src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_avx_256_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_single.h"
  48 #include "kernelutil_x86_avx_256_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_256_single
  52  * Electrostatics interaction: CubicSplineTable
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_256_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrE,jnrF,jnrG,jnrH;
  76     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  77     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  84     real             scratch[4*DIM];
  85     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  89     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  98     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  99     __m256i          vfitab;
 100     __m128i          vfitab_lo,vfitab_hi;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 103     real             *vftab;
 104     __m256           dummy_mask,cutoff_mask;
 105     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 106     __m256           one     = _mm256_set1_ps(1.0);
 107     __m256           two     = _mm256_set1_ps(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm256_set1_ps(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     vftab            = kernel_data->table_elec_vdw->data;
 126     vftabscale       = _mm256_set1_ps(kernel_data->table_elec_vdw->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             /* Load parameters for j particles */
 216             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 217                                                                  charge+jnrC+0,charge+jnrD+0,
 218                                                                  charge+jnrE+0,charge+jnrF+0,
 219                                                                  charge+jnrG+0,charge+jnrH+0);
 220             vdwjidx0A        = 2*vdwtype[jnrA+0];
 221             vdwjidx0B        = 2*vdwtype[jnrB+0];
 222             vdwjidx0C        = 2*vdwtype[jnrC+0];
 223             vdwjidx0D        = 2*vdwtype[jnrD+0];
 224             vdwjidx0E        = 2*vdwtype[jnrE+0];
 225             vdwjidx0F        = 2*vdwtype[jnrF+0];
 226             vdwjidx0G        = 2*vdwtype[jnrG+0];
 227             vdwjidx0H        = 2*vdwtype[jnrH+0];
 228
 229             /**************************
 230              * CALCULATE INTERACTIONS *
 231              **************************/
 232
 233             r00              = _mm256_mul_ps(rsq00,rinv00);
 234
 235             /* Compute parameters for interactions between i and j atoms */
 236             qq00             = _mm256_mul_ps(iq0,jq0);
 237             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 238                                             vdwioffsetptr0+vdwjidx0B,
 239                                             vdwioffsetptr0+vdwjidx0C,
 240                                             vdwioffsetptr0+vdwjidx0D,
 241                                             vdwioffsetptr0+vdwjidx0E,
 242                                             vdwioffsetptr0+vdwjidx0F,
 243                                             vdwioffsetptr0+vdwjidx0G,
 244                                             vdwioffsetptr0+vdwjidx0H,
 245                                             &c6_00,&c12_00);
 246
 247             /* Calculate table index by multiplying r with table scale and truncate to integer */
 248             rt               = _mm256_mul_ps(r00,vftabscale);
 249             vfitab           = _mm256_cvttps_epi32(rt);
 250             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 251             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 252             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 253             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 254             vfitab_lo        = _mm_slli_epi32(_mm_add_epi32(vfitab_lo,_mm_slli_epi32(vfitab_lo,1)),2);
 255             vfitab_hi        = _mm_slli_epi32(_mm_add_epi32(vfitab_hi,_mm_slli_epi32(vfitab_hi,1)),2);
 256
 257             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 258             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 259                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 260             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 261                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 262             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 263                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 264             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 265                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 266             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 267             Heps             = _mm256_mul_ps(vfeps,H);
 268             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 269             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 270             velec            = _mm256_mul_ps(qq00,VV);
 271             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 272             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 273
 274             /* CUBIC SPLINE TABLE DISPERSION */
 275             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 276             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 277             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 278                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 279             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 280                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 281             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 282                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 283             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 284                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 285             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 286             Heps             = _mm256_mul_ps(vfeps,H);
 287             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 288             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 289             vvdw6            = _mm256_mul_ps(c6_00,VV);
 290             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 291             fvdw6            = _mm256_mul_ps(c6_00,FF);
 292
 293             /* CUBIC SPLINE TABLE REPULSION */
 294             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 295             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 296             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 297                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 298             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 299                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 300             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 301                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 302             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 303                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 304             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 305             Heps             = _mm256_mul_ps(vfeps,H);
 306             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 307             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 308             vvdw12           = _mm256_mul_ps(c12_00,VV);
 309             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 310             fvdw12           = _mm256_mul_ps(c12_00,FF);
 311             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 312             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 313
 314             /* Update potential sum for this i atom from the interaction with this j atom. */
 315             velecsum         = _mm256_add_ps(velecsum,velec);
 316             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 317
 318             fscal            = _mm256_add_ps(felec,fvdw);
 319
 320             /* Calculate temporary vectorial force */
 321             tx               = _mm256_mul_ps(fscal,dx00);
 322             ty               = _mm256_mul_ps(fscal,dy00);
 323             tz               = _mm256_mul_ps(fscal,dz00);
 324
 325             /* Update vectorial force */
 326             fix0             = _mm256_add_ps(fix0,tx);
 327             fiy0             = _mm256_add_ps(fiy0,ty);
 328             fiz0             = _mm256_add_ps(fiz0,tz);
 329
 330             fjptrA             = f+j_coord_offsetA;
 331             fjptrB             = f+j_coord_offsetB;
 332             fjptrC             = f+j_coord_offsetC;
 333             fjptrD             = f+j_coord_offsetD;
 334             fjptrE             = f+j_coord_offsetE;
 335             fjptrF             = f+j_coord_offsetF;
 336             fjptrG             = f+j_coord_offsetG;
 337             fjptrH             = f+j_coord_offsetH;
 338             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 339
 340             /* Inner loop uses 73 flops */
 341         }
 342
 343         if(jidx<j_index_end)
 344         {
 345
 346             /* Get j neighbor index, and coordinate index */
 347             jnrlistA         = jjnr[jidx];
 348             jnrlistB         = jjnr[jidx+1];
 349             jnrlistC         = jjnr[jidx+2];
 350             jnrlistD         = jjnr[jidx+3];
 351             jnrlistE         = jjnr[jidx+4];
 352             jnrlistF         = jjnr[jidx+5];
 353             jnrlistG         = jjnr[jidx+6];
 354             jnrlistH         = jjnr[jidx+7];
 355             /* Sign of each element will be negative for non-real atoms.
 356              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 357              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 358              */
 359             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 360                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 361
 362             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 363             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 364             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 365             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 366             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 367             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 368             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 369             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 370             j_coord_offsetA  = DIM*jnrA;
 371             j_coord_offsetB  = DIM*jnrB;
 372             j_coord_offsetC  = DIM*jnrC;
 373             j_coord_offsetD  = DIM*jnrD;
 374             j_coord_offsetE  = DIM*jnrE;
 375             j_coord_offsetF  = DIM*jnrF;
 376             j_coord_offsetG  = DIM*jnrG;
 377             j_coord_offsetH  = DIM*jnrH;
 378
 379             /* load j atom coordinates */
 380             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 381                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 382                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 383                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 384                                                  &jx0,&jy0,&jz0);
 385
 386             /* Calculate displacement vector */
 387             dx00             = _mm256_sub_ps(ix0,jx0);
 388             dy00             = _mm256_sub_ps(iy0,jy0);
 389             dz00             = _mm256_sub_ps(iz0,jz0);
 390
 391             /* Calculate squared distance and things based on it */
 392             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 393
 394             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 395
 396             /* Load parameters for j particles */
 397             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 398                                                                  charge+jnrC+0,charge+jnrD+0,
 399                                                                  charge+jnrE+0,charge+jnrF+0,
 400                                                                  charge+jnrG+0,charge+jnrH+0);
 401             vdwjidx0A        = 2*vdwtype[jnrA+0];
 402             vdwjidx0B        = 2*vdwtype[jnrB+0];
 403             vdwjidx0C        = 2*vdwtype[jnrC+0];
 404             vdwjidx0D        = 2*vdwtype[jnrD+0];
 405             vdwjidx0E        = 2*vdwtype[jnrE+0];
 406             vdwjidx0F        = 2*vdwtype[jnrF+0];
 407             vdwjidx0G        = 2*vdwtype[jnrG+0];
 408             vdwjidx0H        = 2*vdwtype[jnrH+0];
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             r00              = _mm256_mul_ps(rsq00,rinv00);
 415             r00              = _mm256_andnot_ps(dummy_mask,r00);
 416
 417             /* Compute parameters for interactions between i and j atoms */
 418             qq00             = _mm256_mul_ps(iq0,jq0);
 419             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 420                                             vdwioffsetptr0+vdwjidx0B,
 421                                             vdwioffsetptr0+vdwjidx0C,
 422                                             vdwioffsetptr0+vdwjidx0D,
 423                                             vdwioffsetptr0+vdwjidx0E,
 424                                             vdwioffsetptr0+vdwjidx0F,
 425                                             vdwioffsetptr0+vdwjidx0G,
 426                                             vdwioffsetptr0+vdwjidx0H,
 427                                             &c6_00,&c12_00);
 428
 429             /* Calculate table index by multiplying r with table scale and truncate to integer */
 430             rt               = _mm256_mul_ps(r00,vftabscale);
 431             vfitab           = _mm256_cvttps_epi32(rt);
 432             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 433             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 434             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 435             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 436             vfitab_lo        = _mm_slli_epi32(_mm_add_epi32(vfitab_lo,_mm_slli_epi32(vfitab_lo,1)),2);
 437             vfitab_hi        = _mm_slli_epi32(_mm_add_epi32(vfitab_hi,_mm_slli_epi32(vfitab_hi,1)),2);
 438
 439             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 440             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 441                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 442             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 443                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 444             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 445                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 446             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 447                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 448             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 449             Heps             = _mm256_mul_ps(vfeps,H);
 450             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 451             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 452             velec            = _mm256_mul_ps(qq00,VV);
 453             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 454             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 455
 456             /* CUBIC SPLINE TABLE DISPERSION */
 457             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 458             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 459             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 460                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 461             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 462                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 463             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 464                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 465             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 466                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 467             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 468             Heps             = _mm256_mul_ps(vfeps,H);
 469             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 470             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 471             vvdw6            = _mm256_mul_ps(c6_00,VV);
 472             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 473             fvdw6            = _mm256_mul_ps(c6_00,FF);
 474
 475             /* CUBIC SPLINE TABLE REPULSION */
 476             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 477             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 478             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 479                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 480             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 481                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 482             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 483                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 484             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 485                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 486             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 487             Heps             = _mm256_mul_ps(vfeps,H);
 488             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 489             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 490             vvdw12           = _mm256_mul_ps(c12_00,VV);
 491             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 492             fvdw12           = _mm256_mul_ps(c12_00,FF);
 493             vvdw             = _mm256_add_ps(vvdw12,vvdw6);
 494             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velec            = _mm256_andnot_ps(dummy_mask,velec);
 498             velecsum         = _mm256_add_ps(velecsum,velec);
 499             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 500             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 501
 502             fscal            = _mm256_add_ps(felec,fvdw);
 503
 504             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 505
 506             /* Calculate temporary vectorial force */
 507             tx               = _mm256_mul_ps(fscal,dx00);
 508             ty               = _mm256_mul_ps(fscal,dy00);
 509             tz               = _mm256_mul_ps(fscal,dz00);
 510
 511             /* Update vectorial force */
 512             fix0             = _mm256_add_ps(fix0,tx);
 513             fiy0             = _mm256_add_ps(fiy0,ty);
 514             fiz0             = _mm256_add_ps(fiz0,tz);
 515
 516             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 517             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 518             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 519             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 520             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 521             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 522             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 523             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 524             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 525
 526             /* Inner loop uses 74 flops */
 527         }
 528
 529         /* End of innermost loop */
 530
 531         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 532                                                  f+i_coord_offset,fshift+i_shift_offset);
 533
 534         ggid                        = gid[iidx];
 535         /* Update potential energies */
 536         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 537         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 538
 539         /* Increment number of inner iterations */
 540         inneriter                  += j_index_end - j_index_start;
 541
 542         /* Outer loop uses 9 flops */
 543     }
 544
 545     /* Increment number of outer iterations */
 546     outeriter        += nri;
 547
 548     /* Update outer/inner flops */
 549
 550     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*74);
 551 }
 552 /*
 553  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_256_single
 554  * Electrostatics interaction: CubicSplineTable
 555  * VdW interaction:            CubicSplineTable
 556  * Geometry:                   Particle-Particle
 557  * Calculate force/pot:        Force
 558  */
 559 void
 560 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_256_single
 561                     (t_nblist                    * gmx_restrict       nlist,
 562                      rvec                        * gmx_restrict          xx,
 563                      rvec                        * gmx_restrict          ff,
 564                      t_forcerec                  * gmx_restrict          fr,
 565                      t_mdatoms                   * gmx_restrict     mdatoms,
 566                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 567                      t_nrnb                      * gmx_restrict        nrnb)
 568 {
 569     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 570      * just 0 for non-waters.
 571      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 572      * jnr indices corresponding to data put in the four positions in the SIMD register.
 573      */
 574     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 575     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 576     int              jnrA,jnrB,jnrC,jnrD;
 577     int              jnrE,jnrF,jnrG,jnrH;
 578     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 579     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 580     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 581     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 582     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 583     real             rcutoff_scalar;
 584     real             *shiftvec,*fshift,*x,*f;
 585     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 586     real             scratch[4*DIM];
 587     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 588     real *           vdwioffsetptr0;
 589     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 590     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 591     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 592     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 593     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 594     real             *charge;
 595     int              nvdwtype;
 596     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 597     int              *vdwtype;
 598     real             *vdwparam;
 599     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 600     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 601     __m256i          vfitab;
 602     __m128i          vfitab_lo,vfitab_hi;
 603     __m128i          ifour       = _mm_set1_epi32(4);
 604     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 605     real             *vftab;
 606     __m256           dummy_mask,cutoff_mask;
 607     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 608     __m256           one     = _mm256_set1_ps(1.0);
 609     __m256           two     = _mm256_set1_ps(2.0);
 610     x                = xx[0];
 611     f                = ff[0];
 612
 613     nri              = nlist->nri;
 614     iinr             = nlist->iinr;
 615     jindex           = nlist->jindex;
 616     jjnr             = nlist->jjnr;
 617     shiftidx         = nlist->shift;
 618     gid              = nlist->gid;
 619     shiftvec         = fr->shift_vec[0];
 620     fshift           = fr->fshift[0];
 621     facel            = _mm256_set1_ps(fr->epsfac);
 622     charge           = mdatoms->chargeA;
 623     nvdwtype         = fr->ntype;
 624     vdwparam         = fr->nbfp;
 625     vdwtype          = mdatoms->typeA;
 626
 627     vftab            = kernel_data->table_elec_vdw->data;
 628     vftabscale       = _mm256_set1_ps(kernel_data->table_elec_vdw->scale);
 629
 630     /* Avoid stupid compiler warnings */
 631     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 632     j_coord_offsetA = 0;
 633     j_coord_offsetB = 0;
 634     j_coord_offsetC = 0;
 635     j_coord_offsetD = 0;
 636     j_coord_offsetE = 0;
 637     j_coord_offsetF = 0;
 638     j_coord_offsetG = 0;
 639     j_coord_offsetH = 0;
 640
 641     outeriter        = 0;
 642     inneriter        = 0;
 643
 644     for(iidx=0;iidx<4*DIM;iidx++)
 645     {
 646         scratch[iidx] = 0.0;
 647     }
 648
 649     /* Start outer loop over neighborlists */
 650     for(iidx=0; iidx<nri; iidx++)
 651     {
 652         /* Load shift vector for this list */
 653         i_shift_offset   = DIM*shiftidx[iidx];
 654
 655         /* Load limits for loop over neighbors */
 656         j_index_start    = jindex[iidx];
 657         j_index_end      = jindex[iidx+1];
 658
 659         /* Get outer coordinate index */
 660         inr              = iinr[iidx];
 661         i_coord_offset   = DIM*inr;
 662
 663         /* Load i particle coords and add shift vector */
 664         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 665
 666         fix0             = _mm256_setzero_ps();
 667         fiy0             = _mm256_setzero_ps();
 668         fiz0             = _mm256_setzero_ps();
 669
 670         /* Load parameters for i particles */
 671         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 672         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 673
 674         /* Start inner kernel loop */
 675         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 676         {
 677
 678             /* Get j neighbor index, and coordinate index */
 679             jnrA             = jjnr[jidx];
 680             jnrB             = jjnr[jidx+1];
 681             jnrC             = jjnr[jidx+2];
 682             jnrD             = jjnr[jidx+3];
 683             jnrE             = jjnr[jidx+4];
 684             jnrF             = jjnr[jidx+5];
 685             jnrG             = jjnr[jidx+6];
 686             jnrH             = jjnr[jidx+7];
 687             j_coord_offsetA  = DIM*jnrA;
 688             j_coord_offsetB  = DIM*jnrB;
 689             j_coord_offsetC  = DIM*jnrC;
 690             j_coord_offsetD  = DIM*jnrD;
 691             j_coord_offsetE  = DIM*jnrE;
 692             j_coord_offsetF  = DIM*jnrF;
 693             j_coord_offsetG  = DIM*jnrG;
 694             j_coord_offsetH  = DIM*jnrH;
 695
 696             /* load j atom coordinates */
 697             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 698                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 699                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 700                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 701                                                  &jx0,&jy0,&jz0);
 702
 703             /* Calculate displacement vector */
 704             dx00             = _mm256_sub_ps(ix0,jx0);
 705             dy00             = _mm256_sub_ps(iy0,jy0);
 706             dz00             = _mm256_sub_ps(iz0,jz0);
 707
 708             /* Calculate squared distance and things based on it */
 709             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 710
 711             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 712
 713             /* Load parameters for j particles */
 714             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 715                                                                  charge+jnrC+0,charge+jnrD+0,
 716                                                                  charge+jnrE+0,charge+jnrF+0,
 717                                                                  charge+jnrG+0,charge+jnrH+0);
 718             vdwjidx0A        = 2*vdwtype[jnrA+0];
 719             vdwjidx0B        = 2*vdwtype[jnrB+0];
 720             vdwjidx0C        = 2*vdwtype[jnrC+0];
 721             vdwjidx0D        = 2*vdwtype[jnrD+0];
 722             vdwjidx0E        = 2*vdwtype[jnrE+0];
 723             vdwjidx0F        = 2*vdwtype[jnrF+0];
 724             vdwjidx0G        = 2*vdwtype[jnrG+0];
 725             vdwjidx0H        = 2*vdwtype[jnrH+0];
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             r00              = _mm256_mul_ps(rsq00,rinv00);
 732
 733             /* Compute parameters for interactions between i and j atoms */
 734             qq00             = _mm256_mul_ps(iq0,jq0);
 735             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 736                                             vdwioffsetptr0+vdwjidx0B,
 737                                             vdwioffsetptr0+vdwjidx0C,
 738                                             vdwioffsetptr0+vdwjidx0D,
 739                                             vdwioffsetptr0+vdwjidx0E,
 740                                             vdwioffsetptr0+vdwjidx0F,
 741                                             vdwioffsetptr0+vdwjidx0G,
 742                                             vdwioffsetptr0+vdwjidx0H,
 743                                             &c6_00,&c12_00);
 744
 745             /* Calculate table index by multiplying r with table scale and truncate to integer */
 746             rt               = _mm256_mul_ps(r00,vftabscale);
 747             vfitab           = _mm256_cvttps_epi32(rt);
 748             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 749             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 750             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 751             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 752             vfitab_lo        = _mm_slli_epi32(_mm_add_epi32(vfitab_lo,_mm_slli_epi32(vfitab_lo,1)),2);
 753             vfitab_hi        = _mm_slli_epi32(_mm_add_epi32(vfitab_hi,_mm_slli_epi32(vfitab_hi,1)),2);
 754
 755             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 756             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 757                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 758             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 759                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 760             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 761                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 762             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 763                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 764             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 765             Heps             = _mm256_mul_ps(vfeps,H);
 766             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 767             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 768             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 769
 770             /* CUBIC SPLINE TABLE DISPERSION */
 771             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 772             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 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 61 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             /* Load parameters for j particles */
 884             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 885                                                                  charge+jnrC+0,charge+jnrD+0,
 886                                                                  charge+jnrE+0,charge+jnrF+0,
 887                                                                  charge+jnrG+0,charge+jnrH+0);
 888             vdwjidx0A        = 2*vdwtype[jnrA+0];
 889             vdwjidx0B        = 2*vdwtype[jnrB+0];
 890             vdwjidx0C        = 2*vdwtype[jnrC+0];
 891             vdwjidx0D        = 2*vdwtype[jnrD+0];
 892             vdwjidx0E        = 2*vdwtype[jnrE+0];
 893             vdwjidx0F        = 2*vdwtype[jnrF+0];
 894             vdwjidx0G        = 2*vdwtype[jnrG+0];
 895             vdwjidx0H        = 2*vdwtype[jnrH+0];
 896
 897             /**************************
 898              * CALCULATE INTERACTIONS *
 899              **************************/
 900
 901             r00              = _mm256_mul_ps(rsq00,rinv00);
 902             r00              = _mm256_andnot_ps(dummy_mask,r00);
 903
 904             /* Compute parameters for interactions between i and j atoms */
 905             qq00             = _mm256_mul_ps(iq0,jq0);
 906             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 907                                             vdwioffsetptr0+vdwjidx0B,
 908                                             vdwioffsetptr0+vdwjidx0C,
 909                                             vdwioffsetptr0+vdwjidx0D,
 910                                             vdwioffsetptr0+vdwjidx0E,
 911                                             vdwioffsetptr0+vdwjidx0F,
 912                                             vdwioffsetptr0+vdwjidx0G,
 913                                             vdwioffsetptr0+vdwjidx0H,
 914                                             &c6_00,&c12_00);
 915
 916             /* Calculate table index by multiplying r with table scale and truncate to integer */
 917             rt               = _mm256_mul_ps(r00,vftabscale);
 918             vfitab           = _mm256_cvttps_epi32(rt);
 919             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 920             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 921             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 922             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 923             vfitab_lo        = _mm_slli_epi32(_mm_add_epi32(vfitab_lo,_mm_slli_epi32(vfitab_lo,1)),2);
 924             vfitab_hi        = _mm_slli_epi32(_mm_add_epi32(vfitab_hi,_mm_slli_epi32(vfitab_hi,1)),2);
 925
 926             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 927             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 928                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 929             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 930                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 931             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 932                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 933             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 934                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 935             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 936             Heps             = _mm256_mul_ps(vfeps,H);
 937             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 938             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 939             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 940
 941             /* CUBIC SPLINE TABLE DISPERSION */
 942             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 943             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 944             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 945                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 946             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 947                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 948             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 949                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 950             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 951                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 952             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 953             Heps             = _mm256_mul_ps(vfeps,H);
 954             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 955             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 956             fvdw6            = _mm256_mul_ps(c6_00,FF);
 957
 958             /* CUBIC SPLINE TABLE REPULSION */
 959             vfitab_lo        = _mm_add_epi32(vfitab_lo,ifour);
 960             vfitab_hi        = _mm_add_epi32(vfitab_hi,ifour);
 961             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 962                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 963             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 964                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 965             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 966                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 967             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 968                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 969             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 970             Heps             = _mm256_mul_ps(vfeps,H);
 971             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 972             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 973             fvdw12           = _mm256_mul_ps(c12_00,FF);
 974             fvdw             = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
 975
 976             fscal            = _mm256_add_ps(felec,fvdw);
 977
 978             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 979
 980             /* Calculate temporary vectorial force */
 981             tx               = _mm256_mul_ps(fscal,dx00);
 982             ty               = _mm256_mul_ps(fscal,dy00);
 983             tz               = _mm256_mul_ps(fscal,dz00);
 984
 985             /* Update vectorial force */
 986             fix0             = _mm256_add_ps(fix0,tx);
 987             fiy0             = _mm256_add_ps(fiy0,ty);
 988             fiz0             = _mm256_add_ps(fiz0,tz);
 989
 990             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 991             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 992             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 993             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 994             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 995             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 996             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 997             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 998             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 999
1000             /* Inner loop uses 62 flops */
1001         }
1002
1003         /* End of innermost loop */
1004
1005         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1006                                                  f+i_coord_offset,fshift+i_shift_offset);
1007
1008         /* Increment number of inner iterations */
1009         inneriter                  += j_index_end - j_index_start;
1010
1011         /* Outer loop uses 7 flops */
1012     }
1013
1014     /* Increment number of outer iterations */
1015     outeriter        += nri;
1016
1017     /* Update outer/inner flops */
1018
1019     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);
1020 }