsetptr0+vdwjidx0H,
 916                                             &c6_00,&c12_00);
 917
 918             /* Calculate table index by multiplying r with table scale and truncate to integer */
 919             rt               = _mm256_mul_ps(r00,vftabscale);
 920             vfitab           = _mm256_cvttps_epi32(rt);
 921             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 922             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 923             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 924             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 925             vfitab_lo        = _mm_slli_epi32(vfitab_lo,3);
 926             vfitab_hi        = _mm_slli_epi32(vfitab_hi,3);
 927
 928             /* EWALD ELECTROSTATICS */
 929
 930             /* Analytical PME correction */
 931             zeta2            = _mm256_mul_ps(beta2,rsq00);
 932             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 933             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 934             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 935             felec            = _mm256_mul_ps(qq00,felec);
 936
 937             /* CUBIC SPLINE TABLE DISPERSION */
 938             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 939                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 940             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 941                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 942             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 943                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 944             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 945                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 946             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 947             Heps             = _mm256_mul_ps(vfeps,H);
 948             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 949             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 950             fvdw6            = _mm256_mul_ps(c6_00,FF);
 951
 952             /* CUBIC SPLINE TABLE REPULSION */
 953             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 954             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 955             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 956                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 957             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 958                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 959             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 960                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 961             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 962                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 963             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 964             Heps             = _mm256_mul_ps(vfeps,H);
 965             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 966             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 967             fvdw12           = _mm256_mul_ps(c12_00,FF);
 968             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 969
 970             fscal            = _mm256_add_ps(felec,fvdw);
 971
 972             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 973
 974             /* Calculate temporary vectorial force */
 975             tx               = _mm256_mul_ps(fscal,dx00);
 976             ty               = _mm256_mul_ps(fscal,dy00);
 977             tz               = _mm256_mul_ps(fscal,dz00);
 978
 979             /* Update vectorial force */
 980             fix0             = _mm256_add_ps(fix0,tx);
 981             fiy0             = _mm256_add_ps(fiy0,ty);
 982             fiz0             = _mm256_add_ps(fiz0,tz);
 983
 984             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 985             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 986             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 987             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 988             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 989             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 990             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 991             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 992             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 993
 994             /* Inner loop uses 83 flops */
 995         }
 996
 997         /* End of innermost loop */
 998
 999         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1000                                                  f+i_coord_offset,fshift+i_shift_offset);
1001
1002         /* Increment number of inner iterations */
1003         inneriter                  += j_index_end - j_index_start;
1004
1005         /* Outer loop uses 7 flops */
1006     }
1007
1008     /* Increment number of outer iterations */
1009     outeriter        += nri;
1010
1011     /* Update outer/inner flops */
1012
1013     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*83);
1014 }