src/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecCSTab_VdwLJ_GeomW3P1_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_ElecCSTab_VdwLJ_GeomW3P1_VF_avx_256_single
  38  * Electrostatics interaction: CubicSplineTable
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCSTab_VdwLJ_GeomW3P1_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     real *           vdwioffsetptr1;
  75     __m256           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  76     real *           vdwioffsetptr2;
  77     __m256           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  78     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  79     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  80     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  81     __m256           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m256           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  83     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  84     real             *charge;
  85     int              nvdwtype;
  86     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  87     int              *vdwtype;
  88     real             *vdwparam;
  89     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  90     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  91     __m256i          vfitab;
  92     __m128i          vfitab_lo,vfitab_hi;
  93     __m128i          ifour       = _mm_set1_epi32(4);
  94     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  95     real             *vftab;
  96     __m256           dummy_mask,cutoff_mask;
  97     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  98     __m256           one     = _mm256_set1_ps(1.0);
  99     __m256           two     = _mm256_set1_ps(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = _mm256_set1_ps(fr->epsfac);
 112     charge           = mdatoms->chargeA;
 113     nvdwtype         = fr->ntype;
 114     vdwparam         = fr->nbfp;
 115     vdwtype          = mdatoms->typeA;
 116
 117     vftab            = kernel_data->table_elec->data;
 118     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 119
 120     /* Setup water-specific parameters */
 121     inr              = nlist->iinr[0];
 122     iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 123     iq1              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+1]));
 124     iq2              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+2]));
 125     vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 126
 127     /* Avoid stupid compiler warnings */
 128     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 129     j_coord_offsetA = 0;
 130     j_coord_offsetB = 0;
 131     j_coord_offsetC = 0;
 132     j_coord_offsetD = 0;
 133     j_coord_offsetE = 0;
 134     j_coord_offsetF = 0;
 135     j_coord_offsetG = 0;
 136     j_coord_offsetH = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     for(iidx=0;iidx<4*DIM;iidx++)
 142     {
 143         scratch[iidx] = 0.0;
 144     }
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         gmx_mm256_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 162                                                     &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 163
 164         fix0             = _mm256_setzero_ps();
 165         fiy0             = _mm256_setzero_ps();
 166         fiz0             = _mm256_setzero_ps();
 167         fix1             = _mm256_setzero_ps();
 168         fiy1             = _mm256_setzero_ps();
 169         fiz1             = _mm256_setzero_ps();
 170         fix2             = _mm256_setzero_ps();
 171         fiy2             = _mm256_setzero_ps();
 172         fiz2             = _mm256_setzero_ps();
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm256_setzero_ps();
 176         vvdwsum          = _mm256_setzero_ps();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             jnrC             = jjnr[jidx+2];
 186             jnrD             = jjnr[jidx+3];
 187             jnrE             = jjnr[jidx+4];
 188             jnrF             = jjnr[jidx+5];
 189             jnrG             = jjnr[jidx+6];
 190             jnrH             = jjnr[jidx+7];
 191             j_coord_offsetA  = DIM*jnrA;
 192             j_coord_offsetB  = DIM*jnrB;
 193             j_coord_offsetC  = DIM*jnrC;
 194             j_coord_offsetD  = DIM*jnrD;
 195             j_coord_offsetE  = DIM*jnrE;
 196             j_coord_offsetF  = DIM*jnrF;
 197             j_coord_offsetG  = DIM*jnrG;
 198             j_coord_offsetH  = DIM*jnrH;
 199
 200             /* load j atom coordinates */
 201             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 202                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 203                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 204                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 205                                                  &jx0,&jy0,&jz0);
 206
 207             /* Calculate displacement vector */
 208             dx00             = _mm256_sub_ps(ix0,jx0);
 209             dy00             = _mm256_sub_ps(iy0,jy0);
 210             dz00             = _mm256_sub_ps(iz0,jz0);
 211             dx10             = _mm256_sub_ps(ix1,jx0);
 212             dy10             = _mm256_sub_ps(iy1,jy0);
 213             dz10             = _mm256_sub_ps(iz1,jz0);
 214             dx20             = _mm256_sub_ps(ix2,jx0);
 215             dy20             = _mm256_sub_ps(iy2,jy0);
 216             dz20             = _mm256_sub_ps(iz2,jz0);
 217
 218             /* Calculate squared distance and things based on it */
 219             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 220             rsq10            = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
 221             rsq20            = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
 222
 223             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 224             rinv10           = gmx_mm256_invsqrt_ps(rsq10);
 225             rinv20           = gmx_mm256_invsqrt_ps(rsq20);
 226
 227             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 228
 229             /* Load parameters for j particles */
 230             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 231                                                                  charge+jnrC+0,charge+jnrD+0,
 232                                                                  charge+jnrE+0,charge+jnrF+0,
 233                                                                  charge+jnrG+0,charge+jnrH+0);
 234             vdwjidx0A        = 2*vdwtype[jnrA+0];
 235             vdwjidx0B        = 2*vdwtype[jnrB+0];
 236             vdwjidx0C        = 2*vdwtype[jnrC+0];
 237             vdwjidx0D        = 2*vdwtype[jnrD+0];
 238             vdwjidx0E        = 2*vdwtype[jnrE+0];
 239             vdwjidx0F        = 2*vdwtype[jnrF+0];
 240             vdwjidx0G        = 2*vdwtype[jnrG+0];
 241             vdwjidx0H        = 2*vdwtype[jnrH+0];
 242
 243             fjx0             = _mm256_setzero_ps();
 244             fjy0             = _mm256_setzero_ps();
 245             fjz0             = _mm256_setzero_ps();
 246
 247             /**************************
 248              * CALCULATE INTERACTIONS *
 249              **************************/
 250
 251             r00              = _mm256_mul_ps(rsq00,rinv00);
 252
 253             /* Compute parameters for interactions between i and j atoms */
 254             qq00             = _mm256_mul_ps(iq0,jq0);
 255             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 256                                             vdwioffsetptr0+vdwjidx0B,
 257                                             vdwioffsetptr0+vdwjidx0C,
 258                                             vdwioffsetptr0+vdwjidx0D,
 259                                             vdwioffsetptr0+vdwjidx0E,
 260                                             vdwioffsetptr0+vdwjidx0F,
 261                                             vdwioffsetptr0+vdwjidx0G,
 262                                             vdwioffsetptr0+vdwjidx0H,
 263                                             &c6_00,&c12_00);
 264
 265             /* Calculate table index by multiplying r with table scale and truncate to integer */
 266             rt               = _mm256_mul_ps(r00,vftabscale);
 267             vfitab           = _mm256_cvttps_epi32(rt);
 268             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 269             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 270             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 271             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 272             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 273             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 274
 275             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 276             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 277                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 278             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 279                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 280             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 281                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 282             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 283                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 284             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 285             Heps             = _mm256_mul_ps(vfeps,H);
 286             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 287             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 288             velec            = _mm256_mul_ps(qq00,VV);
 289             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 290             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 291
 292             /* LENNARD-JONES DISPERSION/REPULSION */
 293
 294             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 295             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 296             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 297             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 298             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 299
 300             /* Update potential sum for this i atom from the interaction with this j atom. */
 301             velecsum         = _mm256_add_ps(velecsum,velec);
 302             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 303
 304             fscal            = _mm256_add_ps(felec,fvdw);
 305
 306             /* Calculate temporary vectorial force */
 307             tx               = _mm256_mul_ps(fscal,dx00);
 308             ty               = _mm256_mul_ps(fscal,dy00);
 309             tz               = _mm256_mul_ps(fscal,dz00);
 310
 311             /* Update vectorial force */
 312             fix0             = _mm256_add_ps(fix0,tx);
 313             fiy0             = _mm256_add_ps(fiy0,ty);
 314             fiz0             = _mm256_add_ps(fiz0,tz);
 315
 316             fjx0             = _mm256_add_ps(fjx0,tx);
 317             fjy0             = _mm256_add_ps(fjy0,ty);
 318             fjz0             = _mm256_add_ps(fjz0,tz);
 319
 320             /**************************
 321              * CALCULATE INTERACTIONS *
 322              **************************/
 323
 324             r10              = _mm256_mul_ps(rsq10,rinv10);
 325
 326             /* Compute parameters for interactions between i and j atoms */
 327             qq10             = _mm256_mul_ps(iq1,jq0);
 328
 329             /* Calculate table index by multiplying r with table scale and truncate to integer */
 330             rt               = _mm256_mul_ps(r10,vftabscale);
 331             vfitab           = _mm256_cvttps_epi32(rt);
 332             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 333             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 334             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 335             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 336             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 337             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 338
 339             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 340             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 341                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 342             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 343                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 344             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 345                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 346             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 347                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 348             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 349             Heps             = _mm256_mul_ps(vfeps,H);
 350             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 351             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 352             velec            = _mm256_mul_ps(qq10,VV);
 353             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 354             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq10,FF),_mm256_mul_ps(vftabscale,rinv10)));
 355
 356             /* Update potential sum for this i atom from the interaction with this j atom. */
 357             velecsum         = _mm256_add_ps(velecsum,velec);
 358
 359             fscal            = felec;
 360
 361             /* Calculate temporary vectorial force */
 362             tx               = _mm256_mul_ps(fscal,dx10);
 363             ty               = _mm256_mul_ps(fscal,dy10);
 364             tz               = _mm256_mul_ps(fscal,dz10);
 365
 366             /* Update vectorial force */
 367             fix1             = _mm256_add_ps(fix1,tx);
 368             fiy1             = _mm256_add_ps(fiy1,ty);
 369             fiz1             = _mm256_add_ps(fiz1,tz);
 370
 371             fjx0             = _mm256_add_ps(fjx0,tx);
 372             fjy0             = _mm256_add_ps(fjy0,ty);
 373             fjz0             = _mm256_add_ps(fjz0,tz);
 374
 375             /**************************
 376              * CALCULATE INTERACTIONS *
 377              **************************/
 378
 379             r20              = _mm256_mul_ps(rsq20,rinv20);
 380
 381             /* Compute parameters for interactions between i and j atoms */
 382             qq20             = _mm256_mul_ps(iq2,jq0);
 383
 384             /* Calculate table index by multiplying r with table scale and truncate to integer */
 385             rt               = _mm256_mul_ps(r20,vftabscale);
 386             vfitab           = _mm256_cvttps_epi32(rt);
 387             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 388             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 389             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 390             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 391             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 392             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 393
 394             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 395             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 396                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 397             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 398                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 399             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 400                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 401             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 402                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 403             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 404             Heps             = _mm256_mul_ps(vfeps,H);
 405             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 406             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 407             velec            = _mm256_mul_ps(qq20,VV);
 408             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 409             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq20,FF),_mm256_mul_ps(vftabscale,rinv20)));
 410
 411             /* Update potential sum for this i atom from the interaction with this j atom. */
 412             velecsum         = _mm256_add_ps(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             /* Calculate temporary vectorial force */
 417             tx               = _mm256_mul_ps(fscal,dx20);
 418             ty               = _mm256_mul_ps(fscal,dy20);
 419             tz               = _mm256_mul_ps(fscal,dz20);
 420
 421             /* Update vectorial force */
 422             fix2             = _mm256_add_ps(fix2,tx);
 423             fiy2             = _mm256_add_ps(fiy2,ty);
 424             fiz2             = _mm256_add_ps(fiz2,tz);
 425
 426             fjx0             = _mm256_add_ps(fjx0,tx);
 427             fjy0             = _mm256_add_ps(fjy0,ty);
 428             fjz0             = _mm256_add_ps(fjz0,tz);
 429
 430             fjptrA             = f+j_coord_offsetA;
 431             fjptrB             = f+j_coord_offsetB;
 432             fjptrC             = f+j_coord_offsetC;
 433             fjptrD             = f+j_coord_offsetD;
 434             fjptrE             = f+j_coord_offsetE;
 435             fjptrF             = f+j_coord_offsetF;
 436             fjptrG             = f+j_coord_offsetG;
 437             fjptrH             = f+j_coord_offsetH;
 438
 439             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
 440
 441             /* Inner loop uses 145 flops */
 442         }
 443
 444         if(jidx<j_index_end)
 445         {
 446
 447             /* Get j neighbor index, and coordinate index */
 448             jnrlistA         = jjnr[jidx];
 449             jnrlistB         = jjnr[jidx+1];
 450             jnrlistC         = jjnr[jidx+2];
 451             jnrlistD         = jjnr[jidx+3];
 452             jnrlistE         = jjnr[jidx+4];
 453             jnrlistF         = jjnr[jidx+5];
 454             jnrlistG         = jjnr[jidx+6];
 455             jnrlistH         = jjnr[jidx+7];
 456             /* Sign of each element will be negative for non-real atoms.
 457              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 458              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 459              */
 460             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 461                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 462
 463             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 464             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 465             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 466             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 467             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 468             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 469             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 470             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 471             j_coord_offsetA  = DIM*jnrA;
 472             j_coord_offsetB  = DIM*jnrB;
 473             j_coord_offsetC  = DIM*jnrC;
 474             j_coord_offsetD  = DIM*jnrD;
 475             j_coord_offsetE  = DIM*jnrE;
 476             j_coord_offsetF  = DIM*jnrF;
 477             j_coord_offsetG  = DIM*jnrG;
 478             j_coord_offsetH  = DIM*jnrH;
 479
 480             /* load j atom coordinates */
 481             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 482                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 483                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 484                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 485                                                  &jx0,&jy0,&jz0);
 486
 487             /* Calculate displacement vector */
 488             dx00             = _mm256_sub_ps(ix0,jx0);
 489             dy00             = _mm256_sub_ps(iy0,jy0);
 490             dz00             = _mm256_sub_ps(iz0,jz0);
 491             dx10             = _mm256_sub_ps(ix1,jx0);
 492             dy10             = _mm256_sub_ps(iy1,jy0);
 493             dz10             = _mm256_sub_ps(iz1,jz0);
 494             dx20             = _mm256_sub_ps(ix2,jx0);
 495             dy20             = _mm256_sub_ps(iy2,jy0);
 496             dz20             = _mm256_sub_ps(iz2,jz0);
 497
 498             /* Calculate squared distance and things based on it */
 499             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 500             rsq10            = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
 501             rsq20            = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
 502
 503             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 504             rinv10           = gmx_mm256_invsqrt_ps(rsq10);
 505             rinv20           = gmx_mm256_invsqrt_ps(rsq20);
 506
 507             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 508
 509             /* Load parameters for j particles */
 510             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 511                                                                  charge+jnrC+0,charge+jnrD+0,
 512                                                                  charge+jnrE+0,charge+jnrF+0,
 513                                                                  charge+jnrG+0,charge+jnrH+0);
 514             vdwjidx0A        = 2*vdwtype[jnrA+0];
 515             vdwjidx0B        = 2*vdwtype[jnrB+0];
 516             vdwjidx0C        = 2*vdwtype[jnrC+0];
 517             vdwjidx0D        = 2*vdwtype[jnrD+0];
 518             vdwjidx0E        = 2*vdwtype[jnrE+0];
 519             vdwjidx0F        = 2*vdwtype[jnrF+0];
 520             vdwjidx0G        = 2*vdwtype[jnrG+0];
 521             vdwjidx0H        = 2*vdwtype[jnrH+0];
 522
 523             fjx0             = _mm256_setzero_ps();
 524             fjy0             = _mm256_setzero_ps();
 525             fjz0             = _mm256_setzero_ps();
 526
 527             /**************************
 528              * CALCULATE INTERACTIONS *
 529              **************************/
 530
 531             r00              = _mm256_mul_ps(rsq00,rinv00);
 532             r00              = _mm256_andnot_ps(dummy_mask,r00);
 533
 534             /* Compute parameters for interactions between i and j atoms */
 535             qq00             = _mm256_mul_ps(iq0,jq0);
 536             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 537                                             vdwioffsetptr0+vdwjidx0B,
 538                                             vdwioffsetptr0+vdwjidx0C,
 539                                             vdwioffsetptr0+vdwjidx0D,
 540                                             vdwioffsetptr0+vdwjidx0E,
 541                                             vdwioffsetptr0+vdwjidx0F,
 542                                             vdwioffsetptr0+vdwjidx0G,
 543                                             vdwioffsetptr0+vdwjidx0H,
 544                                             &c6_00,&c12_00);
 545
 546             /* Calculate table index by multiplying r with table scale and truncate to integer */
 547             rt               = _mm256_mul_ps(r00,vftabscale);
 548             vfitab           = _mm256_cvttps_epi32(rt);
 549             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 550             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 551             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 552             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 553             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 554             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 555
 556             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 557             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 558                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 559             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 560                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 561             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 562                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 563             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 564                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 565             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 566             Heps             = _mm256_mul_ps(vfeps,H);
 567             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 568             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 569             velec            = _mm256_mul_ps(qq00,VV);
 570             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 571             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 572
 573             /* LENNARD-JONES DISPERSION/REPULSION */
 574
 575             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 576             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 577             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 578             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 579             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 580
 581             /* Update potential sum for this i atom from the interaction with this j atom. */
 582             velec            = _mm256_andnot_ps(dummy_mask,velec);
 583             velecsum         = _mm256_add_ps(velecsum,velec);
 584             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 585             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 586
 587             fscal            = _mm256_add_ps(felec,fvdw);
 588
 589             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 590
 591             /* Calculate temporary vectorial force */
 592             tx               = _mm256_mul_ps(fscal,dx00);
 593             ty               = _mm256_mul_ps(fscal,dy00);
 594             tz               = _mm256_mul_ps(fscal,dz00);
 595
 596             /* Update vectorial force */
 597             fix0             = _mm256_add_ps(fix0,tx);
 598             fiy0             = _mm256_add_ps(fiy0,ty);
 599             fiz0             = _mm256_add_ps(fiz0,tz);
 600
 601             fjx0             = _mm256_add_ps(fjx0,tx);
 602             fjy0             = _mm256_add_ps(fjy0,ty);
 603             fjz0             = _mm256_add_ps(fjz0,tz);
 604
 605             /**************************
 606              * CALCULATE INTERACTIONS *
 607              **************************/
 608
 609             r10              = _mm256_mul_ps(rsq10,rinv10);
 610             r10              = _mm256_andnot_ps(dummy_mask,r10);
 611
 612             /* Compute parameters for interactions between i and j atoms */
 613             qq10             = _mm256_mul_ps(iq1,jq0);
 614
 615             /* Calculate table index by multiplying r with table scale and truncate to integer */
 616             rt               = _mm256_mul_ps(r10,vftabscale);
 617             vfitab           = _mm256_cvttps_epi32(rt);
 618             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 619             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 620             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 621             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 622             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 623             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 624
 625             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 626             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 627                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 628             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 629                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 630             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 631                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 632             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 633                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 634             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 635             Heps             = _mm256_mul_ps(vfeps,H);
 636             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 637             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 638             velec            = _mm256_mul_ps(qq10,VV);
 639             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 640             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq10,FF),_mm256_mul_ps(vftabscale,rinv10)));
 641
 642             /* Update potential sum for this i atom from the interaction with this j atom. */
 643             velec            = _mm256_andnot_ps(dummy_mask,velec);
 644             velecsum         = _mm256_add_ps(velecsum,velec);
 645
 646             fscal            = felec;
 647
 648             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 649
 650             /* Calculate temporary vectorial force */
 651             tx               = _mm256_mul_ps(fscal,dx10);
 652             ty               = _mm256_mul_ps(fscal,dy10);
 653             tz               = _mm256_mul_ps(fscal,dz10);
 654
 655             /* Update vectorial force */
 656             fix1             = _mm256_add_ps(fix1,tx);
 657             fiy1             = _mm256_add_ps(fiy1,ty);
 658             fiz1             = _mm256_add_ps(fiz1,tz);
 659
 660             fjx0             = _mm256_add_ps(fjx0,tx);
 661             fjy0             = _mm256_add_ps(fjy0,ty);
 662             fjz0             = _mm256_add_ps(fjz0,tz);
 663
 664             /**************************
 665              * CALCULATE INTERACTIONS *
 666              **************************/
 667
 668             r20              = _mm256_mul_ps(rsq20,rinv20);
 669             r20              = _mm256_andnot_ps(dummy_mask,r20);
 670
 671             /* Compute parameters for interactions between i and j atoms */
 672             qq20             = _mm256_mul_ps(iq2,jq0);
 673
 674             /* Calculate table index by multiplying r with table scale and truncate to integer */
 675             rt               = _mm256_mul_ps(r20,vftabscale);
 676             vfitab           = _mm256_cvttps_epi32(rt);
 677             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 678             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 679             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 680             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 681             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 682             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 683
 684             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 685             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 686                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 687             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 688                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 689             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 690                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 691             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 692                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 693             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 694             Heps             = _mm256_mul_ps(vfeps,H);
 695             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 696             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 697             velec            = _mm256_mul_ps(qq20,VV);
 698             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 699             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq20,FF),_mm256_mul_ps(vftabscale,rinv20)));
 700
 701             /* Update potential sum for this i atom from the interaction with this j atom. */
 702             velec            = _mm256_andnot_ps(dummy_mask,velec);
 703             velecsum         = _mm256_add_ps(velecsum,velec);
 704
 705             fscal            = felec;
 706
 707             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 708
 709             /* Calculate temporary vectorial force */
 710             tx               = _mm256_mul_ps(fscal,dx20);
 711             ty               = _mm256_mul_ps(fscal,dy20);
 712             tz               = _mm256_mul_ps(fscal,dz20);
 713
 714             /* Update vectorial force */
 715             fix2             = _mm256_add_ps(fix2,tx);
 716             fiy2             = _mm256_add_ps(fiy2,ty);
 717             fiz2             = _mm256_add_ps(fiz2,tz);
 718
 719             fjx0             = _mm256_add_ps(fjx0,tx);
 720             fjy0             = _mm256_add_ps(fjy0,ty);
 721             fjz0             = _mm256_add_ps(fjz0,tz);
 722
 723             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 724             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 725             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 726             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 727             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 728             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 729             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 730             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 731
 732             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
 733
 734             /* Inner loop uses 148 flops */
 735         }
 736
 737         /* End of innermost loop */
 738
 739         gmx_mm256_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 740                                                  f+i_coord_offset,fshift+i_shift_offset);
 741
 742         ggid                        = gid[iidx];
 743         /* Update potential energies */
 744         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 745         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 746
 747         /* Increment number of inner iterations */
 748         inneriter                  += j_index_end - j_index_start;
 749
 750         /* Outer loop uses 20 flops */
 751     }
 752
 753     /* Increment number of outer iterations */
 754     outeriter        += nri;
 755
 756     /* Update outer/inner flops */
 757
 758     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*148);
 759 }
 760 /*
 761  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomW3P1_F_avx_256_single
 762  * Electrostatics interaction: CubicSplineTable
 763  * VdW interaction:            LennardJones
 764  * Geometry:                   Water3-Particle
 765  * Calculate force/pot:        Force
 766  */
 767 void
 768 nb_kernel_ElecCSTab_VdwLJ_GeomW3P1_F_avx_256_single
 769                     (t_nblist * gmx_restrict                nlist,
 770                      rvec * gmx_restrict                    xx,
 771                      rvec * gmx_restrict                    ff,
 772                      t_forcerec * gmx_restrict              fr,
 773                      t_mdatoms * gmx_restrict               mdatoms,
 774                      nb_kernel_data_t * gmx_restrict        kernel_data,
 775                      t_nrnb * gmx_restrict                  nrnb)
 776 {
 777     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 778      * just 0 for non-waters.
 779      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 780      * jnr indices corresponding to data put in the four positions in the SIMD register.
 781      */
 782     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 783     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 784     int              jnrA,jnrB,jnrC,jnrD;
 785     int              jnrE,jnrF,jnrG,jnrH;
 786     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 787     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 788     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 789     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 790     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 791     real             rcutoff_scalar;
 792     real             *shiftvec,*fshift,*x,*f;
 793     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 794     real             scratch[4*DIM];
 795     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 796     real *           vdwioffsetptr0;
 797     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 798     real *           vdwioffsetptr1;
 799     __m256           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 800     real *           vdwioffsetptr2;
 801     __m256           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 802     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 803     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 804     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 805     __m256           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 806     __m256           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 807     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 808     real             *charge;
 809     int              nvdwtype;
 810     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 811     int              *vdwtype;
 812     real             *vdwparam;
 813     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 814     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 815     __m256i          vfitab;
 816     __m128i          vfitab_lo,vfitab_hi;
 817     __m128i          ifour       = _mm_set1_epi32(4);
 818     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 819     real             *vftab;
 820     __m256           dummy_mask,cutoff_mask;
 821     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 822     __m256           one     = _mm256_set1_ps(1.0);
 823     __m256           two     = _mm256_set1_ps(2.0);
 824     x                = xx[0];
 825     f                = ff[0];
 826
 827     nri              = nlist->nri;
 828     iinr             = nlist->iinr;
 829     jindex           = nlist->jindex;
 830     jjnr             = nlist->jjnr;
 831     shiftidx         = nlist->shift;
 832     gid              = nlist->gid;
 833     shiftvec         = fr->shift_vec[0];
 834     fshift           = fr->fshift[0];
 835     facel            = _mm256_set1_ps(fr->epsfac);
 836     charge           = mdatoms->chargeA;
 837     nvdwtype         = fr->ntype;
 838     vdwparam         = fr->nbfp;
 839     vdwtype          = mdatoms->typeA;
 840
 841     vftab            = kernel_data->table_elec->data;
 842     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 843
 844     /* Setup water-specific parameters */
 845     inr              = nlist->iinr[0];
 846     iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 847     iq1              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+1]));
 848     iq2              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+2]));
 849     vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 850
 851     /* Avoid stupid compiler warnings */
 852     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 853     j_coord_offsetA = 0;
 854     j_coord_offsetB = 0;
 855     j_coord_offsetC = 0;
 856     j_coord_offsetD = 0;
 857     j_coord_offsetE = 0;
 858     j_coord_offsetF = 0;
 859     j_coord_offsetG = 0;
 860     j_coord_offsetH = 0;
 861
 862     outeriter        = 0;
 863     inneriter        = 0;
 864
 865     for(iidx=0;iidx<4*DIM;iidx++)
 866     {
 867         scratch[iidx] = 0.0;
 868     }
 869
 870     /* Start outer loop over neighborlists */
 871     for(iidx=0; iidx<nri; iidx++)
 872     {
 873         /* Load shift vector for this list */
 874         i_shift_offset   = DIM*shiftidx[iidx];
 875
 876         /* Load limits for loop over neighbors */
 877         j_index_start    = jindex[iidx];
 878         j_index_end      = jindex[iidx+1];
 879
 880         /* Get outer coordinate index */
 881         inr              = iinr[iidx];
 882         i_coord_offset   = DIM*inr;
 883
 884         /* Load i particle coords and add shift vector */
 885         gmx_mm256_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 886                                                     &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 887
 888         fix0             = _mm256_setzero_ps();
 889         fiy0             = _mm256_setzero_ps();
 890         fiz0             = _mm256_setzero_ps();
 891         fix1             = _mm256_setzero_ps();
 892         fiy1             = _mm256_setzero_ps();
 893         fiz1             = _mm256_setzero_ps();
 894         fix2             = _mm256_setzero_ps();
 895         fiy2             = _mm256_setzero_ps();
 896         fiz2             = _mm256_setzero_ps();
 897
 898         /* Start inner kernel loop */
 899         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 900         {
 901
 902             /* Get j neighbor index, and coordinate index */
 903             jnrA             = jjnr[jidx];
 904             jnrB             = jjnr[jidx+1];
 905             jnrC             = jjnr[jidx+2];
 906             jnrD             = jjnr[jidx+3];
 907             jnrE             = jjnr[jidx+4];
 908             jnrF             = jjnr[jidx+5];
 909             jnrG             = jjnr[jidx+6];
 910             jnrH             = jjnr[jidx+7];
 911             j_coord_offsetA  = DIM*jnrA;
 912             j_coord_offsetB  = DIM*jnrB;
 913             j_coord_offsetC  = DIM*jnrC;
 914             j_coord_offsetD  = DIM*jnrD;
 915             j_coord_offsetE  = DIM*jnrE;
 916             j_coord_offsetF  = DIM*jnrF;
 917             j_coord_offsetG  = DIM*jnrG;
 918             j_coord_offsetH  = DIM*jnrH;
 919
 920             /* load j atom coordinates */
 921             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 922                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 923                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 924                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 925                                                  &jx0,&jy0,&jz0);
 926
 927             /* Calculate displacement vector */
 928             dx00             = _mm256_sub_ps(ix0,jx0);
 929             dy00             = _mm256_sub_ps(iy0,jy0);
 930             dz00             = _mm256_sub_ps(iz0,jz0);
 931             dx10             = _mm256_sub_ps(ix1,jx0);
 932             dy10             = _mm256_sub_ps(iy1,jy0);
 933             dz10             = _mm256_sub_ps(iz1,jz0);
 934             dx20             = _mm256_sub_ps(ix2,jx0);
 935             dy20             = _mm256_sub_ps(iy2,jy0);
 936             dz20             = _mm256_sub_ps(iz2,jz0);
 937
 938             /* Calculate squared distance and things based on it */
 939             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 940             rsq10            = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
 941             rsq20            = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
 942
 943             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 944             rinv10           = gmx_mm256_invsqrt_ps(rsq10);
 945             rinv20           = gmx_mm256_invsqrt_ps(rsq20);
 946
 947             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 948
 949             /* Load parameters for j particles */
 950             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 951                                                                  charge+jnrC+0,charge+jnrD+0,
 952                                                                  charge+jnrE+0,charge+jnrF+0,
 953                                                                  charge+jnrG+0,charge+jnrH+0);
 954             vdwjidx0A        = 2*vdwtype[jnrA+0];
 955             vdwjidx0B        = 2*vdwtype[jnrB+0];
 956             vdwjidx0C        = 2*vdwtype[jnrC+0];
 957             vdwjidx0D        = 2*vdwtype[jnrD+0];
 958             vdwjidx0E        = 2*vdwtype[jnrE+0];
 959             vdwjidx0F        = 2*vdwtype[jnrF+0];
 960             vdwjidx0G        = 2*vdwtype[jnrG+0];
 961             vdwjidx0H        = 2*vdwtype[jnrH+0];
 962
 963             fjx0             = _mm256_setzero_ps();
 964             fjy0             = _mm256_setzero_ps();
 965             fjz0             = _mm256_setzero_ps();
 966
 967             /**************************
 968              * CALCULATE INTERACTIONS *
 969              **************************/
 970
 971             r00              = _mm256_mul_ps(rsq00,rinv00);
 972
 973             /* Compute parameters for interactions between i and j atoms */
 974             qq00             = _mm256_mul_ps(iq0,jq0);
 975             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 976                                             vdwioffsetptr0+vdwjidx0B,
 977                                             vdwioffsetptr0+vdwjidx0C,
 978                                             vdwioffsetptr0+vdwjidx0D,
 979                                             vdwioffsetptr0+vdwjidx0E,
 980                                             vdwioffsetptr0+vdwjidx0F,
 981                                             vdwioffsetptr0+vdwjidx0G,
 982                                             vdwioffsetptr0+vdwjidx0H,
 983                                             &c6_00,&c12_00);
 984
 985             /* Calculate table index by multiplying r with table scale and truncate to integer */
 986             rt               = _mm256_mul_ps(r00,vftabscale);
 987             vfitab           = _mm256_cvttps_epi32(rt);
 988             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 989             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 990             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 991             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 992             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 993             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 994
 995             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 996             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 997                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 998             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 999                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1000             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1001                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1002             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1003                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1004             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1005             Heps             = _mm256_mul_ps(vfeps,H);
1006             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1007             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1008             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
1009
1010             /* LENNARD-JONES DISPERSION/REPULSION */
1011
1012             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1013             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
1014
1015             fscal            = _mm256_add_ps(felec,fvdw);
1016
1017             /* Calculate temporary vectorial force */
1018             tx               = _mm256_mul_ps(fscal,dx00);
1019             ty               = _mm256_mul_ps(fscal,dy00);
1020             tz               = _mm256_mul_ps(fscal,dz00);
1021
1022             /* Update vectorial force */
1023             fix0             = _mm256_add_ps(fix0,tx);
1024             fiy0             = _mm256_add_ps(fiy0,ty);
1025             fiz0             = _mm256_add_ps(fiz0,tz);
1026
1027             fjx0             = _mm256_add_ps(fjx0,tx);
1028             fjy0             = _mm256_add_ps(fjy0,ty);
1029             fjz0             = _mm256_add_ps(fjz0,tz);
1030
1031             /**************************
1032              * CALCULATE INTERACTIONS *
1033              **************************/
1034
1035             r10              = _mm256_mul_ps(rsq10,rinv10);
1036
1037             /* Compute parameters for interactions between i and j atoms */
1038             qq10             = _mm256_mul_ps(iq1,jq0);
1039
1040             /* Calculate table index by multiplying r with table scale and truncate to integer */
1041             rt               = _mm256_mul_ps(r10,vftabscale);
1042             vfitab           = _mm256_cvttps_epi32(rt);
1043             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1044             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1045             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
1046             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
1047             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
1048             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
1049
1050             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1051             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1052                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1053             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1054                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1055             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1056                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1057             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1058                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1059             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1060             Heps             = _mm256_mul_ps(vfeps,H);
1061             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1062             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1063             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq10,FF),_mm256_mul_ps(vftabscale,rinv10)));
1064
1065             fscal            = felec;
1066
1067             /* Calculate temporary vectorial force */
1068             tx               = _mm256_mul_ps(fscal,dx10);
1069             ty               = _mm256_mul_ps(fscal,dy10);
1070             tz               = _mm256_mul_ps(fscal,dz10);
1071
1072             /* Update vectorial force */
1073             fix1             = _mm256_add_ps(fix1,tx);
1074             fiy1             = _mm256_add_ps(fiy1,ty);
1075             fiz1             = _mm256_add_ps(fiz1,tz);
1076
1077             fjx0             = _mm256_add_ps(fjx0,tx);
1078             fjy0             = _mm256_add_ps(fjy0,ty);
1079             fjz0             = _mm256_add_ps(fjz0,tz);
1080
1081             /**************************
1082              * CALCULATE INTERACTIONS *
1083              **************************/
1084
1085             r20              = _mm256_mul_ps(rsq20,rinv20);
1086
1087             /* Compute parameters for interactions between i and j atoms */
1088             qq20             = _mm256_mul_ps(iq2,jq0);
1089
1090             /* Calculate table index by multiplying r with table scale and truncate to integer */
1091             rt               = _mm256_mul_ps(r20,vftabscale);
1092             vfitab           = _mm256_cvttps_epi32(rt);
1093             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1094             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1095             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
1096             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
1097             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
1098             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
1099
1100             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1101             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1102                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1103             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1104                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1105             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1106                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1107             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1108                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1109             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1110             Heps             = _mm256_mul_ps(vfeps,H);
1111             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1112             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1113             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq20,FF),_mm256_mul_ps(vftabscale,rinv20)));
1114
1115             fscal            = felec;
1116
1117             /* Calculate temporary vectorial force */
1118             tx               = _mm256_mul_ps(fscal,dx20);
1119             ty               = _mm256_mul_ps(fscal,dy20);
1120             tz               = _mm256_mul_ps(fscal,dz20);
1121
1122             /* Update vectorial force */
1123             fix2             = _mm256_add_ps(fix2,tx);
1124             fiy2             = _mm256_add_ps(fiy2,ty);
1125             fiz2             = _mm256_add_ps(fiz2,tz);
1126
1127             fjx0             = _mm256_add_ps(fjx0,tx);
1128             fjy0             = _mm256_add_ps(fjy0,ty);
1129             fjz0             = _mm256_add_ps(fjz0,tz);
1130
1131             fjptrA             = f+j_coord_offsetA;
1132             fjptrB             = f+j_coord_offsetB;
1133             fjptrC             = f+j_coord_offsetC;
1134             fjptrD             = f+j_coord_offsetD;
1135             fjptrE             = f+j_coord_offsetE;
1136             fjptrF             = f+j_coord_offsetF;
1137             fjptrG             = f+j_coord_offsetG;
1138             fjptrH             = f+j_coord_offsetH;
1139
1140             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
1141
1142             /* Inner loop uses 128 flops */
1143         }
1144
1145         if(jidx<j_index_end)
1146         {
1147
1148             /* Get j neighbor index, and coordinate index */
1149             jnrlistA         = jjnr[jidx];
1150             jnrlistB         = jjnr[jidx+1];
1151             jnrlistC         = jjnr[jidx+2];
1152             jnrlistD         = jjnr[jidx+3];
1153             jnrlistE         = jjnr[jidx+4];
1154             jnrlistF         = jjnr[jidx+5];
1155             jnrlistG         = jjnr[jidx+6];
1156             jnrlistH         = jjnr[jidx+7];
1157             /* Sign of each element will be negative for non-real atoms.
1158              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1159              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1160              */
1161             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
1162                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
1163
1164             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1165             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1166             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1167             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1168             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
1169             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
1170             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
1171             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
1172             j_coord_offsetA  = DIM*jnrA;
1173             j_coord_offsetB  = DIM*jnrB;
1174             j_coord_offsetC  = DIM*jnrC;
1175             j_coord_offsetD  = DIM*jnrD;
1176             j_coord_offsetE  = DIM*jnrE;
1177             j_coord_offsetF  = DIM*jnrF;
1178             j_coord_offsetG  = DIM*jnrG;
1179             j_coord_offsetH  = DIM*jnrH;
1180
1181             /* load j atom coordinates */
1182             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1183                                                  x+j_coord_offsetC,x+j_coord_offsetD,
1184                                                  x+j_coord_offsetE,x+j_coord_offsetF,
1185                                                  x+j_coord_offsetG,x+j_coord_offsetH,
1186                                                  &jx0,&jy0,&jz0);
1187
1188             /* Calculate displacement vector */
1189             dx00             = _mm256_sub_ps(ix0,jx0);
1190             dy00             = _mm256_sub_ps(iy0,jy0);
1191             dz00             = _mm256_sub_ps(iz0,jz0);
1192             dx10             = _mm256_sub_ps(ix1,jx0);
1193             dy10             = _mm256_sub_ps(iy1,jy0);
1194             dz10             = _mm256_sub_ps(iz1,jz0);
1195             dx20             = _mm256_sub_ps(ix2,jx0);
1196             dy20             = _mm256_sub_ps(iy2,jy0);
1197             dz20             = _mm256_sub_ps(iz2,jz0);
1198
1199             /* Calculate squared distance and things based on it */
1200             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
1201             rsq10            = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
1202             rsq20            = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
1203
1204             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
1205             rinv10           = gmx_mm256_invsqrt_ps(rsq10);
1206             rinv20           = gmx_mm256_invsqrt_ps(rsq20);
1207
1208             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
1209
1210             /* Load parameters for j particles */
1211             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1212                                                                  charge+jnrC+0,charge+jnrD+0,
1213                                                                  charge+jnrE+0,charge+jnrF+0,
1214                                                                  charge+jnrG+0,charge+jnrH+0);
1215             vdwjidx0A        = 2*vdwtype[jnrA+0];
1216             vdwjidx0B        = 2*vdwtype[jnrB+0];
1217             vdwjidx0C        = 2*vdwtype[jnrC+0];
1218             vdwjidx0D        = 2*vdwtype[jnrD+0];
1219             vdwjidx0E        = 2*vdwtype[jnrE+0];
1220             vdwjidx0F        = 2*vdwtype[jnrF+0];
1221             vdwjidx0G        = 2*vdwtype[jnrG+0];
1222             vdwjidx0H        = 2*vdwtype[jnrH+0];
1223
1224             fjx0             = _mm256_setzero_ps();
1225             fjy0             = _mm256_setzero_ps();
1226             fjz0             = _mm256_setzero_ps();
1227
1228             /**************************
1229              * CALCULATE INTERACTIONS *
1230              **************************/
1231
1232             r00              = _mm256_mul_ps(rsq00,rinv00);
1233             r00              = _mm256_andnot_ps(dummy_mask,r00);
1234
1235             /* Compute parameters for interactions between i and j atoms */
1236             qq00             = _mm256_mul_ps(iq0,jq0);
1237             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
1238                                             vdwioffsetptr0+vdwjidx0B,
1239                                             vdwioffsetptr0+vdwjidx0C,
1240                                             vdwioffsetptr0+vdwjidx0D,
1241                                             vdwioffsetptr0+vdwjidx0E,
1242                                             vdwioffsetptr0+vdwjidx0F,
1243                                             vdwioffsetptr0+vdwjidx0G,
1244                                             vdwioffsetptr0+vdwjidx0H,
1245                                             &c6_00,&c12_00);
1246
1247             /* Calculate table index by multiplying r with table scale and truncate to integer */
1248             rt               = _mm256_mul_ps(r00,vftabscale);
1249             vfitab           = _mm256_cvttps_epi32(rt);
1250             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1251             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1252             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
1253             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
1254             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
1255             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
1256
1257             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1258             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1259                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1260             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1261                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1262             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1263                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1264             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1265                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1266             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1267             Heps             = _mm256_mul_ps(vfeps,H);
1268             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1269             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1270             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
1271
1272             /* LENNARD-JONES DISPERSION/REPULSION */
1273
1274             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1275             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
1276
1277             fscal            = _mm256_add_ps(felec,fvdw);
1278
1279             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
1280
1281             /* Calculate temporary vectorial force */
1282             tx               = _mm256_mul_ps(fscal,dx00);
1283             ty               = _mm256_mul_ps(fscal,dy00);
1284             tz               = _mm256_mul_ps(fscal,dz00);
1285
1286             /* Update vectorial force */
1287             fix0             = _mm256_add_ps(fix0,tx);
1288             fiy0             = _mm256_add_ps(fiy0,ty);
1289             fiz0             = _mm256_add_ps(fiz0,tz);
1290
1291             fjx0             = _mm256_add_ps(fjx0,tx);
1292             fjy0             = _mm256_add_ps(fjy0,ty);
1293             fjz0             = _mm256_add_ps(fjz0,tz);
1294
1295             /**************************
1296              * CALCULATE INTERACTIONS *
1297              **************************/
1298
1299             r10              = _mm256_mul_ps(rsq10,rinv10);
1300             r10              = _mm256_andnot_ps(dummy_mask,r10);
1301
1302             /* Compute parameters for interactions between i and j atoms */
1303             qq10             = _mm256_mul_ps(iq1,jq0);
1304
1305             /* Calculate table index by multiplying r with table scale and truncate to integer */
1306             rt               = _mm256_mul_ps(r10,vftabscale);
1307             vfitab           = _mm256_cvttps_epi32(rt);
1308             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1309             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1310             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
1311             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
1312             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
1313             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
1314
1315             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1316             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1317                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1318             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1319                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1320             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1321                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1322             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1323                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1324             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1325             Heps             = _mm256_mul_ps(vfeps,H);
1326             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1327             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1328             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq10,FF),_mm256_mul_ps(vftabscale,rinv10)));
1329
1330             fscal            = felec;
1331
1332             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
1333
1334             /* Calculate temporary vectorial force */
1335             tx               = _mm256_mul_ps(fscal,dx10);
1336             ty               = _mm256_mul_ps(fscal,dy10);
1337             tz               = _mm256_mul_ps(fscal,dz10);
1338
1339             /* Update vectorial force */
1340             fix1             = _mm256_add_ps(fix1,tx);
1341             fiy1             = _mm256_add_ps(fiy1,ty);
1342             fiz1             = _mm256_add_ps(fiz1,tz);
1343
1344             fjx0             = _mm256_add_ps(fjx0,tx);
1345             fjy0             = _mm256_add_ps(fjy0,ty);
1346             fjz0             = _mm256_add_ps(fjz0,tz);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             r20              = _mm256_mul_ps(rsq20,rinv20);
1353             r20              = _mm256_andnot_ps(dummy_mask,r20);
1354
1355             /* Compute parameters for interactions between i and j atoms */
1356             qq20             = _mm256_mul_ps(iq2,jq0);
1357
1358             /* Calculate table index by multiplying r with table scale and truncate to integer */
1359             rt               = _mm256_mul_ps(r20,vftabscale);
1360             vfitab           = _mm256_cvttps_epi32(rt);
1361             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1362             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1363             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
1364             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
1365             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
1366             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
1367
1368             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1369             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1370                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1371             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1372                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1373             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1374                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1375             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1376                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1377             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1378             Heps             = _mm256_mul_ps(vfeps,H);
1379             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1380             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1381             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq20,FF),_mm256_mul_ps(vftabscale,rinv20)));
1382
1383             fscal            = felec;
1384
1385             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
1386
1387             /* Calculate temporary vectorial force */
1388             tx               = _mm256_mul_ps(fscal,dx20);
1389             ty               = _mm256_mul_ps(fscal,dy20);
1390             tz               = _mm256_mul_ps(fscal,dz20);
1391
1392             /* Update vectorial force */
1393             fix2             = _mm256_add_ps(fix2,tx);
1394             fiy2             = _mm256_add_ps(fiy2,ty);
1395             fiz2             = _mm256_add_ps(fiz2,tz);
1396
1397             fjx0             = _mm256_add_ps(fjx0,tx);
1398             fjy0             = _mm256_add_ps(fjy0,ty);
1399             fjz0             = _mm256_add_ps(fjz0,tz);
1400
1401             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1402             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1403             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1404             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1405             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
1406             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
1407             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
1408             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
1409
1410             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
1411
1412             /* Inner loop uses 131 flops */
1413         }
1414
1415         /* End of innermost loop */
1416
1417         gmx_mm256_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1418                                                  f+i_coord_offset,fshift+i_shift_offset);
1419
1420         /* Increment number of inner iterations */
1421         inneriter                  += j_index_end - j_index_start;
1422
1423         /* Outer loop uses 18 flops */
1424     }
1425
1426     /* Increment number of outer iterations */
1427     outeriter        += nri;
1428
1429     /* Update outer/inner flops */
1430
1431     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*131);
1432 }